A New Approach to Adaptively Segment Natural Scene Images for Target Localization for Real-Time Applications

Abstract

This work formulates a new method to segment target areas from natural scene images that have a grey scale average different from that of the background. Most of the available methods suffer from high implementation complexity which makes them unsuitable for real-time applications such as robotics. The presented method is an enhancement to an existing method which can only segment bright targets. This is done by using a new graphical strategy to extract the background class using the analysis of image’s wavelet subband PDF. The new strategy uses the center of masses of the accepted target threshold groups rather than using the globally maximum local minima which has dramatically improved the results. Moreover a novel validation scheme is introduced which simply eliminates the unwanted detections based on their characteristics. The method is applied to a variety of real world cases and is found to give exceptional results.