Multi-object tracking based on behaviour and partial observation

基于行为和部分观测的多目标跟踪(英文)

To cope with multi-object tracking under real-world complex situations, a new video-based method is proposed. In the detecting step, the moving objects are segmented with the third level DWT （discrete wavelet transform ）and background difference. In the tracking step, the Kalman filter and scale parameter are used first to estimate the object position and bounding box. Then, the center-association-based projection ratio and region-association-based occlusion ratio are defined and combined to judge object behaviours. Finally, the tracking scheme and Kalman parameters are adaptively adjusted according to object behaviour. Under occlusion, partial observability is utilized to obtain the object measurements and optimum box dimensions. This method is robust in tracking mobile objects under such situations as occlusion, new appearing and stablization, etc. Experimental results show that the proposed method is efficient.

针对复杂环境下的多目标跟踪问题提出了一种新的跟踪方法.在检测部分,采用第三级离散小波变换和背景差分进行了目标分割.在跟踪部分,首先利用卡尔曼滤波器和缩放因子估计目标在下一帧中的中心位置和矩形框尺寸,然后在中心关联和区域关联的基础上提出了投射率和遮挡率的概念,并结合投射率和遮挡率推断目标行为.最后针对具体目标行为,自适应地调整跟踪方案和卡尔曼参数实现了多目标跟踪.在遮挡情况下利用部分观测调整估计框以获得目标测量值和最优框尺寸.提出的方法对处于遮挡、新出现以及稳定等情况下的运动目标均具有鲁棒的跟踪性能.实验结果表明提出的方法是有效的.