Research And Implementation Of Monocular Visual SLAM Algorithm For Indoor Mobile Robots

Indoor mobile robots require the ability of simultaneous localization and mapping(SLAM)when they perform their tasks autonomously in unknown environments.Combining characteristics of various sensors and image features,this thesis designs a monocular visual SLAM algorithm and implements it on an embedded platform.The main works of the thesis could be summarized as follows:1.Existing SLAM methods often use point features in the image processing part.However,it is often not accurate enough in the case of blurred image features.The line features used in SLAM can improve the robustness of the system in complex environments,but it is time consuming.Therefore,a SLAM algorithm that combines point features and line features is designed to achieve a trade_off between system accuracy and time consuming.Meanwhile,the feature extraction algorithm and data association algorithm are improved to obtain better quality of data that used in pose calculation.The image is divided into different regions to process.As a result of that,features extracted on the image are well-distributed.The Shi-Tomasi score is introduced to improve the quality of the corner extracted by the FAST algorithm.The LSD algorithm is improved to eliminate line features with poor quality.Grid matching method is used in the frame-map association stage to achieve better matching.And the matching method of line features is designed.2.Monocular cameras cannot obtain absolute scale information of the environment,and visual localization is easily to fail in indoor complex scenes.To solve these problems,this thesis designs and implements a monocular vision localization system that combines wheel odometer information.The fusion of odometer information is mainly reflected in:1)In the initialization stage,scene scale information is determined according to the information of wheel odometer.2)In the data association stage,the odometer information provides prediction information for the location of image features in the next frame.3)To eliminate the error of front-end pose estimation,the back-end pose optimization of visual and wheeled odometer information fusion is constructed.4)When the visual localization fails,the system continues to locate and re-localization according to the wheel odometer data,which ensures the integrity and continuity of the localization system.3.The interface of the robot's pose trajectory and the feature map is displayed on PC platform based on Pangolin library.The accuracy and effectiveness of the proposed algorithm are proved by simulation experiments in multiple sets of public datasets.The algorithms is applied to an embedded robot platform with 1.2GHz main frequency and quad-core processors.The experimental results show that the Root-Mean-Squared Error(RMSE)of the algorithm is about 7cm and the processing time is about 90ms/frame.Which means that the designed algorithm satisfies the localization accuracy and time requirements of indoor mobile robots.