A Visual Trajectory-Based Method for Personnel Behavior Recognition in Industrial Scenarios

Area of Science:

• Computer Vision
• Artificial Intelligence
• Industrial Safety

Background:

• Industrial environments present complex challenges for accurate personnel behavior recognition, impacting safety and asset protection.
• Existing methods often struggle with the accuracy required for real-world industrial applications.

Purpose of the Study:

• To develop a novel, lightweight framework for robust and efficient real-time personnel behavior recognition in industrial settings.
• To enhance the accuracy of behavior recognition despite complex environmental conditions.

Main Methods:

• Enhanced YOLOv8n model with Receptive Field Attention Convolution (RFAConv) and Efficient Multi-scale Attention (EMA) mechanisms.
• BOT-SORT algorithm for continuous motion trajectory generation and perspective transformation for bird's-eye view creation.
• Random Forest model for classifying discriminative trajectory features.

Main Results:

• Achieved a 6.9% increase in AP50 and a 4.2% increase in AP50:95 over the baseline YOLOv8n model.
• Generated high-fidelity bird's-eye view trajectories through geometric correction.
• Attained F1-scores exceeding 82% for all behaviors on a proprietary industrial dataset.

Conclusions:

• The proposed framework offers a robust and efficient solution for real-time personnel behavior recognition in challenging industrial environments.
• The integration of attention mechanisms and trajectory analysis significantly improves recognition accuracy.
• Future work will involve exploring advanced algorithms and edge device validation.