AUNet: Adaptive UpdateNet for Dynamic Pedestrian Tracking with Short-Term Occlusion

This paper investigates the pedestrian tracking problem with short-term occlusion and proposes an adaptive UpdateNet (AUNet) framework as the improvement over state-of-the-art. We find that the templates in the original UpdateNet framework are usually ‘corrupted’ when the tracking target is temporarily occluded, and hence, it is very difficult to recover when the target reappears in the following image frames. We therefore construct an evaluation network, which determines whether the current input image frame is occluded or not, and hence adapt the current image frame's influence in the update process of the object tracking template. In this way, we are able to successfully maintain a good set of templates, and hence improve the network's tracking performance when facing short-term occlusions. Experimental results with the canonical dataset (OTB100) and a real world use case show the superior tracking performance of AUNet over state-of-the-art.