Research On Key Technologies For Robot Localization And Navigation Based On Heterogeneous Sensor Data

With the development of big data processing technology and artificial intelligence,the robot's perception and processing capabilities have been continuously enhanced.Mobile robots have been applied to various fields such as industry,service,and military.In various mission scenarios such as detection and patrol,mobile robots need autonomous navigation to complete the specified tasks,and the basis of their motion is robot localization.Since the GPS signal can fail in urban environments or complex electromagnetic interference environments,robots need to seek new localization methods.However,the existing single sensor localization method is not robust,and the robot needs to comprehensively apply a variety of heterogeneous sensing data to improve the localization capability.This paper works on the specific application of robot localization,and proposes a robot initial position estimation method and real-time pose estimation method based on imagepoint cloud heterogeneous data association.Based on this,a large-scale robot navigation system is designed and implemented.The main content and innovations include:· Researching and analyzing the basic theory and basic methods of heterogeneousensor data association(Chapter 2)This thesis systematically expounds the basic theories and method of heteroge-neous sensor data association.According to the different scenarios of heteroge-neous sensor data association,we divide it into three dimensions: multi-sensorsingle-dimension,multi-sensor multi-dimension and multi-location multi-sensor.The thesis also expounds its processing ideas and methods,and provides theoretical guidance for the application of heterogeneous sensor data association to solvespecific applications.· Proposing an initial position estimation method based on image-point cloud hetero-geneous data association(Chapter 3)Considering the robot localization scenario,this thesis realizes an initial positionestimation problem of the robot in the global point cloud map by the monocularcamera without the priori localization information.The convenience of the monocular camera and the robustness of the laser point cloud are fully utilized.We splitand project the global point cloud map into a set of deep images with location in-formation.A single camera image is then predicted using an unsupervised neuralnetwork as a corresponding depth image and retrieved in a depth map set to obtaincorresponding location information.The data set test showed that the positioningaccuracy was 89.3%,and the retrieval performance F1-Score = 80%.· Proposing a real-time pose estimation method based on image-point cloud hetero-geneous data association(Chapter 4)This thesis proposes a real-time pose estimation method based on image-pointcloud data association,which realizes a 6-degree-of-freedom pose estimation of arobot using a monocular camera in laser point cloud map.We use the construc-tion of the monocular odometer to construct a continuous camera image into a 3Dcamera point cloud and then process the global laser point cloud map correspondingly.Finally,the improved point cloud registration method is used to match thecamera point cloud with the laser point cloud map,and the pose of the camera isobtained by the transformation matrix of the point cloud registration.In addition,we have solved the problem of scale uncertainty in monocular vision positioning.The average translational error is 0.16 m,and the average rotation error is 1.42°.· Designing and implementing an autonomous navigation system for robots in thelarge-scale environment in robot operating system(Chapter 5)This thesis designs and implements an autonomous navigation system for robots inlarge-scale environments in robot operating system.The system combines globalGIS spatial data and local SLAM mapping results to achieve hybrid navigation.The robot can directly use the road network data in the GIS database for globaltopological navigation,and can calculate and optimize the local path according tothe local mapping result during the movement.The system realizes autonomousnavigation in large-scale environment,and completes the “sense-decision-action” process.