An Improved Traffic Sign Detection and Recognition Deep Model Based on YOLOv5

In this paper, we aim at the traffic sign detection and recognition in complex road conditions. We proposed a deep model for traffic sign detection and recognition. There are a few difficulties in traffic sign detection task, such as, less recognizable, small target size, easily leading to detected failure and so on. First, for the failure detection, we introduce Coordinate Attention (CA); second, to accelerate the regression of prediction box, we introduce the angle loss into our objective function; third, for the overlapping and occlusion phenomenon of ground truth, a dynamic label assignment strategy- simple Optimal Transport Assignment (SimOTA) is utilized during label assignment; the last and the most important, for the target size problem, we propose a feature fusion network, named hierarchical-path feature fusion network (H-PFANet). Experiments were conducted on two public data sets, the results show that our improved model performed better than YOLOv5s which is the base model and other popular algorithms on precision, recall and mAP. For the difficult samples in data set CCTSDB-2021, the results show that compared to YOLOv5s, the mAP@0.5 is improved by 6.3%, the mAP@0.5: 0.95 is improved by 5.3%, and our method achieved a detection speed of 91 FPS, with better robustness to changes in various traffic scenes, while maintaining the volume of the original YOLOv5s model. On the whole CCTSDB-2021 data set, the precision of our model reached 98.1%, the recall of our model reached 97.6% and the mAP@0.5 reached 98.8% with a speed of 91 FPS. We also compared our method with other current detection algorithms on TT100K data set, the results show that our proposed method performed better, and show the effectiveness of our method.