Stduy On Object Pose Estimation Based On Monocular Vision

With the development of industrial automation and machine vision, 3D object recognition and pose estimation become more important. Especially in the application of industrial automation equipment and sorting robot, how to determine the pose of the object becomes the focus of research. In domestic industrial manufacturing, most of the automation equipment or robots can only complete the simple object recognition in a fixed depth, such as round, square, etc. Some of the workshops even need workers to sort the objects, which is repeating and time consuming. With the development of machine vision and the improvement of industrial production requirement, people pose a higher demand for pose estimation.Aiming at the problems of object recognition and pose estimation, this dissertation proposes a method of pose estimation based on monocular vision. The object can be recognized by using its geometrical model and its input-image, and this method effectively overcomes the depth problem. Compared with the traditional methods, this dissertation makes improvements from two aspects: object model and object recognition. For object model, this dissertation proposes a method of directly using the object CAD model to generate the 3D model. The first step is to calculate the camera internal parameters and distortion coefficients by camera calibration, and then get the projection matrix. The second step is to project the data of CAD model to 2D through a virtual camera after hiding lines, and train the 2D features combined with grid clustering and image pyramid to build the 3D model of the object. For object recognition, this dissertation transforms the recognition process into the displacement process of the edge points, and then establishes the target error function, that is, transforms the recognition problem into the problem of solving the error function. In 2D image, the edge points and gradient features of object are extracted and matched with the 3D model to get the pose. To speed up, this dissertation proposes a search strategy based on the image pyramid and the grid clustering. The first match is at the top of the pyramid, after determining the coarse pose, the recognition occurs on the next pyramid. The search process is rough to fine and finally recognize the object and its pose.The method is released based on Open CV, Open GL on the Windows platform using Visual studio 2013. In this dissertation, the method is verified by using a number of simulate images and some of the test images. The experimental results show that this method can correctly recognize the target object and its pose, and it is robust to rotation, occlusion and scaling.