CUDA-Accelerated Feature-Based Egomotion Estimation

Egomotion estimation is a fundamental issue in structure from motion and autonomous navigation for mobile robots. Several camera motion estimation methods from a set of variable number of image correspondences have been proposed. Seven- and eight-point methods have been first designed to estimate the fundamental matrix. Five-point methods represent the minimal number of required correspondences to estimate the essential matrix. These feature-based methods raised special interest for their application in a hypothesize-and-test framework to deal with the problem of outliers. This algorithm allows relative pose recovery at the expense of a much higher computational time when dealing with higher ratios of outliers. To solve this problem with a certain amount of speedup, we propose in this work, a CUDA-based solution for the essential matrix estimation from eight, seven and five point correspondences, complemented with robust estimation. The mapping of these algorithms to the CUDA hardware architecture is given in detail as well as the hardware-specific performance considerations. The correspondences in the presented schemes are formulated as bearing vectors to be able to deal with all camera systems. Performance analysis against existing CPU implementations is also given, showing a speedup 4 times faster than the CPU for an outlier ratio \(\epsilon =0.5\) which is common for the essential matrix estimation from automatically computed point correspondences, for the five-point-based estimation. More speedup was shown for the seven and eight-point based implementations reaching 76 times and 57 times respectively.