Predictive RANSAC: Effective model fitting and tracking approach under heavy noise and outliers

In this paper, we introduce a robust and efficient algorithm, Predictive RANSAC, to fit and track a model in the presence of a large number of outlier measurements and heavy noise. Our algorithm works in two stages. In the model fitting stage, the algorithm first ranks the measurements from most likely to be inliers to least likely. It searches for the best model fit from the top-ranked to the lower-ranked measurements using a similar, but more effective, procedure to the RANSAC algorithm. In the model tracking stage, we use a Kalman filter to predict the location of the model and generate the confidence interval using the prediction error. To update the Kalman filter, the fitting process begins again with an initial guess using the Kalman filter prediction result and searches for inliers in the confidence interval. We also estimate the parameter of the Kalman filter such as noise covariance using the fitting algorithm and a Gaussian-Uniform mixture model. In this way, the two components work together to improve overall performance. To show the performance of the Predictive RANSAC algorithm, we compare our approach with previous algorithms based on RANSAC frameworks over several synthetic data sets and real world data sets, which include road mark fitting, homography estimation and multiple target tracking. The Predictive RANSAC algorithm shows better results in estimation accuracy, and consumes significantly less computational power.