Purchase this article with an account.
Martin A. Giese; Hierarchical neural model for the recognition of biological motion. Journal of Vision 2001;1(3):356. https://doi.org/10.1167/1.3.356.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The visual system shows an amazing sensitivity for biological motion stimuli. As demonstrated by the classical work of Johansson, such stimuli can be recognized under strongly impoverished stimulus conditions. At the same time, the recognition is highly robust against the presence of occlusions and strong background noise. This raises the question how this robust recognition performance is achieved by the visual cortex. Psychophysical and neurophysiological results about the recognition of complex stationary objects suggest that complex form is analyzed within a hierarchical system of neural detectors which extract increasingly complex features along the dorsal pathway. Very likely the selectivities of the detectors on the highest level of this hierarchy are learned. It seems a reasonable hypothesis to assume that the recognition of complex movement patterns is based on similar neural principles within a neural hierarchical system that encompasses the ventral and the dorsal processing stream. We test if this hypothesis in a computational study by devising a neural model with two hierarchical pathways for the processing of form and motion information. The model neurons reproduce known tuning properties of neurons in the dorsal and the ventral pathway. Along the two pathways the complexity of the extracted features, but also the invariance of the detectors against scaling and translation increases. An additional assumption is that neurons on the recognition level of the hierarchy are laterally connected, giving raise to a recurrent network dynamics that can efficiently associate information over time. We demonstrate that this model reproduces a variety of psychophysical results on biological motion perception, and that it also predicts a number of psychophysical and neurophysiological results. In particular, the model accounts for the view-variance of biological motion recognition and its robustness against reduction of the stimulus quality and clutter. For non-impoverished stimuli, the information conveyed by each individual pathway (dorsal or ventral) alone turns out to be sufficient for a robust recognition of biological motion. The model allows also to predict quantitatively differences between the neural activation in different cortical areas for different stimulus classes. Such predictions are appropriate for comparisons with data from functional imaging experiments (FMRI, PET). Our results support the hypothesis that the recognition biological motion can be accounted for by a hierarchical system of neural detectors for motion and form features with the additional assumption that asymmetric lateral connectivity in the higher visual cortex is exploited to associate information over time.
This PDF is available to Subscribers Only