Object detection with vector quantized binary features

Object Detection with Vector Quantized Binary Features John Krumm Intelligent Systems & Robotics Center Sandia National Laboratories Albuquerque, NM 87185 [email protected] Abstract This paper presents a new algorithm for detecting objects in images, one of the fundamental tasks of computer vision. The algorithm extends the representational efficiency of eigenimage methods to binary features, which are less sensitive to illumination changes than gray-level values normally used with eigenimages. Binary features (square subtemplates) are automatically chosen on each training image. Using features rather than whole templates makes the algorithm more robust to background clutter and partial occlusions. Instead of representing the features with real-valued eigenvector principle components, we use binary vector quantization to avoid floating point computations. The object is detected in the image using a simple geometric hash table and Hough transform. On a test of 1000 images, the algorithm works on 99.3%. We present a theoretical analysis of the algorithm in terms of the receiver operating characteristic, which consists of the probabilities of detection and false alarm. We verify this analysis with the results of our 1000-image test, and we use the analysis as a principled way to select some of the algorithm’s important operating parameters. 1. Overview and Context Detecting objects in images and measuring their location is a fundamental task of computer vision, with applications in manufacturing, inspection, world modeling, and target recognition. Often the scene is inherently cluttered, the object may be partially occluded, and illumination may change. In this case, the algorithm must look at features internal to the objects’ silhouette, and look at them in such a way that missing features and changing illumination are tolerated. Researchers have responded to this need in many ways, including fairly recent, elegant object detection algorithms based on principle components of training images of the object[[6]][[10]]. In particular, Murase and Nayar[[6]] extract templates from training images of the object in different orientations, compute eigenvector principle components of these templates, and recover the object’s orientation in new images by projecting them onto the principle components. They address the problem of illumination changes by taking training images under different lighting