A Novel Framework for Vehicle Detection and Tracking in Night Ware Surveillance Systems

This paper presents a novel framework for vehicle detection and tracking in night-ware surveillance systems, addressing challenges such as low visibility, fog, and low-light conditions. The framework begins with the collection of a vehicle dataset from Google, followed by the use of the Ground Truth Labeler App to annotate the dataset, providing ground truth labels for each vehicle detector. The input images undergo defogging and low-light enhancement processes to improve visual clarity. Enhanced images are stored alongside their corresponding ground truth labels for further processing. A Convolutional Neural Network (CNN) is then designed to extract features, with training options tailored for optimal performance. The YOLOv2 object detector is employed to detect vehicles within the dataset, utilizing custom-designed layers and training options. The system effectively detects vehicles and tracks their movements in real-time, even in challenging night conditions. Additionally, it displays the total count of detected vehicles for monitoring purposes. Extensive testing demonstrates the robustness and accuracy of the proposed framework, making it suitable for applications in night-time surveillance, traffic monitoring, and security systems. The integration of image enhancement techniques with YOLOv2 ensures reliable vehicle detection under challenging conditions, offering a significant improvement in night-ware surveillance.

Index Terms—Low-light sand-dust video, adaptive dynamic brightness correction, rolling guidance filter, dual-threshold interframe detection strategy.