Research Of Monocular Visual Odometry In Illumination Change Scene

When a mobile subject completes autonomous movement in an unknown environment,it needs to construct the surrounding environment and determine its own position at the same time,that is,Simultaneous Localization And Mapping(SLAM).As a sensor of SLAM system,monocular camera is favored by the majority of SLAM researchers because of its low cost and rich application scenarios.Visual odometry is the core part of visual SLAM,and its implementation can be divided into feature point method and direct method.Since the direct method does not need to calculate the descriptor,it can be used in the scene of feature missing.Besides,the direct method has fast computation speed and good real-time performance,and has gradually become the research hotspot of visual SLAM.The realization of the direct method needs to satisfy the Brightingness Constancy Assumption,that is,the pixel gray value of the same space point remains unchanged in each image.However,in the actual environment,due to the camera's automatic exposure or illumination changes,the Brightingness Constancy Assumption is often invalid,which leads to the failure of algorithm positioning in the actual scene.Census transform is a common feature extraction method in the field of stereo matching,and it has good robustness to illumination changes.Therefore,in this paper,the Census transform is used to reduce the sensitivity of direct method to illumination change,and a monocular visual odometry based on the improved Census transform is proposed and implemented.The main research results of this paper are as follows:?Census transform is introduced into the front end of SLAM,changes the form of traditional Census transform so that it can be applied in nonlinear optimization.The form of the traditional Census transform is improved to be expressed in Euclidean space,and the difference measurement is carried out by SSD(Sum of Squared Distance).Meanwhile,the pose estimation based on the improved Census transform is deduced.?The distribution of traditional feature points is uneven,which greatly affects the accuracy of pose estimation.In this paper,an image pyramid model is used to divide the quadtree meshwork of each pyramid image layer,and the threshold reduction is detected for the meshwork with no obvious features,to realize the uniform distribution of features on each image.?The SVO(Semi-Direct Visual Odometry)algorithm is referred,meanwhile,it is improved and extended.In the process of system initialization,the non-plane hypothesis model is added,so that the algorithm can be used when the camera is looking forward.The key frame selection strategy is improved,and the robustness of the algorithm is increased by considering image tracking,camera rotation and running time.Census transform is integrated to obtain accurate camera pose by minimizing Census transform error of map points.Based on Gaussian-uniform distribution probability model,depth filter is constructed to build local map.Finally,experiments on the open datasets of EUROC,New Tsukuba Stereo and TUM show that the proposed algorithm achieves pose estimation under light changes,which verifies the effectiveness,real-time performance and robustness of the algorithm.