Recognition of Lower Limb Movements Based on Multiregional Plantar Pressure and Foot Posture

Lower limb movement is one of the basic features of human locomotion, and its recognition technology has been widely applied in medical, sports, and industrial fields in recent years. Currently, wearable devices based on pressure and inertial sensors have been used to recognize lower limb movement but still have deficiencies in hardware and algorithms. This article proposes a novel and low-cost solution for lower limb movement recognition, which involves embedding a smart insole with 16 pressure sensors under each foot and installing a ten-axis inertial sensor on each ankle. The pressure and inertial data are synchronously collected via a wireless Bluetooth module, enabling efficient and accurate recognition. This multisensor configuration not only enriches the data features, but also effectively addresses the misrecognition issues associated with single-sensor solutions. The experimental data include six common lower limb movements: upstairs, uphill, walking, squatting, downstairs, and downhill. To recognize movements during various locomotion states, this article improves the temporal convolutional network (TCN) by introducing a self-attention mechanism, proposing a self-attention TCN to achieve this purpose. The self-attention TCN model was compared with support vector machine (SVM), K-nearest neighbor (KNN), decision tree (DT), random forest (RF), gated recurrent unit (GRU), long short-term memory (LSTM), convolutional neural network (CNN), bidirectional LSTM (BiLSTM), and LSTM-CNN. The experimental results demonstrate that the self-attention TCN achieves the highest recognition accuracy of 98.29% <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\pm ~0.10$ </tex-math></inline-formula>%.