Abstract
This research introduces a novel, highly precise, and learning-free approach to locomotion mode prediction, a technique with potential for broad applications in the field of lower-limb wearable robotics. This study represents the pioneering effort to amalgamate 3D reconstruction and Visual-Inertial Odometry (VIO) into a locomotion mode prediction method, which yields robust prediction performance across diverse subjects and terrains. The proposed Depth-enhanced Visual-Inertial Odometry (D-VIO) is designed to operate within the computational constraints of wearable configurations while demonstrating resilience against unpredictable human movements and sparse features. Comprehensive experiments across various test conditions—including subjects, scenarios, sensor mounting positions, and camera views—yielded a prediction accuracy rate of 99.00%, confirming the efficacy, generality, and robustness of the proposed method.
Introduction
For lower-limb exoskeletons or prostheses to provide effective assistance, they must anticipate a user's intent. Walking on level ground requires a fundamentally different mechanical strategy than ascending stairs or navigating ramps. Existing solutions typically rely on reactive kinematic recognition, which identifies modes only after a gait cycle has begun, or machine learning (ML) vision models. However, ML approaches are often data-hungry, computationally intensive, and brittle when faced with variations in camera angles, step sizes, or environmental disturbances.
This paper proposes a "learning-free" geometric approach that is computationally light enough for wearable hardware and robust enough to handle the inherent irregularities of real-world human movement.
System Architecture and Methodology
The proposed system utilizes a multi-stage pipeline to transform raw sensor data into actionable locomotion predictions without the need for pre-trained models.
Depth-enhanced Visual-Inertial Odometry (D-VIO)
Building on the VINS-Mono framework, the authors integrated specific depth constraints to maintain accurate tracking. This enhancement allows the system to remain stable even when visual features are sparse (e.g., blank walls) or when human movement is abrupt and erratic.
Environment Mapping and Terrain Reconstruction
The system processes RGB-D images to construct a 3D map using a Truncated Signed Distance Function (TSDF). This mapping technique smooths sensor noise and provides a persistent memory of the environment, allowing the system to maintain predictions even if a moving obstacle temporarily obscures the camera's view. Using Agglomerative Hierarchical Clustering (AHC), the system extracts geometric planes from the map, classifying them as level ground, stairs, or ramps based on slope and relative height.
Locomotion Mode Prediction
By calculating the user's current velocity and gait phase, the system predicts the coordinates of the next footfall. It then queries the reconstructed 3D map at that specific location to determine the upcoming locomotion mode.
Experimental Evaluation
The method was validated using the open-source EuRoC MAV dataset and a custom experimental dataset involving 15 subjects. The custom tests covered 180 condition combinations, including different mounting positions (head vs. chest), walking patterns (wandering, regular, jogging), and various terrain geometries.
Key Results
- High Accuracy: The system achieved a 99.00% accuracy rate across 540 experimental trials.
- Computational Efficiency: The optimized code achieves a processing time of just 31ms per frame, making it viable for real-time onboard processing on wearable devices.
- Robustness: The system proved resilient to changes in viewpoint (helmet vs. chest mount) and walking speed. Because it relies on geometry rather than pattern matching, it requires zero training or parameter tuning for new subjects or environments.
Conclusion
By prioritizing the physics of terrain geometry over visual pattern matching, this research provides a robust, generalizable solution for wearable robotic control. While future work will address "corner cases" during highly unpredictable wandering or jogging, the current system's 99% accuracy and low latency represent a significant step forward in predictive locomotion technology.
Summary updated