YOLOGP: A YOLOv5-Based Lightweight Network for Efficient Vehicle Detection in Autonomous Driving Scenarios

The object detection technology holds paramount significance in realizing autonomous driving and AI-assisted driving systems. Swift and precise object detection is crucial for enhancing the safety of autonomous vehicles. However, for in-vehicle edge computing platforms, colossal models fall short of meeting real-time detection requirements, while lightweight models often compromise on detection accuracy. Addressing this issue, this paper proposes an improved real-time object detection algorithm based on YOLOv5.In the proposed method, we combine the One-Shot Aggregation (OSA) concept with the progressive channel compression idea and introduce GSConv to innovatively propose the GSPCA structure. It aims to improve some of the problems exposed by the original C3 structure, so as to enhance the model efficiency. Secondly, we also apply GSConv to the neck network of YOLOv5 and introduce the Content-Aware ReAssembly of Features (CARAFE) upsampling operator in the FPN structure, which utilizes its spatial perception and large receptive field to improve the quality of the upsampling, thus enhancing the feature fusion performance of the network. Experimental results demonstrate that, compared to the baseline, our proposed model achieves the highest improvement of 5.2% in [email protected]:0.95 on the PASCAL VOC dataset, KITTI dataset, and SODA10m dataset. Furthermore, the model's parameter count and computational load are slightly less than those of the original model.