Using over-sampled single-bit representation for velocity estimation in vision systems

Abstract

Machine vision has long been an important topic of study ever since electronic image sensors were developed. One of the main problems in machine vision is achieving a reliable estimate of velocity for objects of interest within the sensor’s visual field. Traditionally, estimating velocity based on full correlations has been impractical due to hefty computational requirements. As a result, most vision chips have adopted simplified models, such as the Reichardt correlator. With advances in digital microelectronics and mixed signal techniques, there appears to be an opportunity for the development of a velocity estimation chip which performs full correlation-based velocity estimates. This paper presents an algorithmic investigation into the feasibility of such a scheme. In our proposed approach, the image information is encoded using a non-linear, over-sampled single-bit representation. Correlation computations are performed on this over-sampled signal, with the reduced precision of the singlebit representation providing a trade-off against the increased sampling rate. The quality of the achievable velocity estimates are evaluated against correlators operating on conventional image sensors, using rotated natural panoramic image sequences as input. Preliminary results suggest the proposed scheme to provide a reasonable estimations of velocity, with the potential advantage of requiring very simple logical circuits.