We investigate the role of perceptual organization in principal components analysis for vehicle recognition. The principal components analysis is a representative appearance based recognition algorithm that reduces the dimension of search space from the number of pixels to the number of dominant image components. However, its time complexity is high because it performs window operation.
To reduce the search space, we use line junctions with associated quality measures found in perceptual organization methods. Because lines in image or their groupings are inherited from three-dimensional structures, they can form a good initial hypothesis of search process. We define a corner feature as a small image area where the junctions are concentrated, and at the same time, characterize an object. In our application, the corner of the roof inside of the vehicle area is selected, and we name it the corner feature.
By combining principal components analysis and perceptual organization, we could construct an efficient vehicle recognition framework. We also present experimental results using real images of outdoor traffic scenes to demonstrate the efficiency of the algorithm.
