Adaptive Bayesian recognition with multiple evidences

Ensuring robustness in object recognition/pose estimation under a wide variation of environmental parameters, such as illumination, scale, perspective as well as occlusion, is still of a challenge in computer vision. One way to meet this challenge is by using multiple features/evidences that offer their own strengths against particular environmental variations. To this end, methods of how to choose an optimal combination of features/evidences and of how to design an optimal classifier/decision-maker with the assignment of proper weights to the chosen individual features/evidences, for a given environmental parameter reading, are to be addressed. This paper presents a framework of adaptive Bayesian recognition that puts its particular emphasis on addressing the two methods described above while integrating multiple evidences. The novelty of the proposed method lies in 1) an AND/OR graph representation of evidence structure for individual object, representing explicitly a set of combined evidences sufficient for decision, 2) An automatic update of the Bayesian network tables of conditional probabilities based on the current environmental parameters measured, and 3) the incorporation of occlusions into the computation of Bayesian posterior probabilities for decision. The experimental results show that the proposed method is capable of dealing with adverse situations for which conventional methods fail to provide recognition.