A method for crowd separation, the method may include receiving or generating a signature of an image of a first plurality of pedestrians, wherein the image comprises regions, wherein the signature of the image comprises descriptors of the regions, wherein each descriptor of a region is associated with a region and comprises a set of identifiers that identify content included in the region and in a vicinity of the region; detecting, within the descriptors of the regions, unique combinations of identifiers that are indicative of spatial relationships between the regions and bounding boxes that surround pedestrians of the first plurality of pedestrians; wherein the unique combinations are learnt during a supervised machine learning process that is fed with test images of densely positioned pedestrians; and locating the pedestrians of the first plurality of pedestrians based, at least in part, on the spatial relationships related to detected unique combinations and to locations of the regions.