Deep Learning for Human Locomotion Analysis in Lower-Limb Exoskeletons: A Comparative Study

TL;DR

This study evaluates eight deep learning architectures for real-time human locomotion analysis in lower-limb exoskeletons using the CAMARGO 2021 dataset, focusing on ground-condition classification and terrain-parameter estimation (ramp slope, stair height). It demonstrates that IMU data alone outperform EMG, with a three-IMU setup (foot, shank, thigh) providing a minimal yet effective sensor configuration validated by SHAP explainability. LSTM shows superior performance for classification and slope prediction, while CNN-LSTM excels at stair-height estimation, all achieving near real-time inference (~1 ms). The work offers a practical, lightweight approach for adaptive exoskeleton control and provides sensor-reduction guidance through model explainability and LOSO cross-validation.

Abstract

Wearable robotics for lower-limb assistance have become a pivotal area of research, aiming to enhance mobility for individuals with physical impairments or augment the performance of able-bodied users. Accurate and adaptive control systems are essential to ensure seamless interaction between the wearer and the robotic device, particularly when navigating diverse and dynamic terrains. Despite the recent advances in neural networks for time series analysis, no attempts have been directed towards the classification of ground conditions, categorized into five classes and subsequently determining the ramp's slope and stair's height. In this respect, this paper presents an experimental comparison between eight deep neural network backbones to predict high-level locomotion parameters across diverse terrains. All the models are trained on the publicly available CAMARGO 2021 dataset. IMU-only data equally or outperformed IMU+EMG inputs, promoting a cost-effective and efficient design. Indeeds, using three IMU sensors, the LSTM achieved high terrain classification accuracy (0.94 +- 0.04) and precise ramp slope (1.95 +- 0.58°) and the CNN-LSTM a stair height (15.65 +- 7.40 mm) estimations. As a further contribution, SHAP analysis justified sensor reduction without performance loss, ensuring a lightweight setup. The system operates with ~2 ms inference time, supporting real-time applications. The code is code available at https://github.com/cosbidev/Human-Locomotion-Identification.

Figures (6)

• Figure 1: Position of the sensors along the body (Image extracted camargo2021comprehensive)
• Figure 2: Visual Methodology
• Figure 3: Feature importance for the IMU modality, derived by explaining the LSTM model used for the classification task with the SHAP methodology. The labels x-axis are formatted as $B_S^A$, where $B$ denotes the body part (foot: $f$, shank: $s$, thigh: $t$, trunk: $k$), $S$ indicates the sensor type (accelerometer: $A$, gyroscope: $G$), and $A$ specifies the axis (x-axis: $X$, y-axis: $Y$, z-axis: $Z$). The dotted horizontal lines represent the average sensor weights.
• Figure 4: Feature importance for the IMU modality derived using the SHAP methodology, with the upper plot showing the LSTM model for slope prediction and the lower plot showing the CNN-LSTM model for stair height prediction, respectively. The features on x-axis are formatted as $B_S^A$, where $B$ denotes the body part (foot: $f$, shank: $s$, thigh: $t$, trunk: $k$), $S$ indicates the sensor type (accelerometer: $A$, gyroscope: $G$), and $A$ specifies the axis (x-axis: $X$, y-axis: $Y$, z-axis: $Z$). The horizontal lines represent the average sensor weights.
• Figure 5: Accuracy of the LSTM model in the classification task as a function of the number of IMU sensors iteratively added based on their informativeness ranked by the SHAP methodology across the four body sectors. Statistical significance is denoted by asterisks, with * indicating $p \leq 0.05$, ** indicating $p \leq 0.01$, *** indicating $p \leq 0.001$, and **** indicating $p \leq 0.0001$.