Research On Autonomous Navigation Method Of Substation Patrol Robot Based On Fusion Of Machine Vision And Radar Data

With the rapid development of power system and the popularization of intelligent substation,substation patrol robots aiming at improving the efficiency of substation patrol and reducing the pressure of manual patrol are becoming more and more popular.In the process of being more intelligent,the realization of autonomous navigation is the most core and basic technology.Research on the robot with high positioning accuracy,strong environmental adaptability and ability to deal with problems independently is the goal of this paper.Because only using a single sensor for navigation can not meet the needs of substation scenes,a self-directed navigation method based on camera and three-dimensional lidar data is chosen in this paper.The research work of this paper mainly includes sensor calibration technology,road detection,robot positioning technology,obstacle detection and tracking technology and road scene understanding technology.In the part of sensor calibration technology,firstly,the models of camera and lidar are established,and the position relationship between them is determined.Secondly,the selfcalibrations of camera and lidar are carried out,the derivation process of each model is introduced in detail,and the relationship between sensor data and robot is determined by calibration experiments.Then the joint calibration of camera and lidar is carried out from both space and time to realize the fusion of camera and lidar data.Finally,the image captured by camera is transformed by inverse perspective based on IPM,so as to get the top view of the road.As the basis of autonomous navigation,this part is an important bridge to fuse the following chapters.In the part of road detection,the key problem is to improve the robustness of lane detection.Therefore,on the basis of traditional Hough line detection,this paper proposes a method of lane extraction based on HSV color space transformation.This method not only uses Hough transform and constraints to obtain candidate straight line segments,but also uses the threshold of the three channels of HSV color space to extract vehicle marking line regions with specific color features.The final lane is obtained by fusing the two detection results.Experiments show that this method effectively reduces the impact of environmental changes such as illumination and shadows on Lane detection,and improves the accuracy and reliability of lane detection.After extracting the lane lines on both sides,this paper introduces how to use the virtual center line to localize the robot and how to control the robot to follow the virtual center line based on the PID method,so as to achieve the goal of localization and control of the robot.In the part of obstacle detection,firstly,a suitable raster map is constructed to preprocess the point cloud data obtained by the three-dimensional lidar,then the threedimensional point cloud datas of the road environment ahead are transformed into a twodimensional plane raster map.Secondly,the region growing method is used to cluster the grids occupied by the obstacles,and the orientation,distance and size information of the obstacles are obtained,so that the obstacles can be classified by extracting the feature of it.Finally,in order to track the detected obstacles,the extended Kalman filter is used to correct the radar measurement data,which improves the accuracy of obstacle tracking.In the aspect of road scene understanding,in order to highlight the road features ahead and reduce the difficulty of understanding the scene,a road scene sketch based on fusion of vision and radar detection results is presented in this paper.The sketch only uses black and white geometry to describe the road and obstacles in front the robot macroscopically without redundant information.Furthermore,HOG and GLCM features are extracted from the road scene sketch map to increase the distance between different classes.Finally,for the understanding of the road scene,this paper proposes to train a multi-class classifier using support vector machine(SVM)to judge the obstacles in the road by analyzing the characteristics of the substation.Finally this paper verifies the feasibility and validity of the method through experiments.This part is the core of this paper,and also the symbol of the intelligent information processing ability of the robot.