Improving Crowd Detection Under Low-Light Conditions Through Deep Learning-Based Image Enhancement and Multi-Stage Fusion

Crowd detection under low-light conditions remains challenging due to reduced visibility, low contrast, image noise, and overlapping individuals. These factors may affect person detection accuracy and the identification of crowd regions. This study proposes a multi-stage framework that integrates deep learning-based image enhancement, object detection, and prediction fusion to improve detection performance in poorly illuminated environments. The framework combines Retinex-based illumination enhancement and deep curve estimation through pixel-level fusion, followed by adaptive gamma correction. YOLOv8x is then applied independently to the original and enhanced images. Two prediction fusion strategies, maximum-score fusion and probabilistic-sum fusion are evaluated to combine detection information from both image representations. Finally, spatial grouping is applied to person detections to identify candidate crowd regions. Experiments conducted on the ExDark dataset show that image enhancement alone does not necessarily improve detection accuracy. In contrast, maximum-score fusion achieves the best performance, with 13,550 detections, an F1-score of 0.7311, [email protected] of 0.7617, and [email protected]:0.95 of 0.4226, compared with 0.7298, 0.7498, and 0.4184, respectively, for the original-image configuration. Probabilistic-sum fusion also provides modest improvements over the original-image baseline. These findings indicate that combining predictions from original and enhanced images may improve person detection under low-light conditions, providing a basis for subsequent crowd-region identification.