一种基于运动检测的行人多目标跟踪算法  

A MULTI-TARGET PEDESTRIAN TRACKING ALGORITHM BASED ON MOTION DETECTION

作　　者：邹薇[1] 赵勋杰[1] 李权[1] 陆凯[1] 

机构地区：[1]苏州大学物理科学与技术学院,江苏苏州215006

出　　处：《计算机应用与软件》2014年第8期132-135,共4页Computer Applications and Software

基　　金：国家自然科学基金项目(61170124)

摘　　要：针对固定摄像头下的行人跟踪问题,提出一种基于运动检测的多目标跟踪算法。在运动目标检测中,先利用没有目标出现的视频帧建立背景图像,然后计算背景差并计算差分图像的梯度以提高运动区域的检测能力,最后利用区域合并法得到完整的运动目标区域,并间隔一定的时间更新背景模型。在检测到运动目标后转入跟踪。在跟踪目标时,对于有遮挡和没有遮挡的情况分开处理:若目标之间未发生遮挡,基于目标的中心距和加权的颜色直方图特征进行匹配跟踪;当发生遮挡时,用卡尔曼滤波器预测目标的位置。实验证明,相对于传统的基于背景差法的多目标跟踪,该算法能提取更完整、准确的目标区域,对行人这一非刚性目标能实现较好的跟踪。In this paper,we propose a motion detection-based multi-target tracking algorithm in light of pedestrian tracking with fixed camera. In moving target detection,the video frame with no targets appearance is used first to set up the background image; then the background subtraction is calculated together with the calculation of differential image gradient in order to improve motion region detection ability; finally the entire region of moving targets is derived by utilising the region merging method,and the background model will be refreshed at certain interval. After the moving targets are detected,it will turn to tracking. While tracking the targets,the situations will be dealt with separately depending on occlusion or not： When there is no occlusion between the targets,the matched tracking will be carried out based on the centre distances of targets and the weighted colour histogram feature; Otherwise,the Kalman filter will be used to predict the location of the occluded targets. Experiment proves that this algorithm can extract more complete and accurate target regions than the traditional background subtraction-based multi-target tracking algorithm,and has better performance in tracking pedestrians which are the nonrigid targets.

关 键 词：目标检测 边缘检测 多目标跟踪 卡尔曼滤波 加权颜色直方图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]