VLSI implementation of biologically inspired vision system for depth perception

Artificial vision systems are used in many different applications including robotics, manufacturing quality control, transportation safety, and biomedical imaging. Generally, these systems are designed to help or even replace a human operator by providing real-time functionality for immediate diagnostics, adaptation, and navigation. In this dissertation we describe the development of a neuromorphic Analog Very Large Scale Integration (AVLSI) integrated circuit (IC) based system, termed the Depth Through Motion Parallax (DTMP) system, for determining the relative depth of objects in a visual scene. The DTMP system uses an algorithm that is based on psychophysical studies of vision in humans and uses the visual cue known as motion parallax. Motion parallax is an important visual cue that determines the relative depth of objects in the field of view based on their relative movement on the retina.; The DTMP system consists of an AVLSI smart focal plane array and an optical lens, with a laser system for the object. The DTMP IC is implemented using the AMI Semiconductor 1.5 mum process available through the MOSIS IC fabrication service. The DTMP chip and the lens are mounted on a single motion controlled stage to mimic the human head and eye. The object is mounted on a different motion stage and is manually moved to different depths with respect to the lens. The IC computes the apparent velocity of the object from the measured transit times (the time it takes for the image to move from one pixel to the next), which is further used to compute its depth information. Further, the DTMP chip consumes an average power of less than 2 mW.; In the process of depth perception there are cases where ambiguity arises in accurately determining the depth of the objects based on their apparent velocity information. Specifically, this work discusses these cases and presents novel methods that can resolve these issues. Here, experimental results obtained for depth perception of stationary objects using the scenario of the head being stationary and the object moving with a constant velocity, as it is geometrically equivalent to object being stationary and the head moving with a constant velocity are reported. Using the current system with a facilitation time window of 640 ms and an optical lens of focal length 10 cm objects placed beyond 18 cm could not be determined. The relative depth range that could be detected was limited by change in the apparent velocity with distance and the percentage error in the perceived velocity obtained was less than 15%. (Abstract shortened by UMI.)...