Generalized neural trees for outdoor scene understanding

Abstract

A new model of a neural tree, called generalized neural tree (GNT), is presented. In the GNT learning process, the whole tree structure is considered at each learning step, and the entire training set is used to update each node. The main novelty of the proposed approach is that the output obtained when a pattern is presented to the network has a probabilistic interpretation. Experimental tests have been performed by applying the GNT in the context of a visual-based surveillance system for outdoor scenes. In particular, objects moving in the observed scene are firstly classified into 5 different categories. Then, the trajectory of such objects, together with the class information is provided to a second GNT which gives a final interpretation of the scene in terms of presence of dangerous situations.