Object Modeling Using Gaussian Mixture Model for Infrared Image and its Application to Vehicle Detection

In this paper, we propose a new object modeling method for infrared (IR) image. It is based on the modeling method using Gaussian Mixture Model (GMM) that has been originally proposed for general visible image. The original method is one of effective object modeling algorithms that can describe the topological structures of the internal patterns of object. This approach can also eliminate the influences due to small differences between patterns. On the other hand, an IR image is often used instead of visible image in actual applications such as outdoor surveillance. IR images make it easy to extract foreground object regions from background scenes, but their low-contrast makes object modeling difficult. We therefore propose a modeling method using Orientation-Code for IR image. Orientation-Code of each pixel has information about the maximum-gradient orientation of image, not intensity information. Gradient orientation information does not depend on contrast and describes internal pattern structures of objects even in unclear IR images. We also applied proposed method to vehicle detection for outdoor scenes, where it extracts multiple foreground regions as vehicle candidates using background subtraction for IR image, and they are described as models by our method. Models are finally compared with standard vehicle view models pre-memorized to determine which candidate is true vehicle or not. Evaluation tests with actual IR video sequences have proved that our proposed algorithm detects objects robustly.