Human Activity Recognition for People with Knee Abnormality Using Surface Electromyography and Knee Angle Sensors

With the onset of technological advancements, biosensors are being effectively used in a variety of contexts, including the diagnosis of diseases, the promotion of their prevention and rehabilitation, the monitoring of patient health, and human activity recognition (HAR). Recently, HAR research results have been obtained and applied in many commercial applications such as fall detection using smart watch sensors, activity classification using smart home sensors. However, previous HAR models utilizing wearable sensors primarily utilized data from healthy individuals, and such models are frequently inaccurate when applied to individuals with medical mobility impairments. In this work, surface electromyography and goniometer sensors were used to categorize various types of rehabilitation activities based on HAR models developed for individuals with knee abnormalities. To achieve the goal of our study, a deep residual network called ResNeXt was proposed to detect human activities with high performance. To investigate the classification performance of deep learning models, a 5-fold cross-validation technique was applied for both training and testing. Based on the results of our study, the combination of features extracted from the goniometer and electromyograph signals resulted in the highest F1-score (93.31%) and the best accuracy (93.52%).