Low Speed Unmanned Driving SLAM Based On Binocular Vision

With the development of robotics technology,as a kind of special mobile robot,the unmanned vehicle is a multi-functional integrated system integrating control,execution,environment awareness and decision-making.The control and execution of the vehicle can be well implemented in existing vehicles.However,environmental awareness and decision-making is still a worldwide problem.It has attracted the attention of scholars at home and abroad.In addition,due to the huge economic benefits of unmanned vehicles,many high-tech giants at home and abroad have also invested a lot of manpower and material resources.However,it is still far away that the unmanned vehicle wants to achieve a good environment exploration and autonomous navigation.To realize unmanned driving,first of all,vehicles must be able to autonomously locate and construct their environment maps.This is the SLAM(Simultaneous Localization and Mapping,Simultaneous Localization and Mapping)issue that is currently receiving much attention at home and abroad.The visual SLAM is based on its abundant information and low cost and other advantages have gradually become an important direction of development in the field of SLAM.This article relies on the hardware and software platforms of the Automotive Electronics Research Center of the Chinese Academy of Sciences,has systematically studied the visual odometry and construction,closed-loop detection and back-end nonlinear optimization of low-speed unmanned SLAM based on binocular vision.The specific research results are as follows:Firstly,the binocular visual image matching is performed using an improved ORB algorithm.Firstly,feature points with scale invariance are extracted;then feature descriptors are constructed using ORB algorithm;lastly,stereo matching is used to calculate the position sequence of the unmanned vehicle,to calculate the two-dimensional space pose of the unmanned vehicle,at the same time,to complete the construction.Secondly,this dissertation uses the feature descriptors and the BoW(Bag-of-Words)model that have been constructed by the visual odometry and construction to perform closed-loop detection.Firstly,clustering feature descriptors implement class centers;in order to make visual features layered and quantified,recursively generate BoWs for k-skeletal trees;and then use k-fork tree score matching methods that compute similarity increments between images layer by layer;Good candidate closed-loop and culled candidate closed-loop errors in the closed-loop;finally used the time and matching constraints.Through the closed-loop detection of visual SLAM,not only the accuracy of the closed-loop detection is improved,but also the application of the method based on the appearance of the scene in the SLAM system is applied.At the same time,the BoW method in the fields of computer vision and image processing is enriched.Then,a graph-based nonlinear optimization structure of the SLAM rear end of the unmanned vehicle is constructed.The Newton-Gaussian iterative algorithm is used to optimize the error between the pose prediction data and the pose measurement data,calculate the position of the unmanned ’vehicle and the optimal configuration of the landmarks under the minimum error.The vehicle positioning and environment model was optimized to guide the autonomous navigation of unmanned vehicles.Finally,in order to verify the feasibility and effectiveness of the proposed method,pioneer vans purchased in the laboratory were used to conduct experiments in the actual environment.The experimental results show that the three threads of visual odometry and construction,closed-loop detection and back-end nonlinear optimization studied in this paper can run well.