Real time motion analysis using vision based deep learning

The primary aim of this doctoral dissertation is to develop an artificial intelligence–based exercise analysis system capable of accurately evaluating individuals' physical activities in real time using only RGB cameras, without the need for markers or additional sensors. To achieve this goal, a modular software architecture integrating MediaPipe and OpenCV was designed, and time-series skeletal data collected from 103 participants through a dual-camera setup were processed using deep learning models, including LSTM + Attention Mechanism, GRU + Attention Mechanism, and Transformer architectures. The effects of participants' body mass index, sports experience, and performance level on model accuracy were examined through statistical analyses. As a notable contribution of the study, the expert scoring framework was reconstructed based on biomechanical principles, enabling a multi-dimensional assessment of exercise quality through indicators such as postural symmetry, joint-angle consistency, movement stability, and smoothness. These expert-derived scores were incorporated as movement-quality variables during model training, significantly enhancing the networks' real-time evaluation capability. The findings demonstrate that the GRU + Attention Mechanism model achieved the highest performance with an accuracy of 98.9%, and that the integration of the expert scoring system substantially improved model consistency (p < 0.05). The developed system is capable of identifying postural errors without professional supervision and contributes to injury prevention, offering a scalable, customizable, and cost-effective exercise assessment infrastructure for applications in sports, rehabilitation, and public health.