关键词: 尾灯检测, YOLOv3, 特征提取, K-means++

Abstract: In the automatic driving scene, the detection of the front taillights is an extensive and significant problem. Darknet53 is the feature extraction network of YOLOv3. It uses five residual units to extract features from the original image, and uses three scale feature map for fusion prediction. The smaller the size is, the stronger the feature expression ability of large target is. Because taillight detection belongs to small target detection, this paper omits the last residual unit of Darknet53, and increases the repetition times of small-scale feature extraction residual unit. Aiming at the problems of K-means clustering algorithm which is difficult to determine K value and sensitive to the initial clustering center, this paper uses K-means+〖KG-*3〗+ clustering algorithm to obtain anchor value and combines IOU distance measurement index. The experimental results show that the accuracy and speed of taillight detection on the improved YOLOv3 network are higher than those before. The mAP is increased from 79.63% to 89.32%, and the detection time of single image is shorten from 0.014 s to 0.01 s. Compared with other mainstream target detection frameworks, the improved YOLOv3 model has superior detection performance.

Key words: taillight detection, YOLOv3, feature extraction, K-means++