Research On Driverless Vehicle Visual SLAM Technology

The continuous development of intelligent robot technology has enabled more and more various forms of robots to be used in various industries.Letting the robot sense its own state and its surrounding environment is the basis for other functions such as path planning,navigation,and obstacle avoidance.Therefore,Simultaneous Localization and Mapping(SLAM)technology has gradually caused people's concern.As a mobile robot in a broad sense,the unmanned vehicle simplifies the motion form of the humanoid robot,so that it can focus on solving specific problems,which is why most robots are wheeled robots.Compared with the early SLAM technology with laser sensor as data input,the visual-based SLAM technology has become the mainstream in recent years due to its low cost,and because stereo cameras and RGB-D cameras can easily obtain depth information,it is also The focus of research in visual SLAM.Therefore,this thesis mainly studies the unmanned vehicle visual SLAM technology based on RGB-D camera,which mainly includes the following contents:Firstly,the principle of obtaining depth information by RGB-D camera and the imaging model of pinhole camera are studied.How to calibrate the internal parameters and distortion parameters of the camera,and the external parameters between the two cameras,further introduce how to according to the calibration parameters.The collected color image is registered with the depth image.Secondly,the front end visual odometer of visual SLAM is studied,including feature extraction,feature matching and selection,camera motion estimation and local optimization.For the phenomenon that the original ORB features are easy to concentrate on the local regions in the image,this paper proposes using the method of dividing the grid combined with the quadtree to solve it effectively.Firstly,the method of dividing the grid is used to improve the probability that the region lacking texture extracting the FAST corner point,and then the quadtree is used to filter the features,in order to reduce the phenomenon that the feature points are too concentrated.Thereby this method can reduce redundant information of the image and make full use of the entire image.For the mismatch phenomenon that occurs during feature point matching,a random sampling consistency algorithm is used to effectively eliminate the mismatch.In order to estimate the error of camera motion,the combination of EPnP and boundle adjustment is used to reduce the error.Finally,for the phenomenon that the visual odometer produces cumulative error,this paper uses loopback detection combined with global pose gragh optimization to reduce the cumulative error.The loop detection scheme based on the word bag model is studied.The k-ary tree is used to store visual words to greatly improve the access efficiency.The k-means++ clustering algorithm is used to generate the new nodes of the k-ary tree.After detecting the loop,the entire map needs to be globally optimized.At this time,the feature points in the map are ignored,and only the camera pose is optimized,so that the calculation amount can be greatly reduced,and the g2 o library is used for the pose map optimization.The unmanned vehicle software and hardware platform was built to verify the algorithm effect.The TUM data set and the real scene were tested.The trajectory error was analyzed and the sparse feature point map and point cloud map were generated.The experimental results show that the method is effective.