Weighted correlation filters guidance with spatial-temporal attention for online multi-object tracking

In recent years, discriminative correlation filters based trackers have made remarkable achievements for single object tracking, while directly applying these trackers for multi-object tracking may encounter some problem in drifted results caused by occlusion and missing detection from the detector. Thus, we propose a weighted-correlation-filters framework with spatial-temporal attention mechanism for online multi-object tracking to solve the above problems. First, we use the weighted correlation filters with dynamic updating scheme to pre-track each object in the current frame, which helps to filter out the improper detection according to the position of pre-tack for each object and is capable of tracking objects of the false negative. Then, we introduce a spatial-temporal attention mechanism to produce a discriminative appearance model and calculate reliable similarity scores for data association. The proposed online algorithm achieves 48.4% in MOTA on challenging MOT17 benchmark dataset and better performance on MT and ML than some offline methods. (C) 2019 Published by Elsevier Inc.