A Pedestrian Trajectory Prediction Network Based on Trajectory Conditional Visual Attention

Abstract

Pedestrian trajectory prediction task can increase artificial intelligence's understanding of human behavior perception and intention, and is widely used in traffic management, service robots, automatic driving and other fields, which has great research value and significance. Although there is more and more research on pedestrian trajectory prediction tasks, pedestrian trajectory prediction is still a very challenging task due to the uncertainty of pedestrian will and the complexity of interaction in the complex real scene. Most of the research focus on the interaction between the target pedestrian and other pedestrians, which has great computational cost in the crowded pedestrian scene. We proposed CDCNet, by observing the historical track of pedestrians, and using the Conditional Dual Coordinated Attention to obtain the information that pedestrians pay attention to when moving forward, so that the network pays attention to the information around pedestrians instead of the content in the entire video, so that the network can also have a lower computing cost under the crowded pedestrian scene. Our experiments on five publicly available datasets show that CDCNet has good predictive efficiency and advanced performance.