A method for expecting the features of objects and enabling real-time vision processing

This paper presents a mathematical analysis of image processing, algorithms designed according to the results of this analysis, and their implementation. We prove that the search of objects features can be accelerated without loss of precision by using an inhomogeneous density of the sensitive cells the parameters space is composed of. In other words, the visual analysis should be concentrated in the region of the features space around the expected object position. The improvement relative to an uniform cell density is quantified using a cost function corresponding to time and precision optimisation. We show that a Kohonen neural network can be used for efficient image processing, and simulate this strategy. We introduce a simpler algorithm for the case that the object positions are Gauss-distributed around the expected position. This algorithm has been implemented it on a robot guided by a vision system. The robot learned to process images efficiently during the manoeuvres and after that was able to track objects moving in a fast and unpredictable manner