1 Vision - Based Hazard Detection

Advances in sensor and GPS technologies make possible better guidance tools for vehicle/aerial navigation. In the work reported here we look at the detection of hazards, such as aircraft or other objects, on or near the target runway. Our system consists of two modules: 1) regions of interest (ROI) detection, and 2) hazard recognition. One of the harder problems in object recognition is to segment the target object from a cluttered background. In this system we use a “poor man’s” segmentation, by taking advantage of the fact that we have an approximate reference that we can differentiate from. The regions of interest are defined as significant differences between the input image and the reference image. Since this differencing can be complex due to the fact that the images may not be precisely registered, we employed a novel histogram method which is reasonably invariant to spatial transformations for ROI detection. INTRODUCTION Advances in sensor and GPS technologies have opened up the opportunity of developing better guidance tools for aerial/ground navigation. These guidance tools integrate data routinely from different sensors and GPS and then present to the pilot integrated information about the environment for safe navigation and landing. One important part of the guidance tools is hazard/obstacle detection. As the name implies, it searches throughout the scene to detect if there is any possible hazard (aircraft or truck, for example) on or near the runway and determines the type of hazard if present. Our research goal is to develop such hazard detection using a biologically inspired approach. Existing vision-based methods for obstacle detection can be classified into two types: feature based and optical flow based. In feature based methods (Sridhar B. et al., 1993), significant features are detected from the images and matching between features are made from frame to frame using normalized correlation. But since only features are used, the choice of the feature set becomes very critical for the success of the method. Also any transformation of the features due to scaling or rotation would deteriorate its performance. Optical flow methods based on estimation of temporal or spatial disparities from the image sequences have also been suggested (Sull S. and Sridhar B., 1995). A global flow field is first computed from the whole image. A threshold