Deep spatio-temporal graph convolutional network for police combat action recognition and training assessment.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomechanical Analysis

Background:

• Traditional police combat training relies on subjective human evaluation, leading to inconsistent feedback and incomplete analysis of complex movements.
• Existing automated systems lack the robustness to handle diverse environmental conditions and provide comprehensive quality assessments.

Purpose of the Study:

• To develop an enhanced deep spatio-temporal graph convolutional network (ST-GCN) for automated police combat action recognition and quality assessment.
• To improve the objectivity, consistency, and comprehensiveness of police combat training evaluation.

Main Methods:

• Implemented an enhanced ST-GCN framework with adaptive graph topology learning and multi-modal fusion (skeletal and RGB data).
• Incorporated attention-guided feature extraction, curriculum learning, and real-time processing capabilities.
• Developed comprehensive quality assessment algorithms for technique execution evaluation.

Main Results:

• Achieved 96.7% recognition accuracy across twelve police combat action categories.
• Demonstrated real-time processing at 42.8 frames per second.
• Successfully assessed action completion, standardization compliance, and movement fluency.

Conclusions:

• The proposed ST-GCN framework significantly enhances police combat training through automated, objective, and detailed performance evaluation.
• Enables data-driven curriculum development and improves training effectiveness for law enforcement.
• Offers a standardized approach to skill assessment, overcoming limitations of traditional methods.