Feature enhancement based detection of traffic signs in foggy environment
Ruotong Wei, Jinzhao Zhang, Xiao Lu Wang, Ruihong Zou · 2023
Aiming at the problem of missing detection and low accuracy of traffic signs in a foggy environment, the original YOLOv5 target detection algorithm was improved. First, to reduce the loss of traffic signs in the Foggy image in the process of deep convolution network transmission, replace the Backbone end convolution with the Transformer module, established the YOLOv5-Transformer model; Secondly, to enrich the semantic information in the shallow network, by joining the Path Aggregation Networks (PAN) to achieve semantic information fusion with the deep network; Further, the YOLOv5- 4detect model is proposed, which added small size detection head to the original YOLOv5 network model to detect the fused feature maps . The results showed that the YOLOv5-Transformer model and YOLOv5-4 detect model have improved the testing performance of traffic signs by 5.5% and 9.3% in sunny environment; The detection results of traffic signs in foggy environment increased by 9.4% and 15.8%. The experiments illustrate that it is more effective in foggy environments by adding fusion channels than introducing a self-attention mechanism via Transformer. YOLOv5-4detect can significantly improve the performance of YOLOv5 in traffic signs detection.