合肥工业大学硕士学位论文交通视频监控中车辆检测与跟踪的研究姓名：杨强申请学位级别：硕士专业：信号与信息处理指导教师：齐美彬；蒋建国2011-04交通视频监控中车辆检测与跟踪的研究 交通视频监控中车辆检测与跟踪的研究 摘 要 摘 要 智能交通系统(ITS: Intelligent Transportation System)是一种在大范围、全方位发挥作用的，实时、准确、高效的综合运输和管理系统，它通过运用先进的信息技术、数据通讯传输技术、电子传感技术、电子控制技术以及计算机处理等技术，对整个交通运输管理体系进行准确、有效、安全地监控和管理。 车辆的检测与跟踪是智能交通系统中最核心、最关键的技术之一，这项技术的性能优劣直接关系到整个系统能否有效地运行，因此车辆检测和跟踪技术的研究对智能交通系统具有重要的意义和价值。 本文首先采用一种适用于复杂交通场景的多层次背景模型提取算法提取出背景，该算法具有快速准确的特点；然后采用一种用于车辆检测的选择性背景更新方法，可以有效地处理光照变化等因素引起的背景缓慢变化以及背景局部突变的问题，并具有良好的实时性。 其次，对于车辆检测，本文首先去除车辆阴影，然后利用上述算法提取的背景，采用一种改进的对称差分法和背景帧差相融合检测车辆运动区域，并使每一个运动目标成为独立的连通域， 最后采用改进的两次扫描法分割目标车辆，并在此基础上进行种子填充，解决由于车辆表面与路面灰度接近而产生的运动目标“孔洞”问题，进而得到更准确的车辆检测结果。该算法可以适应各种复杂的交通道路场景。 最后进行车辆跟踪。本文使用模板匹配的方法实现对车辆的跟踪，在该算法中引入几个参数描述车辆的整体特征（如运动目标的质心，长宽比等）来建立匹配模板，实现下一帧车辆的匹配，并在匹配完成后实时更新模板，为下一帧的车辆匹配做好准备。在匹配过程中针对车辆间互相遮挡的问题，采用一种基于匹配的方法分割遮挡目标，以使车辆匹配取得更好的效果。 整个算法需实现在DSP上的高效运行，本文最后结合TMS320DM6437嵌入式系统介绍了算法的优化方法，包括编译器选项优化、少使用函数调用、编写汇编代码和EDMA乒乓操作等， 并通过统计算法时间性能的方式测试优化效果。实验结果表明，优化后的本文算法具有较好的实时性和稳定性。 关键词：关键词： 背景更新 车辆检测 两次扫描法 种子填充 车辆跟踪 DSP优化 Research of vehicle object detection and tracing method in Traffic Video Monitoring Abstract Intelligent Transportation System(ITS) is a real-time, accurate and efficient integrated transportation management system which creats a wide range, all-round role.It can accurately, effectively manage and monitor the entire traffic system with integrated advanced information technology, data communication transmission technology, electronic sensor technology, electronic control technology and computer processing technology. The detection and tracking of vehicle is one of the core and most critical technologys, the good or bad performance of the technology directly relates whether the entire system can run effectively, so the research of vehicle detection and tracking technology have great significance and value to ITS. First of all, the title use a kind of way called multi-level background model extraction algorithm which applies to complex traffic scenes to extract background. The algorithm is fast and accurate.Next, it use a selective background update method for vehicle detection, which has good real-time, can effectively deal with the problems that slow changes of background caused by the influence of illumination change and local background mutation. Second, for vehicle detecting, the title remove the shadows of the moving vehicles at first, then utilizing the background extracted by previous algorithm, it use the fusion of improved symmetric difference method and background of the frame-difference-phase to detect vehicle movement area, and to make each of moving target become an independent connected domain, then use seed filling technology based on two Scanning Method, to solve the problem of moving targets “hole“ which is caused by the gray of vehicles surface close to the roads, and achieve more accurate results of vehicles detection. Finally, the title does vehicle tracking. This paper use template matching method to achieve vehicle tracking, the method use several parameters(such as target centroid, aspect ratio, etc) to describe the overall characteristics of the vehicle, and establish the match template of vehicles for matching vehicles in next frame, at last update the template to prepare for the next frame-match. For the problem of vehicles blocking each other in the process of match, the paper use a method base on match to break up target, and achieve better results of vehicle match. The whole algorithm need to achieve efficient operation on the DSP, the title at last introduced optimization algorithm with TMS320DM6437 embedded systems, and the optimization introduced includes compiler options optimization, less use of function calls, writing assembly code and EDMA ping-pong operation.And the title at last tests optimization results with detecting the time of algorithm performance. The last results show that algorithm have good real-time and stability. Keywords：Background update； Vehicle detection； Two scanning method；Seed filling；Vehicle tracking；DSP optimization； 插图清单插图清单 图 1.1 算法的结构流程 .6 图 2.1 生成候选背景 BN .10 图 2.2 三种方法在前 30 帧生成的背景结果.12 图 2.3 三种方法对第 43 帧图像的检测结果.12 图 2.5 初始背景模型和三种背景更新方法的结果.15 图 2.6 提取的 3 帧原始图像.16 图 2.7 三种方法对背景突变区域的更新结果及车辆检测结果 .16 图 3.1 帧间差分法算法过程.19 图 3.2 背景差分法算法过程.20 图 3.3 改进的背景差分法的算法流程图 .