Generic object discrimination for mobile assistive robots using projective light diffusion

A number of assistive robot services depend on the classification of objects while dealing with an increased volume of sensory data, scene variability and limited computational resources. We propose using more concise representations via a seamless combination of photometric and geometric features fused by exploiting local photometric/geometric correlation and employing domain transform filtering in order to recover scene structure. This is obtained through a projective light diffusion imaging process (PLDI) which allows capturing surface orientation, image edges and global depth gradients into a single image. Object candidates are finally encoded into a discriminative, wavelet-based descriptor allowing very fast object queries. Experiments with an indoor robot demonstrate improved classification performance compared to alternative methods and an overall superior discriminative power compared to state-of-the-art unsupervised descriptors within ModelNet10 benchmark.