The adaptability of machine vision solutions for motion capture has generated an interest in determining whether markerless capture methods can achieve a level of accuracy particularly similar to capture techniques with passive or active markers, being an essential factor in the identification and treatment of motor weaknesses in physiotherapeutic fields. In this sense, the aim of this work is to perform a comparison of two motion capture techniques: intrusive and non-intrusive, using the 3D tracking systems OptiTrack and MediaPipe to determine the performance of markerless pose estimation in obtaining the ranges of motion (ROM) of the upper body of a person. The capture is simulated using a virtual environment (VE) developed on the Unity platform, evaluating accuracy and deviation levels using Matlab. The non-intrusive pose estimation solution generated 92.82% average accuracy with respect to the standard trajectory when capturing the ROMs of the limb, proving to be a cost-effective and significantly accurate option for the analysis of upper body kinematics in rehabilitation exercises. The study of the elbow joint indicated similar results in flexion, with 1.382° degrees difference as did the shoulder joint, with 1.237° and 2.969° degrees in flexion and abduction, respectively. Finally, the ROMs captured demonstrate that the integration of a VE decreases the deviation in the trajectory of the limb, guiding the participant in the exercise presented.
