Improved YOLOv5 Network for Real-Time Object Detection in Vehicle-Mounted Camera Capture Scenarios

Abstract

Object detection in the process of driving is a convenient and efficient task. However, due to the complex transformation of the road environment and vehicle speed, the scale of the target will not only change significantly but also be accompanied by the phenomenon of motion blur, which will have a significant impact on the detection accuracy. In practical application scenarios, it is difficult for traditional methods to simultaneously take into account the need for real-time detection and high accuracy. To address the above problems, this study proposes an improved network based on YOLOv5, taking traffic signs and road cracks as detection objects and conducting separate research. This paper proposes a GS-FPN structure to replace the original feature fusion structure for road cracks. This structure integrates the convolutional block attention model (CBAM) based on bidirectional feature pyramid networks (Bi-FPN) and introduces a new lightweight convolution module (GSConv) to reduce the information loss of the feature map, enhance the expressive ability of the network, and ultimately achieve improved recognition performance. For traffic signs, a four-scale feature detection structure is used to increase the detection scale of shallow layers and improve the recognition accuracy for small targets. In addition, this study has combined various data augmentation methods to improve the robustness of the network. Through experiments using 2164 road crack datasets and 8146 traffic sign datasets made by LabelImg, compared to the baseline model (YOLOv5s), the modified YOLOv5 network improves the mean average precision (mAP) result of the road crack dataset and small targets in the traffic sign dataset by 3% and 12.2%, respectively.