Abstract:

In the video multi-target tracking, the joint probability data association (JPDA) algorithm involves a potentially huge number terms, which is weak for the realtime performance, when the number of the target is large. Moreover, targets are often undetected due to occlusion or other detector failures. The classic JPDA often gets the part trajectory of the objects, not the integrity trajectory. To solve these problems, a method based on m-best JPDA and tracklet association is proposed. Firstly, to reduce the computational complexity, the integer linear program is used to find the m-best hypotheses and simplify the JPDA algorithm. After that, the distances between each target trajectory are computed based on the motion evaluation by Kalman filter and the simply linearly extrapolation. The affinity propagation cluster algorithm is used to merge the tracklet of the object and get the fully trajectories. Experiments show that the proposed method still has the effective and real time performance when the number of target is large and occlusion is easy to happen.