TL;DR
This paper introduces a novel framework for real-time detection of key gait events (Initial Contact, Foot Flat, Heel Off, Toe Off) using only kinematic data and a biomechanical model. Three algorithms were developed and validated, demonstrating high precision and low latency, making them suitable for controlling neuroprosthetics and neurorobotics without requiring additional sensors.
What Problem Does This Paper Address?
Controlling neuroprosthetics (NP) and neurorobotics (NR) that assist with walking, such as ankle-foot orthoses (AFOs), requires precise timing of gait events. These events, like initial contact (IC) and toe-off (TO), are crucial for smooth and natural movement. While traditional lab-based methods use force plates and marker data, these are not practical for real-time, wearable applications. Existing wearable solutions often rely on specialized sensors like force-sensitive resistors (FSRs), which have reliability issues, or achieve real-time detection with significant delays. The challenge is to accurately and reliably detect multiple gait events in real-time using only the kinematic data that is already commonly available in NP/NR systems, without adding computational burden or requiring extra hardware.
Methods
The core of the paper is a novel framework that couples kinematic data with a biomechanical model to extract relevant signals for gait event detection. This approach allows for the extraction of features that are typically only available offline or through specialized sensors. Three real-time algorithms (RTGSD-min, RTGSD-B6, and RTGSD-G6) were developed based on this framework.
Key methodological aspects include:
- Biomechanical Model Integration: A generic musculoskeletal model (3DGaitModel2392) was scaled to each subject. This model acts as an intermediary, allowing the extraction of specific kinematic signals (e.g., joint angles, velocities) that are not directly measured but are derived from the overall motion.
- Threshold-Based Detection with Adaptive Thresholds: The algorithms employ a double-layered structure. The first layer uses state-machines to detect events based on real-time signals and pre-defined thresholds. The second layer, an "update layer," adaptively adjusts these thresholds based on the last few detected events. This adaptive policy aims to maintain precision while accommodating variations in gait.
- Decoupled Update and Detection Layers: The update and real-time detection layers were decoupled for computational efficiency and to allow for different sampling frequencies.
- Algorithm Variations: Three specific algorithms were developed, each using slightly different definitions for gait events and requiring different subsets of kinematic data, making them suitable for various NP/NR applications:
- RTGSD-min: Designed for simpler applications like AFOs, requiring minimal data (sagittal shank angular velocity and ankle angle).
- RTGSD-B6: Incorporates definitions from recent literature (Botzel et al.) and is geared towards more complex bilateral systems.
- RTGSD-G6: An ambulatory implementation based on prior work (Ghoussayni et al.), using toe and heel marker velocities (simulated via model segment velocities).
- Secondary TO Detection: A secondary definition for toe-off (TO) was implemented to improve robustness in cases where the primary TO detection might fail.
Datasets
The study utilized data from seven healthy subjects walking on an instrumented split-belt treadmill at three different speeds (3, 4, and 5 km/h). This resulted in a total of 558 steady-state steps analyzed.
- Reference Data: Ground truth for gait events was established using a combination of:
- Force Plates: Used to identify Initial Contact (IC) and Toe Off (TO) based on a 20 N vertical force threshold.
- Motion Capture System (Vicon): Used with reflective markers to identify Foot Flat (FF) and Heel Off (HO) based on a 100 mm/s threshold on toe and heel marker vertical velocities.
The paper does not introduce a new benchmark dataset but rather validates its algorithms against established reference methods.
Key Insights and Ideas
- Biomechanical Model as a Feature Extractor: The most significant insight is the effective use of a biomechanical model to extract features (like joint velocities and accelerations) that are crucial for gait event detection but not directly measurable by standard wearable sensors. This bypasses the need for specialized sensors like FSRs.
- Adaptive Thresholds for Robustness: The adaptive thresholding mechanism is key to handling variations in gait and ensuring consistent performance across different steps and potentially different users, without requiring explicit recalibration for each individual.
- Decoupling for Efficiency: Separating the threshold update from the real-time detection layer is a clever design choice that improves computational efficiency, making the algorithms more suitable for embedded systems in NP/NR.
- HO Detection Remains Challenging: The study highlights that detecting the Heel Off (HO) event in real-time remains a significant challenge, exhibiting higher variability and delays compared to other events, even with advanced methods. This suggests that threshold-based approaches might be insufficient for HO.
Key Results
- High Precision: All three algorithms achieved excellent precision (100% for IC, FF, and TO), meaning no false positives were observed for these events.
- Low Latency for IC, FF, TO:
- IC and TO: Delays were comparable to or better than current state-of-the-art algorithms. RTGSD-G6 showed the smallest mean delay for IC and TO (11.10 ms ± 10.72 ms for IC, 6.40 ms ± 15.49 ms for TO).
- FF: Smaller delays were achieved for FF detection compared to IC and TO.
- HO Detection Challenges: Heel Off (HO) detection showed significant variability across algorithms. RTGSD-min and RTGSD-B6 had small mean delays but very large variability, while RTGSD-G6 had lower variability but a substantial mean delay (-189 ms), falling outside acceptable limits.
- No Additional Sensors Required: The algorithms leverage kinematic data commonly available in NP/NR systems, meaning no extra sensors are needed, thus not impacting cost or reliability.
Limitations and Future Work
- HO Event Variability: The authors acknowledge that HO detection remains problematic due to high variability. They suggest that threshold-based methods might not be optimal for HO and propose exploring probabilistic methods or machine learning to reduce variability.
- Offline Validation: While the algorithms are designed for real-time use, the validation in this study was performed offline using motion capture data. The authors note that real-time applications might introduce slightly higher delays (estimated up to 10 ms) due to the quality of real-time kinematics and filtering.
- Limited Activities: The study focused on walking on a treadmill. Future work should validate the algorithms on NR/NP systems and across a wider range of activities.
- Potential for Improvement: The authors suggest that restoring a stronger coupling between the update and detection layers (as in previous work) could potentially improve TO detection delays, though this would increase computational load.
- Application in Control: Future work will involve applying these detection methods to control an exoskeleton.
Summary updated