Automatic Driving Scene Target Detection Algorithm Based on Improved YOLOv5 Network

Target detection algorithm has become more and more important in the field of automatic driving. It is the most difficult to balance the high precision target detection and high speed processing time in target detection algorithm. Therefore, this paper proposes an improved YOLOv5 algorithm to comprehensively improve the accuracy and operation speed of target detection in automatic driving scenarios. In this paper, DCN network is introduced in the neural network model to replace the CNN layer in YOLOv5, and the fine feature extraction of images is realized in the backbone feature extraction layer. In addition, in the selection of anchor points, K-means++ clustering algorithm is adopted in this paper to replace the K-means algorithm in the original algorithm to achieve faster and more accurate selection of anchor frames. At the same time, DIOU is introduced as the regression loss of the prediction box, which accelerates the convergence speed of the algorithm. The improved algorithm has improved mAP on KITTI dataset and COCO dataset. Therefore, this paper can achieve target detection that is more suitable for automatic driving scenarios.