LoLTV: A Low Light Two-Wheeler Violation Dataset With Anomaly Detection Technique

Abstract

Detecting traffic violations is essential for improving road safety, ensuring rule compliance, and maintaining smooth traffic flow. It also aids in holding violators accountable and supports data-driven decision-making for infrastructure enhancements. To address these challenges, the integration of AI-based methods for automated violation detection is increasingly vital, reducing the need for manual oversight. Lowlight conditions pose additional difficulties, as violations become harder to detect. In this study, we created a novel dataset containing 1032 images with 1475 two-wheeler violations under low-light conditions. We propose a real-time deep learning system using YOLO-v8 for two-wheeler violation detection. Our system addresses the challenge of low-light conditions by incorporating a real-time low-light video enhancement module. Through comprehensive evaluations, our system has achieved an average precision of 98.2%, recall of 97.5%, and an accuracy of 97.05% when tested on our custom dataset. Notably, it successfully detected 172 out of 188 violations in the test dataset and exhibited 60% faster processing compared to other stateof-the-art methods. This suggests that our system not only outperforms existing methods on public datasets but also excels in terms of performance and accuracy when applied to the specifically constructed low-light traffic dataset. Furthermore, our system's practical scalability is evident through its integration with multiple devices and CCTV systems.