Object Recognition Using Stereo Vision and Higher Order Spectra

This paper presents a novel technique to conduct correspondence-free stereo processing. Traditional stereo processing estimates structure through computationally expensive correspondence algorithms resulting in the generation of depth maps. If the purpose of estimating structure is to recognise objects, the depth maps (structural information) need to be further processed to produce features which can then be used for recognition. This paper outlines an alternative technique that bypasses the correspondence algorithms to combine the estimation of depth and feature extraction into one step. This is achieved through the application of Higher Order Spectra (HOS) analysis. Results are presented that demonstrate a Top 2 Rank identification rate of 80% for a small database of human faces.