Road Surface Depression Detection Based on Improved YOLOV7

Abstract

In recent years, with the rapid development of object detection technology, the field of autonomous driving has attracted increasing attention. To address the challenge of enabling autonomous vehicles to quickly detect and avoid potholes on the road surface, a modified YOLOV7 network model has been proposed. The model incorporates an ECA attention module into the backbone and replaces the SPPCSPC in the original head with SPPFCSPC. Furthermore, an improved SIOU target regression loss function is utilized to optimize the model's training convergence speed and enhance the precision of pothole boundary localization. In scenarios where the dataset is limited, the data processing approach employs Mosaic and data augmentation techniques to improve the model's generalization and robustness. Experimental results demonstrate that, on the final test dataset, the model achieves an average precision of 92.6%, recall rate of 85%, [email protected] of 92.2%, and [email protected]:.95 of 73.9%, satisfying the real-time and accuracy requirements of pothole detection on road surfaces. In summary, by introducing the ECA attention module, the improved SIOU target regression loss function, and leveraging data augmentation techniques, the modified YOLOV7 network model significantly enhances performance in the task of pothole detection for autonomous driving applications.