Pedestrian Detection and Behavior Recognition Based on Vision
Huiqin Zhan, Yuan Liu, Zhibin Cui, Hong Cheng · 2019
In the field of autonomous driving, pedestrian detection and pedestrian behavior identification are important in ensuring pedestrian safety and reducing traffic accidents. Facing the complex traffic scene of the city, we propose a vision-based, high-accuracy and real-time pedestrian detection and pedestrian behavior recognition method for the on-vehicle environment. The method firstly uses the YOLOv3-TINY network to accurately and quickly identify pedestrians, and then uses the improved Deepsort algorithm to track the identified pedestrians. An improved Alexnet was then proposed to identify pedestrian behavior and train on data sets adapted to the study based on the mars dataset. After getting the pedestrian behavior, the behavior is corrected by multi-target tracking, and finally the real-time warning level information is given. Compared with similar algorithms, the experimental results show that the algorithm proposed in this paper improves the recognition accuracy and real-time performance, and the frame rate reaches 20frame/s.