Pedestrian Motion Prediction with Improved ADNet Model

Abstract

Motion prediction obtains pedestrian moving direction, which is fundamental control parameters for robot tail following. In this paper, a tracker named ADNet-PMP is proposed for pedestrian motion prediction. The ADNet model is improved with interlace sampling and optimized with model- update mechanism. The network is pre-trained with deep reinforcement learning and supervised learning to track the pedestrian by moving the bounding box sequentially. The movements of bounding box are transformed to actual motion behaviors with a prediction strategy. According to the results on OTB-100 datasets, ADNet-PMP achieves 1.6 times speed enhancement while keeps competitive accuracy against original ADNet. Experiment on pedestrian motion videos validates the effectiveness of motion prediction.