3D Pose Estimation for Planes

Abstract

This paper presents a method to robustly track planes and estimate their 3D poses in a video. A weighted incremental normal estimation method for planes (WINEP) is presented using Bayesian inference. This estimation method guarantees an optimal solution based on all the observations up to the current time, and the computational cost at each time step does not increase with the growing number of past frames. The tracking algorithm integrates boundary information with point feature tracking, which avoids accumulating errors due to intensity changes, image noise, and inaccurate parameter estimation. The tracking algorithm deals with low-textured as well as highly-textured planes. The tracked boundary locations provide the input data for 3D plane pose estimation. Experiments show that our hybrid tracking method using both point and line features is better than using only point features, and our pose estimation algorithm is more robust and accurate than the conventional homography decomposition method, especially under circumstances of noisy observations and low number of input features.