Real-time ground plane segmentation and obstacle detection for mobile robot navigation

In this paper, we propose a real-time ground plane extraction and obstacle detection technique for mobile robot navigation based on a combination of segmentation and optical flow techniques using monocular image sequences. The ground plane, which is captured using a calibrated camera, mounted on a robot platform, has been segmented in a two step process. In the first step which is an offline process, the statistical properties of the ground plane are learned using a robust Gaussian Mixture Model (GMM) based segmentation method. In the online process, the ground plane is segmented using its learned signatures. Planar homography, between the 3D ground plane and the image plane has been used in depth map generation for the segmented images. We have also used the theoretical model of the optical flow field for the real-time moving obstacles detection. Regions of the ground plane, violating this theoretical model indicates the potential obstacle.