Skip to Main Content
A novel structural and statistical approach for model-based visual object recognition using geometric radiance saliencies is presented. The approach acquires accurate high dynamic range images to properly capture complex heterogeneously lighted scenes. Based on these images, the receptive radiance saliency is computed through a Gabor kernel set. This oriented saliency is used to extract and refine the radiance edge-graphs. Subsequently, the combination of two distributions, (i) the topological connectivity and (ii) the spatial arrangement of the subpixel nodes provides a propitious insight into the underlying geometrical composition of the radiance edges. The proposed characterization of the combined distribution profitably unveils and simultaneously segments the geometric edge primitives. Finally, uncovering the complementary geometric patterns and reinforcing the structural regularity is accomplished by the proposed extended perceptual organization. Experimental evaluation with the humanoid robot ARMAR-III is presented.