Gait Classification and Recognition Based on ECA-LSTM Algorithm

• Published in: Chinese Automation Congress (Online) pp. 8439 - 8443
• Main Authors: An, Junlin, Jin, Shanhai
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.11.2023
• Subjects: attention mechanism; Computational modeling; Data models; gait classification; Legged locomotion; Long short term memory; long short-term memory neural network; lower limb exoskeleton; Pattern recognition; Pressure sensors; Robot sensing systems
• ISSN: 2688-0938
• DOI: 10.1109/CAC59555.2023.10450940

The classification and recognition of gait patterns plays a decisive role in whether the exoskeleton robot can provide appropriate auxiliary effects for the lower limbs during the movement. Therefore, this article designs a gait classification method for horizontal walking (HW), going upstairs (GU) and going downstairs (GD) based on the data of inertial sensors and plantar pressure sensors. Compared with only using a single sensor, this method uses multi-source sensor data to reflect the movement pattern of human lower limbs more comprehensively. At the same time, this article suggests a gait classification model utilizing a long short-term memory neural network (LSTM)and efficient channel attention (ECA)module. The model can automatically learn the data representation of gait without manual feature engineering. Firstly, the data acquisition system is employed for gathering information on various walking patterns exhibited by diverse subjects, and the data set is constructed. The ECA module is employed to compute the attention weights for the initial data characteristics, following which the data is weighted and fed into the LSTM model, so that it can better learn the temporal dependence of the features. Comparative experiments reveal that the LSTM classification model achieves an accuracy of 0.858, a recall of 0.860, and a precision of 0. 872. The accuracy, recall and precision of the ECA-LSTM classification model with ECA module are 0.960, 0.961 and 0.961, respectively. This underscores the efficacy of the ECA module in enhancing gait classification recognition accuracy.