Viewpoint-independent object recognition using reduced-dimension point cloud data

Point cloud data offer the potential for viewpoint-independent object recognition based solely on the geometrical information about an object that they contain. We consider two types of one-dimensional data products extracted from point clouds: range histograms and point-separation histograms. We evaluate each histogram in terms of its viewpoint independence. The Jensen-Shannon divergence is used to show that point-separation histograms have the potential for viewpoint independence. We demonstrate viewpoint-independent recognition performance using lidar data sets from two vehicles and a simple algorithm for a two-class recognition problem. We find that point-separation histograms have good potential for viewpoint-independent recognition over a hemisphere.