Spatial Pose Estimation Based On A Single Camera

With the development of science and technology,social productivity levels keep rising.Recently,computer vision systems are increasingly widely used in industrial field.Instead of the traditional manual way,computer vision accomplishes several basic tasks such as image processing,pattern recognition,image understanding and geometric modeling through image sensors and personal computers.A new "Industry 4.0" era dominated by intelligent manufacturing is approaching as for now.The research on computer vision appears more significant under this situation.Served to the industrial project "iPad drop detection",this thesis proposes a fast algorithm which computes spatial pose for the space shape description part in computer vision.Experiments are designed to verify the performance of the proposed algorithm.The algorithm handles the spatial pose estimation problem of the target object by explaining the physical meaning of rotation transformation for rigid body in 3D space.To do this,lens distortion is corrected based on the deep meaning of camera model beforehand.The main contributions of this thesis are summarized as follows:? Firstly,this thesis studies the classic linear camera model,looks in detail at the internal and external parameters of camera,and summarizes the conventional camera calibration method.? In traditional camera calibration algorithms,lens distortion is coupled with intrinsic and extrinsic camera parameters,and thus results in high errors on the camera parameters.To solve this problem,a nonmetric lens distortion calibration method based on collinear vectors is proposed to separate lens distortion from camera parameters.In addition,the coupling between distortion center and distortion coefficients is also discussed,and an evaluation function is then designed to verify the completion of distortion correction.After the distortion is corrected,intrinsic and extrinsic camera parameters can be solved linearly.The proposed method only needs one calibration image to solve all camera parameters,and thus can improve the efficiency of calibration.? By reviewing the classic theory of binocular stereo vision,this thesis points out that the traditional methods still exist problems in spatial pose estimation,including limited detection range,bad stereo matching precision and huge amount of computation.To handle the problem,this paper deeply analyzes rigid body transformation process in 3D space,proposes a new interpretation for the rotation matrix between two coordinate systems and assigns physical meaning for every term in rotation matrix.On this basis,a spatial pose estimation algorithm is proposed.We also extend the rotation relation between two planes into multiple planes.Experimental results confirm that the proposed algorithm keeps fast and effective in the related scenes.The measurement precision for space angles can reach 0.2°.The research result of this thesis has very important significance to enhance the manufacturing industry product quality.