Design Of Localization Algorithm For Sweeping Robot Based On Monocular Vision

In recent years,sweeping robots have been increasingly applied in our daily life.Foreign sweeping robots are capable of visual localization and navigation,which have a remarkable cleaning efficiency,but the price is high.In contrast,sweeping robots in China which can localize by using visual sensor are still under developing.In view of the present situation of the domestic sweeping robots,an indoor localization algorithm based on monocular ceiling vision is designed and implemented by this thesis.It uses a low-cost visual sensor to achieve high accuracy and robust localization.The main works of this thesis are as follows:1.The positioning model is established.The monocular vision positioning mathematical modeling of the robot is performed based on the transformation of the image pixel coordinate system,the camera coordinate system,the robot coordinate system and the world coordinate system.We derive the pose solution function and analyze the function's propagation of error,which determines the threshold of the distance between pixels.2.The algorithms for the image processing of the ceiling are improved.By dividing images into grids,using FAST combined with Shi-Tomasi and adding direction description for the feature points,an improved feature detection algorithm is proposed and implemented.It is suitable for embedded systems with limited processing power,and the extracted feature points are evenly distributed in the image.Besides,data produced by IMU sensors is used to improve feature matching algorithm for ceiling images,which greatly reduces false matching and takes less time.3.Cumulative errors caused by incremental visual odometry are eliminated,and storage computing pressure is relieved.The strategy of Keyframe selection is improved,and redundant Keyframes are deleted,which enhance the efficiency of the system.The Loop Closure Detection algorithm based on Keyframe can detect whether the scene forms a closed loop.In addition,global pose optimization is used to eliminate accumulated errors.Sweeping robot mainly works on plane grounds,thus the local and global plane constraints are added to the pose optimization algorithm,which can improve the accuracy of the positioning algorithm.4.The effectiveness of the algorithm is verified.The algorithm of this thesis is implemented and migrated into the on-board embedded system.Several sets of experiments within the experimental environment of the OptiTrack capture system are designed.In addition,the Keyframes' pose obtained by the monocular visual localization algorithm of this thesis is evaluated by Root-Mean-Square(RMSE)of the absolute trajectory error(ATE).The experimental shows that the RMSE of the ATE is 5.4cm on average,and it takes 32ms for a single-core Vmware-Ubuntu and 640ms for the single-core embedded system to finish the procedure on average,which meets the robot's indoor localization requirements.The results indicate that the proposed algorithm of this thesis works effectively.