Road Condition Detection Based on Deep Learning YOLOv5 Network
Zhaojie Lu, Leiyi Ding, Zhiyuan Wang, Linjie Dong, Zebang Guo
Abstract
Detecting road conditions is crucial for autonomous driving and requires high accuracy and real-time performance, particularly for detecting cracks and vehicles. While YOLOv5 is a general-purpose object detection algorithm, it may not achieve the desired results when applied to crack and vehicle detection due to issues such as sample imbalance and the presence of many small objects in the dataset. This paper presents an enhanced YOLOv5-based model that addresses the issues of sample imbalance and the presence of many small objects in crack and vehicle detection datasets. The proposed approach involves data augmentation to preprocess input images and adjustments to the YOLOv5 network architecture to improve classification performance for large objects. The feature fusion component of YOLOv5 is also modified to better handle small objects. Experimental results demonstrate that the enhanced model achieves a mean average precision (mAP) of 64.5%, compared to 62% for the original YOLOv5 model, while maintaining real-time performance at 150 FPS. Precision and recall also improved, from 69.3% to 71.4% and from 54.5% to 55.6%, respectively. The enhanced YOLOv5 model exhibits superior performance in detecting small objects and improves vehicle detection and resolution while preserving the real-time capabilities of the original YOLOv5 model.