Real-Time Visual SLAM Research On Mobile Robot

SLAM is the key technology for mobile robot to achieve truly autonomous localization and navigation.Compared with the laser radar,visual sensor has the advantages of low price and abundant scene information,suitable for small robots and low-cost solutions.However,visual SLAM is computationally expensive and inefficient in execution.In this thesis,the selection and improvement of the typical SLAM algorithm are used to improve the system realtime and ensure the positioning precision.In this thesis,Kinect is used as a visual sensor to collect color and depth data of indoor scene,and the data is transformed by coordinate system to achieve visual alignment.According to the black hole noise in depth image,deep image denoising algorithm is designed,which can effectively improve the data availability.SLAM front-end is divided into two kinds:direct method and feature point method.In this thesis,feature extraction and matching based on feature point method is implemented,and minimum distance and RANSAC algorithm are used for matching verification.PnP iterative optimization is used to solve the pose-shift between frames.At the same time,the algorithm of sliding window bundle adjustment is used to optimize the local camera trajectory and improve the consistency of the visual odometer trajectory.In view of the large amount of computation of visual SLAM,the previous key frame selection strategy is improved in this thesis,which not only greatly reduces the amount of data,but also appropriately increases the number of key frames to improve the scene information in complex scenes.Based on the theoretical research,the 3D mapping of indoor scene based on Kinect is realized,and the combined filtering scheme is used to filter the scene point cloud,which saves a lot of memory.Further more,the point cloud map is transformed to octomap for the followon navigation.The system has a good positioning accuracy.The root mean square error of the true trajectory is 1.15cm and the maximum trajectory error is 2.76cm in the dataset test.At the same time,this system has a good real-time performance,and the average image processing frame processing frame per second.is about 18 frames,which can be very good on the local scene real-time mapping.Finally,this thesis summarizes the research work carried out,and prospects for future research and direction.