Dynamic Visual SLAM System Design

In recent years,autonomous driving technology has continued to develop vigorously,and unmanned vehicles have emerged one after another.As the core of autonomous driving technology,visual SLAM technology has also been widely used in three-dimensional reconstruction,virtual reality and other fields.The visual SLAM system accuracy and robustness in complex environments have become the research hotspots of the majority of researchers.In order to improve the pose estimation accuracy of visual SLAM in front-end visual odometry,this paper proposes an optimization algorithm for pose estimation based on ICP and local bundle adjustment.The algorithm first uses the ICP algorithm to obtain the pose of the camera from the point cloud in the camera coordinates and the map points in the world coordinates.Then,in order to improve the accuracy of the pose estimation and avoid the phenomenon of trajectory "drift",the camera pose obtained by the ICP algorithm is used as the initial pose of the bundle adjustment,and the local binding adjustment is performed to optimize the camera pose.Using the initial camera pose provided by the ICP algorithm can avoid the local bundle adjustment algorithm from falling into the local optimal problem,and improve the accuracy of the pose estimation and the efficiency of iteration.In addition,in view of the problems of dynamic object detection based on semantic segmentation and detection of dynamic objects based on geometric methods in the current dynamic vision SLAM,this paper proposes a priori dynamic objects static point recall algorithm.The algorithm first uses the Mask-RCNN neural network to perform a priori dynamic object segmentation,and uses the multi-view combination method to further detect dynamic feature points in the environment,and at the same time judge the real movement of the prior dynamic object,and compare the static features in the prior object recall points to increase the number of effective feature points and improve camera tracking accuracy.Experimental results show that the prior objects static point recall algorithm can fully combine the advantages of semantic segmentation and multiview combination methods to ensure that the dynamic feature points in the scene are eliminated,and it can also ensure that the static feature points of the prior dynamic object are not random sculling.Finally,this paper uses the pose optimization algorithm and a priori dynamic objects static point recall algorithm to complete the design of a dynamic vision SLAM system.The test results on the TUM data set show that the trajectory error of the dynamic vision SLAM system designed in this paper is small and can reach the centimeter level;the error on some image libraries can even reach the millimeter level,which is of practical value.