Research and application of smart insole assisted gait recognition technology Research and application of smart insole

• Published in: The Journal of supercomputing Vol. 81; no. 6
• Main Authors: Yuan, Yan, Xu, Jinwei, Wei, Fansen, Mu, Jingsong, Chen, Pin, Wang, Yong
• Format: Journal Article
• Language: English
• Published: New York Springer US 29.04.2025
• Subjects: Compilers; Computer Science; Interpreters; Processor Architectures; Programming Languages; CNN; LSTM; EEMD; Smart insoles; Gait recognition
• ISSN: 1573-0484
• DOI: 10.1007/s11227-025-07226-6

As a nascent biometric technology, gait recognition has demonstrated considerable potential for application across various fields, including security monitoring, medical diagnosis, and human–computer interaction. Wearable devices offer significant advantages in gait recognition, including real-time tracking, high portability, and independence from specific environments. They can continuously collect gait data in various settings, including daily life and complex scenes. Although existing studies have made some progress in gait recognition, there are still challenges regarding the generalization ability of the models and the lack of recognition accuracy. In this study, a lightweight and wireless smart insole based on flexible materials and integrated with pressure and inertial sensors is developed. The monitoring system is constructed using a combination of ensemble empirical mode decomposition (EEMD), convolutional neural network (CNN), and long short-term memory network (LSTM). The initial stage of the process involved data preprocessing, which included data cleaning, the exclusion of nonspecific actions, feature extraction, and the organization of the data. Secondly, EEMD extracted the crucial intrinsic mode functions (IMFs) encompassing essential gait characteristics. The principal IMF was then extracted from the power spectrogram to train the CNN-LSTM model. To address the overfitting issue in the classification model, the max-pooling and dropout techniques are adopted. Subsequently, the model was trained and evaluated to construct a multidimensional evaluation system. This study included twenty healthy subjects. Their gait data were collected in six real-life scenarios while wearing insoles, and the performance of CNN, CNN-LSTM, and K-nearest neighbor (KNN) models was compared using leave-one-subject-out cross-validation. The results show that the proposed CNN-LSTM model exceeds a high level of 96% for a variety of evaluation metrics for the recognition of gait patterns and phases, with a Matthews correlation coefficient of 0.96 and root-mean-square errors of 0.15 and 0.17, which show significant superiority compared to other methods.