Vision-Based Vehicle Behavior Analysis: A Structured Learning Approach via Convolutional Neural Networks

With the maturity of vision-based vehicle detection and tracking, vision-based behavior analysis of on-road vehicles has emerged as an active research field, which sheds light on the environmental perception of autonomous driving and intelligent traffic monitoring. In this paper, we are committed to predicting vehicle behavior by incorporating the structure information of vehicle behavior into the learning process. Inspired by structured learning, the structure information is extracted from the detected vehicle as its corresponding structured label, which visually expresses the vehicle behavior as contrast to the discrete numeral label. With the structured label, a structured convolutional neural networks (SCNN) method is constructed to predict the vehicle behavior. As for performance evaluation, recognition accuracy and contour similarity are used. Experimental results demonstrate that the proposed method achieves better recognition accuracy than traditional methods of discrete labels, while the learned structured labels have the implication of semantic interpretation of on-road vehicle behavior.