Learning-Based Image Ground Segmentation Using Multiple Cues

Image ground segmentation is an important task in the area of computer vision for the robot navigation because the ground region is often taken as the traversable terrain. In this paper, we propose a learning-based method for image ground segmentation which applies the Adaboost learning method to combine multiple cues for detection of the ground region in an image. Firstly, an image is segmented into a number of small regions. Color, texture, location and shape cues are then extracted for the representation of each image region. Finally, the ground classifier is designed using Adaboost learning method and is applied to label the image regions as ground or non-ground. This approach not only can detect the grounds with different appearances for long range perception, but also identify obstacles which have the similar appearance as the ground. Experimental results are presented to show the effectiveness of the proposed algorithm for the ground segmentation of images in a wide range of scenes.