Multidimensional Feature-Guided Cross-Population Human Activity Recognition and Prediction.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Medicine

Background:

• Increasing demand for human behavior recognition in rehabilitation and human-machine collaboration.
• Challenges in cross-population gait analysis due to individual variability.
• Difficulties in model generalization caused by coupled pathological and normal gait features.

Purpose of the Study:

• To propose a novel SG-LSTM framework for cross-population lower-limb activity recognition and prediction.
• To address variability in gait features and improve model generalization.
• To enhance adaptability by jointly optimizing gait prediction and classification.

Main Methods:

• Developed a dual-branch SG-LSTM framework with symmetric LSTM (S-LSTM) and grouped LSTM (G-LSTM) networks.
• S-LSTM models spatiotemporal symmetry in normal gait; G-LSTM models abnormal motion coupling in pathological gait.
• Implemented a dynamically weighted multi-task loss function for joint optimization of prediction and classification.

Main Results:

• The proposed framework achieved state-of-the-art (SOTA) performance in cross-population human behavior prediction and gait recognition.
• Demonstrated superior performance compared to existing approaches on multiple datasets.
• Showcased potential clinical application value in gait analysis.

Conclusions:

• The SG-LSTM framework effectively handles gait variability and pathological features for improved recognition and prediction.
• The dual-branch architecture and multi-task learning enhance model adaptability and performance.
• The method holds significant potential for advancing rehabilitation medicine and human-machine collaboration through accurate gait analysis.