Distributed multiple view fusion for two-arm distance estimation

We propose an approach for estimating the distance of two moving robot arms based on fusion of multiple vision data. The images are fused by simple concatenation of all images followed by a projection of the high-dimensional visual input data into an appropriate low-dimensional subspace. We extended the well known principal component analysis to the so-called output relevant features and present a distributed online computation algorithm performing the projection in parallel. We show that complex sensor data can be efficiently compressed if the robot motions are constrained to a local scenario. The second component of our model is an adaptive B-spline neuro-fuzzy controller whose input space is the constructed subspace and whose output is the estimated distance. Our experimental setup is a two-arm robot system with four uncalibrated cameras. Experiments with a complex circular motion show that the method works even if no robust geometric features can be extracted from the sensor pattern.