Recognition of occluded polyhedra from range images

Abstract

Occlusion remains a major hindrance for automatic recognition of 3-D objects. In this paper, we address the occlusion problem in the context of polyhedral object recognition from range data. A novel approach is presented for object recognition based on sound occlusion-guided reasoning for feature distortion analysis and perceptual organization. This type of reasoning enables us to maximize the amount of information extracted from the scene data, thus leading to robust and efficient recognition. The proposed approach is based on a multi-stage matching process, which attempts to recognize scene objects according to their order in the occlusion hierarchy (i.e., an object is recognized before those that are occluded by it). Such a strategy helps in resolving some occlusion-induced ambiguities in feature distortion analysis. Furthermore, it leads to verification of object/pose hypotheses with greater confidence. Matching is based on a hypothesize-cluster-and-verify approach. Hypotheses are generated using an occlusion-tolerant composite feature, a fork, which is a pair of non-parallel edges that belong to the same surface. Generated hypotheses are then clustered and verified using a robust pixel-based technique. Indexing is performed using distortion-adaptive bounds on a rich set of viewpoint-invariant fork attributes, for high selectivity even in the presence of heavy occlusion. Performance of the system is demonstrated using complex multi-object scenes.