A Preliminary Study of Deep Learning-Based Intention Recognition for Sit-to-Stand Cycle Movements Using Foot Pressure*

The objective of this study was to assess the reliability of a Sit-to-Stand cycle motion intention recognition method that integrates foot pressure data and deep learning algorithms, with the aim of replacing the Electromyography-based Sit-to-Stand cycle motion intention recognition method in myoelectric-controlled FES systems in subsequent studies. Therefore, a foot pressure dataset for human sit-stand motion was constructed and a lightweight network model T-2CNN with organic fusion of Transformer and CNN was proposed. In order to validate the performance advantages of the proposed model, we replicated LeNet, VGG16, and the LCNN model that had been deployed on an embedded platform for action intention recognition and compared them as benchmarks. The experimental results demonstrated that the T-2CNN model attained 100% maximum and average accuracy on the test set. In contrast, the LeNet network model achieved a maximum accuracy of 94.04% and an average accuracy of 90.64±7.69% on the test set, which were lower than the T-2CNN model's 5.96% and 9.36%, respectively. The highest accuracy and average accuracy of VGG16 on the test set were 99.32% and 97.80±2.83%, respectively, which were lower than the 0.68% and 2.20% achieved by the T-2CNN model. The highest accuracy and average accuracy of LCNN on the test set were 99.15% and 97.53±1.19%, respectively, which were lower than the 0.85% and 2.47% achieved by the T-2CNN model.