Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

IMU-based Real-Time Crutch Gait Phase and Step Detections in Lower-Limb Exoskeletons

Anis R. Shakkour, David Hexner, Yehuda Bitton, Avishai Sintov

TL;DR

This work tackles real-time gait phase and step detection for lower-limb exoskeletons using a minimal hardware setup by mounting a single low-cost IMU on the crutch grip and adopting a $5$-phase gait model that includes a non-locomotor auxiliary state. It benchmarked three architectures (TCN, LSTM, Transformer) and augmented their outputs with a Finite State Machine using a threshold $α=0.6$ to enforce biomechanical consistency, achieving high accuracy with low latency on both PC and embedded hardware. The Temporal Convolutional Network emerged as the best performer, delivering strong phase and step detection (including a reported $94\%$ crutch-step accuracy) and demonstrated notable zero-shot generalization to a paralyzed user trained only on healthy subjects. The approach offers a cost-effective, scalable pathway for real-time exoskeleton control without force sensing, with practical potential for clinical deployment and safe, synchronized assistance in mobility-impaired individuals.

Abstract

Lower limb exoskeletons and prostheses require precise, real time gait phase and step detections to ensure synchronized motion and user safety. Conventional methods often rely on complex force sensing hardware that introduces control latency. This paper presents a minimalist framework utilizing a single, low cost Inertial-Measurement Unit (IMU) integrated into the crutch hand grip, eliminating the need for mechanical modifications. We propose a five phase classification system, including standard gait phases and a non locomotor auxiliary state, to prevent undesired motion. Three deep learning architectures were benchmarked on both a PC and an embedded system. To improve performance under data constrained conditions, models were augmented with a Finite State Machine (FSM) to enforce biomechanical consistency. The Temporal Convolutional Network (TCN) emerged as the superior architecture, yielding the highest success rates and lowest latency. Notably, the model generalized to a paralyzed user despite being trained exclusively on healthy participants. Achieving a 94% success rate in detecting crutch steps, this system provides a high performance, cost effective solution for real time exoskeleton control.

IMU-based Real-Time Crutch Gait Phase and Step Detections in Lower-Limb Exoskeletons

TL;DR

This work tackles real-time gait phase and step detection for lower-limb exoskeletons using a minimal hardware setup by mounting a single low-cost IMU on the crutch grip and adopting a -phase gait model that includes a non-locomotor auxiliary state. It benchmarked three architectures (TCN, LSTM, Transformer) and augmented their outputs with a Finite State Machine using a threshold to enforce biomechanical consistency, achieving high accuracy with low latency on both PC and embedded hardware. The Temporal Convolutional Network emerged as the best performer, delivering strong phase and step detection (including a reported crutch-step accuracy) and demonstrated notable zero-shot generalization to a paralyzed user trained only on healthy subjects. The approach offers a cost-effective, scalable pathway for real-time exoskeleton control without force sensing, with practical potential for clinical deployment and safe, synchronized assistance in mobility-impaired individuals.

Abstract

Lower limb exoskeletons and prostheses require precise, real time gait phase and step detections to ensure synchronized motion and user safety. Conventional methods often rely on complex force sensing hardware that introduces control latency. This paper presents a minimalist framework utilizing a single, low cost Inertial-Measurement Unit (IMU) integrated into the crutch hand grip, eliminating the need for mechanical modifications. We propose a five phase classification system, including standard gait phases and a non locomotor auxiliary state, to prevent undesired motion. Three deep learning architectures were benchmarked on both a PC and an embedded system. To improve performance under data constrained conditions, models were augmented with a Finite State Machine (FSM) to enforce biomechanical consistency. The Temporal Convolutional Network (TCN) emerged as the superior architecture, yielding the highest success rates and lowest latency. Notably, the model generalized to a paralyzed user despite being trained exclusively on healthy participants. Achieving a 94% success rate in detecting crutch steps, this system provides a high performance, cost effective solution for real time exoskeleton control.
Paper Structure (10 sections, 3 figures, 1 table)

This paper contains 10 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Experiments
  4. Dataset
  5. Models Architectures
  6. Models Evaluation
  7. Model Accuracy
  8. Computational Efficiency
  9. Data Efficiency
  10. Conclusions

Figures (3)

  • Figure 1: Confusion matrix for detecting the five crutch gait phases using the TCN and without using the FSM.
  • Figure 2: Real versus predicted steps of TS2 (exoskeleton user) based on observing crutch gait cycles.
  • Figure 3: Classification success rate with regard to the number of subjects included in the training data, evaluated on the test data with the TCN model.