基于视频多目标跟踪的高度测量算法  

Height Estimation Algorithm Based on Visual Multi-Object Tracking

作　　者：姜明新[1,2] 王培昌[2] 王洪玉[1] 

机构地区：[1]大连理工大学信息与通信工程学院,辽宁大连116024 [2]大连民族学院信息与通信工程学院,辽宁大连116600

出　　处：《电子学报》2015年第3期591-596,共6页Acta Electronica Sinica

基　　金：国家自然科学基金项目(No.61172058;No.61403060);中央高校自主基金(No.DC110321);辽宁省教育厅资助项目(No.L2012476);中国博士后科学基金资助项目(No.2014M551081);大连民族学院人才引进项目(No.20136212)

摘　　要：本文提出了一种基于视频多目标跟踪的高度测量算法.首先,采用码本模型检测前景,利用图割理论实现对多目标的跟踪.然后,提取每一帧中目标的头部特征点和脚部特征点,根据投影几何的约束关系计算每个目标的高度.最后,融合多帧的测量结果进行数据优化.本文提出的算法不需要对相机进行完全标定,只需计算摄像机的灭点和地平面的灭线,降低了计算的复杂度.实验结果表明,本文提出的算法具有较高的测量精度,对遮挡和运动状态变化具有较强的鲁棒性,同时,能够满足实时性要求.In this paper, we propose a height estimation algorithm based on visual multi-object tracking. Firstly, the fore- ground is detected by using codebook modeling algorithm, and multi-object tracking is performed by using graph cuts. Then, the head feature points and the feet feature points of objects are extracted in each frame,and the height of every object is computed ac- cording to projective geometry constraint. Finally, data optimization is performed by fusing multi-flame measurement results. This method does not require full camera calibration, but require computing the vanishing point of camera and the vanishing line of ground. Using this method,the computational complexity can be reduced. The experimental results dernousWate that our method has higher measurement accuracy, and is robust to occlusion and the changing of motion state. Meanwhile, it can also achieve real-time performance.

关 键 词：高度测量 视频多目标跟踪 投影几何约束 图割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]