Predictive-RANSAC: An effective data fitting and tracking method with application in curve tracking in video
Yingmao Li, Nicholas Gans · 2016
In this paper, we propose the Predictive RANSAC algorithm to efficiently fit measurement data in the presence of outliers and track the model with a Kalman filter. We model the measurement data by a mixture of Gaussian-Uniform random processes. We estimate the covariance of the inlier data and the weighting factor of the mixture once the best fitting is found. A Kalman filter predicts the location of the model and generates the confidence interval using the prediction error. Each RANSAC iteration begins with an initial guess using the Kalman filter prediction result and searches for inliers in the confidence interval. The Kalman filter makes updates based on the model produced by RANSAC , and also tunes parameters such as noise covariance using the RANSAC algorithm. In this way, the two components work together to improve overall performance. This algorithm is designed to be computational efficient and can be implemented on embedded systems hardware. We apply the Predictive RANSAC algorithm to detect road boundaries in video. The experimental results on a PC show that it is able to run much faster than real-time. The accuracy of our approach is proved using simulations and manually labeled video clips of lanes under a variety of shapes and road conditions.