Hardware-Software Trade-Offs in Robot Vision

Abstract

Visual perception or, better, vision processing gives rise to a great deal of computational problems which usually have many special properties that, in some cases, make difficult the use of general purpose processors (computers). That’s probably one of the main reasons why image processing hardware and robotic vision have received particular attention during the last decade. In this paper we would like to review some of the hardware-software approaches which have been developed during the last years, mostly for what is called low-level vision. A question arises when we think about the hardware-software trade-off needed to solve this sort of vision processing tasks. How much computational power is required to deal with such kind of tasks?. Obviously this will depend on several factors such as the image size, the complexity of the operations to be performed, in terms of computations, the expected and desired representation and the response time among others.

Since a machine vision system must produce a description of what is imaged, probably one of the most constraining factors will be how to guarantee that such a description will be supplied on time to be useful. Here arises another very important question. What is meant by “real time performance”?. It seems quite clear that the concept of “real time” will change depending on the application. Moreover we cannot talk about a “universal” vision system but about dedicated vision processors.

So, the information expected from a vision system which has to analize aerial imagery won’t be the same as the information needed to direct a robot arm to pick parts off a conveyor belt.

We will try to focus on the robotic vision problem although it isn’t so easy to establish a border or to separe robot vision from computer vision in general.