This thesis describes the design and development of the web application "AI Fitness Coach", which uses artificial intelligence to take on the role of a digital personal trainer for home workouts. Exercising at home has become increasingly common, but the absence of professional supervision often leads to incorrect exercise execution and an increased risk of musculoskeletal injuries. The application addresses this problem by monitoring the user through a webcam and providing real-time feedback on exercise form. As a starting point, I used an open-source project that extracts 33 anatomical body landmarks from the video stream using the MediaPipe framework and recognises four fitness exercises and counts repetitions with a bidirectional BiLSTM neural network. I extended this foundation into a complete multi-user web application: I developed a Next.js client, a FastAPI backend with asynchronous data streaming over the WebSocket protocol, and a custom form-analysis layer that detects errors based on biomechanical joint angles and distances, classifies them by severity, and translates them into understandable coaching instructions. I deployed the system to a production Ubuntu Linux environment and evaluated it in terms of classification accuracy, processing speed, user experience, and server resource consumption. Measurements showed an average latency of 73.7 ms, ensuring immediate feedback below the threshold of human perception, and confirmed that modern web technologies combined with artificial intelligence can replace basic expert supervision in home workouts without requiring software installation or dedicated hardware.
|
