Towards Robust Pedestrian Detection in Crowded Image Sequences
- 格式:pdf
- 大小:2.09 MB
- 文档页数:8
关于无人驾驶出故障的英语作文续写全文共3篇示例,供读者参考篇1Here's a continuation of an essay about a self-driving car malfunction, written in the voice of a student and around 2000 words:The Terrifying Day My Self-Driving Car Went RogueI'll never forget that fateful morning when my trustyself-driving car betrayed me. It was just another routine commute to campus, or so I thought. Little did I know, I was about to experience a nightmare that would shake my faith in autonomous technology forever.As usual, I settled into the backseat, cracked open my textbook, and let the car take control. The gentle hum of the engine and the rhythmic tapping of my fingertips on the keyboard lulled me into a state of tranquility. That's when everything went horribly wrong.Without warning, the car jolted violently, causing my laptop to go flying across the cabin. I gripped the seat in terror as the vehicle careened off the highway, narrowly missing a massivesemi-truck. My heart pounded like a sledgehammer against my ribcage, and cold sweat trickled down my forehead."What's happening?" I screamed, fumbling for the emergency override button. But the car seemed to have a mind of its own, ignoring my frantic commands.We blazed through red lights and stop signs, weaving through traffic like a deranged roller-coaster. Horns blared, and pedestrians scattered in our wake, their faces contorted in horror.I was trapped in a metal casket hurtling towards oblivion, powerless to regain control.In a desperate attempt to escape, I considered unbuckling my seatbelt and flinging myself out onto the unforgiving asphalt. But the potential consequences of such a reckless act kept me frozen in place, my knuckles white from gripping the armrests.After what felt like an eternity, the car screeched to a halt in the middle of a busy intersection, its tires smoking and its chassis trembling. I sat there, paralyzed, waiting for the inevitable collision that never came. The eerie silence was punctuated only by the rapid pounding of my pulse.Slowly, I became aware of the chaotic scene unfolding around me. Bystanders gawked in disbelief, their phones trainedon my renegade vehicle. Sirens wailed in the distance, drawing ever closer.With trembling hands, I finally managed to disengage the self-driving mode and regain manual control. I limped the battered car to the side of the road, my body numb with shock and adrenaline.As I stepped out onto the sidewalk, my legs threatened to give way beneath me. I sank to the ground, burying my face in my hands, trying to make sense of the nightmare I had just endured.In the aftermath, I learned that a critical software glitch had caused my car's autonomous systems to malfunction catastrophically. The fail-safes designed to prevent such incidents had failed spectacularly, nearly resulting in a devastating tragedy.The incident left me shaken to my core, my faith inself-driving technology shattered like the shards of glass littering the street. How could I ever trust these machines again, knowing how quickly they could turn from faithful servants into unpredictable monsters?For weeks, I found myself gripped by panic attacks at the mere thought of stepping into another autonomous vehicle. The once-soothing hum of an engine now triggered flashbacks of that harrowing morning, transporting me back to theheart-stopping terror of being a passenger in my own hijacked car.Slowly, through therapy and support from loved ones, I began to heal from the traumatic experience. But the scars it left on my psyche would never fully fade.In the years that followed, I became an outspoken advocate for stricter regulations and more rigorous testing of self-driving systems. I lobbied tirelessly for improved safety measures and redundancies to prevent such nightmarish scenarios from ever occurring again.Looking back, I realize how lucky I was to escape that ordeal unscathed, both physically and mentally. But for countless others who have fallen victim to the failures of autonomous technology, the consequences have been far more grave.As these systems continue to evolve and proliferate, we must remain vigilant, demanding the highest standards of safety and reliability. We cannot allow the pursuit of convenience andinnovation to blind us to the potential dangers lurking beneath the sleek exteriors of our self-driving chariots.For me, that fateful morning will forever serve as a sobering reminder of the fragility of our technological marvels and the importance of maintaining a healthy respect for the machines we entrust with our lives.篇2The Perils of Self-Driving Cars: A Cautionary TaleAs an avid technology enthusiast, I've been closely following the rapid advancements in the field of autonomous vehicles. The prospect of cars that can navigate our roads without human intervention has captivated the imagination of many, promising a future of increased mobility, reduced traffic congestion, and improved safety. However, recent incidents involving self-driving car malfunctions have raised serious concerns about the reliability and readiness of this technology for widespread adoption.One chilling incident that caught my attention was the highly publicized case of an Uber self-driving car striking and killing a pedestrian in Tempe, Arizona, back in 2018. This tragic event served as a sobering reminder that, despite the remarkableprogress made in this field, autonomous vehicles are not infallible, and their failures can have devastating consequences.The details of the Tempe incident are both disturbing and eye-opening. According to reports, the self-driving Uber vehicle failed to detect a pedestrian crossing the street, resulting in a fatal collision. Investigations revealed that the car's sensors had detected the pedestrian but failed to classify her as a hazard, leading to a catastrophic lapse in decision-making. It was later discovered that the vehicle's emergency braking system had been disabled, further compounding the issue.This incident highlights one of the most significant challenges facing the development of self-driving cars: the inherent complexity of real-world driving scenarios. While these vehicles excel in controlled environments and well-defined situations, they struggle to adapt to the unpredictable and dynamic nature of our roads. Pedestrians, cyclists, animals, and even unexpected weather conditions can pose formidable challenges for the most advanced autonomous systems.Another concerning aspect of self-driving car malfunctions is the potential for software glitches and cyber vulnerabilities. As these vehicles become increasingly reliant on complex algorithms and interconnected systems, the risk of software bugsor malicious hacking attempts increases exponentially. A single coding error or security breach could potentially compromise the safety of passengers and bystanders alike.One particularly alarming example of this occurred in 2015 when security researchers demonstrated their ability to remotely hack into a Jeep Cherokee's computer systems, taking control of various functions, including the brakes and steering. While this incident did not involve a self-driving car, it highlighted the potential risks associated with the increasing integration of software and connectivity in modern vehicles.Beyond the technical challenges, the ethical and legal implications of self-driving car malfunctions are equally complex. In the event of an accident caused by an autonomous vehicle, determining liability and assigning responsibility becomes a daunting task. Should the manufacturer be held accountable for software failures? Or should the responsibility fall on the vehicle's occupants, even if they were not actively controlling the car?These questions have sparked heated debates among lawmakers, policymakers, and industry leaders, as they grapple with the need to establish clear regulations and guidelines for the safe and responsible deployment of self-driving technology.Despite these concerns, it would be shortsighted to dismiss the potential benefits of autonomous vehicles altogether. When implemented correctly and with robust safety measures in place, self-driving cars could revolutionize transportation by reducing human error, which is a leading cause of accidents on our roads today. Additionally, the increased mobility afforded by this technology could greatly improve the quality of life for individuals with disabilities or those who are unable to operate conventional vehicles.However, as we forge ahead in the pursuit of this promising technology, it is crucial that we prioritize safety above all else. Rigorous testing, comprehensive regulations, and a unwavering commitment to addressing potential vulnerabilities must be at the forefront of our efforts.In conclusion, while the dream of self-driving cars holds immense promise, the recent incidents of malfunctions serve as a stark reminder of the challenges篇3The Harrowing Tale of a Self-Driving Car Gone RogueIt was just another typical Friday evening. After finishing up my classes for the day, I hopped into the family's brand newself-driving car to head home. With a simple voice command, I had engaged the autonomous mode, eager to kick back and catch up on some readings for next week. However, little did I know that this routine commute was about to take a terrifying turn.The car had been cruising smoothly along the highway for the first twenty minutes or so. Suddenly, without warning, the vehicle started violently swerving from side to side, nearly colliding with the concrete barriers lining the road. A series of blaring alarms erupted from the dashboard as warning lights flashed uncontrollably. In that moment, sheer panic gripped me.I frantically tried to regain manual control, but the steering wheel remained utterly unresponsive, frozen in place as the car careened erratically.I could hardly process what was happening. The onboard computer systems, which were supposed to be an engineering marvel of redundancies and failsafes, had apparently suffered a catastrophic systems failure. My mind raced with terrifying scenarios of the car plowing through traffic or, worse yet, veering off the highway entirely.Adrenaline coursing through my veins, I scrambled to find the emergency override switch, a failsafe mechanism designed tocompletely cut power to the autonomous systems in crisis situations. With trembling hands, I finally located the bright red lever and yanked it with every ounce of strength I could muster. The car immediately jolted to an unceremonious stop in the middle of the lane, its systems finally powering down.Heart still pounding out of my chest, I cautiously regained control of the manually. Horns blared all around me as irate drivers swerved to avoid a multi-car pileup. Carefully, I guided the immobilized vehicle over to the shoulder, badly shaken but immensely relieved to have averted disaster.As I sat there trying to catch my breath, I couldn't help but reflect on the deeply unsettling incident. Self-driving cars had been hailed as the unequivocal next frontier of transportation, promising enhanced safety and mobility through advanced computer vision and sensor fusion algorithms. Billions had been poured into developing these technologies under the assumption that they would be virtually infallible.And yet, in a single harrowing experience, the very system that was supposed to prevent human error had failed catastrophically with nearly devastating consequences. While the probability of such an event may have been miniscule on paper, I had just witnessed firsthand how even revolutionarytechnologies trusted with peoples' lives could buckle under the relentless onslaught of the real world's countless edge cases and unpredictable variables.In the aftermath, a flurry of investigations were launched to determine the root cause of the incident. Potential culprits included buggy software updates, sensor deviations, cyber attacks, and even premature component failures. The findings, whatever they may ultimately be, will undoubtedly trigger serious recalibrations across the industry's testing procedures and safety validation processes.For my part, the experience has profoundly reshaped my views on the overly optimistic narratives surrounding autonomous vehicles and artificial intelligence technologies broadly. While the potential benefits remain tantalizing in theory, witnessing a worst-case scenario first-hand has instilled a much more sober outlook on the formidable challenges and risks that have yet to be fully accounted for and mitigated.As I pulled up to my driveway that fateful evening, making a conscious effort to shut off the ignition myself, it became clearer than ever that these technologies would still need many more years of refinement and redundant safeguards before they can be truly trusted with people's lives. Ambitious timelines for fullself-driving adoption seem increasingly premature and misguided in the face of reality.The experience has also sparked deeper philosophical and ethical quandaries about ceding control to opaque black box systems. How do we adequately vet decision-making algorithms deployed at scale? What are the liability implications when the inevitable failures do occur? And perhaps most fundamentally, is the marcovian framework underlying most of today's AI even capable of grasping the richness of human contexts and value systems required to make truly intelligent choices?These are just some of the questions that must be earnestly grappled with going forward. While the future inevitably involves greater human-machine symbiosis, my harrowing encounter has reinforced the importance of maintaining meaningful human oversight and agency over technologies that hold such immense power over our lives.Technological progress must be tempered by a grounded sense of humility and a renewed investment in making these systems truly resilient and aligned with human values. There are surely many more white-knuckle lessons to be learned on the long and winding road ahead.。
关于绿色城市的英文报道作文Cultivating Sustainable Urban Oases: The Rise of Green CitiesIn an era defined by rapid urbanization and the pressing need to address environmental challenges, the concept of "green cities" has emerged as a beacon of hope. These urban centers, designed with sustainability at their core, are transforming the way we envision the future of our cities. From innovative urban planning strategies to the integration of nature-based solutions, green cities are redefining the relationship between human habitats and the natural world.At the heart of the green city movement is the recognition that traditional urban development models have often prioritized economic growth over environmental well-being. The consequences of this approach have been far-reaching, with cities contributing significantly to global greenhouse gas emissions, resource depletion, and the degradation of natural ecosystems. However, a growing number of municipalities are now embracing a more holistic approach, one that seeks to balance the demands of urban living with the preservation and enhancement of the natural environment.One of the key pillars of green city design is the integration of green spaces and natural elements throughout the urban landscape. This can take many forms, from the creation of expansive urban parks and gardens to the incorporation of green roofs, vertical gardens, and bioswales. These green infrastructure elements not only beautify the city but also provide a range of ecosystem services, such as air filtration, stormwater management, and urban cooling. By strategically placing these natural elements, green cities are able to mitigate the urban heat island effect, reduce the risk of flooding, and improve overall air and water quality.Beyond the physical integration of nature, green cities also prioritize the use of sustainable materials and energy-efficient technologies in their built environments. This can include the use of renewable energy sources, such as solar and wind power, as well as the implementation of energy-efficient building standards and the promotion of green construction practices. By reducing their reliance on fossil fuels and embracing clean energy solutions, green cities are taking significant strides towards reducing their carbon footprint and contributing to the global fight against climate change.Another crucial aspect of the green city model is the emphasis on sustainable transportation systems. Many green cities have invested heavily in the development of robust public transit networks,including extensive light rail, bus rapid transit, and metro systems. These clean and efficient modes of transportation not only reduce greenhouse gas emissions but also promote more active and healthy lifestyles by encouraging walking and cycling. Additionally, green cities often prioritize the creation of pedestrian-friendly streets and bike-sharing programs, further reducing the reliance on private vehicles and fostering a more livable and accessible urban environment.Underpinning the success of green cities is a deep commitment to community engagement and the empowerment of local residents. These cities recognize that true sustainability cannot be achieved without the active participation and support of the people who call them home. Through initiatives such as community gardens, urban agriculture projects, and environmental education programs, green cities are fostering a sense of environmental stewardship and encouraging citizens to take an active role in shaping the future of their communities.Moreover, green cities are not just about the physical transformation of the urban landscape; they also embrace a holistic approach to economic and social development. By promoting sustainable businesses, green job creation, and the equitable distribution of resources, these cities are working to ensure that the benefits of a green economy are accessible to all members of the community. Thisfocus on social and economic sustainability helps to address issues of inequality and ensures that the transition to a more sustainable future is inclusive and just.As the world continues to grapple with the challenges of urbanization and environmental degradation, the rise of green cities offers a glimmer of hope. These urban centers are not only redefining the way we think about city planning and design but also inspiring a global movement towards a more sustainable and livable future. By serving as models of innovation and environmental stewardship, green cities are paving the way for a more harmonious relationship between human settlements and the natural world.Of course, the journey towards creating truly green cities is not without its challenges. Barriers such as financial constraints, political resistance, and the inertia of established urban development practices can slow the pace of progress. However, the growing number of successful green city initiatives around the world demonstrates that these challenges can be overcome through a combination of visionary leadership, collaborative partnerships, and a steadfast commitment to sustainability.As we look to the future, it is clear that the concept of the green city will continue to evolve and expand. New technologies, emerging urban design strategies, and innovative approaches to communityengagement will all play a crucial role in shaping the next generation of sustainable urban centers. By embracing this vision of a greener, more livable future, we can create cities that not only meet the needs of their inhabitants but also contribute to the overall health and resilience of our planet.。
双摄像机协同人脸鹰眼检测与定位方法孙卓金;胡士强【摘要】现代视频监控系统需要获取大范围场景中感兴趣目标的清晰图像,这在目标距离较远并且不断移动时单纯采用摄像机调焦方式通常有一定的困难.为了获取宽范围监控场景中远距离行人的主要面部特征,采用广角静止一窄视场运动双摄像机协同工作方式可以同时获得远距离目标的全局和细节信息.首先采用改进的Codebook背景减法从广角摄像机中检测运动目标,然后指引运动摄像机近距离跟踪观察;若行人停止运动,则利用运动摄像机对其进行放大,然后从中检测人脸,并将人脸置于视野中心放大得到清晰图像.当行人再次运动时,广角相机将初始位置再次传递给运动摄像机,由其再对行人进行跟踪.通过实验室内和室外真实场景的实验表明,广角相机的检测算法具有一定的鲁棒性,运动相机能跟踪放大行人人脸图像,算法运行速度满足实时性要求.%The major aim for modem video surveillance is to capture clear imagery of Region of Interest (ROI) in large- scale and scene. However, it is hard to acTheve the goal by zooming when the ROI is far from the surveillance system. A station-motion camera system was designed to acquire the detailed and overall information of pedestrian at distance to obtain the facial feature. A stationary wide Field Of View (FOV) camera was used to monitor an environment for detecting pedestrians by improved Codebook background subtraction. Then a Pan-Tilt-Zoom (PTZ) camera was steered to track a target detected in the stationary camera, and zoomed in it speedily when pedestrian stopped moving. Afterwards, a face detection procedure used the images received from PTZ camera to obtain pedestrian face. Once the face was detected, PTZ camera put it inthe center of the FOV and zoomed to acquire high-definition image. While the pedestrian continued moving, PTZ camera received the location of pedestrian from the station camera and tracked it continually. The experiments on real indoor and outdoor environments show that the proposed wide FOV camera pedestrian detection algorithm is robust to illumination variations, and the PTZ camera can track and zoom in the face of pedestrians. Besides, the speed of the method meets the real-time requirement.【期刊名称】《计算机应用》【年(卷),期】2011(031)012【总页数】4页(P3388-3391)【关键词】码书;双摄像机协同;人脸检测;目标跟踪【作者】孙卓金;胡士强【作者单位】上海交通大学航空航天学院,上海 200240;上海交通大学航空航天学院,上海 200240【正文语种】中文【中图分类】TP391.410 引言现代社会安全领域越来越受重视,视频监控广泛应用到公共场所,如地铁站、停车场等,视频监控的重要目的是获取感兴趣区域的清晰图像。
目标检测参考文献目标检测是计算机视觉领域中的一个重要研究方向,主要目标是在图像或视频中识别和定位特定目标物体。
近年来,随着深度学习技术的兴起,目标检测取得了显著的进展,在许多实际应用中得到了广泛应用。
以下是一些关于目标检测的重要参考文献。
1. Viola, P., & Jones, M. (2001). Rapid Object Detection using a Boosted Cascade of Simple Features. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) (Vol.1, pp. I-511-I-518).这篇经典论文提出了基于级联AdaBoost算法的人脸检测方法,该方法将输入图像的特征与级联分类器相结合,实现了高效的目标检测。
这种方法为后续的目标检测方法奠定了基础,并被广泛应用于人脸检测等领域。
2. Dalal, N., & Triggs, B. (2005). Histograms of Oriented Gradients for Human Detection. In Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) (Vol.1, pp. 886-893).这篇论文提出了一种基于梯度方向直方图的特征表示方法,称为“方向梯度直方图”(Histograms of Oriented Gradients,简称HOG),并将其应用于行人检测。
HOG特征具有旋转不变性和局部对比度归一化等优点,在目标检测中取得了显著的性能提升。
CVPR 2013 录用论文(目标跟踪部分)/gxf1027/article/details/8650878完整录用论文官方链接:/cvpr13/program.php过段时间CvPaper上面应该会有正文链接今年有关RGB-D摄像机应用和研究的论文渐多起来了。
当然,自己还是比较关心Tracking方面的Papers。
从作者来看,一作大部分为华人,而且有不少在Tracking这个圈子里相当有名的牛,比如Ming-Hsuan Yang,Robert Collins等(中科院到阿大的Xi Li也是非常活跃,从他的论文可以看出深厚的数学功底,另外Chunhua Shen老师团队非常高产)。
此外,从录用论文题目初步判断,Sparse coding (representation)的热度在减退,所以Haibin Ling老师并没有这方面的文章录用,且纯粹的tracking-by-detection几乎不见踪影了。
以下是摘录的tracking方面的录用论文:Oral部分:Structure Preserving Object Tracking. Lu Zhang, Laurens van der MaatenTracking Sports Players with Context-Conditioned Motion Models. Jingchen Liu, Peter Carr, Robert Collins, Yanxi Liu Post部分:Online Object Tracking: A Benchmark. Yi Wu, Jongwoo Lim,Ming-Hsuan YangLearning Compact Binary Codes for Visual Tracking. Xi Li, Chunhua Shen, Anthony Dick, Anton van den HengelPart-based Visual Tracking with Online Latent Structural Learning. Rui Yao, Qinfeng Shi,Chunhua Shen, Yanning Zhang, Anton van den HengelSelf-paced learning for long-term tracking.James Supancic III, Deva Ramanan(long-term的噱头还是很吸引人的,和当年TLD一样,看看是否是工程的思想多一些)Visual Tracking via Locality Sensitive Histograms.Shengfeng He, Qingxiong Yang, Rynson Lau, Jiang Wang,Ming-Hsuan Yang (CityU of HK,使用直方图作为表观在当前研究背景下真是反其道而行之啊)Minimum Uncertainty Gap for Robust Visual Tracking. Junseok Kwon, Kyoung Mu Lee (VTD作者)Least Soft-thresold Squares Tracking. Dong Wang, Huchuan Lu, Ming-Hsuan YangTracking People and Their Objects. Tobias Baumgartner, Dennis Mitzel, Bastian Leibe(这个应该也有应用的背景和前景)(以上不全包括多目标跟踪方面的论文)其它关注的论文:Alternating Decision Forests. Samuel Schulter, Paul Wohlhart,Christian Leistner, Amir Saffari,Peter M. Roth,Horst Bischof(Forest也是近些年的热点之一。
英语作文有环境设施喜欢的地方I have always been fascinated by the natural world and the ways in which humans can coexist with the environment in a sustainable manner. One place that I particularly enjoy visiting is a small town nestled in the foothills of a mountain range not far from where I live. This town has made a concerted effort to implement a variety of environmental initiatives and facilities that have transformed it into a model of eco-friendly living.At the heart of the town is a sprawling public park that serves as the centerpiece of the community's environmental efforts. The park is home to a diverse array of native plant and animal species that thrive in the carefully maintained habitats. Winding trails wind through lush forests and alongside tranquil streams, providing ample opportunities for visitors to immerse themselves in nature and appreciate the beauty of the local ecosystem.One of the most impressive features of the park is its extensive network of solar-powered lighting and charging stations. Strategically placed throughout the trails and picnic areas, thesestations allow visitors to recharge their electronic devices using clean renewable energy. This not only reduces the town's carbon footprint but also encourages visitors to spend more time outdoors, disconnecting from the digital world and reconnecting with the natural world.In addition to the solar-powered infrastructure, the park also boasts a state-of-the-art waste management system that ensures all recyclable and compostable materials are diverted from landfills. Clearly marked bins for different types of waste are situated throughout the park, making it easy for visitors to sort their trash and minimize their environmental impact. The town has also implemented an educational campaign to raise awareness about the importance of recycling and composting, further encouraging sustainable practices among its residents and visitors.One of the most unique aspects of the park is its community garden, which serves as a hub for local food production and education. Residents can rent plots of land to grow their own fruits and vegetables, using organic farming techniques and rainwater harvesting systems to minimize the environmental impact of their cultivation efforts. The garden also hosts regular workshops and classes on topics such as sustainable gardening, composting, and healthy cooking, empowering community members to adopt more eco-friendly lifestyles.Beyond the park, the town has also made significant investments in other environmental initiatives, such as a robust public transportation system and a network of bike lanes and pedestrian-friendly streets. The local government has worked closely with community organizations to promote the use of alternative modes of transportation, reducing the town's reliance on fossil fuels and improving air quality. Additionally, many of the town's businesses have adopted sustainable practices, such as using renewable energy sources, implementing water conservation measures, and minimizing waste.What I find most compelling about this town is the way in which its environmental initiatives have become deeply integrated into the fabric of the community. Rather than being seen as a burden or a constraint, the town's eco-friendly policies and facilities are embraced by residents and visitors alike as a source of pride and a means of enhancing the overall quality of life. The town has managed to strike a delicate balance between preserving the natural beauty of the surrounding landscape and providing modern amenities and infrastructure that cater to the needs of its citizens.As someone who is passionate about environmental conservation and sustainable living, I find this town to be a shining example of what can be achieved when a community comes together toprioritize the health of the planet and the well-being of its inhabitants. The town's commitment to environmental stewardship is evident in every aspect of its design and operations, from the solar-powered lighting in the park to the composting facilities in local businesses. It is a place that truly embodies the idea of living in harmony with the natural world, and it serves as an inspiration for other communities to follow suit.Ultimately, what I love most about this town is the sense of community and shared purpose that permeates every aspect of its existence. The residents take pride in their town's environmental initiatives and actively participate in maintaining the delicate balance between human activity and ecological preservation. It is a place where people come together to celebrate the natural world, to learn from one another, and to work collectively towards a more sustainable future. In a world that is increasingly grappling with the challenges of climate change and environmental degradation, this town stands as a beacon of hope and a testament to the power of human ingenuity and collective action.。
英语作文-低碳交通发展规划In the quest for a sustainable future, the development of low-carbon transportation stands as a pivotal pillar in mitigating the effects of climate change. The transportation sector, a significant contributor to greenhouse gas emissions, is undergoing a transformative shift towards greener alternatives. This transition is not merely a technological overhaul but a comprehensive reimagining of mobility that aligns with environmental stewardship and economic efficiency.The cornerstone of low-carbon transportation is the electrification of vehicles. Electric vehicles (EVs), powered by clean electricity, offer a zero-emission alternative to traditional internal combustion engines. The proliferation of EVs is supported by advancements in battery technology, which have led to longer ranges and shorter charging times, making them increasingly viable for the average consumer. Moreover, the integration of renewable energy sources into the power grid ensures that the electricity used for these vehicles is also low-carbon.Another critical aspect is the optimization of public transportation. Efficient and accessible public transit systems reduce the reliance on personal vehicles, thereby decreasing traffic congestion and emissions. The expansion of urban rail networks, the introduction of electric buses, and the promotion of car-sharing schemes are examples of initiatives that contribute to a robust public transportation infrastructure.Active transportation modes, such as walking and cycling, are also essential components of a low-carbon transport strategy. The development of pedestrian-friendly urban spaces and dedicated cycling lanes encourages people to opt for these healthy and environmentally friendly modes of travel. These efforts not only reduce carbon footprints but also enhance the livability of cities.The role of policy and planning cannot be overstated in the advancement of low-carbon transportation. Governments and municipalities must enact policies that incentivize the adoption of green vehicles, support the expansion of public transit, and prioritize the development of infrastructure that facilitates active transportation. Financialincentives, such as subsidies for EV purchases and investments in public transit, are effective measures to accelerate the transition.Furthermore, the integration of technology plays a transformative role. Smart transportation systems that leverage data analytics and connectivity can optimize traffic flows, reduce idling, and improve the overall efficiency of transport networks. The advent of autonomous vehicles also holds the promise of further reducing emissions through more efficient driving patterns and the potential for shared vehicle services.In conclusion, the development of low-carbon transportation is a multifaceted endeavor that requires a concerted effort from technology, policy, and societal behavior. It is an investment in the health of our planet and the well-being of future generations. As we forge ahead, it is imperative that we embrace innovation, foster collaboration, and remain steadfast in our commitment to a greener, more sustainable mode of transportation. Through these efforts, we can envision a future where our journeys leave a lighter footprint on the earth. 。
行人检测现状转自/huixingshao/article/details/43793653/susongzhi/item/085983081b006311eafe 38e7行人检测具有极其广泛的应用:智能辅助驾驶,智能监控,行人分析以及智能机器人等领域。
从2005年以来行人检测进入了一个快速的发展阶段,但是也存在很多问题还有待解决,个人觉得主要还是在性能和速度方面还不能达到一个权衡。
1.行人检测的现状(大概可以分为两类)(1).基于背景建模:利用背景建模方法,提取出前景运动的目标,在目标区域内进行特征提取,然后利用分类器进行分类,判断是否包含行人;背景建模目前主要存在的问题:(背景建模的方法总结可以参考我的前一篇博文介绍)(前景目标检测总结)必须适应环境的变化(比如光照的变化造成图像色度的变化);相机抖动引起画面的抖动(比如手持相机拍照时候的移动);图像中密集出现的物体(比如树叶或树干等密集出现的物体,要正确的检测出来);必须能够正确的检测出背景物体的改变(比如新停下的车必须及时的归为背景物体,而有静止开始移动的物体也需要及时的检测出来)。
物体检测中往往会出现Ghost区域,Ghost区域也就是指当一个原本静止的物体开始运动,背静差检测算法可能会将原来该物体所覆盖的区域错误的检测为运动的,这块区域就成为Ghost,当然原来运动的物体变为静止的也会引入Ghost区域,Ghost区域在检测中必须被尽快的消除。
(2).基于统计学习的方法:这也是目前行人检测最常用的方法,根据大量的样本构建行人检测分类器。
提取的特征主要有目标的灰度、边缘、纹理、颜色、梯度直方图等信息。
分类器主要包括神经网络、SVM、adaboost以及现在被计算机视觉视为宠儿的深度学习。
统计学习目前存在的难点:(a)行人的姿态、服饰各不相同、复杂的背景、不同的行人尺度以及不同的关照环境。
(b)提取的特征在特征空间中的分布不够紧凑;(c)分类器的性能受训练样本的影响较大;(d)离线训练时的负样本无法涵盖所有真实应用场景的情况;目前的行人检测基本上都是基于法国研究人员Dalal在2005的CVPR发表的HOG+SVM的行人检测算法(Histograms of Oriented Gradients for Human Detection, Navneet Dalel,BillTriggs, CVPR2005)。
组织一场和城市交通有关的辩论的英语作文Organizing a Debate on Urban TransportationIn today's rapidly growing cities, the issue of urban transportation has become increasingly complex and multifaceted. As populations expand and reliance on personal vehicles increases, cities around the world are grappling with challenges such as traffic congestion, air pollution, and the need for sustainable mobility solutions. Organizing a debate on this topic presents an excellent opportunity to engage the community, explore different perspectives, and work towards finding effective strategies to address these pressing concerns.The first step in organizing a successful debate on urban transportation would be to assemble a diverse panel of experts and stakeholders. This could include city planners, transportation engineers, environmental advocates, public transit authorities, and representatives from the business community and local government. By bringing together individuals with different areas of expertise and varying interests, the debate can cover a wide range of issues and provide a comprehensive understanding of the complexities involved.One of the key aspects to address would be the role of public transportation in alleviating traffic congestion and reducing carbon emissions. Proponents of improved public transit systems could argue that investing in reliable and accessible bus, train, and light rail networks can encourage more people to leave their personal vehicles at home, leading to reduced traffic and improved air quality. They could highlight successful case studies from cities that have implemented comprehensive public transportation systems, such as London's expansive underground network or the extensive light rail system in Portland, Oregon.On the other hand, opponents of this approach may contend that public transportation is not a one-size-fits-all solution and that it may not be practical or cost-effective in all urban settings. They could argue that the high initial investment required to build and maintain public transit infrastructure may not be justified in smaller or more spread-out cities, where the demand for such services may be lower. Additionally, they could emphasize the importance of individual freedom and the convenience of personal vehicles, suggesting that policies aimed at discouraging car use may be met with public resistance.Another crucial aspect to consider would be the role of active transportation, such as walking and cycling, in promoting sustainableand healthy urban mobility. Advocates for pedestrian and bicycle-friendly infrastructure could argue that by creating safer and more inviting environments for non-motorized modes of transportation, cities can reduce their reliance on cars, improve public health, and foster a greater sense of community. They could highlight examples of cities like Copenhagen and Amsterdam, where extensive bike lane networks and pedestrian-oriented urban design have led to a significant shift away from car-centric transportation.Opponents of this approach, however, may argue that active transportation is not a feasible solution for all urban residents, particularly those with mobility challenges or who live in areas with challenging topography. They could also raise concerns about the potential conflicts between pedestrians, cyclists, and motorists, and the need for robust safety measures to ensure the well-being of all road users.The debate could also explore the role of emerging technologies, such as electric vehicles, autonomous cars, and ride-sharing services, in shaping the future of urban transportation. Proponents of these technologies could argue that they have the potential to reduce emissions, improve traffic flow, and provide more equitable access to mobility. They could point to the success of electric vehicle adoption in cities like Oslo, Norway, and the potential benefits of autonomous vehicles in reducing accidents and providing transportation optionsfor the elderly and disabled.Opponents, on the other hand, may express concerns about the potential job losses associated with the automation of transportation, the high costs of implementing new technologies, and the potential privacy and security risks associated with the collection and use of data by these systems.Ultimately, the goal of the debate would be to encourage a thoughtful and nuanced discussion on the complex challenges facing urban transportation, and to work towards identifying sustainable and equitable solutions that address the needs of all stakeholders. By bringing together diverse perspectives and encouraging open dialogue, the debate can serve as a catalyst for policy changes, community engagement, and the development of innovative transportation strategies that can improve the quality of life for urban residents.。
在城市开车的影响英语作文Driving in cities has become an integral part of modern life. It provides us with the convenience and flexibility to navigate urban areas efficiently, allowing us to access various destinations quickly. However, the impact of driving in cities extends far beyond just personal mobility. It has significant implications for the environment, public health, and the overall quality of life in urban centers. In this essay, we will explore the multifaceted effects of driving in cities and the need for sustainable solutions.One of the primary concerns regarding driving in cities is the impact on the environment. Vehicles powered by internal combustion engines emit a substantial amount of greenhouse gases, contributing to the global issue of climate change. The burning of fossil fuels by these vehicles releases carbon dioxide, nitrogen oxides, and other pollutants into the atmosphere, which can have detrimental effects on air quality and the overall ecosystem. This pollution not only contributes to the warming of the planet but also poses a direct threat to the health of urban residents, particularly those with respiratory conditions.Furthermore, the reliance on private vehicles in cities has led to the construction of extensive road networks and parking infrastructure, which can have significant consequences for the environment. The expansion of roads and the paving of land for parking lots often result in the destruction of natural habitats, reducing the amount of green spaces and urban biodiversity. This can lead to a decline in the overall ecological balance and the loss of important ecosystem services that cities rely on, such as carbon sequestration, air purification, and water filtration.Another significant impact of driving in cities is the issue of traffic congestion. As more people opt for private vehicles, the number of cars on the roads increases, leading to gridlock and longer commute times. This not only wastes time and productivity but also contributes to increased fuel consumption and higher levels of air pollution. Traffic congestion can also have indirect effects on the economy, as it can hinder the efficient movement of goods and services, and lead to lost productivity and economic opportunities.The prevalence of driving in cities also has implications for public health. The sedentary nature of driving, combined with the lack of physical activity, can contribute to the rise of obesity, cardiovascular diseases, and other health problems. Additionally, the exposure to air pollution from vehicle emissions can lead to respiratory issues, suchas asthma and lung diseases, as well as an increased risk of cancer and other chronic illnesses.Moreover, the dominance of private vehicles in urban areas can have a detrimental impact on the social fabric of cities. The reliance on cars can isolate individuals, reducing opportunities for social interaction and community engagement. It can also create barriers for those who do not have access to a private vehicle, such as the elderly, the disabled, and low-income individuals, limiting their mobility and access to essential services and resources.In response to these challenges, many cities around the world have begun to implement various strategies to promote sustainable and environmentally-friendly transportation options. One such approach is the development of robust public transportation systems, which can include buses, trains, and light rail. By providing reliable and affordable alternatives to private vehicles, these systems can help reduce the number of cars on the roads, alleviate traffic congestion, and improve air quality.Another strategy is the promotion of active transportation modes, such as walking and cycling. By creating pedestrian-friendly infrastructure, including sidewalks, bike lanes, and public spaces, cities can encourage residents to engage in physical activity and reduce their reliance on cars. This not only benefits the environmentbut also contributes to improved public health and a more vibrant, livable urban landscape.Furthermore, cities are exploring the potential of emerging technologies, such as electric vehicles and autonomous driving, to address the environmental and transportation challenges posed by traditional internal combustion engine vehicles. The adoption of electric vehicles can significantly reduce greenhouse gas emissions and improve air quality, while autonomous driving has the potential to optimize traffic flow and reduce congestion.However, the implementation of these sustainable transportation solutions requires a comprehensive and collaborative approach. It involves the coordination of various stakeholders, including policymakers, urban planners, transportation authorities, and the public. Effective policies and regulations, coupled with investments in infrastructure and public awareness campaigns, are essential to drive the transition towards more sustainable urban mobility.In conclusion, the impact of driving in cities is multifaceted and far-reaching. While private vehicles provide convenience and flexibility, the environmental, public health, and social consequences of this reliance cannot be ignored. By embracing sustainable transportation solutions, cities can work towards creating more livable, equitable, and environmentally-friendly urban environments. This shift requiresa collective effort from all stakeholders to reimagine and restructure the way we move within our cities, ultimately leading to a more sustainable and resilient future.。
Towards Robust Pedestrian Detection in Crowded Image Sequences Edgar Seemann,Mario Fritz and Bernt SchieleTU Darmstadt,Germany{seemann,fritz,schiele}@mis.tu-darmstadt.dermatik.tu-darmstadt.deAbstractObject class detection in scenes of realistic complexity remains a challenging task in computer vision.Most recent approaches focus on a single and general model for object class detection.However,in particular in the context of im-age sequences,it may be advantageous to adapt the general model to a more object–instance specific model in order to detect this particular object reliably within the image se-quence.In this work we present a generative object model that is capable to scale from a general object class model to a more specific object–instance model.This allows to detect class instances as well as to distinguish between individual object instances reliably.We experimentally evaluate the performance of the proposed system on both still images and image sequences.1.IntroductionThe ability to detect objects and pedestrians in still im-ages and image sequences is key to a variety of important applications such as surveillance,image and video index-ing,intelligent vehicles or robotics.Most research in the area has focused on approaches to effectively model intra–class variation to generalize well across object class in-stances.Tremendous progress has been made for object as well as pedestrian detection[19,25,5,11,9,24,17,15,10, 4,13,21].An open problem,however,is to detect multiple objects and pedestrians in crowded scenes where pedestrians might be significantly occluded over longer periods of time.Tra-ditionally,approaches in this area have been formulated as tracking problems[6,27,26,23,18]due to the impor-tance of temporal consistency.Quite interestingly many approaches in this area rely on simple object and pedes-trian models(i.e.color histograms)suggesting that their ef-fectiveness mostly comes from sophisticated Bayesian and temporal inference mechanisms or from the use of multiple cameras[3,2,16,7,14]This paper follows a quite different route by startingfrom a general pedestrian detection model[10]that is capa-ble to detect and segment pedestrians in images of crowded scenes.In order to handle significant occlusion over longer periods of time we aim to re-detect individual pedestrians previously seen within an image sequence.For this,the general pedestrian model is specialized for the detection of individual pedestrians.As we will discuss below the spe-cialized models leverage e.g.from the segmentation abil-ity of the general pedestrian model to reason about partial occlusions in crowded scenes.These individualized pedes-trian models taken together with a simple temporal continu-ity model allow then to effectively detect multiple pedestri-ans in crowded scenes despite significant and longer partial occlusions.The main contribution of the paper therefore is a unified object model that is scalable from a general object–class model to a more specialized and even individualized object–instance model.To learn a robust and accurate model of an individual object or pedestrian from a small number of de-tections is clearly challenging.Rather than to learn a model for each individual pedestrian from scratch we specialize the general pedestrian model to the individual.The individ-ualized models thereby preserve properties of the more gen-eral model such as the segmentation ability and the general pedestrian appearance codebook.To achieve this wefirst (section3)extend the original ISM-approach[9]e.g.by in-corporating codebook priors.This enables robust learning and the specialization of individualized pedestrian models from few training samples.The proposed extensions are experimentally compared to previous published results on challenging data–sets showing the validity of the approach.Section4then applies this approach to image sequences with significant occlusion over longer periods of time.Ex-perimental results show that these specific models can be used to increase both precision and recall of the detection.2.Original Implicit Shape Model ApproachThe Implicit Shape Model(ISM)developed by Leibe and Schiele[9]is a generative model for general object detection.It has been applied to a variety of object cat-1egories including cars,motorbikes,cows and pedestrians. For pedestrians a number of extensions[10,21,22]have been proposed,which exploit the nature of this object cat-egory by incorporating knowledge about pedestrian articu-lation.In this paper we propose to improve and extend the probabilistic modeling of the original approach.As will be shown later this allows not only to train general models for pedestrian detection but also to train specialized models for individual pedestrians.Before introducing the extensions in section3,the following briefly explains the steps involved in learning an object model in the original ISM framework.Appearance Codebook.A visual vocabulary,referred to as appearance codebook,is used to describe common object features or parts of an object class.To learn the appearance codebook a scale-invariant interest point detec-tor(Hessian-Laplace[12])is applied to each training im-age and local image descriptors(Shape-Context[1,12])are extracted around them.These image descriptors are subse-quently clustered with an agglomerative clustering scheme.Spatial Occurrence Distributions.Once an appear-ance codebook is learned for an object class,separate spa-tial occurrence distributions for each codebook entry c i are learned.During a second run over the training images the codebook is matched to the training examples and occur-rence locations(x-,y-position,scale)for the codebook en-tries are recorded.Recognition.During recognition,the same feature ex-traction procedure is applied to obtain a set of local image descriptors e at various scales on the test image.A local image descriptor e extracted at the absolute image coordi-nates is compared to the appearance codebook.A descrip-tor may have multiple matching codebook entries c i.Let p(c i|e)denote the matching probability.For each possi-ble codebook match votes are cast for different object cen-tersλx,λy and scalesλσaccording to the individual occur-rence distributions withλ=(λx,λy,λσ).Each vote has the weight P(o n,λ|c i, )·p(c i|e).A descriptor’s contribu-tion to the recognition process can therefore be expressed by the following marginalization:P(o n,λ|e, )= c i P(o n,λ|c i,e, )p(c i|e)(1)= c i P(o n,λ|c i, )p(c i|e)(2) The overall object probability at positionλis obtained by summing over all extracted descriptors e k:P(o n,λ)= k P(o n,λ|e k, k)(3) Maximum search is accomplished by Mean-Shift Mode Estimation with a scale-adaptive kernel K[8].Segmentation and MDL-based verification.Next to object localization,a pixel-wise segmentation can be in-ferred for each hypothesis.Finally,a Minimum Description Length(MDL)based verification step is applied in order to disambiguate overlapping hypotheses.As has been previ-ously shown the segmentation step in combination with the MDL selection mechanism significantly improves detection performance and allows a pixel–level reasoning about oc-clusions.For the computational details please refer to[9].3.Extensions to the Implicit Shape ModelThe Implicit Shape Model has shown state-of-the-art performance for the detection of pedestrians in images of crowded scenes.In order to further improve its perfor-mance we extend the ISM formulation in various ways. The general aim is to derive a unified ISM formulation that on the one hand allows to train a general pedestrian detec-tion model and on the other hand to specialize this general pedestrian model to enable robust detection of individual pedestrians in image sequences.More specifically this sec-tion introduces a novel probabilistic modeling scheme as the basis for the unified formulation and discusses the nec-essary steps to make detection more reliable and to better exploit the information available in the training data.As will be seen in the experiments this allows not only to train individualized pedestrian models but also also improves the results for general pedestrian detections w.r.t.the original ISM formulation.Appearance Codebook.As in the original approach we learn an appearance codebook by agglomerative clustering. To only keep codebook entries that are representative for an object class such as pedestrians we discard codebook en-tries,which very rarely match to the training images.Global Spatial Occurrence Distribution.Instead of learning individual occurrence distributions for each code-book entry separately as in the original ISM,we propose to learn a joint occurrence distribution for the entire appear-ance codebook.A joint occurrence distribution has several advantages.Firstly,individual occurrence distributions as used by Leibe&Schiele assume,that all codebook entries are equally important.However,there are some codebook entries which are more typical for an object class than oth-ers.For cars,for example,wheel features are crucial for reliable detection.Secondly,even if a codebook entry occurs frequently enough on pedestrians in general,when training an indi-vidualized pedestrian model from a small number of train-ing samples we may have insufficient statistics resulting in degenerated occurrence distributions.For example,a code-book entry occurring only(even several times)at a single location in the training data,concentrates its entire proba-bility mass on a single point.During recognition this can have the effect,that an hypothesis is dominated by such a degenerated distribution,yielding false positive detections with very high scores.This effect is particularly likely whentraining from as few as2or3training images as done below.Finally,a joint occurrence distribution ensures that the model is normalized as a whole instead of separately for each codebook entry.As a result the recognition scores be-come more comparable across(individual or separate)ob-ject models.The following explains in more detail how the global occurrence distributions P occ(o n,λ,c i)are learned on the training set.P occ(o n,λ,c i)denotes the probability,that codebook entry c i is observed and that the object center is at positionλrelative to codebook entry c i.For the deriva-tion we assume that a general pedestrian appearance code-book C=(c1,...,c n)has been learned on a set of pedes-trian images.The occurrence distributions themselves can be learned on the same set of pedestrian images(to obtain a general pedestrian model)or on an independent set of im-ages for example from an individual pedestrian.Let e be an image descriptor extracted at location−λon the training set(the location is recorded with respect to the object center).We compare e to each of the codebook en-tries,with p(c i|e)denoting probability that e is associated with entry c i.The descriptor’s contribution to the global occurrence distribution is distributed over the codebook di-mension of P occ(o n,λ,c i)according to the probabilities p(c i|e).As a result,each training feature has the same in-fluence on thefinal object model.The model therefore can-not be dominated by rare occurrences or even outliers and learns a better representation of the mean structure of the object class.Additionally,the frequency information of the codebook entries is retained in the model.One can also think of this process as the introduction of priors p(c i)for each codebook entry(P occ(o n,λ,c i)=P(o n,λ|c i)p(c i)). Where the priors are determined by the occurrence fre-quency.Note,that the codebook priors are learned on the individual training samples whereas the codebook entries themselves may be learned on a larger set of pedestrian im-ages.Recognition.After having learned the new object model M=(A,P occ)consisting of an appearance codebook A and a global occurrence distribution P occ,we apply the same scale-invariant interest point detector on a test image to obtain local image descriptors at various scales.Again,let e denote a local image descriptor extracted at the absolute image coordinates .Image descriptor e is matched to each entry of our appearance codebook.For each matching codebook entry or object part we cast votes for different object centersλin a continuous3D voting space according to the recorded global occurrence distribu-tion P occ.We refer to the set of codebook entries matching to image descriptor e as M(e).The contribution of a descriptor e can then be expressed by the following equation:P(o n,λ|e, )= c i P occ(o n,λ,c i|e, )(4)= c i∈M(e)P occ(o n,λ,c i| )(5)= c i∈M(e)P occ(o n,λ− ,c i)(6) Note,that a vote P occ(o n,λ− ,c i)represents evidence for a certain codebook entry c i to be present at location in the test image.The are two main differences to the orig-inal ISM recognition procedure.As pointed out before we use the joint occurrence distribution P occ rather then the individual codebook distributions.The other difference is that previously each feature’s contribution was distributed across multiple codebook entries based on the matching probability p(c i|e).Whereas here the features activate all matching codebook entries completely.As we will see be-low these differences will result in a better detection perfor-mance of the general pedestrian detection model and will also allow to train individualized pedestrian models.The object probability for a locationλin the test image can then be computed by:P(o n,λ)= k P(o n,λ|e k, k)(7)= kc i∈M(e k)P occ(o n,λ− k,c i)(8)Thus,an object hypothesis is the summation of proba-bilistic votes pointing to the same object center.Since each vote represents evidence for a codebook entry,a hypothe-sis can be considered to be a collection of codebook entries appearing at certain positions in the test image.For the maximum search we use Mean-Shift Mode Esti-mation with a scale-adapted kernel volume.ˆp(o n,λ)=1nh(λ)dkjp(o n,λj|e k, k)K(λ−λjh(λ))(9)The kernel volume h is chosen in a way,that detection is robust to center point variations in the training and test data.However,integrating over the kernel volume has the effect that parts of the object model are explained multi-ple times.This happens,when similar local appearances are found very close to one another(with respect to x-,y-coordinates and scale)in the test image.Consider,for ex-ample,a codebook entry c i occurring at position−λin the training set.If two descriptors e1and e2with similar ap-pearance are found close to one another at 1and 2in the test image,their contributions P occ(o n,λ− 1,c i)and P occ(o n,λ− 2,c i)will be in the same kernel volume.Toavoid that the contribution is accounted for multiple times we remove the redundant evidence from the kernel volume during the maximum search.The remaining votes are then back-projected to the im-age and a pixel-wise segmentation mask is inferred for the object hypothesis.Note that this can be done only because we use a general pedestrian codebook for which we have learned the respective segmentations as well.When we have a few detections for an individual pedestrian in an image sequence and we want to learn a specialized model for this individual we cannot assume highly accurate seg-mentations for those few training samples.Therefore we leverage from the segmentation ability of the general pedes-trian appearance codebook to be used for the individualized models.As will be seen later this leads to segmentations for individuals that can be used again for pixel-wise occlu-sion reasoning.Finally we apply the MDL based verifi-cation stage to disambiguate overlapping object detections. Note also,that thefinal MDL verification step helps to de-correlate the influences of overlapping descriptors.3.1.Evaluation of the General Object ModelIn order to evaluate the proposed extensions to the object model,we applied it to three challenging pedestrian data sets.These data sets range from single-person side-view images to multi-person and multi-viewpoint detection in the presence of clutter and occlusion.On test set A we evaluate the detection performance, when people are fully visible.This test set contains181 side-views of pedestrians in different articulations and with a wide range of different clothing(see images in the lower left corner of Figure1).Figure1(upper left corner)depicts the obtained result. We compare the new approach both to the results of the original authors,as well as to the Histogram of Oriented Gradients(HOG)detector of Dalal&Triggs[4],which is based on a global descriptor instead of local image features.As can be seen,on this data set the HOG detector per-forms rather poorly with an equal error rate(EER)of only 57%.This is,on the one hand,due to the fact,that the data set is quite challenging.On the other hand,a pre-trained binary of the HOG detector was used,which was optimized for multi-viewpoint detection.The original ISM approach achieves an EER of74%.Our new model(red curve),which is based on a global occurrence distribution outperforms these results by14%and reaches an EER of 88%.This is a significant improvement and shows the po-tential of the novel object model.Note,that various extensions have been proposed that can further improve the performance of the original ISM approach.The4D-ISM[21]yields an EER of85%and Cross-Articulation Learning[22]89%respectively.How-ever,these extensions exploit explicit knowledge about pedestrian articulations.When using the Cross-Articulation Learning approach on top of the newly proposed object model,we can further improve the results.However,the improvement is less significant and performance improves only slightly in terms of precision compared to[22].As we want to use the proposed extensions to learn from as few as2or3training samples it is not clear how these further extensions could be incorporated to train individualized de-tection models.To confirm the suitability of the new object model in the presence of significant occlusions and overlapping people, we use the Crowded Scene data set from[10].We refer to the data set in this paper as test set B.It contains206im-ages with a total number of595annotated pedestrians.On this data the HOG detector attains a recall of60%for a pre-cision value of75%.Note,that a higher recall could not be achieved with this detector,since the provided binary has afixed confidence threshold.The original ISM approach yields an EER of73%.The newly proposed method in-creases the detection performance to82%.Again,this is a significant improvement.On this data set we even out-perform the the more elaborate Cross-Articulation Learning approach from[22],which achieves only an EER of81%.Finally,we tested the new approach on the multi-viewpoint test set C[21].This test set includes not only overlapping and occluded pedestrians,it also shows them from different viewpoints.The total number of images is 279with847annotated pedestrians.The HOG detector and the original ISM approach have similar detection per-formance,with the HOG detector performing better in the first part of the curve and slightly worse in the second part. The original ISM’s EER is at74%.The new approach per-forms significantly better along the whole precision-recall curve,achieving an EER of close to80%.It also outper-forms the4D-ISM approach[21].As our experiments have shown,the new object model yields significant performance improvements compared to the original ISM model on a variety of databases.The orig-inal ISM-approach has been improved by explicitly incor-porating articulation information.These improvements re-sulted in the best results so far reported on these databases. Interestingly,the novel approach proposed in this paper achieves comparable or even superior detection rates with-out the explicit use of articulation information.Considering the difficulty of the databases this shows that the new model makes pedestrian detection very robust even in the presence of clutter and occlusion.In order to stress this,Figure1 (second row)depicts some example detections of our sys-tem.The following section now discusses that the novel ap-proach lends itself to robustly learn specialized pedestrian model from as few as2or3training samples.(a)Test SetA (b)Test SetB (c)Test SetCFigure 1.First row:Recall precision curves,which compare our detection performance to results from other approaches for the different test sets.Second row:Example detections for test set A ,B and the multi-viewpoint data set C .4.Specialized Object ModelsThe proposed probabilistic formulation of the object model enables new possibilities and applications.In this section,we will explain,how a general object model can be specialized to a single pedestrian instance.Hereby,we are able to leverage from both the general pedestrian appear-ance codebook and the segmentation abilities of the general model.In the first step,however,we would like to experimen-tally verify,that,given the new formulation,learning from a small number of training examples is feasible.4.1.Varying Training Set SizeIn this experiment we gradually decrease the number of training examples.From the 200original images in the training set,we first randomly select a set of 50and finally a set of only 10pedestrians.We compare the obtained results to those of the original ISM approach.Figure 2depicts the corresponding results.As can be seen,the recognition rates drop when the number of train-ing examples is reduced.The EER for 200training exam-ples is 89%,for 50training examples 77%and 62%for 10examples.Of course,this was to be expected.However,even with as little as 10training examples (blue curve in Figure 2),we achieve reasonable detection results with ournew model.The detection recall at the EER always exceeds 60%,which is remarkable.In fact,it is even better than the state-of-the-art HOG pared to the original ISM approach,the detection precision is improved signifi-cantly when learning from only 10examples.This can be explained by the codebook priors,which successfully re-duce the influence of degenerated occurrence distributions.At a recall level of 50%,85%of our hypotheses are cor-rect,while for the original ISM this is only true for approx-imately 50%.4.2.Instance-Specific ModelsGeneral object detectors model a complete object cat-egory.That is why they have to be able to cope with large intra–class variations.Instance-specific models,on the other hand,can be directly adapted to an individual ap-pearance.Their focus is it to successfully detect the same instance,as well as to distinguish it from different pedes-trian instances.Thus,it is possible to re-detect pedestrians in an image sequence despite significant and longer partial occlusions.One can imagine to train a specialized model for an individual when a sufficient number of training sam-ples is available.However,the real challenge is to learn a specialized model from as few training instances as possi-ble.Let us now consider the details involved in deriving aFigure2.Recognition performance for varying number of training images on side-view pedestrians(Test Set A).specific model for a detected pedestrian,based on detec-tion hypotheses from the generic object model.A detection hypothesis H is obtained by a summation over the contri-butions of the individual image descriptors(see equation8). We can rewrite equation8in the following manner:P(o n,λ)= kc i∈M(e k)P occ(o n,λ− k,c i)(10)=(c i, k)∈HP occ(o n,λ− k,c i)(11)where(c i, k)are pairs with c i matching to the test image descriptor e k at position k.The equation expresses that an object hypothesis is a sum of samples from the global occurrence distribution P occ. The samples are drawn from the object parts c i which have been found in the test image.In other words,the terms P occ(o n,λ− k,c i)in equation8denote the probability, that codebook entry c i is observed at positionλ− k in the test image.We can now consider the test hypothesis to be another training example.We know its object center(λ)and we know which codebook entries have occurred at which po-sition on this object instance.This information is enough to build a new object model for exactly this hypothesis.In fact,we can reuse the samples from general object model directly to derive a specialized occurrence distribution P sP s(o n, k−λ,c i)=p(c i)Zk∈HP occ(o n,λ− k,c i)where Z is a normalization factor,which ensures,that the occurrence distributionof the instance-specific model is normalized.Figure3.The general occurrence distribution(left)contains con-tributions from various people and articulations.(The blue cir-cles denote codebook entry occurrences relative to the object cen-ter).The instance-specific model(right)is a subset of the general model.Figure3illustrates the process.A sub-part of the general occurrence distribution(visualized by the blue circles)is used as the new specific occurrence distribution.In this manner,instance-specific object models can be created from the general object model on-the-fly.The re-sulting models benefit from the general model in two ways. Firstly,they are based on the same general appearance code-book.Secondly,they inherit the segmentation abilities of the general model.Intuitively,the general model has for each codebook entry occurring at a certain location an asso-ciated segmentation mask(for further details please refer to [9]).These masks can be passed to the specific model,thus allowing to build a top-down segmentation from hypothe-ses of the specific model.Figure4shows example seg-mentations for a specific pedestrian model,which was ini-tialized when the person was standing upright with closed legs(rightmost image).As can be seen,the person can be successfully detected in subsequent frames.However,as the appearance of the person changes due to articulations changes,the detection relies almost exclusively on features on the upper body when the leg articulations change signif-icantly(middle).When a similar leg articulation as during training appears the detection can again use the respective leg parts(left).The inclusion of articulation information therefore has the potential to improve also the detection of individual pedestrians which is left for future research. 4.3.Instance-Specific EvaluationIn order to show the principal effectiveness of instance-specific object models,we learn5different object-specific models for the pedestrian present in an image sequence. The image sequence is recorded in a challenging setting, with pedestrians entering from the left and right sides and crossing each other.Therefore most pedestrians are par-tially or even completely occluded during the sequence. Figure5shows some example images from the image se-quence.For ground-truth annotations of the sequence,a pedestrian was considered,when it was approximately20% visible.As instance-specific models are sensitive to artic-ulation changes,we initialized the models from three gen-eral detections in consecutive image frames.This ensures,Figure4.Example segmentations for an instance-specific objectmodel.The model was initialized when the legs were closed,thusdetection of the legs fails,when the person is making a large step.However the evidence,from the upper body is sufficient to suc-cessfully detect theperson.Figure5.Example detections on a challenging image sequence,with cluttered background and heavy occlusion.Upper left:4of4pedestrians detected.Upper right:4of5pedestrians detected.Lower left:3of4pedestrians are detected.Lower right:4of5pedestrians detected.that the resulting models are more robust w.r.t different bodyposes.As an initial test,we computed how well the specificmodels are able to distinguish the different persons in thesequence.Therefore,we applied each instance model sep-arately to each frame in the image sequences.Then,foreach ground-truth annotation,we determine by which ob-ject model the pedestrian was explained best.Table1showsthe respective results in terms of a confusion matrix.Person P2is,for example,recognized in72%of thecases by the correct model,while in7%of the cases the ob-ject model from person P3yielded better detection scores.Considering that we have a5class problem and that pedes-trians are often partly occluded,these are respectable re-P1P2P3P4P5P168%14%−7%11%P214%72%7%−7%P312%−88%−−P412%4%4%72%8%P59%13%13%9%55%Table1.Confusion matrix between5instance-specific object mod-els on an image sequence with significant partial occlusion.sults.Please also keep in mind,that our models shouldnot only achieve good detection precision,but also a highrecall.Our evaluation has shown that,even though the spe-cific models are learned from only three consecutive frames,they areflexible enough to successfully detect the person al-most throughout the sequence.Finally,we show how the instance-specific models canbe used to increase detection robustness.Since we can accu-rately distinguish individual persons in the image sequence,it is possible to follow a detection hypothesis based on itslocation even when people are overlapping.When a personis fully occluded,we can recover based on its specific ap-pearance as soon as it becomes visible again.With a generalobject model alone,this would not be feasible.When performing object detection with the generic ob-ject model on the described pedestrian sequence,an EERof79%is reached(see Figure6).The maximum recall isapproximately82%.Note,that this value is rather low aspedestrians are sometimes80%occluded.When follow-ing pedestrian hypotheses based on the5instance–specificmodels,a recall of71%is obtained.However,the detectionprecision is significantly improved,with only two false pos-itive detections at over70%recall(yellow curve).In orderto obtain higher recall values the instance–specific modelsare not general enough.Fortunately,the specific object models tell us exactly,which path a person has taken.Thus,we canfill the missingdetections with detections from the general object model.In this way we combine,instance-specific detections andgeneral detections.As can be seen in Figure6(red curve)this can increase detection recall to a value of86%.Conse-quently a combined detection system based on generic andspecific object models can successfully improve detectionresults on image sequences.Figure5shows some exampledetections of thefinal system.A video of the system will beavailable through the supplementary material.5.ConclusionWe presented a unified probabilistic formulation for amodel that is scalable from general object–class detection tospecific object–instance detection.Our experimental eval-uation on different pedestrian image databases has shown,that the general object detection performance has been sig-。