Abstract: Object detection algorithm has an important application in the field of intelligent transportation, but it has two problems: low accuracy of small targets and loss of multi-scale information. To solve these problems, an improved YOLOv3＿4 d algorithm is proposed in this paper, which can achieve recognition of multi-scale targets in complex traffic scenes. Firstly, in order to solve the problem of information loss of different targets in the common algorithm, a multi-scale detection network is constructed by adding 128128 scale to the detector to mine information with different scales. Secondly, the attention residual unit and feature enhancement module are designed. The attention residual unit extracts semantic feature information of small targets, and the feature enhancement module splices multi-scale feature maps to obtain rich fusion information. Finally, the GIoU and Focal functions were introduced to accelerate the convergence speed and enhance the robustness of the algorithm. The experimental results of the BDD 100 K and VOC 2012 show that the mAP is improved by 9.2% and 4.0%, and the mAP_S of small targets is improved by 2.3% and 3.2%, which fully confirms the feasibility of the proposed algorithm.