Accelerated Object Detection for Autonomous Driving in CVIS Based on Pruned CNN Architecture

While processing delay has become the primary consideration for future Cooperation Vehicle Infrastructure Systems (CVIS), different enhanced methodologies based on pruned network has widely studied. More efficient pruned architecture can be accomplished by considering prior information like motion feature prediction (optical flow, etc.) to achieve optimized feature computing. With consideration of continuous features backward propagation methods, an accelerated neural network pruning method is proposed in this paper, to achieve real-time pruning based on prior information in CNN inference time. Furthermore, proposed method can fully utilize the information from CVIS to maximize pruning ratio in feature space from the CVIS monitoring stream, with significant improvement for recall index, a more suitable performance evaluation index for object detection in autonomous driving. With comparison to general detection framework, like YOLOv3 based on darknet-52, more than 20% processing speed gain can be significantly achieved in our proposed method, with minimum loss of recall index. At the same time, the basic recall rate has also been increased by 13.7%.