GPU accelerated elevation map based registration of aerial images

Abstract

This paper proposes a lower latency implementation of the georegistration algorithm proposed by [5]. The algorithm has been modified to mitigate the registration errors and has been parallelized to map to a Graphical Processor Unit (GPU). Also, the target image offset and the painting value computations have been combined to a single loop to eliminate the use of shared memory. Modifications to a current widely used algorithm are proposed. The proposed modified algorithm has been implemented in compute unified device (CUDA) architecture to reduce latency. A fixed coordinate system is used to represent the image, focal, and projection planes. Experimental results show that the proposed algorithm is capable of generating accurate georegistered images for high flying airborne vehicles. While this method has been tested using aerial photographs, it can be extended to Satellite images as well as other image data. A speedup of over 10x has been achieved over the CPU version.