GPU-Accelerated Cross-Modal SAR-EO Image Homography Estimation

This article proposes a massively-parallel approach for solving the difficult problem matching high-altitude image pairs from different domains, e.g., Synthetic Aperture Radar (SAR) and visible light imagery (EO). The through-weather measurement capability of SAR allows this technology to yield vehicle position fixes in inclement weather and during either night or daytime for image-aided navigation. This work focuses on developing capabilities to match across a large range of variations in the unknown parameters of the homography that brings these image pairs into correspondence. This is a problem that is not well-solved by any existing approaches and is important in practice as cross-domain imagery from aerial platforms often exhibits large variations in scale, keystone, rotation and translation effects that can be different in the x and y axes. Our approach for cross-modal image matching uses a mutual information loss function and applies a massively-parallel search procedure in CUDA to detect and explore the loss function to find satisfactory homographies to match the image pairs. Experiments are performed using simulated image telemetry obtained by flying a fixed wing aircraft in a virtual environment with image data derived from Google Maps and RADARSAT Google Earth Engine image databases. Results show a comparison rate of 12.79 Gpixel/sec and has a search rate of 1.8M matches/sec allowing for exhaustive search solutions. Our approach is found to yield accurate homography values according to our normalized corner error metric for 68% of the image database pairs.