基于异步隐马尔可夫的视频多目标数据关联  

Data Association for Video Multi-target Based on Asynchronous Hidden Markov

作　　者：秦鹏 陈莹[1,2] 

机构地区：[1]江南大学轻工过程先进控制教育部重点实验室,江苏无锡214122 [2]上海交通大学系统控制与信息处理教育部重点实验室,上海200240

出　　处：《控制工程》2016年第7期1076-1079,共4页Control Engineering of China

基　　金：国家自然科学基金(61104213);江苏省自然科学基金(BK2011146)

摘　　要：针对视频中由于丢帧或遮挡造成的不完整目标轨迹,利用异步隐马尔可夫模型(AHMM)进行多目标数据关联。通过在HSV颜色空间采用背景差分法获取视频中多目标位置,并从中提取目标位置及方向角特征作为观测值以提高关联算法对于目标位移扰动的鲁棒性。将多目标的数据关联问题转化为轨迹的识别分配问题,通过异步隐马尔可夫模型的新增时间标示和时变转移矩阵,提高了对不完整轨迹的建模和数据关联的准确性。实验结果表明,当视频跟踪过程中发生数据丢失时,相对于经典的联合数据关联算法和隐马尔可夫算法,所提出的多目标数据关联算法具有较低的关联错误率。For multi-target data association, the asynchronous hidden Markov model is used to deal with the incomplete trajectory caused by the video frame loss or occlusion. Multi-target trajectory locations are obtained using the background subtraction method in which the background model is generated in HSV colour space. To improve the robustness of the association algorithm for target shift perturbation, the target position and direction angle are extracted as the observation value. Then, the data association problem is converted to the trajectory recognition and allocation problem. The accuracy of the incomplete trajectory＇s modeling and data association is improved by the additional new timestamp and the time-varying transition matrix of the asynchronous hidden Markov model. Experiments indicate that when data is lost during the video tracking, the proposed method has lower association error than the classic joint data association algorithm and the hidden Markov model.

关 键 词：空、丢帧 遮挡 数据关联 异步隐马尔可夫模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]