Research On Key Technology Of Mobile Robot Location Based On SLAM Technology

In recent years,with the rise of the artificial intelligence industry,traditional industries have gradually embraced new technologies.Such as the use of drones for photogrammetry,the use of autonomous navigation unmanned vehicles for high-precision map mapping at the city level and so on,These emerging technologies have brought tremendous changes in traditional industries.Owing to the drawbacks of GPS,inertial navigation,etc.it is especially important to find an effective way to solve the problem of positioning and navigation.Simultaneous Localization and Mapping is the key technology for robots to achieve autonomous positioning and navigation.The camera mounted on it is favored by researchers because of its many advantages.Therefore,this paper focuses mostly on the application of monocular vision SLAM technology in unmanned mobile vehicles,based on the ORB-SLAM algorithm.Under the analysis of relevant literature at domestic and foreign,the following works was carried out:1)For the theoretical basis of SLAM,the rotational translation representation of 3D Euclidean space is discussed.The relationship of Lie group Lie algebra is discussed in manifold.The imaging model of pinhole camera is studied.Meanwhile,the frontend data association of visual SLAM system,backend graph optimization model and closed loop detection module are discussed.Lastly,for the pinhole camera model,a partial scene reconstruction experiment was performed with the depth map.2)Aiming at the problem of feature extraction in the existing system,a certain degree of improvement is carried out.Firstly,the superior corner calculation and description algorithm is analyzed.A combined OFAST+BRISK feature extraction algorithm is studied.For the no rotation invariance of BRIEF descriptor,this method can effectively alleviate this problem.Experiments show that the correct matching rate of this method is above 80%.At the same time,in order to make the feature distribution more uniform,and the features are extracted in grids and divided into quadtrees,which effectively avoids the phenomenon of clustering.3)Analyze the overall framework of the system.Firstly,parallel acceleration optimization analysis is carried out for feature extraction of front-end.Secondly,based on the principle of multi-view geometry,the geometric relationship between two frames is analyzed.Thirdly,the back-end optimization based on the factor graph model is analyzed,and the experiment is carried out using the g2 o framework.At the same time,the similarity detection experiment and analysis are carried out based on the visual bag of word.Finally,the optimize experiment of closed loop based on Sim(3)is performed using the TUM data set.4)Experimental verification on the overall system performance.Firstly,the calibration experiment was carried out on the sensor.At the same time,the effect of improved front-end feature extraction was discussed,it is more reasonable than the existing system and the number of features is relatively stable.Secondly,the multi-threading technology is used to accelerate and optimize the module of front-end feature processing,and the acceleration ratio is about 1.3.Finally,the BRISK-SLAM studied in this paper is applied to indoor and outdoor public data sets,in which the positioning accuracy of indoor data set is improved by about 0.27 cm,and the outdoor is about 0.845 m.In addition,the system has been field tested on the autonomous unmanned mobile car.Based on Lidar,IMU,Camera,it can locate and avoid obstacles under the experimental scene.Since there is no ground truth of car trajectory,then the one estimated by hdl_localization with higher positioning accuracy is taken as reference.After calculating the scale of trajectory estimated by BRISK-SLAM with respect to it estimated by Lidar,the trajectory error ATE is about 1.51 m.The experimental results show that the SLAM system studied in this paper can be effectively positioning and mapping.