Traditional YOLOv3 algorithm has the problem of inaccurate detection frame positioning. In this paper, an improved YOLOv3 algorithm was proposed to re-estimate the detection frame position and improve the positioning accuracy of smart cars in hazy traffic environment. Firstly, the collected images were preprocessed using image dehazing algorithms. Then, to improve positioning accuracy, estimated test box position was selected by constructing positioning confidence as the reference to replace classification confidence, and the non-maximum suppression (NMS) algorithm was improved. Finally, the soft-NMS was introduced, and the coordinates were updated using the weighted average. The experimental results show that compared with the original YOLOv3 algorithm, the mAP(mean average precision) of improved YOLOv3 algorithm increased by 0.44%, which could detect more targets and locate the detected vehicles more accurately in real-time detection.