Campus risk detection using the S-YOLOv10-SIC network and a self-calibrated illumination algorithm

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Campus safety requires accurate and efficient risk detection systems.
• Existing object detection models face challenges in low-light environments and computational efficiency.

Purpose of the Study:

• To develop an enhanced YOLOv10 algorithm for intelligent campus risk detection.
• To improve detection accuracy, especially under low-light conditions, and reduce computational complexity.

Main Methods:

• Integration of a self-calibrated illumination algorithm with YOLOv10.
• Optimization of the loss function using an auxiliary bounding box.
• Enhancement of the network structure with StarNet and the Convolutional Block Attention Module.

Main Results:

• Reduced classification loss by ~20% and feature point loss by ~16%.
• Decreased Parameters, Gradients, and GFLOPs by over 80%.
• Improved Precision (+0.99%), Recall (+3.31%), F1 (+2.15%), and mAP (+1.23%).

Conclusions:

• The improved YOLOv10 algorithm demonstrates superior performance in campus risk detection, particularly in low-light scenarios.
• The model offers significant improvements in efficiency and accuracy, supporting the development of smarter campus environments.