Online Pedestrian Multiple-Object Tracking with Prediction Refinement and Track Classification

The performance of pedestrian multiple object tracking (MOT), which is based on the tracking-by-detection framework, is exceedingly susceptible to the quality of detection, especially suffering from detection missing or inaccuracy caused by occlusion. Several studies aimed at alleviating the problem continue to perform poorly in scenarios with frequent heavy occlusions. In this study, a novel online pedestrian MOT method is proposed for targets with severe occlusion. First, a regression network is employed to refine the predicted position of the target to obtain a precise bounding box and visibility score. Considering the visibility score and the overlap between these refined bounding boxes globally, the targets that are heavily occluded are categorised into the following two types: (1) targets occluded by a non-pedestrian object and (2) targets occluded by other pedestrians. Then, these occluded targets are handled in different ways, which reduces the number of false negatives (FNs) and false positives (FPs). Finally, to enhance the precision of the prediction, a motion model that combines the Kalman filter and camera motion compensation is developed. The tracking results applied to three widely used pedestrian MOT benchmark datasets demonstrates the state-of-the-art performance of the proposed method.