A multimodal framework based on deep belief network for human locomotion intent prediction.

Area of Science:

• Robotics and Human-Computer Interaction
• Biomedical Engineering
• Artificial Intelligence

Background:

• Accurate prediction of human locomotion intent is crucial for the safe and seamless operation of lower limb exoskeletons.
• Existing methods may struggle with recognizing diverse locomotion modes and predicting transitions between them, especially across different terrains.

Purpose of the Study:

• To develop a multimodal framework using a deep belief network (DBN) for recognizing human locomotion modes and predicting transition tasks.
• To evaluate the effectiveness of different data fusion strategies (data, feature, and decision levels) within the DBN framework.

Main Methods:

• A deep belief network (DBN) was designed as a multimodal framework to process and integrate various sources of locomotion data.
• Three distinct fusion strategies were investigated to optimize the network's performance in recognizing locomotion intent.
• The model was tested on public datasets for both steady-state and transition locomotion states.

Main Results:

• The DBN-based multimodal framework achieved high prediction accuracy: 97.64% (user-dependent) and 96.80% (user-independent) for steady-state locomotion.
• The system demonstrated exceptional performance during transitions, accurately predicting all transitions with 96.37% (user-dependent) and 95.01% (user-independent) accuracy.
• Optimal network performance was achieved through the exploration of different fusion strategies.

Conclusions:

• The developed multimodal framework based on DBN accurately predicts human locomotion intent, including complex transitions between different gaits.
• The findings highlight the potential of this DBN model for volition control in lower limb exoskeletons, improving user safety and interaction.
• The proposed method shows promise for real-world applications in assistive robotics and human augmentation.