A Gesture Recognition System Using Electrical Impedance Tomography With Improved Electrode Layout and Classification Techniques

Abstract

Accurate and reliable gesture recognition using electrical impedance tomography (EIT) holds significant potential for human-computer interaction and assistive technologies, yet ensuring consistent performance across multiple sessions remains challenging due to factors such as system noise, electrode shifts, and frequency-dependent signal variation. To address these issues, we propose an optimized EIT-based gesture recognition system featuring a dual-ring electrode configuration, an enhanced classification algorithm, and a high-frame-rate data acquisition approach. By systematically examining the similarity evaluation index (SEI) at various frequencies, we identified 10 kHz as the optimal operating frequency, achieving an SEI of 16.5%, substantially exceeding the baseline SEIL value. Our improved neural network architecture, PEU-SFU-ResNet50, further enhances feature extraction and classification robustness, attaining 88.18% accuracy in intersession tests—approximately 12% higher than the baseline model—and demonstrating 98% accuracy in single-session scenarios, outperforming standard ResNet50 and artificial neural network (ANN). Ablation experiments and cross-validation validated the efficacy and robustness of our proposed system, underscoring its potential for multisession gesture recognition applications.