Visibility Prediction Based on Landmark Detection in Foggy Weather

Abstract

Visibility prediction has important application value in transportation, aviation, military, agriculture and other fields. Traditional visibility monitoring instruments have the disadvantages of complex operation, expensive price and low accuracy. At the same time, the existing visibility prediction methods based on video images also have the disadvantages of low accuracy. In this paper, a visibility prediction method in foggy weather based on deep learning is proposed. The proposed method uses target detection network YOLOv5 (YOU ONLY LOOK ONCE V5) to detect the ground landmarks in the video, then establishes a camera imaging model to calculate distance of the ground landmarks, and finally obtains the visibility value according to the distance of the farthest landmark that can be detected. In the proposed method, CIOU _Loss is selected as the loss function of YOLOv5 to improve the convergence speed and prediction accuracy. The experimental results show that thanks to yolov5's powerful and fast detection capabilities, after the training of 200 epoches, the proposed method can detect landmarks in foggy images with a 100% recall rate, which has the advantages of low cost and high accuracy.