Validation of marker-less pose estimation for 3D kinematics during upper limb reaching

Kinematic analysis of movement following brain damage is key for diagnosing motor impairments and for recovery assessment. Advances in computer vision offer novel marker-less tracking tools that could be implemented in the clinic due to their simple operation and affordability. An important question that arises is whether marker-less technologies are sufficiently accurate compared to well established marker-based technologies. This study aims to perform validation of kinematic assessment using two high-speed cameras and a 3D pose estimation model. Four participants performed reaching movements with the upper limb between fixed targets, in different velocities. Movement kinematics were simultaneously measured using the DeepBehavior model and marker-based optical motion capture (QTM), as a gold standard. The differences in corresponding joint angles, estimated from the two different methods throughout the analysis, are presented as a mean absolute error (MAE) of the elbow angle. Quantitatively, the MAE of all movements was relatively small across velocity and joints (~2{degrees}). In a condition where the movements were made towards the DeepBehavior cameras, and the view of the elbow was occluded in one of the cameras, the errors were higher. In conclusion, the results demonstrated that marker-less motion capture is a valid alternative to marker-based motion capture. Inaccuracies of the DeepBehavior system could be explained by occlusions of key-points and are not associated with failure of the pose estimation algorithm.