Analog VLSI Circuits for the Perception of Visual Motion

4

Visual Motion Perception Networks

Chapter 2 proposed a basic outline for a visual motion perception system. I will now formulate the computational goal of that system as a constraint satisfaction problem, and then derive analog electronic network architectures that find the optimal solution which represents the estimated optical flow.

These networks are considered physical systems rather than methods or algorithms for the estimation of visual motion. A system is an entity built of individual parts that together implement a function. Typically, the function describes a dynamical relation between the input the system receives, and the output that it generates. Also, the function should be defined over the whole range of combinations of input, output, and internal states it can possibly encounter. Otherwise, a supervisor would be required to carefully monitor all states and detect those conditions for which the system does not work correctly; that is, those for which it is not defined. Under these conditions, the supervisor must bypass the system (possibly even reset it) and notify receivers of the output accordingly. It is obvious that such supervision is expensive to maintain and implement when dealing with the sparsely constrained real-world visual environment. Therefore, it is a basic requirement that these networks do not need a supervisor, hence that they are well-conditioned and do not need to be reset.

4.1 Model for Optical Flow Estimation

The brightness constancy constraint ...