A Gait Phase Classifier using a Recurrent Neural Network

Won Ho Heo, Euntai Kim, Hyun Sub Park, Jun-Young JungJournal of Institute of Control Robotics and Systems 2015 · DOI

Abstract

This paper proposes a gait phase classifier using a Recurrent Neural Network (RNN). Walking is a type of dynamic system, and as such it seems that the classifier made by using a general feed forward neural network structure is not appropriate. It is known that an RNN is suitable to model a dynamic system. Because the proposed RNN is simple, we use a back propagation algorithm to train the weights of the network. The input data of the RNN is the lower body's joint angles and angular velocities which are acquired by using the lower limb exoskeleton robot, ROBIN-H1. The classifier categorizes a gait cycle as two phases, swing and stance. In the experiment for performance verification, we compared the proposed method and general feed forward neural network based method and showed that the proposed method is superior.

Introduction

For lower-limb exoskeleton robots to assist users effectively, they must accurately identify the user's current gait phase—specifically, whether the foot is in the "swing" phase (in the air) or the "stance" phase (on the ground). While many systems rely on physical sensors like pressure-sensitive insoles (FSRs) or load cells, these can be cumbersome and prone to mechanical wear.

An alternative approach uses kinematic data, such as joint angles and velocities. However, walking is a dynamic system where a specific leg pose might appear identical at different points in a step. Standard Feed-Forward Neural Networks (FFNNs) lack "memory," treating each data point as an isolated event. Because a gait phase depends heavily on the preceding state (e.g., "toe-off" must follow "heel-strike"), the authors argue that a network with a feedback loop is required to model human walking accurately.

Methodology

The researchers implemented a simplified Jordan-type Recurrent Neural Network to capture the temporal context of walking.

Network Architecture

The proposed RNN consists of an input layer, a single hidden layer with 20 neurons, and an output layer. The defining characteristic of this architecture is the feedback loop: the output from previous time steps is fed back into the input layer. Specifically, the model uses a moving average of the last 20 outputs as a "dynamic neuron" input, providing the network with a smoothed history of recent states to prevent jitter caused by sensor noise.

Comparison of standard Feed-Forward (a) and Recurrent (b) Neural Network architectures.

Data Collection and Training

Data was collected using the ROBIN-H1 exoskeleton worn by four healthy subjects. The model utilizes four primary kinematic inputs: the joint angles and angular velocities of the hip and knee. To establish ground truth, subjects walked over Zebris and KISTLER force plates, which provided the "gold standard" for actual ground contact. The network was trained using a standard backpropagation algorithm to produce a binary signal: 0 for swing and 1 for stance.

Key Results and Insights

The RNN significantly outperformed the standard FFNN across all performance metrics, particularly in timing accuracy. The primary advantage of the RNN is its ability to use temporal context to anticipate transitions, such as the subtle kinematic changes preceding a heel-strike.

  • Error Rates: The RNN achieved an average error rate of 3.75%, a substantial improvement over the 8.89% error rate of the FFNN.
  • Timing Accuracy: The RNN was much faster at recognizing phase transitions. Its timing error for heel-strike was only 0.036 seconds, compared to 0.122 seconds for the FFNN. Similarly, it detected toe-off with an error of 0.044 seconds versus 0.063 seconds for the FFNN.
  • Stability: The RNN showed fewer "Type 2" errors (detecting a phase change too early), which is critical for the safety and stability of exoskeleton control.
Comparison of gait phase detection between the proposed RNN classifier and a standard FFNN against the ground truth.

Conclusion and Future Work

By incorporating a feedback loop, the Jordan-type RNN effectively models the physical constraints of human biomechanics, providing a more stable and accurate classification than static models. However, the researchers noted a "first step" problem where performance drops during the initial movement from a standstill, as the kinematics differ from steady-state walking. Future work will focus on resolving this initialization error and expanding the model to classify more granular sub-phases of the gait cycle, such as mid-stance and terminal swing.

Summary updated

Abstract

This paper proposes a gait phase classifier using a Recurrent Neural Network (RNN). Walking is a type of dynamic system, and as such it seems that the classifier made by using a general feed forward neural network structure is not appropriate. It is known that an RNN is suitable to model a dynamic system. Because the proposed RNN is simple, we use a back propagation algorithm to train the weights of the network. The input data of the RNN is the lower body's joint angles and angular velocities which are acquired by using the lower limb exoskeleton robot, ROBIN-H1. The classifier categorizes a gait cycle as two phases, swing and stance. In the experiment for performance verification, we compared the proposed method and general feed forward neural network based method and showed that the proposed method is superior.
BibTeX
@article{heo2015gait,
  author = {Won Ho Heo and Euntai Kim and Hyun Sub Park and Jun-Young Jung},
  title = {{A Gait Phase Classifier using a Recurrent Neural Network}},
  journal = {Journal of Institute of Control Robotics and Systems},
  year = {2015},
  doi = {10.5302/j.icros.2015.15.9024},
}
Figure