Visual Simultaneous Localization And Mapping Based On Temporal-spatial Consistency

SLAM(Simultaneous Localization and Mapping)technology has been successfully used in service robots,unmanned vehicles,autonomous driving,augmented reality,virtual reality and 3D reconstruction.The main work of SLAM technology is to estimate the pose of the sensor and the reconstruction of the map.The sensor includes monocular camera,binocular camera,depth camera,laser radar and so on.The classical SLAM algorithm includes two parts:the front-end part and the back-end part.The front-end is to estimate the initial pose of the sensor,and the back-end receives and optimizes the initialization pose provided by the front-end.Every pose and map estimation of SLAM inevitably brings about errors.There are two ways to reduce the error,one is based on the filtering method,the other is the graph-based optimization method.In the case of the same time complexity,the graph-based optimization method is more accurate than the filtering-based method,so this paper uses the graph-based optimization method to reduce the error.In this paper,we study the characteristics of feature points in depth and consider that the stable feature points have the following two characteristics:they appear in multiple continuous frames and are observed by cameras from different viewpoints.We describe these two characters by temporal consistency and spatial consistency,that is temporal-spatial consistency.Stereo odomerty based on temporal-consistency feature selection and monocular slam based on temporal-spatial consistency.Temporal consistency decides the timing of keyframe insertion,and spatial consistency filters the 3D points.Experiments show that our method is not only more efficient and robust,but also effectively reduces error accumulation,which improves estimation accuracy of stereo odometry and monocular slam.