Study And Realization Of Visual Inertial Simultaneous Localization And Mapping

Viusal inertial SLAM is one of the main techniques in the fields of robot positioning and automatic driving.The technology establishes a mathematical model of the trajectory in the world.The embedded sparse visual inertial SLAM system is discussed.On an embedded platform,feature extraction and matching module output feature matches based on the image stream.The system tracks the rough pose of the object using acceleration and angular velocity collected by inertial measurement unit(IMU),and constructs a sparse point cloud map after the initialization.Finally,the system uses graph optimization method to optimize poses based on constraints between poses,map and feature matches.The basic visual inertial SLAM is built based on VINS.The module of feature extraction and matching,initialization and graph optimization have been improved in an updated visual inertial SLAM system named Fusion＿VISLAM.Finally,the system is accelerated for the embedded platform through the following procedures.First,on the basis of the traditional FAST feature extraction method,the uniform distribution of features is realized by dividing the image to different patches to provide more effective image information for the system.In the difficult environment with weak texture or fast camera motion,feature mismatches exist.After the initialization,a more accurate constraint model is calculated by combining IMU data and system poses to detect and remove the error,which improves the accuracy of feature matches.Second,to improve the initialization accuracy of a system using both the visual and IMU information,we divide the trajectory into visual and IMU segments according to the magnitude of camera motion.Then two parts of the trajectory are used to calculate more accurate initial parameters of the system.So that the negative impact of tightly coupled accelerometer bias and gravity acceleration to the trajectory has been removed.Finally,through reducing outliers in the point cloud,we improve the accuracy of map and pose.Furthermore,we manage to achieve a real-time performance of the embedded system by controling the number of image features.The evaluation on the Eu Ro C dataset shows that the absolute trajectory error of the improved system is relatively reduced by 40.17% compared to VINS at the expense of part of calculation time.On the Fire Fly-RK3399 embedded platform,after combining with the acceleration scheme,the system runs at 13 FPS with an absolute trajectory error of approximately 10 cm based on the Eu Ro C dataset.