Sensor modeling, Probabilistic Hypothesis Generation, and Robust Localization for Object Recognition

Abstract

In an effort to make object recognition efficient and accurate enough for applications, the authors have developed three techniques; sensor modeling, probabilistic hypothesis generation, and robust localization-which form the basis of a probabilistic object recognition algorithm. To minimize recognition time, these techniques exploit prior knowledge to reduce the number of verifications (the most expensive and critical part of the algorithm) required during recognition. The approach utilizes statistical constraints generated by modeling the entire sensing process, resulting in more accurate constraints on matches. Hypotheses are pruned by a probabilistic algorithm which selects matches based on image evidence and prior statistical constraints. The reliability of the verification decision is increased by robust localization. The authors have implemented these techniques in a system for recognizing polyhedral objects in range images. The results demonstrate accurate recognition while greatly limiting the number of verifications performed.