Evolving Object Detectors with a GPU Accelerated Vision System

Abstract

Abstract. Using GPU processing, it is now possible to develop an evolutionary vision system working at interactive frame rates. Our system uses motion as an important cue to evolve detectors which are able to detect an object when this cue is not available. Object detectors consist of a series of high level operators which are applied to the input image. A matrix of low level point operators are used to recombine the output of the high level operators. With this contribution, we investigate, which image processing operators are most useful for object detection. It was found that the set of image processing operators could be considerably reduced without reducing recognition performance. Reducing the set of operators lead to an increase in speedup compared to a standard CPU implementation. MotivationIn the field of evolutionary computer vision, evolutionary algorithms are used to search for optimal or approximately optimal solutions for computer vision problems [2]. A programmer developing a computer vision algorithm needs to decide in what sequence well known image processing operators have to be arranged to obtain a desired result. In evolutionary computer vision, Genetic Programming [1,13] is used to arrange different image processing operators to obtain a particular output or to perform a given task such as object recognition. This approach is particularly interesting for problems for which the solution is not readily apparent to those skilled in the art. Unfortunately, most experiments in evolutionary computer vision require enormous computational resources because multiple algorithms have to be evaluated over several generations to find an appropriate solution. However, the graphics processing unit (GPU) of a PC can be used for speeding up image processing tasks [7,8]. The GPU is ideal for speeding up image processing as the same operation usually needs to be computed for each image pixel.We have created a GPU accelerated evolutionary image processing system which is able to learn how to detect a user-specified object in an image [3,4].