High-Accuracy Early Recognition of Upper-Limb Motions for Exoskeleton-Assisted Mirror Rehabilitation

Upper-limb early motion recognition (EMR) significantly enhances human-computer interaction and skill transfer in exoskeleton-assisted mirror rehabilitation. However, achieving early and accurate recognition of upper-limb motion remains a challenge, limiting the transfer of natural movements from the healthy to the affected side. To address these challenges and limitations, this study introduces a novel high-accuracy upper-limb EMR method and implements it within an exoskeleton-assisted mirror rehabilitation system. Specifically, a new architecture is designed to model and parameterise upper-limb motion, transforming it from a three-dimensional Cartesian coordinate system into a four-dimensional parametric space. The parameterized results were then evaluated with 11 algorithms, using public datasets, P-BTBS dataset, and Arm-CODA. Experimental results show that the proposed method achieves over 99% recognition accuracy for both full and first 30% upper-limb motion sequences while saving at least 80% of recognition time. Comparative analysis identifies RF, XGBoost, KNN, and deep learning as the most promising algorithms, with bidirectional encoder representations from transformers (BERT) pioneering advancements in upper-limb motion recognition. These findings indicate that the proposed architecture enables high-accuracy upper-limb EMR at the earliest possible stage (30%), offering a new paradigm for human-computer interaction, personalised medicine, and mirror motion rehabilitation.