Multi-Object Detection and Tracking of Foggy Vehicles Based on iMproved YOLOV5s and Optimized Deepsort

In view of the problems of ID false detection and missing detection often occur in the multi-target detection and tracking of the vehicle in front of the autonomous vehicle in hazy days, this paper proposes a new de-fog algorithm based on dark channel prior to improve the quality of the input image. CBAM-Net was embedded into the CSPDarknet53 backbone network of YOLOv5s model to improve the feature extraction capability and optimize the detection effect. A new CMTIOU loss algorithm is proposed to replace the IOU in Deepsort, and comprehensively consider the factors of time, feature and location geometry to improve the tracking effect and HOTA index. The experimental results show that the MOTA and HOTA of the proposed algorithm are 77.6% and 78.8% respectively compared with the traditional algorithm.