A Computationally Efficient Algorithm for Building Statistical Color Models

Abstract

Though widely used in surveillance systems of human or fire detection, statistical color models suffer from long training time during parametric estimation. To solve this low-dimension huge-number density estimation problem, we propose a computationally efficient algorithm: weighted EM, which learns the parameters of finite mixture distribution from the histogram of training data. Thus by representing data with a small number of parameters, we significantly reduce long-time storage costs. At the same time, estimating parameters from the histogram of relatively small size ensures the computational efficiency. The algorithm can be readily applied to any mixture model which can be estimated by EM and its online learning form is also given in our paper. In the experiment of skin detection, the algorithm is tested in a database of nearly half a billion training samples, and the results show that our algorithm can do density estimation accurately and enjoys significantly better computational and storage efficiency.