Multi-view cooperative matching single target tracking based on fusion of detection and prediction information

Abstract:
TBD (Tracking-by-detection) is one kind of important single target tracking methods. However, the current TBD for single target tracking still has shortcomings in the following two aspects: (1) The corresponding target cannot be tracked effectively when the detector fails to detect due to its strong dependence on the detection results; (2) Its robustness is poor in complex application scenarios since the tracking strategy usually extracts the features only from a single perspective. Aiming at these problems, we improve the traditional TBD single target tracking method from the following two aspects: (1) The target prediction is first introduced to reduce the dependence on the detection results and increase the abilities of the system to screen objects; (2) A multi-view collaborative matching (MCM) mechanism is proposed to screen targets, which can extract features from different perspectives to enrich the information of targets. Comparative experiments on the large public data sets OTB100 demonstrate that the proposed method has achieved the best performance compared with several classical tracking methods.