Computational architecture of a visual model for biological motions segregation

Abstract

ABSTRACTThis paper outlines a neural model inspired by the dorsal stream of the visual system for motion recognition. Two areas are considered: the primary visual area (V1) and the middle temporal area (MT). In model area, V1 neurons are organized to detect eight local motion directions. MT is modelled using classical receptive field (CRF), where the cells respond to wide-field motion. The biological motion can be identified through spatial motion dynamics for the limbs and body. In this article, we propose spatio-temporal sampling detectors, where a set of circular masks over motion scenario are utilized to detect the motion dynamics. Two alternative mechanisms, Max-pooling and Sum-pooling, are used to extracting spatio-temporal descriptors from motion energy occupied by the circular masks. To improve the classification results, centroid kinematics is added to the feature vectors, where this feature contributes substantially to characterizing the motion pattern of an action. We evaluate our model by using two challenging datasets: the Weizmann biological action dataset and the KTH biological motion dataset. Our results reflect the potential of spatio-temporal sampling detectors in describing the biological motion of body and limbs using only short video frames (snippets). In addition, the centroid kinematic feature improves the recognition rate and refines the action classification.