Moving vehicle detection and tracking under hazy environment for traffic surveillance system

Vehicle location is crucial for transportation and computer vision. Bounding boxes distinguish vehicles, which is crucial for real-time movement estimation applications requiring precise area data. This study presents an adaptive approach for accurate vehicle detection and tracking in challenging scenarios such as heavy traffic, poor visibility, and adverse weather conditions. The proposed method integrates fuzzy subtraction and gradient partial equation (FGPE) techniques for background subtraction, overcoming fluctuations and shadows. It uses energy and histogram-oriented gradient features, chosen through recursive feature elimination, to improve discrimination capability. Further, a normalisation-based attention module (NAM) is integrated into the Enhanced YOLOv5 model for vehicle detection. The Multi-Object-based DeepSORT algorithm for vehicle tracking enhances the feature extractor. Deployment on edge devices achieves a traffic flow detection accuracy of 0.98%. Evaluation metrics including multiple-object tracking algorithm (MOTA) and multiple-object tracking precision (MOTP) validate the effectiveness of the proposed model for real-world traffic surveillance systems.