Enhanced Human Crawling Phase Recognition Based on Kinematic Synergies and Machine Learning

Abstract Background: Hands-and-knees crawling, an effective rehabilitation method for children with motor impairments, requires precise phase detection for optimizing assistive devices. However, research on phase detection in human crawling remains limited. The research explores whether multi-joint kinematic synergy features provide better accuracy than traditional time-domain features. Methods: Nine healthy adults participated in the study, where accelerometers and pressure sensors were used to capture motion data during crawling. The data were pre-processed and used to define four distinct phases of crawling and kinematic synergy features were extracted using Singular Value Decomposition (SVD)-based Principal Component Analysis (PCA). Machine learning models, including Classification and Regression Trees (CART), K-Nearest Neighbors (KNN), and Error-Correcting Output Codes Support Vector Machines (ECOC-SVM), were trained to recognize the crawling phases. Their performance was compared to that of time-domain features. Results: The phase recognition method based on multi-joint kinematic synergies achieved an average accuracy of 89.37%. Specifically, the accuracy for CART was 88.41%, for KNN was 85.51%, and for ECOC-SVM was 94.20%. In comparison, the phase recognition using traditional time-domain features yielded lower accuracy, with overall accuracies of 78.98% for CART, 76.09% for KNN, and 85.51% for ECOC-SVM Conclusion: The findings demonstrate that using kinematic synergy features significantly improves the accuracy of crawling phase recognition compared to traditional time-domain features. This research provides valuable insights into the design and control of rehabilitation robots based on human kinematic synergies