AI-based ACL Exercises Recognition System Using Wearable Multi-Sensor Data Fusion

Human activity recognition has attracted researchers’ attention in the last two decades. Anterior Cruciate Ligament (ACL) exercises are example of these activities that have to be performed correctly to ensure efficient knee joint recovery. Hence, Machine and Deep Learning algorithms have been employed to classify ACL exercises and evaluate its correctness. This study investigates the accuracy of five machine learning algorithms, SVM, Decision Tree, Random Forest, Gradient boosting and KNN, with CNN in terms of their ability to classify ACL exercises. The data of seven ACL exercises, performed by four subjects, were collected using Accelerometer and gyroscope sensors, then these data were used to train the algorithms. Results showed that both CNN and Random Forest models performed well and achieved higher accuracy among the other algorithms with real accelerometer and gyroscope data. However, Random Forest model outperformed other models when relying on real accelerometer data only or with synthesized data. Moreover, it is also found that gyroscope data are essential for such systems to train the algorithms efficiently and excluding such data leads to downgrade the classification performance.