3-Point RANSAC for fast vision based rotation estimation using GPU technology

In many sensor fusion algorithms, the vision based RANdom Sample Consensus (RANSAC) method is used for estimating motion parameters for autonomous robots. Usually such algorithms estimate both translation and rotation parameters together which makes them inefficient solutions for merely rotation estimation purposes. This paper presents a novel 3-point RANSAC algorithm for estimating only the rotation parameters between two camera frames which can be utilized as a high rate source of information for a camera-IMU sensor fusion system. The main advantage of our proposed approach is that it performs less computations and requires fewer iterations for achieving the best result. Despite many previous works that validate each hypothesis for all of data points and count the number of inliers for it, we use a voting based scheme for selecting the best rotation among all primary answers. This methodology is much more faster than the traditional inlier based approach and is more efficient for parallel implementation of RANSAC iterations. We also investigate parallel implementation of the proposed 3-point RANSAC using CUDA technology which leads to a great improvement in the processing time of estimation algorithm. We have utilized real datasets for evaluation of our algorithm and also compared it with the well-known 8-point algorithm in terms of accuracy and speed. The results show that the proposed approach improves the speed of estimation algorithm up to 150 times faster than the 8-point algorithm with similar accuracy.