High frame rate action classification and state tracking in competitive sports using SlowFast dual-stream mesh...

Area of Science:

• Computer Vision
• Machine Learning
• Sports Analytics

Background:

• High frame rate sports videos present challenges for motion recognition due to rapid movements and state transitions.
• Existing methods often separate semantic and dynamic features, limiting performance in complex action classification and continuous state modeling.

Purpose of the Study:

• To propose a spatiotemporal fusion method for accurate motion recognition and phase-state modeling in high frame rate table tennis videos.
• To address the limitations of current methods by integrating semantic and dynamic information effectively.

Main Methods:

• A dual-stream grid structure (SlowFast network) is employed, with Slow branch for semantic features and Fast branch for dynamic features.
• A grid feature fusion module aligns spatial information and applies channel attention.
• A Gated Recurrent Unit (GRU) models temporal state sequences for unified learning.

Main Results:

• The proposed model achieved 94.5% action classification accuracy on a primary benchmark, outperforming TimeSformer and I3D.
• Temporal consistency and state transition stability metrics showed significant improvements over the MS-TCN baseline.
• The model demonstrated robustness in multi-view cross-domain evaluations, maintaining high accuracy across different camera viewpoints.

Conclusions:

• The spatiotemporal fusion method effectively captures both semantic and dynamic information for high frame rate sports video analysis.
• The proposed model offers superior performance in motion classification and continuous state modeling compared to existing approaches.
• The grid structure and GRU-based state modeling contribute to robust and accurate recognition of complex actions and their temporal evolution.