An Integrated Deep Learning Framework for Occluded Pedestrian Tracking

Abstract

Numerous object-tracking and multiple-person-tracking algorithms have been developed in the field of computer vision, but few trackers can properly address the issue of when a pedestrian is partially or fully occluded by other objects or persons. In order to achieve efficient pedestrian tracking in various occlusion conditions, a pedestrian tracking framework is proposed and developed based on the deep learning networks. First, a pedestrian detector is trained as a tracking mechanism based on the Faster R-CNN, which narrows the search range and efficiently improves accuracy, as compared with the traditional gradient descent algorithm. Second, in the process of target matching, a color histogram and scale-invariant feature transform are combined to provide the target model expression, and a full convolution network (FCN) is trained to extract the pedestrian information in the target model, based on an FCN image semantic segmentation algorithm that can remove background noise effectively. Finally, the extensive experiments on a commonly used tracking benchmark show that the proposed method achieves better performance than the other state-of-the-art trackers in various occlusion situations.