Research on application of improved YOLO V3 algorithm in road target detection

Abstract

Due to the lack of average accuracy and missed detection in the process of real road scene target detection through YOLO V3 network, the improvement scheme is put forward. The K-means clustering algorithm is used to replace the K-means clustering algorithm in the original network to analyze the anchor number and aspect ratio of the Udacity data set, in order to make the obtained parameters more suitable; in addition, in order to improve the performance of the road target detection algorithm, the existing network output is upgraded, and a 104 x104 feature detection layer is added, and the feature map output by 8 times sampling can be output by 2 times up sampling, and 4 the feature maps of down-sampling are stitched together, and the 104 x 104 feature maps obtained can effectively reduce the disappearance of features. Through the experimental results, we can see that compared with the improved YOLOV3 algorithm, the average detection accuracy of the improved algorithm for road target detection is quite high to 3.17%, and the missing detection rate is reduced by 5.62%.