A lightweight object detection algorithm based on YOLOv3 for vehicle and pedestrian detection

摘要

YOLOv3, the third version of the YOLO family, performs significantly well on object detection. Nevertheless, using YOLOv3 for real-time vehicle and pedestrian detection on unmanned vehicles with limited computing resources is still a very big challenge due to the high computational complexity of YOLOv3. In this paper, a new network architecture for vehicle and pedestrian detection based on YOLOv3 is proposed which is named as Lightweight-YOLOv3. Three improvements are presented in Lightweight-YOLOv3. Firstly, to reduce the model size and computing complexity, channel and layer pruning is proposed by introducing L1 regularization on the batch normalization layer. Thus, unimportant channels and layers are recognized and removed. Secondly, to reduce the missed detection in crowded scenes and locate targets better, the MergeSoft-NMS which merges the bounding boxes with high overlap is designed based on Soft-NMS. Thirdly, considering the obvious aspect ratio of vehicle and pedestrian, the anchor boxes which are designed based on multi-class is redesigned for better vehicle and pedestrian matching and localization in Lightweight-YOLOv3. In the experiment, compared with YOLOv3 and YOLOv3-tiny, LightweightYOLOv3 which performs well on detection accuracy and speed is effective and compact for vehicle and pedestrian detection.