| With the rapid development of technologies such as driverless vehicles,augmented reality and virtual reality,as a key technique,simultaneous localization and mapping(SLAM)has become a research hotspot.According to the sensor equipped with,SLAM can be divided into laser SLAM and visual SLAM.Compared with the expensive laser lidar,visual sensors have low cost and richer information,and visual SLAM based on such sensors has been widely recognized and applied.Visual SLAM is mainly divided into feature-point method and direct method.The feature-point method relies on feature extraction and matching to track the camera pose,it is insensitive to photometric changes,and has good robustness.However,it is time-consuming to extract and match features,and it is easy to fail in the face of feature missing scenes.The direct method directly uses pixel information to estimate the camera pose and build a map by minimizing the photometric error.It eliminates the time-consuming feature extraction and matching process,making the algorithm more efficient.However,since it is based on the gray-scale invariant assumption,it is sensitive to scene photometric changes and requires high image quality.At the same time,due to the unobservability of the monocular-vision SLAM,the algorithm will produce scale drift.In view of the above problems,this thesis presents a relevant research focusing on the monocular vision SLAM technology based on direct method.First,a robotic platform for SLAM testing was built.The platform consists of a mobile robot Turtlebot2,a monocular camera and a laptop computer.As a traveling carrier,the Turtlebot2 robot was controlled to move through the ROS(Robot Operating System)running on the laptop,to provide a test platform for the monocular-vision SLAM algorithm designed in this thesis.Then,a monocular-visual odometry based on the direct method was designed and implemented.The direct method uses the gray-scale invariant assumption,and factors such as camera exposure parameters,lens vignetting attenuation,and camera response function will destroy this assumption to some extent.This thesis draws on the DSO(Direct Sparse Odometry)algorithm to calibrate the camera to establish a more sophisticated camera imaging model.A 7-dimensional SSD(Sum of Squared Distance)residual template is used to improve the utilization of pixel information.The camera pose and point-cloud map parameters are obtained by minimizing the photometric error by the Gauss-Newton method.At the same time,the sliding window is used to marginalize the redundant data by using the sparseness of the Hessian matrix to ensure the real-time performance of the algorithm.The visual odometry in this thesis was evaluated by public data sets and achieved good results.Finally,based on the visual-odometry obtained above,the monocular-visual simultaneous localization and mapping algorithm based on the direct method is designed and implemented.The algorithm uses the Bag-of-Words model to detect the closed loop,establishes the pose constraint based on the similar transform group,and adds the constraint to the global pose graph optimization to obtain more accurate camera pose estimation,which reduces the cumulative errors of the algorithm.The algorithm implemented is evaluated by the public dataset.The results show that the cumulative error of the algorithm is much smaller than that of the monocular-visual odometry without closed-loop detection.The effectiveness of the proposed algorithm in real-world scenarios is verified by the SLAM tests on the Turtlebot2 mobile robotic platform. |
