下载文档原格式
前言：最近由于工作的关系，接触到了很多篇以前都没有听说过的经典文章，在感叹这些文章伟大的同时，也顿感自己视野的狭小。
想在网上找找计算机视觉界的经典文章汇总，一直没有找到。
失望之余，我决定自己总结一篇，希望对 CV领域的童鞋们有所帮助。
由于自
己的视野比较狭窄，肯定也有很多疏漏，权当抛砖引玉了
1990年之前
1990年
1991年
1992年
1993年
1994年
1995年
1996年
1997年
1998年
1998年是图像处理和计算机视觉经典文章井喷的一年。
大概从这一年开始，开始有了新的趋势。
由于竞争的加剧，一些好的算法都先发在会议上了，先占个坑，等过一两年之后再扩展到会议上。
1999年
2000年
世纪之交，各种综述都出来了
2001年
2002年
2003年
2004年
2005年
2006年
2007年
2008年
2009年
2010年
2011年
2012年。
基于ARM Cortex-A53的图像处理系统研究摘要：随着科技的不断进步，数字图像处理的应用已经普及到千家万户，使人们的生活变得越来越便捷。
早期实现的图像处理系统均是基于桌面PC机，而其图像处理系统存在较多约束，主要依赖于软件操作，且便携性较差。
本文采用基于Cortex-A53架构的S5P6818处理器以及Linux实时操作系统，构建实时图像处理系统，该系统能够进行实时图像采集、图像处理以及运动目标检测与跟踪等功能，同时可以实现联网，易于携带。
论文在图像处理系统及图像处理算法方面的研究工作如下：（1）根据系统功能以及开发成本的需求，对各硬件模块选型。
系统以X6818bv3开发板为核心连接外围硬件模块，构建系统硬件平台。
（2）图像处理系统移植了uboot、Linux操作系统、根文件系统以及视频服务器等软件，建立了系统运行软件平台，保证系统应用程序的正常执行，实现图像采集和图像处理的功能。
（3）图像处理系统的应用程序可分为图像采集和图像处理两部分。
在Qt Creator 集成开发环境中编写图形界面程序，实现视频图像显示、jpeg图像保存、图像增强以及边缘检测等功能，并且增添了远程监控功能。
（4）本文针对传统的ViBe算法在前景检测过程中存在光照变化敏感以及检测率较低等问题，提出了基于ViBe算法的改进算法，并选取样本进行对比，算法在视频图像运动目标的检测过程中能够有效解决检测率低的问题，且对不同的光照强度具有较好的抑制性。
图像处理系统包括软硬件平台的搭建以及应用程序的设计。
论文的研究内容为基于ARM图像处理系统平台的搭建及其应用奠定了坚实的理论和技术基础。
图41幅，表1个，参考文献64篇。
关键词：数字图像处理；Cortex-A53；Linux；Qt中图分类号：TP391Research on Image Processing SystemBased on ARM Cortex-A53Abstract:With the constant progress of science and technology, the application of digital image processing has been widely used in thousands of households, making people's life more and more convenient. The early implementation of image processing system is based on desktop PC, but there are many constraints in PC-based processing system, which mainly depend on software operation, and the portability is poor.This paper uses S5P6818 processor based on Cortex-A53 architecture and Linux real-time operating system to construct a real-time image processing system. The system can perform real-time image acquisition, image processing, moving object detection and tracking, and so on. At the same time, network processing can be realized and it is easy to carry. The research work on image processing system and image processing algorithm is as follows:(1)According to the requirement of system function and development cost, each hardware module is selected. The system takes the X6818bv3 development board as the core to connect the peripheral hardware module, and constructs the system hardware platform.(2)The image processing system transplants software such as uboot, Linux operating system, root file system and video server, and establishes a system running software platform to ensure the normal execution of the system application and realize the functions of image acquisition and image processing.(3)The application program of image processing system is divided into two parts: image acquisition and image processing. The graphic interface program is written in Qt Creator integrated development environment to realize video image display, image saving in jpeg format, image enhancement and edge detection, etc. and the remote monitoring function is added.(4)Aiming at the problems of traditional ViBe algorithm in foreground detection, such as sensitivity to illumination change and low detection rate, an improved algorithm based on ViBe algorithm is proposed in this paper, and samples are selected for comparison. The algorithm can effectively solve the problem of low detection rate in the process of moving object detection in video images. The algorithm has high real-time performance and has better inhibition to different illumination intensity.The image processing system includes the construction of software and hardware platform and the design of application program. The research content of this paper lays a solid theoretical and technical foundation for the establishment and application of ARM-based image processing system platform. There are 41 graphs, 1 table and 64 references.Meng-long Huang (Electrical Engineering)Directed by Yun-hong Li Key words: Digital image processing, Cortex-A53, Linux, QtClassification: TP391目次1绪论 (1)1.1研究背景及意义 (1)1.2国内外研究现状及发展趋势 (1)1.3论文章节结构 (3)2系统总体设计 (5)2.1系统功能分析 (5)2.2系统整体开发流程 (5)2.3系统硬件选型 (6)2.4系统软件选型 (7)2.4.1嵌入式操作系统的选择 (7)2.4.2引导加载程序的选择 (8)2.5本章小结 (8)3系统硬件平台搭建 (9)3.1 X6818bv3开发板简介 (9)3.2 X6818bv3开发板硬件资源 (10)3.3系统外围电路 (11)3.3.1电源管理模块 (11)3.3.2串口电路 (11)3.3.3 SD卡模块 (11)3.3.4 USB接口电路 (12)3.3.5 LCD液晶屏与触摸屏接口 (13)3.4摄像头选型 (13)3.5本章小结 (14)4运动目标的检测与跟踪 (15)4.1预处理 (15)4.2运动检测 (15)4.2.1 ViBe背景建模方法 (15)4.2.2 ViBe算法的改进 (17)4.2.3前景检测后处理 (18)4.3运动目标跟踪 (19)4.4运动动态模型 (19)4.5卡尔曼滤波跟踪 (20)4.6本章小结 (22)5系统软件平台搭建与应用程序设计 (23)5.1系统软件平台搭建流程 (23)5.2 tftp及ckermit安装配置 (23)5.3构建交叉编译环境 (24)5.4 uboot移植 (25)5.5 Linux内核配置与移植 (26)5.6根文件系统制作与nfs服务器搭建 (27)5.7 Qt的安装与移植 (29)5.7.1宿主机Qt的安装 (29)5.7.2 ARM版本Qt的编译与移植 (29)5.8 OpenCV函数库移植 (31)5.9 mjpg-streamer视频服务器移植 (32)5.9.1 mjpg-streamer简介 (32)5.9.2视频服务器与客户端的通信 (33)5.9.3 mjpg-streamer移植 (34)5.10系统应用程序设计 (35)5.10.1应用程序总体框架设计 (35)5.10.2主窗口设计 (35)5.10.3 Qt多线程编程 (37)5.10.4视频图像显示 (39)5.10.5视频图像的保存 (41)5.11图像处理的实现 (42)5.11.1图像灰度化 (43)5.11.2图像二值化处理 (43)5.11.3图像平滑 (43)5.11.4图像增强 (44)5.11.5图像的边缘检测 (45)5.12宿主机应用程序测试 (46)5.13本章小结 (47)6系统测试 (49)6.1系统自启动 (49)6.2视频图像采集测试 (50)6.3图像处理算法测试 (50)6.4本章小结 (52)7总结与展望 (53)7.1总结 (53)7.2展望 (53)参考文献 (55)作者攻读学位期间发表学术论文清单 (59)致谢 (61)1 绪论1绪论1.1研究背景及意义数字图像处理技术经历了百年的发展，其理论和方法已经形成了完整的体系，数字图像处理系统包括图像信号的采集、数字图像处理、图像压缩等部分，并且在科学研究及社会生活等众多领域得到了越来越广泛的应用[1]，因此对数字图像处理系统的研究有着重要的现实意义以及很高的应用价值。
Video - DINION IP starlight 6000 HDDINION IP starlight 6000 HDu Excellent low-light performanceu Built-in Essential Video Analytics to trigger relevant alerts and quickly retrieve data u Intelligent Streaming and Intelligent Dynamic Noise Reduction for low network strains and storage costsu Extended Dynamic Range mode to see details in bright and dark areas simultaneously u Auto back focus for fast installationThis camera provides clear images 24/7 – even at night or under low-light conditions.The exceptional starlight sensitivity enables thiscamera to work with a minimum of ambient light. The extended dynamic mode provides detailed images in scenes with challenging lighting.The camera is available in 1080p or 720p resolution versions and provides up to 60 images per second.There is a selection of high quality lenses separately available.FunctionsExceptional low-light performanceThe latest sensor technology combined with the sophisticated noise suppression results in an exceptional sensitivity in color. The low-lightperformance is so good that the camera continues to provide excellent color performance even with a minimum of ambient light.Fast performanceThe 60 frames per second mode provides for optimum performance in fast action scenes that ensures no critical data is lost.High Dynamic RangeThe camera has High Dynamic Range. This is based on a multiple-exposure process that captures more details in the highlights and in the shadows even in the same scene. The result is that you can easily distinguish objects and features, for example, faces with bright backlight.The actual dynamic range of the camera is measured using Opto-Electronic Conversion Function (OECF)analysis according to IEC 62676 Part 5. This method is used to provide standardized measurements, which can be used to compare different cameras.Essential Video AnalyticsThe built-in video analysis reinforces the Intelligence-at-the-Edge concept and now delivers even more powerful features. Essential Video Analytics is ideal for use in controlled environments with limited detection ranges.The system reliably detects, tracks, and analyzes objects, and alerts you when predefined alarms are triggered. A smart set of alarm rules makes complex tasks easy and reduces false alarms to a minimum.Metadata is attached to your video to add sense and structure. This enables you to quickly retrieve the relevant images from hours of stored video. Metadata can also be used to deliver irrefutable forensicevidence or to optimize business processes based on people counting or crowd density information.Calibration is quick and easy – just enter the height of the camera. The internal gyro/accelerometer sensor provides the rest of the information to precisely calibrate the video analytics.Intelligent Auto ExposureFluctuations in backlight and front light can ruin your images. To achieve the perfect picture in every situation, Intelligent Auto Exposure automatically adjusts the exposure of the camera. It offers superbfront light compensation and incredible backlight compensation by automatically adapting to changing light conditions.Intelligent Dynamic Noise ReductionQuiet scenes with little or no movement require a lower bitrate. By intelligently distinguishing between noise and relevant information, Intelligent Dynamic Noise Reduction reduces bitrate by up to 50%. Because noise is reduced at the source during image capture, the lower bitrate does not compromise on video quality.Intelligent Dynamic Noise Reduction adjusts spatial and temporal filtering (3DNR) based on intelligent analysis of the scene content. Motion compensated temporal filtering (MCTF) reduces motion blur normally associated with standard temporal filtering. This maintains image quality of fast moving objects while still optimizing bitrate.With Intelligent Dynamic Noise Reduction, our focus is to significantly reduce storage costs, and lessen network strain by only using bandwidth when needed. Area-based encodingArea-based encoding is another feature which reduces bandwidth. Compression parameters for up to eight user-definable regions can be set. This allows uninteresting regions to be highly compressed, leaving more bandwidth for important parts of the scene. Bitrate optimized profileThe average typical optimized bandwidth in kbits/s for various image rates is shown in the table.Scene modesThe camera has a very intuitive user interface that allows fast and easy configuration. Nine configurable modes are provided with the best settings for a variety of applications. Different scene modes can be selected for day or night situations.Multiple streamsThe innovative multi-streaming feature delivers various H.264 streams together with an M‑JPEG stream. These streams facilitate bandwidth-efficient viewing and recording as well as integration with third-party video management systems.The camera can run multiple independent streams that allows to set a different resolution and frame rate on the first and second stream. The user can also choose to use a copy of the first stream.The third stream uses the I-frames of the first stream for recording; the fourth stream shows a JPEG image at a maximum of 10 Mbit/s.Regions of interest and E-PTZRegions of Interest (ROI) can be user defined. The remote E-PTZ (Electronic Pan, Tilt and Zoom) controls allow you to select specific areas of the parent image. These regions produce separate streams for remote viewing and recording. These streams, together with the main stream, allow the operator to separately monitor the most interesting part of a scene while still retaining situational awareness.Intelligent Tracking can follow objects within the defined regions of interest. Intelligent Tracking can autonomously detect and track moving objects or the user can click on an object which the tracker will then follow.Easy installationPower for the camera can be supplied via a Power-over-Ethernet compliant network cable connection. With this configuration, only a single cable connection is required to view, power, and control the camera. Using PoE makes installation easier and more cost-effective, as cameras do not require a local power source.The camera can also be supplied with power from+12 VDC power supplies.To increase system reliability, the camera can be simultaneously connected to both PoE and +12 VDC supplies. If one power source fails, the other source takes over without a reboot so providing power redundancy.The auto-focus lens wizard makes it easy for an installer to accurately focus the camera for both day and night operation. The wizard is activated from the web browser or from the on-board camera push button making it easy to choose the workflow that suits best. The automatic motorized back focus adjustment with 1:1 pixel mapping ensures the camera is always focused accurately.Automatic image rotationThe integrated gyro/accelerometer sensor automatically corrects the image orientation in steps of 90° if the camera is mounted at right angles or upside down. The sensor image can also be rotated manually through steps of 90°.To efficiently capture details in long hallways without loss of resolution, mount the camera at right angles. The image is displayed upright at full resolution on your monitor.Hybrid operationA surge-protected analog video output allows full hybrid operation. This means that high resolution IP video streaming and an analog video output areavailable simultaneously. The hybrid functionality offers an easy migration path from legacy CCTV to a modern IP-based system.Storage managementRecording management can be controlled by the Bosch Video Recording Manager or the camera can use iSCSI targets directly without any recording software.Edge recordingInsert a memory card into the card slot to store up to 2 TB of local alarm recording. Pre-alarm recording in RAM reduces recording bandwidth on the network, and extends the effective life of the memory card. Cloud-based servicesThe camera supports time-based or alarm-based JPEG posting to four different accounts. These accounts can address FTP servers or cloud-based storage. Video clips or JPEG images can also be exported to these accounts.Alarms can be set up to trigger an e-mail or SMS notification so you are always aware of abnormal events.True day/night switchingThe camera is a true day/night camera with a mechanical filter for vivid daytime color and exceptional night-time imaging while maintaining sharp focus under all lighting conditions. The filter can be switched remotely, or automatically via a light level sensor or contact input.Data securitySpecial measures are necessary to ensure the highest level of security for device access and data transport. On initial setup, the camera is only accessible over secure channels. You must set a service-level password in order to access camera functions.Web browser and viewing client access can be protected using HTTPS or other secure protocols that support state-of-the-art TLS 1.2 protocol with updated cipher suites including AES encryption with 256 bit keys. No software can be installed in the camera, and only authenticated firmware can be uploaded. A three-level password protection with security recommendations allows users to customize device access. Network and device access can be protected using 802.1x network authentication with EAP/TLS protocol. Superior protection from malicious attacks is guaranteed by the Embedded Login Firewall, on-board Trusted Platform Module (TPM) and Public Key Infrastructure (PKI) support.The advanced certificate handling offers:•Self-signed unique certificates automatically created when required•Client and server certificates for authentication •Client certificates for proof of authenticity •Certificates with encrypted private keys Complete viewing softwareThere are many ways to access the camera’s features: using a web browser, with the BVMS, with the free-of-charge Bosch Video Client or Video Security Client, with the video security mobile app, or via third-party software.Video security appThe Bosch video security mobile app has been developed to enable Anywhere access to HD surveillance images allowing you to view live images from any location. The app is designed to give you complete control of all your cameras, from panning and tilting to zoom and focus functions. It’s like taking your control room with you.This app, together with the separately available Bosch transcoder, will allow you to fully utilize our dynamic transcoding features so you can play back images even over low-bandwidth connections.System integration and ONVIF conformanceThe camera conforms to the ONVIF Profile S , ONVIF Profile G , ONVIF Profile M , and ONVIF Profile T specifications.Third-party integrators can easily access the internal feature set of the device for integration into large projects. Visit the Bosch Integration Partner Program (IPP) website () for more information.Regulatory information* Chapters 7 and 8 (mains voltage supply requirement) are not applicable to the camera.However, if the system in which this camera is used needs to comply with this standard, then any power supplies used must comply with this standard.Installation/configuration notesControlsTechnical specificationsOrdering informationNBN-63013-B Fixed camera 1MP HDRHigh-performance IP box camera for intelligent HD surveillance in low light and with hybrid IP/analog operation.720pNDAA compliantOrder number NBN-63013-B | F.01U.314.349 NBN-63023-B Fixed camera 2MP HDRHigh-performance IP box camera for intelligent HD surveillance in low light and with hybrid IP/analog operation.1080pNDAA compliantOrder number NBN-63023-B | F.01U.314.805NBN-65023-B Fixed camera 2MP HDR 24V1080pHigh-performance IP box camera for intelligent HD surveillance in low light and with hybrid IP/analog operation.NDAA compliantOrder number NBN-65023-B | F.01U.349.538AccessoriesLVF-5003C-P2713 Varifocal lens, 2.7-13mm, 3MP, CS mountVarifocal P-iris megapixel IR corrected lens with 1/2.7 inch sensor and CS-mountOrder number LVF-5003C-P2713 | F.01U.381.550LVF-5005C-S0940 Varifocal lens, 9-40mm, 5MP, CS mountVarifocal SR megapixel IR corrected lens with 1/2.5" sensor and CS-mountOrder number LVF-5005C-S0940 | F.01U.274.352LVF-5005C-S1803 Varifocal lens, 1.8-3mm, 5MP, CS mountVarifocal SR megapixel IR corrected lens with 1/2.5" sensor and CS-mountOrder number LVF-5005C-S1803 | F.01U.274.354LVF-5005C-S4109 Varifocal lens, 4.1-9mm, 5MP, CS mountVarifocal SR megapixel IR corrected lens with1/1.8” sensor and CS-mountOrder number LVF-5005C-S4109 | F.01U.297.770LVF-5005N-S1250 Varifocal lens, 12-50mm, 5MP, C mountVarifocal megapixel IR corrected lens with 1/1.8" sensor max and C-mountOrder number LVF-5005N-S1250 | F.01U.305.567UPA-1220-60 Power supply, 120VAC 60Hz,12VDC 1A outPower supply for camera. 100-240 VAC, 50/60 Hz In;12 VDC, 1 A Out; regulated.Input connector: 2-prong, North American standard (non-polarized).Order number UPA-1220-60 | F.01U.076.155UPA-1220-50 Power supply, 220VAC 50Hz, 12VDC 1A outPower supply for camera. 100-240 VAC, 50/60 Hz In;12 VDC, 1 A Out; regulated.Input connector: 2-prong, European Europlug standard (4 mm / 19 mm).Order number UPA-1220-50 | F.01U.076.158TC9210U Camera mount, 6", indoorA universal 6-inch wall/ceiling grid with off-white finish for 4.5 kg (10 lb) max load, incl. T-Bar ceiling clip and wall/ceiling mount flange.Order number TC9210U | F.01U.143.373UHO-HBGS-11 Outdoor housing, 24VAC, feed-through Outdoor housing for (24 VAC / 12 VDC) camera with24 VAC power supply, blower and feed-through cabling. Order number UHO-HBGS-11 | F.01U.302.304UHO-HBGS-51 Outdoor housing, blower, 230VAC/35W Outdoor housing for (230 VAC / 12 VDC) camera with 230 VAC power supply, blower and feed-through cabling. Order number UHO-HBGS-51 | F.01U.302.310UHO-HBGS-61 Outdoor housing, blower, 120VAC/35W Outdoor housing for (120 VAC / 12 VDC) camera.120 VAC power supply; blower; feed-through cabling Order number UHO-HBGS-61 | F.01U.302.311LTC 9210/01 Column mount, 8", 9KG/20lb loadFeed-through column mount for 20 cm (8 in), 5 kg (11 lb) maximum load; light gray finish; for indoor use. Order number LTC 9210/01 | F.01U.027.057LTC 9215/00 Wall mount with cable feed through, 12" Wall mount for camera housing, cable feed-through,30 cm (12 in); for outdoor use.Order number LTC 9215/00 | 4.998.137.651LTC 9215/00S Wall mount for UHI/UHOWall mount for camera housing, cable feed-through,18 cm (7 in); for indoor use.Order number LTC 9215/00S | F.01U.503.621LTC 9219/01 Feed through J mountJ-mount for camera housing, 40 cm (15 in); for indoor use.Order number LTC 9219/01 | F.01U.503.623LTC 9213/01 Pole mount adapter forLTC9210,9212,9215Flexible pole mount adapter for camera mounts (use together with the appropriate wall mount bracket). Max.9 kg (20 lb); 3 to 15 inch diameter pole; stainless steel strapsOrder number LTC 9213/01 | F.01U.009.291NBN-MCSMB-03M Cable, SMB to BNC, camera-cable, 0.3m0.3 m (1 ft) analog cable, SMB (female) to BNC (female) to connect camera to coaxial cableOrder number NBN-MCSMB-03M | F.01U.291.564NBN-MCSMB-30M Cable, SMB to BNC, camera-monitor/DVR3 m (9 ft) analog cable, SMB (female) to BNC (male) to connect camera to monitor or DVROrder number NBN-MCSMB-30M | F.01U.291.565VJT-XTCXF VIDEOJET XF TRANSCODERHigh-performance video transcoder. H.264; CF card slot; ROI; max resolution 1080p; 2 channelsOrder number VJT-XTCXF | F.01U.261.015NPD-5001-POE Midspan, 15W, single port, AC inPower-over-Ethernet midspan injector for use with PoE enabled cameras; 15.4 W, 1-portWeight: 200 g (0.44 lb)Order number NPD-5001-POE | F.01U.305.288NPD-5004-POE Midspan, 4 port x 15W, AC inPower-over-Ethernet midspan injector for use with PoE enabled cameras; 15.4 W, 4-portsWeight: 620 g (1.4 lb)Order number NPD-5004-POE | F.01U.305.289UHO-POE-10 POE outdoor housing, heater, blower Outdoor camera housing with PoE+ power supply. Order number UHO-POE-10 | F.01U.300.502Represented by:Europe, Middle East, Africa:Germany:North America:Asia-Pacific:Bosch Security Systems B.V.P.O. Box 800025600 JB Eindhoven, The Netherlands Phone: + 31 40 2577 284/xc/en/contact/ Bosch Sicherheitssysteme GmbHRobert-Bosch-Ring 585630 GrasbrunnTel.: +49 (0)89 6290 0Fax:+49 (0)89 6290 1020****************************Bosch Security Systems, LLC130 Perinton ParkwayFairport, New York, 14450, USAPhone: +1 800 289 0096Fax: +1 585 223 9180*******************.comRobert Bosch (SEA) Pte Ltd, Security Systems11 Bishan Street 21Singapore 573943Phone: +65 6571 2808Fax: +65 6571 2699/xc/en/contact/Data subject to change without notice | 202306221252 | V39 | June 22, 2023© Bosch Security Systems 2023。
现代电子技术Modern Electronics TechniqueMay 2023Vol.46No.102023年5月15日第46卷第10期多目标跟踪是计算机视觉中的一个重要领域，多应用在视频监控和自动驾驶中。
而遮挡问题是多目标跟踪面临的一个关键挑战，针对这一问题，人们对多目标跟踪算法提出了许多优化方案。
有仅依靠检测器跟踪目标的算法[1]、IoU 跟踪器[2]以及基于核相关滤波器（KCF ）的跟踪器[3]等。
随着深度学习逐渐成熟，目前主要出现了两种目标跟踪优化方向，分别是检测跟踪范式（Tracking⁃by⁃Detection ）的跟踪器[4]和检测与跟踪联合学习的跟踪器[5]。
第一种分离式算法通常利用目标的外观特征、运动特征和前后帧信息来提高目标跟踪的鲁棒性和稳定性，从而解决遮挡问题；第二种联合式算法通过神经网络和深度学习将目标的表观特征和运动特征联DOI ：10.16652/j.issn.1004⁃373x.2023.10.032引用格式：李伟，颜旒.应对遮挡问题对DeepSORT 进行轨迹拟合优化[J].现代电子技术，2023，46（10）：173⁃180.应对遮挡问题对DeepSORT 进行轨迹拟合优化李伟，颜旒（重庆交通大学机电与车辆工程学院，重庆400074）摘要：检测跟踪范式是多目标跟踪的主要研究方向，也是自动驾驶汽车的主要应用框架。
完善的检测跟踪范式可以在提高多目标跟踪精度的同时有效降低跟踪框的失真率。
然而在现有的先进方法中，这种范式通常会遇到多目标重叠后的ID 交换问题，严重影响跟踪精度和轨迹判断。
为解决该问题，文中基于经典的DeepSORT 算法提出改进方案。
首先，在卡尔曼滤波器中添加跟踪框进行置信度的预测和更新，并按降序对更新后的置信度进行排列，在后续匹配中优先匹配预测置信度更高的跟踪目标；然后，比较预测置信度和检测置信度之间的差异，识别出置信度突变的目标，以进行跟踪轨迹的优化和剪枝；最后，使用余弦相似度和交并比（IoU ）识别重叠目标，并对重叠目标中置信度最高的目标消失后的轨迹进行多项式轨迹拟合，以纠正错误的ID ，完成精确匹配。
毕业设计（论文）原创性声明和使用授权说明原创性声明本人郑重承诺：所呈交的毕业设计（论文），是我个人在指导教师的指导下进行的研究工作及取得的成果。
尽我所知，除文中特别加以标注和致谢的地方外，不包含其他人或组织已经发表或公布过的研究成果，也不包含我为获得及其它教育机构的学位或学历而使用过的材料。
对本研究提供过帮助和做出过贡献的个人或集体，均已在文中作了明确的说明并表示了谢意。
作者签名：日期：指导教师签名：日期：使用授权说明本人完全了解大学关于收集、保存、使用毕业设计（论文）的规定，即：按照学校要求提交毕业设计（论文）的印刷本和电子版本；学校有权保存毕业设计（论文）的印刷本和电子版，并提供目录检索与阅览服务；学校可以采用影印、缩印、数字化或其它复制手段保存论文；在不以赢利为目的前提下，学校可以公布论文的部分或全部内容。
作者签名：日期：学位论文原创性声明本人郑重声明：所呈交的论文是本人在导师的指导下独立进行研究所取得的研究成果。
除了文中特别加以标注引用的内容外，本论文不包含任何其他个人或集体已经发表或撰写的成果作品。
对本文的研究做出重要贡献的个人和集体，均已在文中以明确方式标明。
本人完全意识到本声明的法律后果由本人承担。
作者签名：日期：年月日学位论文版权使用授权书本学位论文作者完全了解学校有关保留、使用学位论文的规定，同意学校保留并向国家有关部门或机构送交论文的复印件和电子版，允许论文被查阅和借阅。
本人授权大学可以将本学位论文的全部或部分内容编入有关数据库进行检索，可以采用影印、缩印或扫描等复制手段保存和汇编本学位论文。
涉密论文按学校规定处理。
作者签名：日期：年月日导师签名：日期：年月日注意事项1.设计（论文）的内容包括：1）封面（按教务处制定的标准封面格式制作）2）原创性声明3）中文摘要（300字左右）、关键词4）外文摘要、关键词5）目次页（附件不统一编入）6）论文主体部分：引言（或绪论）、正文、结论7）参考文献8）致谢9）附录（对论文支持必要时）2.论文字数要求：理工类设计（论文）正文字数不少于1万字（不包括图纸、程序清单等），文科类论文正文字数不少于1.2万字。
华中科技大学硕士学位论文摘要目标检测和跟踪有很多的应用领域，包括智能交通网、视频内容分析和目标行为理解等。
MMDTS（Mulitple Moving Object Detection and Tracking System）是我们研究的一个对于多个运动目标检测和跟踪的系统。
它采用混合高斯模型建立背景模型，通过连通成分分析和前景区域上的角点检测分别获得团特征和角点特征，接着采用一个多层次级别的跟踪策略来稳定的跟踪独立目标。
最后，扩展的应用系统还具备目标跟踪轨迹自动记录，人车流量统计，车辆逆向行驶警告等功能。
本文所做的研究工作主要有MMDTS系统的设计与实现，和在此基础上的三个扩展应用功能的实现。
MMDTS系统包括背景建模，特征提取，特征跟踪和预测模型部分。
特征提取和预测模型选用了较为常用和成熟的技术来实现，背景建模和特征跟踪是本文研究的重点。
在背景建模部分，本文在经典的混合高斯背景建模方法的基础上，加入了光线矫正和HSV颜色空间中阴影去除的考虑。
在特征跟踪部分，本文设计了一个多层次级别的跟踪策略，分别在团特征、独立目标特征、角点特征上的匹配跟踪。
三层级别的跟踪信息彼此交互和利用，使得跟踪能够适应实际的各种情况，具备真正意义的实际应用。
MMDTS已经在多个数据库上测试，在检测和跟踪目标上均有较好的表现，能适应场景光线的变动，在目标与目标之间出现交叉时也依然能跟踪独立目标。
然而系统也有不足之处，例如对于首次出现就交叠在一起的几个独立目标，系统会将它们作为一个整体跟踪直到它们分离。
最后，扩展的应用系统还具备目标轨迹自动记录，人车流量统计，车辆逆向行驶警告等功能。
关键词：混合高斯阴影检测目标检测角点检测目标跟踪华中科技大学硕士学位论文AbstractObject detection and tracking has many applications in some area, including intelligent transport network, video content analysis and object behavior understanding. MMDTS is our system to detect and track multiple moving objects. In our systrem, a mixture gaussian is used to model the background. Then, blob features are grouped by using connected component label method and corner features are obtained by corner detection in the foreground. Finally, a multi-level feature tracking strategy is used to track independent object stably. Also, our extended system has three applications, object tracking recording, people and vehicle counting and vehicle reverse traveling warning.This paper's main research work is to design and implement the MMDTS system and three extended applications. MMDTS system is content of background model, feature extraction, feature tracking and prediction model. We select common and mature technology for the feature extraction and prediction model. Background model and feature tracking are the focus of the paper. Based on Mixed Gaussian Model we have two cosiderations, luminous correction and shadow removal in HSV color space to modeling the background. In the feature tracking part, we design a multi-level tarcking straegy to tracking independent object by matching blob feature, independent object feature and corner feature. Three levels of tracking information are used interact with each other, and it can adapt variety situations in practical. This means it has real application.MMDTS has been tested in multiple databases, and it has well performance in both detection and tracking objects. It can adapt luminous change and handle objects crossing. However, the system also has some shortcomings, such as objects overlaped at first time would be tracked as one independent object. Also, our extended system has three applications, object tracking recording, people and vehicle counting and vehicle reverse traveling warning.Keyword: Mixture Gaussian Model Shadow Detection ObjectDetection Corner Detection Object Tracking独创性声明本人声明所呈交的学位论文是我个人在导师指导下进行的研究工作及取得的研究成果。
Ubhale Apurva .S, Pusdekar P. N., International Journal of Advance Research, Ideas and Innovations in Technology.ISSN: 2454-132XImpact factor: 4.295(Volume3, Issue3)Available online at System to Detect Human Being Buried Under the Rubble duringDisasterApurva S. UbhaleP. R. Pote (Patil)Welfare & Education Trust’s Group of Institutions College of Engineering & Management, Amravati**********************Prof. P. N. PusdekarP. R. Pote (Patil)Welfare & Education Trust’s Group of Institutions College of Engineering & Management, AmraAbstract: Death of lots of people occurs as a cause of the earthquake. Such news comes in the newspaper after the unavoidable casualty. Because of this, unlimited numbers of people die. It occurs as a result of disasters such as tunnel dropping, snowfall, and fall of landslides. The effect of fall of landslides occurs in June 2013, due to heavy rain came in Uttarakhand. Such worst incident happened in 2013, because of this, lost precious lives. When many people are buried under rubble during a disaster, at that situation the most important question come in mind that how to enter the area using rescue teams.The microwave life detection system is developed to detect subject or object buried under the rubble of collapsed building during the earthquake or other disasters. The object or subject includes human being/ victims. The proposed motion detection system uses microwave frequency electromagnetic signal which is able to detect motion of moving object. This system decides whether the object is in motion or not. The Doppler frequency shift of the wave is the operational principle of motion detection system. The motion detection system uses microwave Doppler radar sensor sense these waves reflected back from the object if the object is in motion and present below the ground level. Once the motion is detected, then able to decide whether the object is a human being or not. Then system using microwave test bench is used to decide whether the object is a human being or not. Also, the system operates using microwave test bench used to detect the breathing and heartbeat signals of the subject. The Matlab Simulink model show heartbeat and breathing signal. By using all these systems, able to decrease world death rate to a greater extent.Keywords: Life under Rubble, Doppler Shift, Dual Antenna System, Modulation Due To Body Oscillations, MicrowaveLife Detection System.I.INTRODUCTIONWhen we read a newspaper or watch the news on TV, many times we read or heard the news of earthquake or other disasters like landslides happened. Such disasters cause losses of many human lives every year as a human being are buried under rubble. There are some existing methods used for detecting human being are the utilization of dogs, some optical devices, acoustic life detector and rescue robot. Such existing devices or method are ineffective in recovering victims lying much below few feet and also for cases where victims are trapped completely or very weak to respond to the signal sent by rescue to rescuers. Then in such cases, life detection system used for detecting human being trapped under rubble.So, save more lives due to detection of victims. Depending upon this fact “A Revolutionary System to Detect Human Being Buried under the Rubble” [1] used to trap the buried victims under earthquake rubble or collapsed buildings.The system has been designed by the utilization of microwave frequency electromagnetic signal and microcontroller. The basic block diagram of motion detection system as shown below:Fig.1: Basic block diagram of motion detection system [14]In fig. [1], microwave module connects with microcontroller and microcontroller connects with LCD display. InAdditional, we provide buzzer along with programming counter. When the system detects motion of moving objects, LCD display ‘Alive ‘instruction and buzzer become on. Otherwise, buzzer become off when LCD display ‘No Response’ instruction.II.LITERATURE SURVEYSometimes unavoidable casualty occurs where we live or where we work. Images of such accidents serve as constant reminders of the vulnerability of these places. Emergency forces required correct information about the exact position of a human trapped or buried under rubble.It also needs information about collapse building as well as standardized intervention procedures along with information on the state of victim’s health. When an earthquake or other disasters occur across the world then man-made structure like building, houses or bridges become collapse with different magnitude. In such situation, people try to survive as much as possible. There are two principle research directions including moving human tracking and life-sign detection [8-9]. Life-sign detection is applied to determine the condition of the stationary human with the life sign such as breathing [9-10].The concept of microwave life detection system emerged with the development of the systems for the rescue operation. In the initial phase, dogs were used to detect the presence of human then acoustic detectors and robot radar comes into existence for detection of buried victims. But these systems are having some major drawbacks. K. M. Chen who brings out the concept of detection of buried victims using microwave beam in 1986[1].A rescue radar system is proposed by M. Donelli in 2011[2]. A microwave life detection system operated on radio frequency was proposed in 1991 [3]. The phase change of a reflected microwave signal will provide the precious information about the buried victims [4]. An X-band experimental set up for life signs detection introduced using simple strategy [11]. W. S. Haddad gave the idea of Rubble Rescue Radar (RRR) [12] for the detection system of trapped human personnel. M. Bimpas gave the concept of three band radar system [13].PROPOSED SYSTEMThe microwave frequency electromagnetic signal is sent from the oscillator and used to detect human being become alive or not. This signal having the characteristic to pass through barriers and would reflect back from some objects (human being).The system utilizes 10GHz microwave frequency electromagnetic signal.When system operates using microwave test bench used to detect the breathing and heartbeat signals of living subjects. But microwave detection system can work on a different range of frequency from 2GH (L-band) to 10GHz(X-band) [3].A microprocessor-controlled automatic clutter cancellation subsystem, consisting of a programmable microwave attenuator and a programmable microwave phase shifter controlled by a microprocessor-based control unit, has been developed for a microwave life-detection system (L-band 2 GHz or X-band 10 GHz) which can remotely sense breathing and heartbeat movements of living subjects [6]. The proposed system made up of 3 parts:1)Motion detection system 2)Whether object or subject is a human being or not, detect using microwave test bench system. 3) Simulink model show waveform for heartbeat or breathing signals.1) Motion detection systemSystem ArchitectureThe following figure shows system architecture for motion detection system:Fig.2: System Architecture of life detection systemThe microwave life detection system has two important parts. These two parts are as given below:∙ Microwave circuit which generates, amplifies and distributes microwave signals to different microwave components.∙ A dual antenna system, which consists of two separate antennas energized sequentially.1. Microwave Doppler Radar Sensor for Motion:This sensor can detect motion or speed of moving objects through Doppler principle. It transmits a 10 GHz microwave frequency electromagnetic signal and waits for the signal to receive back and monitors the shift in frequency signal [16]. The block diagram of Doppler radar as shown below:Fig.3: Block diagram of Doppler radar [15-16]2. DOPPLER EFFECTThe Doppler Effect is a shift in frequency perceived by a receiver from a signal source due to relative movement of the source and/or receiver. Doppler radar system, a known frequency signal is transmitted from an antenna which is pointed at a reference object. A separate antenna is used to receive the signal that is reflected back from the reference to measure the Doppler shift of the signal.3. PIC16F877A DEVICEHere a microcontroller used is ‘Peripheral Interface Controller’PIC16F877A devices are available in 40 packages. It has 5 input output ports and 3 timers/counters. It has 15 interrupts and 8 A/D input channels. The Parallel Slave Port is implemented only on the 40 pin devices. Its operating speed is DC – 20 MHz clock input DC – 200 ns instruction cycle. It has 8K x 14 words of Flash Program Memory and 368 x 8 bytes of Data Memory (RAM).Its pinout compatible to other 28-pin or 40/44-pin.Hardware set upThe fig. 4 shows hardware model for motion detection of moving object:Fig.4: Hardware model for motion detection systemA.Working Principles of motion detection systemThe principle of detection is first, 10 GHz microwave frequency electromagnetic signal is sent from the oscillator and pass through rubble to detect vital signs of life. The microwave is having the property to penetrate through barriers and would reflect back from some objects. These objects include humans. So, the reception of signals shows the presence of a live human under the rubble. When sensor sense motion, buzzer become on, counter become start. Programming counter time period is from 0 to 6.If the sensor detects motion during the counter time period, it increments and again starts from 1 up to 6 along with LCD display ‘Alive’ instruction. During this counter time period, once the motion is not detected then buzzer become off and LCD display ‘No Response’ instruction, counter become stop to 0. Thus in order to maintain a high sensitivity for this application, the wave reflected from the rubble or the surface of the ground has to be canceled as thoroughly as possible.2) System using microwave test benchHardware set upThe fig.5 show hardware model for system using microwave test benchFig.5: Hardware model for system using microwave test bench (pyramidal horn antenna)B.Working Principles for system used to detect whether object or subject is human being or not, using microwave testbenchWhen microwave life detection system operates using microwave test bench then it is used to detect the breathing and heartbeat signals of living subjects. The phase shift between transmitted and received signals is the operational principle of heartbeat or breathing signal detection system. When microwave beam detects any vibration or oscillation as it penetrates up to it. Then the beam was reflected back from that oscillating surface with phase shift after detection of oscillation. On the basis of amplitude, decide whether the subject is human or not. In this part, various antennas are used such as pyramidal horn, E-plane section horn, H-plane section horn and parabolic dish antennas to see the pattern on CRO. The microwave life detection system using microwave test bench as shown below in fig.6:Fig.6: Microwave life detection system using microwave test bench [7]3)INTRODUCTION TO SIMULINKSimulink is model-based design for dynamic and embedded systems. It provides an interactive graphical environment and customizable set of block libraries that design, simulate and implement. It tests a variety of time systems, including communications, controls, signal processing, video processing, and image processing. Simulink software provides huge library functions used for developing the Simulink model. Simulink model shows a simulation of life detection system based on detection of the heartbeat using Doppler frequency [5].The Simulink model is shown in figure 7:Fig.7: Simulink Model [14]The Flow Chart for motion detection system as shown belowCFig.8: Flow Chart for motion detection systemIII. EXPERIMENTAL RESULTAnalysis of Result1) The waveform for the output of motion detection sensor display on CRO, shown in fig. a fig. b and fig this microwave Doppler radar sensor (motion detection sensor) detect whether the human being or any non-living object having motion or not. Also, it senses the speed of moving object. The speed of moving object given by more number of oscillations. If speed is more than a number of oscillations is more, shown in fig. a and fig. c. If there is no object and having no speed then it displays straight horizontal line, shown in fig. b.2) Pattern using the microwave test bench system display on CRO, shown in fig. d and fig. e. It detects whether the object is a human being or non-human being. Detection of a human being or non-human being depend on amplitude while the object moves in a horizontal and vertical direction. If the amplitude is more then it detects this object is a human being, shown in fig. d. If the amplitude is less then it detects this object is not- a human being and shown in fig. e.3) The waveform obtained from simulink model display in scope block on the laptop screen, shown in fig. f and fig. g. The waveform for breathing or heartbeat signals for alive, shown in fig. f. Both transmitting and reflected waves are present and shown in fig. f. When an object is not alive then fig. g shows only transmitting wave and reflected wave become absent.For all these 3 parts, the waveform for the output of each model for analysis of result as shown below:Is counter up to 6 Display it on LCD Is 1st ADCO/P > 65 Initialize ADC of PIC Controller Select 1st channel of ADC1)Waveform for output of motion detection sensor on CROFig. a: Waveform for Fig. b: Waveform for human being Fig. c: Waveform for non-living objectHuman having motion absent having motionFig. a, b and c show waveform for an object having motion or not. Fig. a shows a number of oscillations and having some amplitude.It means the object is present, has some speed and object has motion. So, the sensor senses it and display ‘Alive’instruction on LCD and buzzer become ON. Fig. b shows no amplitude and no oscillations. It means the object is absent and having no speed and no motion. It displays ‘No Response’ instruction on LCD and buzzer become OFF only when programming counter become off. Fig. c shows waveform for a non-living object having some oscillations and some amplitude. It means the object is in motion and has some speed. From all these waveforms of fig. a, b and c, it is observed that when waveform possess some amplitude and oscillations then it means that object having some motion as well as speed. When the distance between object and sensor become increases then the amplitude of waveform decreases and vice-versa. It means distance and amplitude of wave having an inversely proportional relation. When the speed of object increases then a number of oscillations also increases and when the speed of object decreases then the number of oscillations also decreases. It means oscillations and speed having proportionate.2)Waveform pattern for various antenna using microwave test bench system on CROFig. D: Waveform for pyramidal horn antenna for human being Fig. e: Waveform for pyramidal horn antenna for non-human beingFig. d and fig. e shows waveform for pyramidal horn antenna for the subject (living, non-living). Fig. d shows waveform for human being and fig. e shows waveform for non-living object/subject when object either move in a horizontal or vertical direction. From all these waveforms of fig. d and fig. e, it is observed that amplitude become increases for human being and amplitude become decreases for non-human being. It means that amplitude depends on the object (living/non-living). The object is used as a reference while either move in a horizontal or vertical direction. The amplitude varies according to the object (either object is a human being or non-human being).3)Waveform obtained from simulink modelFig .f: Waveform for breathing or Fig g: Waveform for breathing orheartbeat signal heartbeat signal absentFig. f shows waveform for breathing or heartbeat signal of alive person in the scope2 block. At some distance, using dist. from alive block and fm modulator block are used to display these waveforms mention in fig. f. Both block (dist. from alive and fm modulator) shown in fig. 7 of simulink model. Fig. f shows both transmitting and reflected waves. Upper row show transmitting waves and below scale (row) show reflected waves for breathing or heartbeat signal. Fig. g show waveform for breathing or heartbeat signal in scope block when the object is not present. Fm mod2 (2nd modulator) block is used to display this waveform mention in fig. g. Fm mod2 block shown in fig. 7 of simulink model. Fig. g display only transmitting wave in an upper row and reflected wave is absent in below row when a person is not alive. From this waveform of fig. f and fig. g, it is observed that simulink model waveform output depends on reflected waves.CONCLUSIONMicrowave life detection system able to detect a human being buried under rubble most efficiently and as possible in short time. When system operates using microwave doppler radar sensor, then it is used to detect whether the object is in motion or not. For motion detection system is conclude that distance between the sensor, object, and amplitude of wave having an inversely proportional relation. When system operates using microwave test bench, used to detect the breathing and heartbeat signals of living or non-living subjects. It means this system conclude that amplitude depends on the object. The amplitude varies according to the object (either object is a human being or non-human being). Matlab simulink model shows heartbeat or breathing signals in scope block on laptop or computer. It is conlude that simulink model waveform output depends on reflected waves. So, all these systems used to detect victims easily and successfully. Therefore, save precious lives and reduce world death rate to a greater extent.In future, develop such system which is used to detect more number of victims at a time. Also, give a graphical representation of victims in what position victims which are buried under rubble during disasters.ACKNOWLEDGEMENTThe author would like to thank my guide Prof. P. N. Pusdekar for the idea of such life detection system and HOD Prof. R. D. Ghongade for providing the guidance. We are grateful for acknowledging the support of all the IEEE paper on life detection system using a various model which is used to detect victims.REFERENCES[1] K. M. Chen, D. Misra, H. Wang, Huey- Ru Chuang, E. Postow, “An X-band M/W life-detection system,” IEEE Trans. Biomedical Eng., Vol. BME-33, No. 7, July 1986.[2] M. Donelli, “A rescue radar system for the detection of victims trapped under rubble based on the independent component analysis algorithm,” Progress in Electromagnetics Research, M, Vol. 19, 173-181, 2011.[3] Chen KM, Huang Y, Zhang J, Norman A, “Microwave life detection systems for searching human subjects under earthquake rubble or behind a barrier,” IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, Vol. 27, No.1, January 2000.[4] Wu, C. W., and Z. Y. Huang,”Using the Phase Change of a Reflected Microwave to Detect a Human Subject Behind a Barrier,” IEEE Transaction Biomedical Engg, Vol. 55, No. 1, January 2008.[5] A. Izadi, Z. Ghatan, B. Vosoughi Vahdat, and F. Farzaneh, “Design and Simulation of Life Detection System Based on the detection of the Heart Beat Using Doppler Frequency,” IEEE International Symposium on Signal Processing and Information Technology, 2006.[6] Huey-Ru Chuang, Y.-F. Chen, Kun-Mu Chen,”Automatic Clutter-Canceler for Microwave Life-Detection Systems,”IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 40, NO. 4, AUGUST 1991.[7] S. C. Deshmukh, B. B. Godbole, S. M. Single,” Detection of Live Human behind the Wall,”Journal of IJIRAE, Volume 1, Issue11,November 2014.[8] J. Goodwin, M. Tobias, J. Pozderac, and J. Peabody, “Detection algorithm implementation and measured results for a real-time,through-wall radar system using a TDM MIMO antenna array,” 2012 IEEE Radar Conference, pp. 240-246, 2012.[9] Y. Xu, S. Wu, C. Chen, J. Chen, and G. Fang, “A novel method for automatic detec tion of trapped victims by ultra- wideband radar,” IEEE Trans. on Geoscience and Remote Sensing, vol. 50, pp. 3132-3142, 2012.[10] Y. Zhou, L. Kong, G. Cui, and J. Yang, “Remote sensing of the human body by a stepped-frequency continuous wave,” 3rd International Conference on Bioinformatics and Biomedical Engineering, pp. 1-4, 2009.[11] M. D'Urso, “A SIMPLE STRATEGY FOR LIFE SIGNS DETECTION VIA AN X-BAND EXPERIMENTAL SET-UP” Centro Ricerche Giugliano, SELEX Sistemi Integrati SpA Via Circumvallazione Esterna di Napoli, zona ASI, Giugliano, I-80014, Italy, Progress In Electromagnetics Research C,Vol. 9, 119-129, 2009.[12] W. S. Haddad, “The Rubble Rescue Radar (RRR): A Low Power Hand-Held Microwave Device for the Detection of Trapped Human Personnel”, Work performed under the auspices of the US. Department of Energy by the Lawrence Livermore National Laboratory under Contract W-7405-Eng-48, April 10, 1997.[13] M. Bimpas, N. Paraskevopoulos, K. Nikellis, D. Economou and N. Uzunoglu, “Development of a three band radar system for detecting trapped alive humans under building ruins” Progress in Electroma gnetics Research, PIER 49, 161–188, and 2004.[14] R. Kantute, A. Kubde, M. Kitukale,”Detecting Alive Humans Using Embedded System,” Journal of IJECSCSE, IETE 46th Mid Term Symposium ”Impact of technology on skill development”MTS-2015.[15] Skolnik M., INTRODUCTION TO RADAR SYSTEMS, New Delhi,3rd edition,2001.[16] 。
第36卷第10期 光电工程V ol.36, No.10 2009年10月Opto-Electronic Engineering Oct, 2009文章编号：1003-501X(2009)10-0001-05运动摄像机下快速提取运动目标的新方法向桂山( 浙江科技学院信息与电子工程学院，杭州 310023 )摘要：本文研究摄像机和目标同时运动情况下的实时目标提取问题。
首先运用背景差方法，检测出静止摄像机下的运动区域，为了克服连通域分析法耗时长的不足，提出重心偏移迭代法快速获得感兴趣运动目标。
在改进Camshift跟踪算法中，提出采用Bayesian概率法则在由Kalman滤波器预测的感兴趣区域(ROI)内获取颜色概率密度分布图像(CPDDI)，引入即时背景(IB)以抑制背景特征。
提出依据跟踪结果进行目标提取的方法，即结合CPDDI 特征，并辅以适当的形态学滤波策略，从跟踪结果中提取出运动摄像机下的运动目标，解决目标被动态背景干扰的问题。
实验结果表明，提出的算法能够较稳定和完整地提取出运动摄像机下的运动目标，对复杂动态背景的适应性较强，且算法完全达到了实时的运行速度。
关键词：Camshift；重心偏移迭代方法；Bayesian概率法则；感兴趣区域；即时背景中图分类号：TP391 文献标志码：A doi：10.3969/j.issn.1003-501X.2009.10.001A Novel Method for Moving Object Extraction with an Active CameraXIANG Gui-shan( School of Information and Electronic Engineering,Zhejiang University of Science & Technology, Hangzhou 310023, China )Abstract: How to extract moving object in the case of moving camera is researched in this paper. At first, the background subtraction is implemented to detect the moving region in the case of motionless camera, and then the interested moving object is fast segmented from the detected moving region with the proposed method of barycenter shift iteration instead of the time-consuming method of connected components analysis. In the enhanced Camshift tracking algorithm, the Color Probability Density Distribution Image (CPDDI) in the Region of Interest (ROI) predicted by the Kalman filter is obtained, and Instantaneous Background (IB) is introduced to suppress background features. A novel method is suggested to extract the moving object from the tracking result combining CPDDI with proper morphologic filter strategy in the case of moving camera, which can resist the disturbance from the background. Experiments show that the proposed method can extract intact moving objects stably in the case of moving camera, and can well deal with dynamic complex background. The system is computationally efficient and can run in real-time speed.Key words: Camshift; barycenter shift iteration method; Bayesian law; region of interest; instantaneous background0 引 言智能视频监控是机器视觉的一个重要研究分支和应用领域，准确而快速地从监控视频图像中提取出运动目标是视频监控高层处理如目标分类与识别、行为理解预测及语义描述等的重要基础研究内容。
