Target Recognition And Grasping Location Based On Binocular Stereo Vision

Grasping or operation of the object is a task for the intelligent robot that leaves the lab, and enters the family environment to service for people. Target object recognition and localization for the intelligent robot is the precondition that successfully grab objects, however, intelligent robot is not as humans can easily identify and locate the target objects in complex environment. There are many kinds of sensors to help the robot percept environment information, binocular stereo vision sensor is widely used in target recognition and location for its abundant environment information and the perception of scene depth information.In this paper, we research methods and processes involved by target recognition and location based on binocular stereo vision for the purpose of grasping, in the stage of target recognition, we use image processing techniques to extract the local invariant descriptor that can represent the target object for the target recognition; In the target location stage, the center of the object target contour is obtained by image segmentation method, and combining with 3 d reconstruction principle to solve the contour center three-dimensional coordinates, then we estimate posture for the target object based on the relationship of the target object template and the actual scene objects. In this paper, we summarized as follows:Firstly, the camera calibration process, we analyze four coordinates transformation relation and linear imaging model and nonlinear imaging model in the camera imaging process, then Zhang Zhengyou’s camera calibration method used for internal and external parameters of camera calibration, the camera calibration accuracy is verified by experiment.Secondly, in the stage of target recognition, we select local invariant feature descriptor SIFT that represent the target object to target recognition based on the research of image features. In order to identify the target from different angles, we also establish a target object image database. After completing target recognition, we estimate target object area using the homographic matrix between template matching with the actual scene images.Finally, in the target location stage, we extract the target contour by image segmentation Grab Cut algorithm, then combining SIFT algorithm and Grab Cut algorithm to solve the problem that Grab Cut algorithm need artificial initialization. Due to the left and right scene around target contour center mismatching, so we use the template matching algorithm based on gray to look for matching point of the left and right scene images for 3 d reconstruction of contour cente. Finally, we estimate posture by the corresponding relationship between matching the images and scene images, and verify the accuracy of the estimation.