Research on defect detection method of automotive running lights based on keypoint identification

As a novel type of automotive signal light, running lights are gradually being utilized in automobiles. In view of the defect characteristics of automotive running lights, this paper proposes a method for accurately locating the defective parts through the detection of keypoints. Based on YOLOv7, a keypoint detection network called YOLO-kpt specialized for automotive running lights is proposed in this paper. Firstly, the network incorporates PConv to replace some conventional convolutions, effectively reducing the number of parameters. Secondly, a new convolutional calculation module called ResPC is proposed, inspired by the advantages of residual networks, which integrates PConv. Lastly, YOLO-kpt employs two SPPCSPC modules to further enhance the network’s feature extraction capability. Experiments carried out using a self-built dataset show that the YOLO-kpt proposed in this paper is able to accurately detect the defective parts of automotive running lights with an average Euclidean distance of only 1.8383 pixels for the keypoints error.