A video of selected movements is provided on the accompanying CD and by the following link.
Video 2.1: Appendices/FMC Video/Nancy 2005.AVI
2.1.4.
Discussion
The results indicate that it is possible to use IMUs to capture some types of motion. It is not possible to determine absolute accuracy quantitatively from this experiment, but qualitative analysis of the motion is possible. For slow, short movements the captured motion appeared to resemble the actual motion. The slow dance movements were successfully captured, (Figure 2.2 and Video 2.1). The method successfully mapped each sensor to the subject‟s correct body segment and converted the numerical data into to a real time animation.
The system unfortunately failed to accurately capture the motion of rapid movements. Of concern with regard to ski racing motion capture, are the results from the rotating chair. The centripetal acceleration experienced by the distal limb segments while rotating on the chair caused the erroneous motion in the animation. The hypothesis that centripetal acceleration is an issue is corroborated by the fact that the limb segments became erroneously more elevated when they experienced more centripetal acceleration as a result of the chair being rotated faster.
The algorithm provided by the vendor was unable to adapt to the dynamic movements of the subject, possibly because of incorrect assumptions made in the vendor‟s algorithm. In the vendor‟s algorithm the acceleration, measured by the IMU accelerometers, was assumed to approximate the global Z-axis (reaction to Earth‟s gravity) and was used to correct for the pitch and roll drift of the IMU, (XSens, 2004b). Unfortunately if an IMU is attached to a body segment undergoing continuous rotation, the acceleration, measured in the IMU local reference frame does not always converge to the global Z-axis.
For example, for the distal limb segments in the rotating chair test, the measured acceleration was a combination of reaction to gravity and centripetal acceleration and the resultant acceleration vector was deflected into the centre of rotation. Consequently, the acceleration measured in the IMU reference frame did not approximate the global Z-axis, which may have caused large errors in the calculated orientation. The vendor‟s algorithm was provided as a „black box‟, so it is not possible to confirm if this is the underlying reason for the high errors in IMU orientation observed.
In the walking tests the captured motion initially appeared realistic, but over time the solution deteriorated. It is not known why this occurred but possibly the long-term drift was a result of small magnetic disturbances in the laboratory, or instabilities in the motion capture system. In skiing there are large accelerations of four to six times the gravitational acceleration when the athlete turns at speed. In skiing there may also be magnetic disturbances from the athlete‟s equipment and ski area infrastructure and a ski race is over a minute in duration. The system as purchased is therefore unsuitable for the biomechanical analysis of body segment orientation in skiing. The poor results were a serious setback for the project and, as a consequence, further investigations into the properties of the sensors were required.
2.2. The Pendulum
After the discovery that the IMU based motion capture system purchased was unlikely to provide accurate motion capture of skiing, two questions emerged:
Could more be found out about the accuracy of IMUs in dynamic situations? Could better data fusion algorithms for dynamic motion be designed?
In order to attempt to answer these questions a simple pendulum experiment was devised. The motion of a simple pendulum is repeatable, predictable and continuous. If the pendulum swing orientation could be measured accurately with an IMU, then motion capture systems based on IMUs might be feasible for biomechanical research. Further information about this experiment is provided on the accompanying CD (Brodie, et al., 2008a).
Figure 2.3: The pendulum set up
IMUs produce raw data in a local reference frame, and typically a Kalman filter is used to process the raw data to obtain orientation and measurements in the global reference frame. (Brodie, et al., 2008a, 2008d; Giansanti, Maccioni, Benvenuti, & Macellari, 2007; Luinge & Veltink, 2005; Luinge, Veltink, & Baten, 2007; Pfau, Witte, & Wilson, 2005; Waegli, Skaloud, Ducret, & Roland, 2007; XSens, 2004a)
The IMUs used in the experiments reported here were MT9 sensors from XSens Technologies. XSens provide no details about the internal workings of their Kalman filter supplied with the IMUs except that dynamic orientation accuracy is „less than 3º RMS‟ (XSens, 2004a). This is consistent with the results of other research (Luinge & Veltink, 2005):
…„Although the problem of integration drift around the global vertical continuously increased in the order of 0.5ºs−1, the inclination estimate was accurate within 3º RMS.‟…
The vendor specification also came with the following fine print: “…may depend on the type of motion measured”. More research was required therefore to determine what „type of motion‟ could be measured accurately. It was assumed that the Kalman filter implementation was similar to that used by Luinge and Veltink with the addition that the magnetometer data were used to prevent heading drift (integration drift around the global vertical).