Moving posture reconstruction from perspective projections of jointed figure motion
Jianmin Zhao, Norman I. Badler · ScholarlyCommons (University of Pennsylvania) · 1993
Our goal is to reproduce a human figure's motion with a computer simulated human figure: Given a sequence of perspective projections of a set of feature joints of the moving figure, we tried to recover the original 3D postures through an accurate human figure model and the continuity requirement (temporal coherence) in the sequence. Our approach follows two clues: Given the human figure model, the responsible posture for a frame is constrained by the projections of all the feature joints and, in this limited set of postures we can choose one based on the postures in the previous frames and the temporal coherence, unless there occurs a critical condition, when the projection ray of a feature is perpendicular to the link of which the feature is the distal end. Owing to the fast inverse kinematics algorithm we developed to solve the spatial constraints, we were able to exploit the temporal coherence in projection sequences of frequencies as high as 100 Hz. We used finite state automata to detect critical conditions, and developed various strategies to overcome special difficulties around critical frames. Furthermore, we investigated the impact of errors in linear measurements of body parts on the reconstruction process. Based on mathematical analysis, we proposed some heuristics to discover and recover from the possible modeling errors. To test the theory, we implemented an experimental system. By imposing the temporal coherence constraint whenever possible, this system responds to the incoming images almost linearly: Since the error-prone critical conditions are detected and handled at the very early stage, the system is able to do away with endless recursive backtracking so that only one level of roll-back is needed to handle a limited number of critical conditions whose chances of occurrence are independent of the sampling rate. The system admits generic human motion. It has been tested on synthesized images from actual 3D human motions. Since we knew the original motion, we were able to evaluate results quantitatively. It turned out that the reconstructed motions agreed with the original ones not only in general but also in fine details.