Probabilistic 3D object recognition from 2D invariant view sequence based on similarity

In this paper, a scheme of probabilistic 3D object recognition from 2D view sequence is presented. For the tradeoff between information simplicity and sufficiency, based on similarity measure, proper invariant characteristic set is first acquired in a model view learning procedure. A discrete, nonparametric probabilistic strategy then estimates the overall functional form of posterior probability distribution P(On|I)P(On|I) in several simplifications and modifications, which captures the joint statistics of local pattern and position as well as the statistics of local pattern in the visual world at large, by counting the occurrence frequency of patterns over various objects. The decision is made on maximum a posteriori (MAP) estimation in Bayes decision rule. An evidence accumulation mechanism is finally introduced for effect improvement. Simulation experiment has demonstrated promising results, achieved omnidirectional information, and proved effective, superior and feasible in the approach proposed.