Algorithm for Fast Registration of Radar Images

Abstract

Radar imagery provides an all-weather and 24 h coverage, making it ideal for critical defence applications. In some applications, multiple images acquired of an area need to be registered for further processing. Such situations arise for battlefield surveillance based on satellite imagery. The registration has to be done between an earlier (reference) image and a new (live) image. For automated surveillance, registration is a prerequisite for change detection. Speed is essential due to large volumes of data involved and the need for quick responses. The registration transformation is quite simple, being mainly a global translation. (Scale and rotation corrections can be applied based on known camera parameters). The challenge lies in the fact that the radar images are not as feature-rich as optical images and the image content variation can be as high as 90 per cent. Even though the change on the ground may not be drastic, seasonal variations can significantly alter the radar signatures of ground, vegetation, and water bodies. This necessitates a novel approach different from the techniques developed for optical images. An algorithm has been developed that leads to fast registration of radar images, even in the presence of specular noise and significant scene content variation. The key features of this approach are adaptability to sensor/terrain types, ability to handle large content variations and false positive rejection. The present work shows that this algorithm allows for various cost-performance trade-offs, making it suitable for a wide variety of applications. The algorithm, in various cost-performance configurations, is tested on a set of ERS images. Results of such tests have been reported, indicating the performance of the algorithm for various cost-performance trade-offs.