Towards Markerless Motion Estimation of Human Functional Upper Extremity Movement

Markerless motion capture of human movement is a potentially useful approach for providing movement scientists and rehabilitation specialists with a portable and low-cost method for measuring functional upper extremity movement. This is in contrast with optical and inertial motion capture systems, which often require specialized equipment and expertise to use. Existing methods for markerless motion capture have focused on inferring 2D or 3D keypoints on the body and estimating volumetric representations, both using RGB-D. The keypoints and volumes are then used to compute quantities like joint angles and velocity magnitude over time. However, these methods do not have sufficient accuracy to capture fine human motions and, as a result, have largely been restricted to capturing gross movements and rehabilitation games. Furthermore, most of these methods have not used depth images to estimate motion directly. This work proposes using the depth images from an RGB-D camera to compute the upper extremity motion directly by segmenting the upper extremity into components of a kinematic chain, estimating the motion of the rigid portions (i.e., the upper and lower arm) using ICP or Distance Transform across sequential frames, and computing the motion of the end-effector (e.g., wrist) relative to the torso. Methods with data from both the Microsoft Azure Kinect Camera and 9-camera OptiTrack Motive motion capture system (Mocap) were compared. Point Cloud methods performed comparably to Mocap on tracking rotation and velocity of a human arm and could be an affordable alternative to Mocap in the future. While the methods were tested on gross motions, future works would include refining and evaluating these methods for fine motion.