Object Detection in Infrared Images

This paper describes algorithms for detecting and classifying objects such as tanks and trucks in forward-looking infrared (FLIR) imagery. It summarizes research conducted in the course of a two-year project in the areas of image modeling, preand post-processing, segmentation, feature extraction, and classification. 1. Image models The work on image modeling conducted under this project was concentrated in three main areas: i) Modeling of the joint (gray level, edge value) statistics of FLIR scenes, as a basis for defining threshold selection techniques. 2) Modeling of thresholding and edge detection responses to background regions, as a basis for predicting false alarm rates. 3) Modeling edges in images as a basis for defining optimal edge detection operations and for evaluating edge detector output. This work is briefly summarized in the following subsections. References are given to earlier project reports [1-4] in which detailed treatments can be found. i.I Model-based threshold selection An approach to modeling FLIR imagery has been developed, based on the simplifying assumption that targets appear as homogeneous hot regions within a homogeneous cooler surround. This model describes the joint probability density of gray level and edge strength in such images, for various edge-detecting operations [1,2]. In brief, the model predicts that for low edge values (corresponding to points in the interiors of objects and background), there should be two relatively well separated probability peaks, of different sizes, representing the gray levels of object and background interiors, respectively. For higher edge values, corresponding to points on object/background borders, these peaks should merge together and become a single peak representing the