Active perception system for recognition of 3D objects in image sequences

The authors describe an active 3D object recognition system that can learn complex 3D objects completely unsupervised and that can recognize previously learnt objects from different views. First a decision of which is the best next view is taken. The system developed for this task is an iterative active perception system that executes the acquisition of several views of the object, builds a stochastic 3D model of the object and decides is the best next view to be acquired, based on an entropy measure. In this paper, we are focusing on a module for the recognition of objects in image sequences. We evaluate the optical flow in the sequence and extract a set of invariant features. As a pattern recognizer we suggest the cellular neural network (CNN) architecture and generate an associative memory. The CNN paradigm is considered as a unifying model for spatio-temporal properties of the visual system.