Occlusion-aware pedestrian detection and tracking

Abstract

This paper proposes an occlusion-aware mechanism, used on a framework for detecting and tracking pedestrians in videos acquired from surveillance cameras, which includes the extraction of trajectory points, estimation of walking velocities, detection of groups, and projection of the final trajectories into a 2D plan. The occlusion-aware mechanism is introduced in order to manage irregularities in the pedestrian trajectory data derived from occlusions. This mechanism is able to identify the parts of the human body that are occluded, using skeleton data generated by human pose estimation algorithms, and adjust the dimensions of the bounding boxes of the occluded pedestrians.