Improving Performance of DeepCC Tracker by Background Comparison and Trajectory Refinement

Abstract

DukeMTMCT is the largest and most completely labeled dataset in Multi-Target Multi-Camera Tracking (MTMCT). We investigate a state-of-the-art work on DukeMTMCT named DeepCC, and dig out two main problems. The first problem is that the openpose is prone to false detection, which seriously affects performance. The second problem is that two different persons may be assigned with the same ID. According to the corresponding problems, we not only propose a method to measure the similarity between detected bounding box and its original background avoiding false detection caused by OpenPose, but also design a strategy to correct the tracking trajectories which are affected by the unreliability of the correlation matrix clustering method proposed by DeepCC. Our method outperforms the state-of-the-art on DukeMTMCT.