| Abstract [eng] |
This master’s thesis focuses on the creation of upper body rehabilitation system using virtual reality and research of artificial intelligence models tailored for such a system. During this project, a system consisting of virtual reality, artificial intelligence and rehabilitation specialist interface subsystems was designed and developed. Hand tracking data was collected using only virtual reality hardware. For this reason, no additional specialised equipment was needed thus simplifying the process. It was found that the hand representation of 26 bones is redundant for the use case of identifying upper body impairments. Similarly, positional features made it harder to classify the impairments. Classification experiments used “LSTM”, “XGBoost”, “ST-GCN” and “ST-GIN” models. “XGBoost” achieved the best performance when identifying the impairment combinations (𝐹1 ̅ = 0,6100). However, unlike “LSTM”, the model was too slow to run real-time. The proposed “LSTMXGBoost” model manages to maintain high accuracy (𝐹1 ̅ = 0,6128) while cutting the computational time in half. |
