Extracting viewpoint invariance relations using fuzzy sets

An autonomous vehicle equipped with a vision system constantly analyzes 2D images to derive scene descriptions as it navigates around the 3D environment. Deriving scene descriptions requires searching and matching the extracted image features with the model base. The number of possible matches between image features and model features is virtually combinatorially large due to instability of features caused by viewpoints. A crucial issue is the type of features used for matching. The authors propose using viewpoint invariance, relations (VIRs) resulting from the grouping process by perceptual organization. Extracting and measuring VIR in a noise-free image is relatively straightforward. This is not the case in practice due to uncertainty inherents in image formation process and imprecision of information obtained from the process. The authors use fuzzy set theory to deal with this source of uncertainty. Each type of VIR is treated as a fuzzy set characterized by its associated membership function. The membership function is defined by geometric parameters defining the nonfuzzy VIR. The degree to which an image feature is a member of a VIR is determined by the associating membership function. The measured degree of membership for each set of image features is aggregated to obtain an overall degree of measure which signifies the degree of evidence to which the unknown is an instance of an object model