CNN-LSTM Model for Recognizing Video-Recorded Actions Performed in a Traditional Chinese Exercise

Area of Science:

• Computer Vision
• Artificial Intelligence
• Rehabilitation Technology

Background:

• Human action recognition from video is crucial for intelligent rehabilitation assessment.
• Traditional methods using geometric features struggle with complex scenarios and lack robustness.
• Accurate motion feature extraction and pattern recognition are key challenges.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for recognizing complex human actions.
• To apply the model to the specific task of recognizing Baduanjin, a traditional Chinese exercise.
• To compare the proposed model's performance against traditional geometric feature-based methods.

Main Methods:

• A combined Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model was developed.
• The CNN module extracted abstract image features from video frames.
• The LSTM model processed sequential data for action recognition, with comparisons to OpenPose-based geometric features.

Main Results:

• The CNN-LSTM model achieved a high accuracy of 96.43% on the testing dataset.
• Traditional methods using manually extracted geometric features achieved only 66.07% accuracy.
• CNN-extracted features significantly improved the LSTM model's classification performance.

Conclusions:

• The proposed CNN-LSTM model offers a robust and accurate solution for complex human action recognition.
• This method shows significant potential as a valuable tool in intelligent rehabilitation and exercise analysis.
• Deep learning approaches, particularly CNN-LSTM, surpass traditional methods in recognizing intricate physical activities.