Research on small target detection technology on assisted driving road based on Yolov3

Abstract

Aiming at the huge difference in the size of the bounding box of different types of targets on the road in natural traffic scenes. The existing original target detection algorithm yolov3 cannot well balance the detection accuracy of large and small targets, so the target detection algorithm based on yolov3 is redesigned. Firstly, the detection module is improved and designed, and a new feature output module for small targets is added to obtain a new road target detection method yolov3 with four detection scales_ T. At the same time, focal loss is added to the loss function to reduce the result error caused by the imbalance of positive and negative samples [1]. The results show that the improved yolov3_ The average detection accuracy of T algorithm on the data set mixed with bdd100k and Kitti is 0.465, which is 0.048 higher than that of the original yolov3, especially for small targets.