GSC-YOLO: A Pedestrian Detection Method for Low-Light Security Surveillance Scenarios

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Low-light conditions degrade visible-light images, causing issues like noise and loss of detail, which hinders pedestrian detection.
• Existing multimodal solutions (e.g., RGB-infrared) are costly and complex, limiting their use in resource-constrained environments.
• Efficient pedestrian detection in monocular RGB low-light scenarios is crucial for intelligent perception systems.

Purpose of the Study:

• To develop an efficient and structurally optimized pedestrian detection model for low-light monocular RGB scenarios.
• To enhance feature representation and fusion techniques to overcome limitations of low-light imaging.
• To provide a practical solution for cost-sensitive and lightweight deployment in intelligent perception.

Main Methods:

• Proposed GSC-YOLO model built upon YOLOv13, incorporating GhostNetV3 as the backbone for improved multi-scale feature representation.
• Introduced a Semantic-Spatial Alignment (SSA) module to enhance information compensation and noise suppression during feature fusion.
• Integrated C2f_Faster into the high-level semantic branch to optimize information flow and reduce computational overhead.

Main Results:

• GSC-YOLO achieved superior performance on LLVIP and KAIST datasets compared to the YOLOv13 baseline.
• Achieved mAP@0.5:0.95 of 57.70% (LLVIP) and 66.61% (KAIST), with Recall values of 89.93% and 90.49%, respectively.
• Demonstrated effective improvement in pedestrian perception in low-light RGB scenes with favorable real-time inference capabilities.

Conclusions:

• The proposed GSC-YOLO model effectively enhances pedestrian detection in low-light RGB conditions.
• The model offers a practical and efficient solution for scenarios with hardware or cost limitations.
• Results suggest GSC-YOLO can serve as a valuable reference for future low-light vision sensing research.