Enhanced Pedestrian Detection in Low-Light Conditions Using Dual-Path Networks and Noise-Resilient Techniques

Abstract

Pedestrian detection in low-light environments presents unique challenges for applications such as autonomous driving and unmanned aerial vehicle (UAV) surveillance. Traditional detection methods often fall short under poor visibility, noise, and variable lighting conditions, resulting in inaccuracies and inefficiencies. This study introduced DuoLightNet, a novel dual-path framework designed to enhance pedestrian detection under low-light conditions. The primary objective was to develop a robust model capable of effectively identifying pedestrians in challenging lighting scenarios. DuoLightNet integrates two specialized networks: GlowEdgeNet, which improves edge clarity and adjusts for lighting variations, and NoiseResilientNet, which focuses on reducing noise while preserving the essential details. The model was thoroughly evaluated using the Exclusively Dark (ExDark) dataset, achieving a mean average precision (mAP) of 89.0% and processing speed of 33.5 frames per second (FPS) under balanced lighting conditions. These results indicate a significant improvement over the existing methods, with up to 15% higher mAP in low-light scenarios. In addition, DuoLightNet maintains real-time processing capabilities, emphasizing its suitability for time-sensitive applications. However, challenges remain in handling extremely low-light conditions and partially occluded pedestrians. Future research should focus on enhancing the generalization of the model across broader low-light scenarios and optimizing it for deployment in resource-constrained environments. The findings of this study contribute to the advancement of pedestrian detection technologies and offer practical solutions for improving the safety and operational efficiency of autonomous and surveillance systems