A genetically optimized graph-based people extraction method for embedded transportation systems real conditions

In this paper, we present a new method for people extraction in complex transport environments. Many background subtraction methods exist in the literature but don't give satisfactory results on complex images acquired in moving trains that include several locks such as fast brightness changes, noise, shadow, scrolling background, etc. To tackle this problem, a new method for people extraction in images is proposed. It is based on an image superpixel segmentation coupled with graph cuts binary clustering, initialized by a state-of-the-art foreground detection method. The proposed strategy is composed of four blocks. A pre-processing block that uses filters and colorimetric invariants to limit the presence of artifacts in images. A foreground detection block that enables to locate moving people in images. A post-treatment block that removes shadow regions of no-interest. A people extraction block that segments the image into SLIC superpixels and performs a graph cut binary clustering to precisely extract people. Tests are realized on a real database of the BOSS European project and are evaluated with the standard F-measure criteria. Since many state-of-the-art methods can be considered in our three first blocks along with many associated parameters, a genetic algorithm is used to automatically find the best methods and parameters of the proposed approach.