Efficient nonparametric kernel density estimation for real time computer vision

Many problems in computer vision, such as recognition, detection and segmentation, involve obtaining the probability density function describing an observed random quantity. While classical parametric densities are mostly unimodal, practical computer vision problems involve multivariate multimodal densities. In general, the forms of the underlying density functions are not known. Kernel density estimation techniques are quite general and powerful methods for this problem. In this dissertation, kernel density estimation techniques are utilized for building statistical representations for the appearance of objects. Such representations are used to facilitate different computer vision tasks such as moving objects detection and target tracking.; We describe an algorithm for background subtraction based on building a statistical representation of the scene background using kernel density estimation techniques. The model can handle situations where the background of the scene is cluttered and not completely static but contains small motions such as tree branches and bushes. We also describe how to build statistical representations of the foreground (moving objects). We present an approach to model the colors of homogeneous image regions using the kernel density estimators to estimate the color probability distributions. Modeling the color distribution of a homogeneous region has a variety of applications for object tracking and recognition. We use this approach to segment foreground regions corresponding to multiple people in order to track them through occlusion. We present a general probabilistic framework that uses maximum likelihood estimation to estimate the best arrangement for people in terms of 2D translation and depth arrangement that yields a segmentation for the foreground regions. We also use kernel density estimators to represent the joint feature-spatial distributions and use this representation to track the targets.; Kernel density estimators have a significant disadvantage in that they are computationally intensive. The dissertation also presents an efficient computational framework for kernel density estimation based on Fast Multipole methods that facilitates the use of this powerful statistical tool in real-time computer vision. We show the application of this algorithm for color modeling and tracking.