Research on Traffic Target Detection Method Based on Improved YOLOv3
Xinyu Zhang, Sheng Ding · 2021
In order to solve the problem that it is difficult to balance the real-time and accuracy of the existing traffic target recognition methods in the field of automatic driving, an improved YOLOv3 target detection algorithm YOLO-R is proposed. The original three feature scales of YOLOv3 are increased to four to reduce the miss detection rate of small objects; K-means target frame clustering is used to get a new target detection candidate frame, which improves the detection accuracy; Through sparse training and pruning the unimportant channels in the model after sparse training, the model size can be reduced, the detection speed can be accelerated and over fitting can be prevented. On the Udacity data set used in the automatic driving algorithm competition, the experimental results show that under NVIDIA GTX 1080 Ti, the FPS of the new method is 42.42 frames/s, which is 1.44 frames/s higher than that of YOLOv3; the detection accuracy (mAP) is 93.61%, which is 3.43 % higher than that of YOLOv3. Compared with the original YOLOv3, the proposed YOLO-R has higher detection speed and accuracy in traffic environment.