Evaluation Metrics and Results of Human Arm Movement Imitation. Pomplun, M. & Matarić, M. 2000.
abstract   bibtex   
We present a psychophysical study of human arm movement imitation,and an approach to analyzing the resulting data, which can be applied to human or humanoid movement analysis. We describe a joint-space based segmentation and comparison algorithm that allow us to evaluate the performance of 11 different subjects on a series of arm movement imitation tasks. The results provide analytical evidence for the strong interference effects of simultaneous rehearsal during observation. Additionally, the results also demonstrate that repeated imitation in these tasks did not affect the subjects� performance.
@Conference{Pomplun2000,
  author    = {Pomplun, M. and Matari\'{c}, M.},
  title     = {Evaluation Metrics and Results of Human Arm Movement Imitation},
  booktitle = {Proceedings of IEEE/RAS International Conference on Humanoid Robotics},
  year      = {2000},
  abstract  = {We present a psychophysical study of human arm movement imitation,and an approach to analyzing the resulting data, which can be applied to human or humanoid movement analysis. We describe a joint-space based segmentation and comparison algorithm that allow us to evaluate the performance of 11 different subjects on a series of arm movement imitation tasks. The results provide analytical evidence for the strong interference effects of simultaneous rehearsal during observation. Additionally, the results also demonstrate that repeated imitation in these tasks did not affect the subjects� performance.},
  groups    = {STAT841},
  keywords  = {motion segmentation},
  review    = {Data input: FastTrak motion tracking

Wanted to look at effects of rehearsal in human arm motion imitation (the original motion was video-recorded and replayed for each subject). The more significant part is the comparison algorithm between the sample and the subject. They used RMS on distance, but different velocity meant distance traveled over some time would be different between the two, so they did time scaling on the data points so both sets of motion were executed over the same amount of time. 

The imitated motion consists of several smaller actions, which was segmented based on "velocity turning points" (since velocity of a limb segment would approach zero in order to change direction). In the event that some segment was just slow and the algorithm splits that into several segments, the segments were combined by testing combinations of segments and see which one lead to the smallest total temporal deviations between the sample and the subject. They also had a "minimum time", the shortest possible duration for a segment.

Anyway, ANOVA shows that there is a significant difference between rehearsed and non-rehearsed actions.



Pomplun and Matari\'{c} \cite{Pomplun2000} employed zero-velocity crossings (ZVCs) to identify points where the velocity value changes sign, denoting when a joint segment direction change, as segment points. If multiple joints are examined simultaneously, segment points can be declared by thresholding the sum of squares of the velocities. A minimum threshold for segments was included, to prevent spurious ZVCs from creating large numbers of false positives. Although a fast algorithm, ZVC tends to over-segment, particularly with noisy data or with increasing number of DOFs. Since ZVC does not consider motion templates, it is difficult to determine which crossing points can be safely ignored. In addition, the ZVC algorithm does not provide a method for motion identification. Pomplun suggests that a distance metric can be used, but these metrics are sensitive to spatial and temporal variations, and may not provide reliable movement labels. The algorithm was used to assess a 11-subject study on human imitation learning, where video clips of arm motions were shown to the participants. The participants were either instructed to practice the observed motion before data collection, or not. The collected data was segmented and compared against the demonstrator's motions. Since the study was on the impact of rehearsing on the ability to accurately reproduce a motion, segmentation accuracy of the algorithm was not explicitly reported.},
  timestamp = {2009.09.11},
}

Downloads: 0