Optimization Of Visual SLAM Algorithm Based On Embedded System

Solving the problem of positioning the robot in the location environment is an important part of the realization of robot intelligence.Existing embedded simultaneous localization and mapping system have problems such as poor real-time performance and poor mapping effect.The reason is mainly because the system itself is too long for image processing and the performance of embedded hardware is poor,resulting in a slow calculation rate,Even if some algorithms can be embedded,it will be limited by the platform.As a result,there is a big gap between the operation effect and the PC platform.In response to this problem,this paper optimizes the visualization module and the ORB "dictionary" of the simultaneous localization and mapping system,and built a SLAM system with embedded development board TX1 as the host computer..The research work in this paper is divided into the following three parts:(1)The components of visual simultaneous localization and mapping system are studied.In the visual odometer part,the feature point method and the direct method visual odometer were studied,Combining the random sampling consistency(RANSAC)algorithm,the extraction and matching experiments of three characteristic points of SURF,SIFT and ORB are completed and compared.Compares Kalman filtering-based backend and graph-based optimization-based backend.The word bag model commonly used in loop detection is studied.(2)The system framework of the ORB-SLAM2 algorithm is completely analyzed,and the way of visualizing the module and loading the ORB dictionary is improved.The function of the visualization module in the original system is to monitor the normal operation of the program,Among them,the library files Pangolin and OpenCV,which are used to implement map display and feature display functions,consume a lot of computing resources in operation.In response to this deficiency,this paper replaces the original algorithm visualization module by using the command-end input function,so that the program updates the key frame and map point information with the command port during normal operation.Improved,the plain text output port reduces the amount of computation compared to the visualization module,while the ability to monitor program execution is preserved.In the back end of the original system,the ORB "dictionary" is stored and loaded in the form of a TXT file,but this method affects the efficiency of map construction.Therefore,this article further replaces the TXT file with a binary file.With this improvement,the memory occupied by the file is reduced by 100 M,and the speed of closed-loop detection is accelerated,and the initialization time of the program is also shortened.(3)Independently designed and built an experimental robot for simultaneous localization and mapping with embedded development board Jetson TX1 as the host computer.The TUM,Kitti and EuRoc data sets were compared,the initial system test was completed on the TUM data set,(the TUM dataset is a data file used to test simultaneous localization and mapping),and the error of the output trajectory was analyzed.Finally,the robot is used to test the field in the laboratory environment,analyze the output trajectory,and complete the 3D reconstruction of the map on the PC side,comparing the map with the real environment.Through the above experimental part,it is proved that the simultaneous localization and mapping system can complete the task of positioning and mapping.