Scene Reconstruction And Recognition Based On 3D Laser And Monocular Vision

For mobile robot system working in the dynamic indoor environment, it has been a great challenge to make them possessing the ability of information acquisition, cognition and understanding of the environment, and it is considered to be an important topic for researchers at home and abroad. Recognizing the scene where the robot placed in and recovering from accidental location failure efficiently serves an essential task for service robots to work autonomously.Laser range finder and vision sensor are used separately to handle different condition for indoor scene recognition and we puts particular emphasis on using laser range finder. Vision sensor is used in good illumination condition while laser range finder is used when light changes greatly, or there is even no light at all which cause the vision sensor failure. Meanwhile, an information fusion method between camera and laser range finder is proposed in this paper, which make the remote personnel who monitor and control the robot have a good perception of the scene where the robot works in.This paper focuses on the indoor scene recognition using laser range finder. As the construction of feature database for known scenes and the scene recognition experiment all require acquiring laser data at the entrance of the door, we first pick out the door by feature extraction and using semantic scene technique. By processing of 3D laser data, we present the global features and the local SIFT features to describe each scene. Global spatial features are first extracted from laser data to give a general description of the area of a scene and speed up scene recognition. Then 3D laser is transferred into a picture called Bearing Angle image. SIFT features extracted from Bearing Angle image is acted as the local features of the scene. By feature matching between query scene and the ones in database, transfer matrixes are worked out. This paper also introduces a new evaluate function in the evaluate process to pick out the best transfer matrix. Good result is obtained by using the angle neighborhood of laser point and Gaussian weight function. Experimental results and the analysis of data show the efficiency of our method.Besides, static and dynamic disturbances faced by mobile robot frequently are tested, which demonstrates the strong anti-interference ability of our method.