An Improved Lightweight Network Based on YOLOv5s for Object Detection in Autonomous Driving

Abstract

Object detection with high accuracy and fast inference speed based on camera sensors is important for autonomous driving. This paper develops a lightweight object detection network based on YOLOv5s which is one of the most promising object detection networks in the current literature. Our proposed network not only strengthens the object positioning ability in Cartesian coordinates without using space transformation, but also reinforces the reuse of feature maps from different scales of receptive fields. To show its robustness, different driving datasets are used to evaluate the effectiveness of our proposed network. The detection performances on objects with different sizes and in different weather conditions are also examined to show the generalizability of our proposed network. The results show that our proposed network has superior object detection performances in the conducted experiments with a very high running speed (i.e., 75.2 frames per second). This implies that our proposed network can serve as an effective and efficient solution to real-time object detection in autonomous driving.