A Bio-Inspired Lightweight Human Action Recognition Method Based on Human Keypoint Detection.

Area of Science:

• Computer Vision
• Biomimetic Systems
• Industrial Safety

Background:

• Human action recognition in industrial settings is challenging due to complex environments and computational demands.
• Accurate monitoring of worker postures is crucial for preventing accidents in power-grid safety.
• Resource-constrained edge devices require lightweight and efficient recognition systems.

Purpose of the Study:

• To develop a bio-inspired lightweight human action recognition framework for industrial safety monitoring.
• To address limitations in existing methods regarding biomechanical invariance, adaptability to occlusion, and edge deployment trade-offs.
• To integrate principles of human musculoskeletal motion perception into a computational model.

Main Methods:

• Proposed a framework combining an improved YOLO-Pose model with a gated recurrent unit (GRU) network.
• Incorporated bio-inspired mechanisms: polar-coordinate encoding for rotation invariance, three-stage filtering for error correction, and GRU gating for selective information propagation.
• Integrated anatomical structural constraints into the computational pipeline, moving beyond generic computer vision approaches.

Main Results:

• Achieved 95.04% accuracy on the self-constructed SKPose dataset in underground power-grid environments.
• Outperformed state-of-the-art methods ST-GCN (by 3.67%) and 2s-AGCN (by 1.94%).
• Demonstrated efficient inference speed of 48 FPS with only 8.7 million parameters, suitable for edge devices.

Conclusions:

• The proposed bio-inspired framework offers a significant advancement in human action recognition for industrial safety.
• The method provides a viable solution for reliable monitoring on resource-constrained edge devices.
• This work supports the development of biomimetic perception systems and enhances industrial safety monitoring capabilities.