Object Recognition And Localization Based On RGB-D Data

In recent years, with the accelerating process of intelligence, the service robotic has got rapid development. As the necessary function, object recognition and localization for intelligent service robot has become research focus. The capabilities of accurate object recognition and localization is a precondition for a robot to perform specified tasks, which has extensively applied in daily life, such as pouring water, cooking, mopping the floor and so on.Most of traditional object recognition and localization methods are based on color image, and the object is recognized and located by the means of extracting some related features from color images. Although these methods have achieved good results, both the shape recognition and the pose estimation of the 6-DOF need plenty of space information because the natural characteristics of the object is a three-dimensional, Therefore, in this thesis, the two-dimensional RGB image data in combination with 3 D point cloud data, the method of object recognition and location based on RGB-D data is proposed.In this thesis, the process of establishing objects model database is introduced firstly by using KinectV2 sensors to collect two-dimensional and three-dimensional data. Then, according to the two-dimensional RGB image data, SURF features are extracted for origin recognition of scene object. Thereafter, mapping the result of origin recognition to three-dimensional point cloud data, and using the method of Min-Cut segment object point cloud from the scene point clouds. Finally, using viewpoint feature histogram descriptor for accurately recognize object and estimate 6DOF pose. Furthermore, in order to get a more accurate localization, a novel method is proposed which combines 6-DOF pose estimation with point clouds registration. The experimental results show that this method is effective and feasible.