基于HEVC压缩域的车辆行为事件检测研究

发布时间：2018-05-26 04:49

本文选题：压缩域 + HEVC　；参考：《南京邮电大学》2017年硕士论文

【摘要】：目前,人们对车辆行为事件检测的研究主要集中在像素域,与运动目标检测与运动目标跟踪的技术研究也主要集中像素域。虽然有些学者基于早期的视频编码标准进行了压缩域运动目标分割检测的研究,但也都是套用像素域的运动目标检测处理技术。本文对最新的HEVC视频编码标准进行了分析,基于压缩码流中的运动矢量、编码单元模式以及预测残差等信息,提出了一套基于HEVC压缩域的车辆行为事件检测方法。论文的研究工作分为三个部分:一、运动车辆目标检测。以运动矢量、编码单元划分和预测残差数据作为车辆目标特征信息,提出一种基于阈值的运动车辆目标检测算法。算法首先对HEVC视频压缩码流进行解析,提取码流中的运动矢量、编码单元划分模式、预测残差等信息,然后对运动矢量场进行预处理和阈值处理,进而检测出运动车辆目标,实现运动车辆目标检测。运动矢量场预处理主要包括帧内编码单元虚拟运动矢量分配、运动矢量标准化、中值滤波等预处理过程。论文提出一种使用极坐标中值算法为帧内编码单元分配虚拟运动矢量的方法,减少了帧内编码单元被分配错误的虚拟运动矢量。实验证明,在摄像机相对固定的道路监控视频中,提出的算法能够较好的检测出运动车辆目标。二、运动车辆目标跟踪。以HEVC压缩域运动矢量作为车辆目标的运动信息和特征信息,提出一种基于区域匹配的目标车辆跟踪算法。算法使用当前帧运动矢量前向映射方法,对相邻视频中目标车辆位置进行预测,然后选取运动矢量方向角直方图作为匹配特征,在预测区域内进行车辆目标区域匹配跟踪,最终实现目标车辆的跟踪。实验证明,在背景相对固定的道路监控视频中,提出的算法能够较有效的实现运动车辆目标的跟踪。三、车辆行为事件检测。论文使用基于区域匹配的运动车辆目标跟踪结果提取出连续视频帧中车辆目标的中心坐标,通过车辆行驶轨迹拟合分析,提出了压缩域中车辆行为事件检测判断标准,包括车辆运动方向检测、逆行检测和掉头检测等车辆行为事件的检测。另外,提出了一种基于HEVC码流的停车事件检测算法,该算法不需要进行车辆目标检测与跟踪,仅仅利用编码单元的运动矢量信息进行停车事件判断。实验证明,提出的车辆行为事件检测能够有较好的实现效果,而基于HEVC码流的停车事件检测算法也能够较好的检测出监控视频中包含的停车事件。
[Abstract]:At present, the research on vehicle behavior event detection is mainly focused on pixel domain, and the technology of moving target detection and moving target tracking is also focused on pixel domain. Although some scholars have studied the segmentation and detection of moving targets in compressed domain based on the earlier video coding standards, they are also applied in pixel domain to detect and process moving targets. In this paper, the latest HEVC video coding standards are analyzed. Based on the information of motion vector, coding unit mode and prediction residuals in compressed bitstream, a set of vehicle behavior event detection method based on HEVC compressed domain is proposed. The research work is divided into three parts: first, moving vehicle target detection. Based on motion vector, coding unit partition and prediction residual data as vehicle target feature information, a threshold-based moving vehicle target detection algorithm is proposed. The algorithm firstly parses the HEVC video compressed bitstream, extracts the motion vector, the coding unit partition mode, the prediction residual information and so on, then preprocesses the motion vector field and the threshold processing, then detects the moving vehicle target. The target detection of moving vehicle is realized. The preprocessing of motion vector field mainly includes the preprocessing process of virtual motion vector allocation, standardization of motion vector, median filter and so on. In this paper, a method of allocating virtual motion vectors to intra coding units using polar coordinate median algorithm is proposed, which reduces the virtual motion vectors that the coding units are assigned incorrectly. Experimental results show that the proposed algorithm can detect moving vehicle targets well in the relatively fixed video of road surveillance. Second, moving vehicle target tracking. A vehicle tracking algorithm based on region matching is proposed by using the motion vector of HEVC compressed domain as the motion information and feature information of vehicle targets. The algorithm uses the current frame motion vector forward mapping method to predict the vehicle position in the adjacent video, and then selects the direction histogram of the motion vector as the matching feature to track the vehicle target region in the prediction region. Finally, the target vehicle tracking is realized. Experimental results show that the proposed algorithm can effectively track moving vehicle targets in road surveillance video with relatively fixed background. Third, vehicle behavior event detection. In this paper, the moving vehicle target tracking results based on region matching are used to extract the center coordinates of the vehicle target in the continuous video frame, and the vehicle behavior event detection and judgment standard in the compressed domain is proposed by fitting and analyzing the vehicle trajectory. Including vehicle motion direction detection, retrograde detection and reverse detection of vehicle behavior events detection. In addition, a parking event detection algorithm based on HEVC code stream is proposed. The algorithm does not need to detect and track the vehicle object, but only uses the motion vector information of the coding unit to judge the parking event. Experimental results show that the proposed vehicle behavior event detection can achieve better results, and the parking event detection algorithm based on HEVC code stream can also detect the parking events contained in the surveillance video.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【相似文献】