Hybrid Deep Neural Network Framework Combining Skeleton and Gait Features for Pathological Gait Recognition
Kooksung Jun1,2, Keunhan Lee3, Sanghyub Lee2
1Robocare, Seongnam 13449, Republic of Korea.
Bioengineering (Basel, Switzerland)
|October 28, 2023
Summary
Combining skeleton data, joint angles, and gait parameters significantly improves pathological gait recognition. This hybrid deep learning model enhances diagnostic support for physicians by outperforming single-input methods.
Area of Science:
- Biomechanics
- Computer Science
- Medical Imaging
Background:
- Pathological gait recognition aids physician diagnosis.
- Existing methods often use raw skeleton data or extracted gait features.
- Combining diverse data types may enhance recognition accuracy.
Purpose of the Study:
- To develop a deep neural network (DNN) model for improved pathological gait recognition.
- To investigate the efficacy of combining skeleton sequences, joint angles, and gait parameters.
- To create a hybrid DNN framework for effective multi-modal data fusion.
Main Methods:
- A hybrid DNN framework integrating Graph Convolutional Networks (GCN), Recurrent Neural Networks (RNN), and Artificial Neural Networks (ANN).
- GCN for skeleton sequences, RNN for joint angle sequences, and ANN for gait parameters.
- Feature fusion from all three input types before final classification.
Main Results:
- The proposed hybrid model demonstrated improved pathological gait recognition performance on two datasets (simulated and vestibular disorder).
- Multi-modal input significantly outperformed single-input models.
- The model achieved state-of-the-art results for skeleton-based action recognition.
Conclusions:
- Integrating skeleton data, joint angles, and gait parameters enhances pathological gait recognition.
- The hybrid DNN framework offers a robust approach for multi-modal gait analysis.
- This method shows promise for improving diagnostic decision-making in clinical settings.
