Object detection in robot binocular vision image processing based on YOLOv4 algorithm

Abstract

To improve the recognition accuracy of pavement crack target detection robots based on image processing, this research proposes a road crack recognition model built on the improved YOLOv4 algorithm. The enhanced YOLOv4 model, which utilizes optimized K-means and attention mechanisms for improved detection, is combined with a binocular vision calibration method for accurate camera setup. Feature points are extracted and matched using an accelerated robust feature algorithm to measure crack widths, resulting in a novel binocular vision-based road crack detection system. The experiment demonstrated that the detection accuracy of the proposed enhanced YOLOv4 exceeded 97% on datasets, with recall, F1, AUC, and precision values of 0.973, 0.986, 0.985, and 0.989, respectively. The accuracy of the research model exceeded 98%, and the fit with the sample data exceeded 96%. The research methods can effectively achieve intelligent and accurate detection of road cracks, which has positive significance for road maintenance work.