Research On Visual Odometry Based On Light Field Camera

With the rapid development of light field technology and the continuous improvement of computer processing capabilities,the research related to light field camera has caught more and more people's attention.As a light field acquisition device,a light field camera can record the position and direction information of light through a single exposure.A single frame of light field image can be used to recover scene depth information without an additional visual sensor.Therefore,light field cameras can be widely used in computer vision research,such as depth estimation,simultaneous positioning and mapping.In addition,compared with the depth camera,the light field camera has the advantage on size,which makes it have wider application scenarios in the visual odometry system,such as unstructured and mobile scenarios.However,most of the current researches on visual odometry are mature systems that directly use ordinary cameras,while the research based on light field cameras is relatively small.In this regard,this paper mainly focuses on visual odometry based on light field cameras.Due to the low spatial resolution of current light field image,the visual odometry system of ordinary cameras cannot directly obtain satisfactory results.In this regard,this article designs and implements a set of visual odometry system based on the feature method for light field cameras.Based on the characteristics of the light field camera that can collect the geometric information of the scene,it is equivalent to an RGB-D camera,and the camera movement is directly obtained by the alignment(registering)of two frames of point clouds,thereby simplifying pose estimation of camera in the visual odometer system.Specifically,for the camera pose estimation in this system,According to the actual application scenario,this paper proposes two light field camera pose estimation algorithms based on threedimensional point cloud alignment with different algorithm complexity to restore the camera pose,which are three-dimensional point cloud alignment based on single sub-aperture image and multiple sub-aperture image 3D point cloud alignment pose estimation algorithm.The experimental results show that the proposed light field camera pose estimation algorithm can accurately estimate the camera pose,and the visual odometry system based on the light field camera can also effectively recover the camera's motion trajectory.The pose estimation of the camera is an important part in the visual odometry.Considering that the existing studies about visual odometry based on light field camera have directly used the pose estimation algorithm of ordinary cameras,this approach does not take full advantage of the characteristics of the light field camera,and the performance of the algorithm needs to be further improved.Therefore,according to the characteristics of the light field camera,it is extremely important to research the light field camera's pose estimation algorithm for the light field domain and restore the camera's pose with an accurate solution method.The light field camera pose estimation algorithm based on the corresponding light field points was completed earlier,which can accurately and robustly restore the position of the light field camera.Based on this,this paper proposes three solving algorithms to solve the correspondence of light field points,which can more accurately estimating the light field camera pose.Experimental results show that all three algorithms can effectively recover the pose of the light field camera.Through comparative experiments on the stability of the three solving algorithms,different numbers of feature point pairs,and different noise levels,this article also analyzes the advantages and disadvantages of each solving algorithm in order to find the most suitable scenario for each algorithm.