ABSTRACT Object detection under low‐light conditions is a critical perceptual capability for autonomous robotic systems performing real‐world tasks such as autonomous driving, nighttime surveillance, drone inspection, and underwater exploration. However, image degradation in such environments severely disrupts visual feature extraction, leading to frequent false positives and missed detections. To address this challenge, we propose a novel Multi‐scale Adaptive Network (MANet) for robust object detection in low‐light scenarios. MANet comprises two main components: a cascaded feature extraction network built upon our proposed multi‐scale feature extractor, and an adaptive fusion network that integrates our adaptive feature extractor and a fast normalized fusion module. Additionally, we introduce a joint loss function to further improve classification performance in complex lighting conditions. Experimental results show that MANet achieves an mA p @0.5 of 0.718 and an mA p @0.5:0.95 of 0.451 on the ExDark data set, and also delivers competitive performance on DARKFACE, DUO, and TrashCan. In addition, MANet demonstrates strong cross‐scene generalization under real‐world low‐light conditions. These results validate the effectiveness of MANet in reducing false positives and missed detections, enhancing detection robustness, and laying a solid foundation for robotic perception and decision‐making in real‐world outdoor low‐light environments.
Zhang et al. (Tue,) studied this question.