Computational image processing for a computer vision system using biomimetic sensors and eigenspace object models

Abstract

Two challenges to an effective, real-world computer vision system are speed and reliable object recognition. Traditional computer vision sensors such as CCD arrays take considerable time to transfer all the pixel values for each image frame to a processing unit. One way to bypass this bottleneck is to design a sensor front-end which uses a biologically-inspired analog, parallel design that offers preprocessing and adaptive circuitry that can produce edge maps in real-time. This biomimetic sensor is based on the eye of the common house fly (<i>Musca domestica</i>). Additionally, this sensor has demonstrated an impressive ability to detect objects at subpixel resolution. However, the format of the image information provided by such a sensor is not a traditional bitmap transfer of the image format and, therefore, requires novel computational manipulations to make best use of this sensor output. The real-world object recognition challenge is being addressed by using a subspace method which uses eigenspace object models created from multiple reference object appearances. In past work, the authors have successfully demonstrated image object recognition techniques for surveillance images of various military targets using such eigenspace appearance representations. This work, which was later extended to partially occluded objects, can be generalized to a wide variety of object recognition applications. The technique is based upon a large body of eigenspace research described elsewhere. Briefly described, the technique creates target models by collecting a set of target images and finding a set of eigenvectors that span the target image space. Once the eigenvectors are found, an eigenspace model (also called a subspace model) of the target is generated by projecting target images on to the eigenspace. New images to be recognized are then projected on to the eigenspace for object recognition. For occluded objects, we project the image on to reduced dimensional subspaces of the original eigenspace (i.e., a “subspace of a subspace” or a “sub-eigenspace”). We then measure how close a match we can achieve when the occluded target image is projected on to a given sub-eigenspace. We have found that this technique can result in significantly improved recognition of occluded objects. In order to manage the combinatorial “explosion” associated with selecting the number of subspaces required and then projecting images on to those sub-eigenspaces for measurement, we use a variation on the A* (called “A-star”) search method. The challenge of tying these two subsystems (the biomimetic sensor and the subspace object recognition module) together into a coherent and robust system is formidable. It requires specialized computational image and signal processing techniques that will be described in this paper, along with preliminary results. The authors believe that this approach will result in a fast, robust computer vision system suitable for the non-ideal real-world environment.