Analog VLSI Circuits for the Perception of Visual Motion

7

Extended Network Implementations

In this chapter, I introduce two sensors that are extensions of the smooth optical flow chip. Following the recurrent network architectures described in Chapter 4, the motion segmentation chip and the motion selection chip provide the dynamic adaptation of the local smoothness and bias conductances, respectively. The pixel schematics of both sensors are based on the schematics of the optical flow unit. Yet, each sensor follows a different strategy to implement the recurrent network extensions. The motion segmentation chip incorporates the second network within the same focal-plane array on-chip, and so has a very complex pixel design. The motion selection chip, however, only adds the requisite interface and programmability to the optical flow network so that individual bias conductances can be dynamically set in one of two states. The idea is that the recurrent feedback loop is formed with a second, external processor. This digital/analog hybrid approach provides more flexibility in emulating different recurrent network architectures. Both sensors are described in detail and their performance is illustrated with several real-world visual experiments.

7.1 Motion Segmentation Chip

The motion segmentation chip is an attempt to implement the motion segmentation network architecture proposed in Chapter 4 (see Figure 4.13). As shown there, this architecture recurrently connects the optical flow network with a second, two-layer motion discontinuity network ...