This paper describes an innovative algorithm that estimates the tilt angle of the camera and detects single or multiple ground planes with a RGB-D camera. All of the 3D points given by the RGB-D camera are projected onto a normal vector defined by an assumed tilt angle of the camera to obtain the height distribution. The probability density of it is calculated with the Kernel Density Estimator, which is used to select the ground plane reliably. The assumed camera tilt angle is refined with the height distribution, which is used to improve the accuracy of the ground detection. Our algorithm is evaluated on several challenging data sets. The experimental results show it is able to find parallel and unparalleled planes, and hold high accuracy and processing speed over the traditional methods.
Yankun LANG, Haiyuan WU, Toshiyuki AMANO, QianCHEN
Faculty of System Engineering, Wakayama University