Occlusion-robust online multi-object visual tracking using a GM-PHD filter with CNN-based re-identification

We propose a novel online multi-object visual tracker using Gaussian mixture Probability Hypothesis Density (GM-PHD) filter and deep appearance learning. The GM-PHD has linear complexity with the number of objects observations while estimating states cardinality time-varying objects, however, it is susceptible to miss-detections does not include identity objects. use visual-spatio-temporal information obtained from object bounding boxes deeply learned representations perform estimates-to-tracks data association for target labeling as well formulate an augmented likelihood then integrate into update step filter. also employ additional unassigned tracks prediction after overcome susceptibility towards caused by occlusion. Extensive evaluations on MOT16, MOT17 HiEve benchmark sets show that our significantly outperforms several state-of-the-art trackers in terms tracking accuracy identification.