A Fast Floor Segmentation Algorithm for Visual-Based Robot Navigation

Abstract

We present a novel technique that robustly segments free-space for robot navigation purposes. In particular, we are interested in a reactive visual navigation, in which the rapid and accurate detection of free space where the robot can navigate is crucial. Contrary to existing methods that use multiple cameras in different configurations, we use a downward-facing monocular camera to search for free space in a large and complicated room environment. The proposed approach combines two techniques. First, we apply the Simple Linear Iterative Clustering super-pixel algorithm to the input images. Then, by relying on particular characteristics of floor superpixels, we use a simple classification method based on a normalized SSD similarity measure to group together those superpixels that belongs to the floor (considered as free space). The method intermittently examines low resolution images (80 × 60) in the CIE Lab color model. Experimental results show that our segmentation approach is robust, even in the presence of severe specular reflections and allows for real-time navigation.