Development of a Digital Coaching Application with Automated Mistake Identification using a Multi-Sensor Configuration

Sammanfattning: Home-based exercise is a popular physical activity of maintaining fitness, health andwellness in general. However, without proper supervision and basic knowledge of theexercises in the workout plan, there is an increased risk of injury. Considering that noteveryone is willing to attend crowded gyms or schedule professional personal trainingsessions, in this study, a novel feedback system is proposed, in the form of a mobileapplication. Accelerometer and gyroscope data were collected from 10 volunteersperforming 3 exercises, squats, lunges and bridges, with inertial sensors attachedto their back lumbar region, on both shanks and on both thighs. Each participantperformed 5 repetitions of the correct technique and 5 repetitions of 4 mistakes foreach exercise. The accuracies of 3 classifiers, a SVM, a RF and DT were comparedwith the SVM performing the best across all 3 exercises. The best location and numberof sensors was determined by examining the accuracy of a SVM model for 15 uniquemulti-sensor configurations. The best performing setup, being the configuration with 2sensors, one at the lumbar area and one at the shank, was used in exploring the efficacyof different data processing techniques. Time-domain statistical features, sensor angletimeseries and the filtered signal timeseries were evaluated as input to a NN. The timedomainfeatures performed the best achieving the highest accuracy in all 3 exercises,with an accuracy of 67% for the squats, 87% for the lunges and 75% for the hip bridges.Overall, the final model demonstrated promising capabilities of classifying exercisetechnique of basic lower-body exercises, with a real-time feedback implementationbeing a feasible solution for self-efficient fitness.