Vehicle Multi-target Detection in Foggy Scene Based on Foggy env-YOLO Algorithm

The images in foggy scenes exhibit poor contrast, reduced saturation, tonal shift and loss of detail, resulting in low accuracy and poor real-time detection of autonomous vehicles, which in turn seriously affects autonomous driving safety and road safety. Therefore, this paper proposes a foggy env-YOLO detection method, which pre-processes the dense fog images with stylisation, highlights the detection target features of the dense fog images, extracts its edge feature map and sends it to the detection model for training as an input along with the light and medium fog images. In addition, the Convolutional Block Attention Module (CBAM) is added to the Cross Stage Partial (CSP) structure of Backbone feature extraction network of Yolov5 to improve the feature extraction ability and thus enhance the detection ability of fog-weather model. Experiments show that in the mixed fog concentration datasets, the improved detection method achieves a detection rate of 93.6% for Mean Average Precision (MAP) and 96.2% for the most common bus, which is better than the detection effect of the same type of Faster-Rcnn, YOLOv5 network under foggy conditions and has better applicability.