Deep learning for pedestrian behavior understanding

Visual perception and road scene understanding are critical components of autonomous driving (AD). Detecting objects and predicting surrounding road user behavior are important tasks for robust and safe driving systems. With the rise of Machine Learning (ML) and Deep Learning (DL) methods using convolutional neural networks (CNN) , research on road scene understanding and motion recognition have made significant progress. In this paper, a real-time recognition of pedestrian’s intention to cross is presented as a step towards autonomous driving. Our approach is constructed using a real-Time visual pedestrian tracker based on a one-stage object detector that gave a fast and robust learning produces dynamic and static features as inputs of a linear classifier for predicting crossing behavior. We apply our approach to public real-life benchmark - JAAD dataset "Joint Attention in Autonomous Driving." The experimental results have shown that our proposed method is efficient and competitive with other approaches. Overall, our system achieves an accuracy of 92.88%.