Optimization of the dynamics of vision-based control

Abstract

The tracking performance of visual fixation control systems is investigated and the optimal system configuration found. The performance measure is the maximum change in velocity of the target that can be tracked. The configurations evaluated are serial or parallel image acquisition and processing, and pipeline processing. These are compared with the effect of processing vision data on the fly. The basis of the evaluation is the design of an optimal controller which yields an equivalent transient response behavior for different latencies within the visual feedback system. Applying this controller a relation between system latency and maximum pixel error is derived. This maximum pixel error defines the window size necessary for not loosing the target. The relationship between window size and processing time is then used to find the dynamic performance for all the system configurations. Processing in a pipeline obtains highest velocity due to high cycle rate of the system showing a point of maximum tracking performance. The results obtained for fixation are theoretically expanded to the general visual-serving task. The performance for fixating a target with circular motion is evaluated. Both a feedback control and a more sophisticated feedforward control structure are applied. For the latter an accurate estimation and prediction of the target pose is of importance. For this purpose a Kalman filter is applied which is being tuned in order to obtain high dynamic filtering and prediction quality. The dynamic properties of the steady- state Kalman filter are compared with the time-varying filter. Different dynamic models of the object motion with respect to the camera are considered. A time-varying Kalman filter which is switched from a constant velocity to a constant acceleration model after a motion dependent number of sampling intervals gives best results.