5. MC-RANSAC: A Pre-processing Model for RANSAC using Monte Carlo method implemented on a GPU
Guide: Dr. K Muthunagai, Vellore Institute of Technology, Chennai, India
Probability and Statistics Course Project, Chennai, India
Published at the Student Research Symposium, ICACCI 2013, Mysore, India
RANSAC is a repeating hypothesize-and-verify procedure for parameter estimation and filtering of noise or outlier data. In the traditional approach, this algorithm is evaluated without any prior information on the set of data points which leads to an increase in the number of iterations and compute time. In this work, we implemented a GPU based RANSAC algorithm with pre-processing of the assumed sample set of hypothetical inliers by Monte Carlo method. Based on our implementation and results using the Point Cloud Library and NVIDIA CUDA framework for data intensive tasks we obtain significant improvement in the performance of plane segmentation algorithm over the randomly sampled subset of hypothetical inliers. The final consensus set is formed with less number of iterations using our pre-processing model. We can conclude that a pre-processed sample set of hypothetical inliers results in a faster determination of the consensus set consisting of maximum inliers
The results of MC-RANSAC on GPU were compared with the traditional RANSAC and a significant change in the number of iterations was observed. With significant advances in accelerators, our model performed better than the existing RANSAC extensions proposed on the CPU. In order to scale the proposed algorithm, better hardware capabilities would sufficiently enhance the performance. Also, enhanced sampling techniques can further increase the determination of the consensus set with minimum number of iterations
References: Fisher, M., Bolles, R.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Comm. of the ACM 24(6), 381-395 (1981).