A multi-target tracking algorithm based on feature point trajectories

In a continuous video stream, the multi-target tracking task is to determine the positions of the concerned targets in each frame. However, the tracking algorithm suffer from many challenging issues, such as appearance variation, lighting change, occlusion and cluttered background. Especially, occlusion has the most negative impact on tracking performance. Therefore, a tracking algorithm is proposed based on feature point trajectory to solve the tracking problem where multiple targets may occlude each other. The main idea of the proposed tracking algorithm is to introduce the delay during tracking, and acquire future N frame images in advance when processing the current frame; extract feature points from the obtained frame images and connect them to form feature trajectories, and estimate the positions of targets after N frames according to the obtained trajectories. After predicting the future positions of the targets, the motion of targets can be analyzed so as to precisely determine their locations at the current frame. Experiments show the this algorithm can effectively deal with occlusion. Moreover, the complexity of the proposed algorithm is lower than that of many traditional algorithms, which guarantees real-time tracking on the low-end processor in actual applications.