Real-time object recognition using relational dependency based on graphical model

Abstract

This paper proposes a real-time object recognition using the relational dependency among the objects that is represented by the graphical model. When we recognize the objects, it is effective to use the relational dependency in which several different objects co-exist each other. The relational dependency has been modeled by the transition matrix in the graphical model. The transition matrix precisely represents the conditional probability of object's existence at time t, given the existence of others at time t - 1 . We use a very fast cascaded adaboost detector in order to detect all object candidates in the image. Then, the existence probability of the object from a given object candidate is estimated by a logistic regression using the softmax function. The estimated existence probability is updated by the trained transition matrix to reflect the relational dependency of the objects. The object's existence is determined by the threshold level. Experiment results validate that the proposed method is a very fast and effective way of recognizing the objects in terms of high recognition rate and low false alarm rate.