Small target detection in coal mine underground based on improved RTDETR algorithm.

Area of Science:

• Computer Vision
• Machine Learning
• Mining Safety Technology

Background:

• Low accuracy in detecting small safety equipment (helmets, self-rescuers) in complex underground coal mine environments poses a significant risk.
• Existing object detection models struggle with the scale and contextual challenges of mine imagery.

Purpose of the Study:

• To develop a novel small object detection method tailored for underground coal mines.
• To improve the accuracy and efficiency of detecting critical safety gear in challenging mining conditions.

Main Methods:

• Proposed an improved Real-Time Detection Transformer (RTDETR) model, termed PDP-RTDETR.
• Integrated a BasicBlock-Partial Convolution (PConv) module to reduce model parameters and computation.
• Incorporated Deformable Attention to enhance feature extraction capabilities.
• Added a P2 small object detection layer to focus on shallow feature map details.

Main Results:

• Achieved a Mean Average Precision (mAP) of 56.6% on a custom dataset for small target detection.
• Demonstrated superior performance compared to Yolov5s, Yolov7-Tiny, Yolov8n, and the base RTDETR model, with improvements ranging from 5.2% to 12.1%.
• Reduced the parameter count by 2.6 million compared to the base RTDETR model.

Conclusions:

• The improved PDP-RTDETR model effectively enhances small object detection accuracy in mine scenes.
• The method provides a valuable reference for applying small object detection in safety-critical industrial environments.
• The optimized model offers improved efficiency with fewer parameters, suitable for real-world deployment.