Tracking Algorithm Research Based On The Object Detection

With the rapid development of object detection and recognition, the tracking algorithms based on detection has become state-of-art tracking algorithms. We propose compressive tracking based on a particle filter framework to solve the drifting problem. On the other hand, we propose a multi-target tracking algorithm based on object detection and the algorithm is good at handling occlusion problem.We first introduce some classic tracking algorithms and through the experimental analysis we find most of online learning algorithms are suffered from drifting problem and the causes are analyzed.We propose to incorporate the particle filter as the motion model of compressive tracking. Take use of the advantages of particle filter, we can find object again when drifting problem happens. We consider the correct rate of the classifier corresponding to each feature to improve the discriminative power of the naive Bayes classifier.We do research on multi-target tracking based on object detection. The new appearance and motion model are proposed. A greedy data association algorithm is proposed. A two-steps association strategy based on tracklet confidence is adopted and the definition of the tracklet confidence is modified. We prove the competitive performance of our algorithm when faced with occlusion, missing and inaccurate detections.