Multi-level Features Extraction And Matching In Changing Illumination Environment

Simultaneous localization and mapping(SLAM)plays an important role in the autonomous positioning and navigation of smart devices.SLAM solves the problem of locating the robot's own position while building a map of an unknown environment.The existing SLAM technology is widely used in self-driving cars,unmanned aerial vehicles,autonomous underwater vehicles,planetary robots,emerging domestic robots,and even inside the human body.Among them,Visual SLAM has become a hot research branch in the field because of the low cost and light weight of visual sensors that are easy to carry on mobile smart devices.However,one of the key technologies of current mainstream VSLAM algorithms,which is data association,is based on the appearance of the image.Among them,the direct-based method can hardly adapt to simple illumination changes due to the brightness constancy assumption can rarely hold in real world,and the feature method which extracts the features in the local area of the image,such as gradient distribution of pixel values,is more robust to illumination changes though.But they generally can only deal with even illumination changes under sufficient illumination.That how to make the VSALM algorithm adapt to the various dynamic illumination in reality is still an important technical problem that limits its popularization and application.Therefore,in view of the above-mentioned problems in existing VSLAM algorithms,taking into account that in the dynamic lighting environment,although the gray value of the image will change drastically with the change of illumination,so methods based on image appearance will be damaged.But the 3D information of corresponding Points is theoretically unchanged when the pixels in the image are converted to real world.In addition,point distribution,density variation and points miss due to the change of illumination can be solved by combining different point cloud features.After obtaining point clouds in VSLAM,we then uses a multi-level features which integrates local features and semi-global features of point cloud to determine whether there are enough overlapping parts between any two observations,and the similarity change matrix between camera poses is calculated using these correlation.So the impacts on the accuracy of the data correlation of the VSLAM and the relative positioning accuracy between the cameras caused by illumination changes can be considerably mitigated.The main results of this article include:1.For the problem that changing image appearance caused by complex illumination changes are difficult to be associated accurately,a multi-level features combining local features and semi-global features is proposed,which makes camera position tracking more robust to dynamic illumination.This method expands the method based on image domain to point clouds domain and uses multi-level point clouds features for data association.In detail,a modified local feature(CCVFH,Compact CVFH)which improves CVFH descriptor and a semi-global descriptor(IOSEF,Inter-object Ensemble Shape Function)which borrows the idea of a global feature called ESF descriptor are used.CCVFH describes low-level features within objects,and IOESF describes high-level features between objects.Those multi-level features are used instead of a single feature,and the features used are accordingly modified so that it can effectively reduce the influence of illumination changes on the point cloud obtained from images.2.For the difficulty of balancing the high matching degree and pose calculation accuracy based on high-level feature matching,a pose tracking method based on multi-level feature matching and point cloud subdivision is proposed to ensure correct data association and obtain more accurate relative positioning results between cameras still at the same time.After multi-level feature matching,this method can ensure that the matching objects are correct.Then the objects are subdivided based on octree and the subdivided points(voxels)are selected object-by-object.Finally the SVD-ICP method is used to calculate the relative transformation matrix between the two point clouds.Compared with the commonly used SVD-ICP based on brute force search,correct and more accurate relative positioning results can be obtained.3.According to the method proposed in this paper,a verification experiment was designed to obtain data under different illumination directions,illumination intensities,global illumination and local illumination.And the algorithm which implemented in C++ using PCL proved the better robustness to various illuminations.Especially,stable data associations can still be achieved under darker,non-global uniform lighting changing.