Fliter Based Object Pose Estimation

Object pose estimation methods play significant roles in the field of computer vision,Augmented Reality,and robotics.Existing methods for pose estimation rely on prior information of the object such as shape,size,and trajectory.In addition,given no prior,estimation error accumulates when a camera is the only sensor to perform pose estimation.An extended study is carried out in this thesis,and three methods are proposed to solve the above problems.A sliding window filter based pose estimation method capable of estimating the 6 degree of freedom motion of an arbitrary rigid body traveling on arbitrary trajectory is proposed in the thesis.The algorithm obtains the color and depth information of the feature points by using a RGB-D camera,and perform pose and structure estimation simultaneously for an object moving in SE(3)space.Gaussian-Newton iterative optimization is used to refine the pose and structure of the object in every window step.The initial value of the Gaussian-Newton iteration is generated by using RANSAC and OPnP algorithm.Simulations shows that the proposed method can obtain a reliable set of results without the knowledge of prior information.Experiments have been conducted to verify the effectiveness and robustness for 4 objects moving under different illumination and on different trajectories with the help of the real video data from standard RGB-D object tracking datasets.As a further development of the proposed sliding window filter,a Kalman filter based structural filter algorithm is proposed.The algorithm performs filtering for the structure by constructing a motion model under rigid body assumption,and by using an observation model obtained by Gaussian-Newton method.The filtering results are used to update the point cloud model.Simulation results show that the structural filter method is capable of suppressing error accumulation,and drastically increase the accuracy of pose estimation.A third pose estimation algorithm based on IEKF is proposed to alleviate the computational load of the sliding window filter implemented in a real time machine.The algorithm needs the constant velocity assumption but does not rely on any prior information of the object.Different from the existing IEKF algorithms,the propose algorithm in this thesis includes the object pose and structure as the states to estimate them at the same time.An IEKF scheme for an object moving on SE(3)is derived based on perturbation.By performing multiple iterations on the observation model,the algorithm reduces the error generated by the inconsistency between the constant velocity model and real object motion.Experiments show that the proposed algorithm can achieve almost same accuracy level of sliding window filter and OPnP algorithm but with a significant boost on computational speed.Meanwhile,the inconsistency of motion model has litter impact on pose estimation.