Wide-area recognition using hybrid motion stereo: Outlier rejection for motion stereo using sequential data

Robots working in complex environment require accurate perception of a wide field of view. Using cameras, Hybrid Motion Stereo, which combines the computations by stereo vision and motion stereo, is capable of acquiring positional information of the whole field of view. However, the technique generates errors in computation by motion stereo when tracking feature points fails. This paper describes a method to reject the outliers to avoid erroneous recognition. The method uses computation results from several previous images, computing the spatial deviation and temporal deviation of points, and rejecting those considered to be deviated. The evaluation function is composed with a weighted average of the sequential results, each with a weight of the total number of points existing in the neighboring space, which represents the spatial deviation. Experimental results using the humanoid robot HRP-2 denote the effectivity of the method.