Development of an integrated automatic image registration scheme

Abstract

Photogrammetric applications in mapping and earth observation often result into series of overlapping image pairs that must be accurately fused together to obtain mosaics, which provide holistic and stereoscopic view of the geographical scene of interest, and also ensure accurate spatial details extraction. Manual approach to this fusion is time consuming, rigorous, and very prone to mismatch, hence, the need for the complete automation of the entire image registration process. This study presents the development of an integrated automatic image registration scheme which is aimed at improving the overall speed and accuracy of an image registration process. The Modified Harris and Stephens Corner Detector (MHCD), Speeded Up Robust Features (SURF) and Scale Invariant Feature Transform (SIFT) were independently implemented as the feature descriptors in the development of the image registration scheme. These feature descriptors were then individually integrated with epipolar constraints in the development of hybrid algorithms for fast and accurate automatic image registration. Other algorithms integrated in the development of the image registration scheme are the Random Sampling Consensus (RANSAC), which is a model fitting algorithm for exclusion of outliers from the extracted corresponding features, Sum of Absolute Difference (SAD) as the feature matching metric, Bilinear interpolation for image resampling and the eight-point algorithm for the estimation of fundamental matrix. The developed image registration algorithms were tested in two different registration campaigns using images extracted from Google Earth and image pairs acquired using a DJI Phantom 4 Unmanned Aerial vehicle (UAV), for the first and second campaigns respectively. The robustness of the image registration scheme was tested using the estimated speed, and the accuracy (using signed distances and Root Mean Square Errors(RMSE)) of the algorithms. The result of the first registration campaign shows that the integrated MHCD-epipolar algorithm, with RMSE of , was more accurate and 4 times faster (with a speed of 463 milliseconds) than the conventional MHCD-based image registration algorithm while the integrated SURF-epipolar algorithm (with a speed of 1153 milliseconds) and SIFT-epipolar algorithm (with a speed of 1630 milliseconds) were 2.5 times and 2 times faster than the conventional SURF and SIFT-based image registration algorithms respectively. For the second registration campaign, the integrated MHCD-epipolar algorithm with a RMSE of also proved to be more accurate and approximately 4 times faster (with a speed of 1770 milliseconds) than the conventional MHCD algorithm (with a speed of 6649 milliseconds) while the integrated SURF-epipolar algorithm was more accurate (with a RMSE of ) and approximately 2 times faster (with a speed of 7055 milliseconds) than the conventional SURF algorithm (with a speed of 13109 milliseconds). The integrated SIFT-epipolar algorithm was also more accurate (with a RMSE of ) and approximately 2 times faster (with a speed of 5923 milliseconds) than the conventional SIFT algorithm. The result obtained shows that the developed integrated epipolar correlation image registration algorithms outperformed the conventional feature descriptors in terms of speed and accuracy of the image registration. The research recommends that the integrated MHCD-epipolar correlation image registration algorithm should be adopted when both speed and accuracy of the image registration is of specific interest to Photogrammetrists and image registration analysts.