Problem: The analysis of optic flow has been a popular topic in psychophysics, neurophysiology and computer vision for many years. Most research has focused either on the low-level vision problem how local motion information can be integrated into smooth optic flow fields that allow an extraction of information about, typically rigid, moving objects, or on the analysis of global optic flow parameters that are useful for the analysis of ego-motion, heading and navigation. A related topic that has not received much attention is the analysis of the complex movement patterns that are generated by articulated biological objects. In spite of its biological relevance and the fact that since Johansson (1973) "biological motion perception" is a classical topic in psychology our knowledge about the neural and computational mechanisms underlying the perception of such complex optic flow patterns is very rudimentary. In my view, research on the neural mechanisms of articulated motion recognition has gained new actuality for the following reasons: (1) Neural principles underlying the recognition of complex stationary shape have been successfully studied within the last few years. Such principles and the experimental methods to study them might be relevant also for the recognition of complex movement patterns. (2) Tracking and recognition of complex movement patterns from articulated objects is presently a popular topic in computer vision. Much clearer ideas about the underlying computational problems exist now than several years ago.

First results: We have developed a neural model for the recognition of biological motion that is compatible with the existing neurophysiological and psychophysical data. The model is based on the following neural principles: (1) Processing in two pathways that analyze form and motion that consist of hierarchies of neural feature detectors that model the properties of cortical neurons. (2) Invariance is achieved by nonlinear pooling of the responses of non-invariant detectors along the hierarchies. (3) Learning of the selectivity of the neural detectors from prototypical example movement patterns. (4) Dynamic neural networks with asymmetric lateral connections account for the temporal association of information on the recognition level.

It is shown that the model achieves a robust recognition of biological motion and reproduces a variety of psychophysical results. Also, the model provides an elegant account for the view-dependence of biological motion perception. The model makes a number of predictions that can be tested in psychophysical, neurophysiological and FMRI experiments. In particular, the model predicts the activity distributions over the relevant areas in the dorsal and ventral pathway for different stimulus classes. Such predictions can be directly tested in FMRI experiments.

Perspectives: The proposed model shows that an explanation of biological motion perception in terms if a hierarchical neural system that is trained with prototypical example movement patterns is computationally feasible. The model and the underlying assumptions lead to a variety of questions that must be clarified experimentally: (1) What are the spatial and temporal generalization properties for learned biological motion patterns? (2) What happens during learning? (3) What are the roles of the ventral and dorsal pathway during the recognition of complex motion patterns. (4) What are the neural mechanisms that associate information over time, and how do they relate to the computational approaches to the same problem in computer vision?

MG is supported by the DFG, Germany, and Honda R & D, Americas. CBCL is supported by ONR and NSF.