Visual Motion Processing

Produktbeschreibung

The capability to recognize biological motion, i.e. gestures, human actions or face movements is crucial for social interactions, for predators, prey or artifcial systems interacting in a dynamic environment.
In this thesis an artifcial feed-forward neural network for biological motion recognition is proposed. Like its natural counterpart, it consists of multiple layers organized in two streams, one for processing static and one for processing dynamic form information. The key component of the proposed system is a novel unsupervised learning algorithn, called VNMF, that is based on sparsity, non-negativity, inhibition and direction selectivity.
In the frst layer of the dorsal stream, the VNMF is modifed to solve the optical flow estimation problem. In the subsequent layer the VNMF algorithm extracts prototypical patterns, such as optical flow patterns shaped e.g. as moving heads or lim parts. For the ventral stream the VNMF algorithm learns distict gradient structures, resembling edges and corners. All these patterns represent simple cells of the feed-forward hierachy. The classifcation performance of the feed
foward neural network is analyzed on three real world datasets for human action recognition and one face expression recognition dataset, achieving results comparable to current computer vision approaches.

Keywords: Human Action Recognition, Computational Neuro-science, Computer Vision, Machine Learning, Deep Learning, Human Action Recognition, Computational Neuro-science, Computer Vision, Machine Learning, Deep Learning