Simultaneous Localization And Dense Mapping Using A Monocular Camera

In recent years,human beings are coming to an age of intelligence for mobile robots,such as unmanned vehicles,UAVs and augmented reality with the improvement of hardware computing capabilities and software algorithms.SLAM algorithm,which can provide autonomic location and environmental awareness,has become a basic and key technology for realizing intellectualization.Monocular SLAM technology has become an important research area due to its low cost and wide application prospects.At present,monocular SLAM is not mature in localization and mapping.In terms of localization,the high complexity of the algorithm causes the system to run slowly and cannot be run on the mobile device in real time.In the aspect of mapping,only sparse maps composed of feature point depths or semi-dense maps composed of high-gradient point depths are generally available without GPU acceleration.It brings great limitations on robots' obstacle avoidance,path planning,and interaction in augmented reality applications due to the lack of map information.This thesis focuses on the design of a monocular SLAM system that is fast and accurate on pose estimation while creating a dense map.In terms of localization,we propose a semidirect visual odometry system.The algorithm selects some high-gradient points with reliable depth,and then uses a depth-constrained pyramidal optical flow algorithm to find matching points and solve the pose by optimizing reprojection errors.In the aspect of mapping,we propose an algorithm for low-texture area reconstruction based on plane model.The algorithm first extracts the plane area on the image,and then through the matching filtering,multi plane segmentation,and the filtering and fusion of the reconstruction results,we can estimate a consistent and accurate map of the weak texture.Combine with the reconstruction of high gradient points,a dense map is finally obtained.Experimental results on public datasets show that our two algorithms proposed in this thesis have greatly improved in performance compared with other related algorithms.Finally,we build a AR application based on the location and dense mapping algorithm proposed in this thesis.