An Algorithm for Identifying Unsafe Behaviors of Miners Based on the Improved AlphaPose

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mining Safety Engineering

Background:

• Video surveillance is crucial for coal mine safety but struggles with complex underground environments, leading to low accuracy in identifying unsafe miner behaviors.
• Challenges include chaotic backgrounds and personnel occlusion, which hinder human pose estimation and feature extraction.

Purpose of the Study:

• To develop an improved miner unsafe behavior recognition algorithm (RS-AlphaPose) that enhances accuracy in complex underground mining conditions.
• To address limitations of existing methods in pose estimation and feature extraction within challenging environments.

Main Methods:

• The RS-AlphaPose algorithm integrates an improved real-time detection Transformer (RTDETR) for enhanced target detection in complex scenes.
• It incorporates sliding window and channel attention mechanisms to improve skeleton extraction accuracy, especially with occlusion.
• A spatio-temporal graph convolution network is utilized to capture temporal features of dynamic behaviors from skeleton sequences.

Main Results:

• The algorithm achieved an average posture estimation accuracy of 72.5% on the COCO2017 dataset, surpassing the basic AlphaPose model by 2.2%.
• On a custom miner behavior dataset, RS-AlphaPose reached 94.5% average recognition accuracy for unsafe actions like climbing and crossing, a 4.5% improvement over the baseline.
• Demonstrated effectiveness in mitigating interference from complex underground environments.

Conclusions:

• The proposed RS-AlphaPose algorithm significantly improves the accuracy of dynamic unsafe behavior recognition for miners in challenging underground settings.
• It offers a reliable technical solution for enhancing safety in coal mine production through advanced video surveillance analysis.