Research On Real-time Positioning And Temporary Road Detection Algorithm For Unmanned Vehicle

With the development of economy and continuous progress of science and technology,driverless vehicles have become one of the most important research directions in the automotive field.Unmanned vehicle positioning technology is an important part of environment perception,and also the basis of subsequent vehicle path planning and decision control.At present,the scheme of autonomous vehicle positioning using a single sensor has some shortcomings.However,the positioning technology based on multi-sensor information fusion has become a new direction of positioning technology research because of its advantages of high precision,high stability and high generalization ability.Among them,camera sensor and IMU sensor are widely favored due to their low hardware cost and strong complementarity.Therefore,this paper proposes a positioning method based on visual inertial fusion.The measurement information of tightly coupled camera and IMU can realize the high-precision real-time positioning of unmanned vehicles.As the road conditions of unmanned vehicles are complex and changeable,the temporary road built by traffic cones,as one of the special road conditions,is often used in road maintenance and maintenance,traffic accidents and temporary road control,etc.However,there are few research results on the automatic driving strategy of such temporary roads,so this paper takes this scene as the research object.The road detection of unmanned vehicles in this scene is studied.The main research contents are as follows:(1)Design of front-end preprocessing and initialization algorithm for visual inertia odometer.The kinematic model of the camera was built,and the feature extraction algorithm based on FAST corners was constructed,and the patch-based image pyramid optical flow tracking algorithm was proposed to accurately solve the pose of the image frame.The IMU measurement model and IMU pre-integral model are constructed to reduce the front-end calculation.Combining the camera kinematics model and the IMU measurement model,the external parameter calibration of the camera and IMU,the zero bias of the IMU gyroscope,the gravity vector,the initial velocity and the displacement scale parameter estimation are realized.(2)Design of back-end optimization algorithm for visual inertia odometer.A vision inertial tight coupling method based on nonlinear optimization was proposed,and the vision measurement residual model,IMU measurement residual model and prior residual model were constructed,and the state variables in all residual models were jointly optimized.A loopback detection algorithm based on the word bag model was constructed to reduce the global pose error and improve the robustness of the system positioning through image frame similarity matching.(3)Design of temporary road detection algorithm based on improved Delaunay.For the temporary road guided by traffic cone bucket,an improved Delaunay triangulation algorithm was proposed.The Delaunay triangulation network was constructed while the traffic cone bucket was detected.Through triangulation filtering and optimization,realtime road detection of unmanned vehicles in this special scenario was realized.(4)The feasibility and performance of the algorithm were verified through open source data set and real road test.First,the visual inertial fusion positioning experiment was carried out under the Eu Ro C open data set,and the performance of the proposed algorithm and other algorithms was compared and analyzed.The experimental results show that,in most sequence scenes,the positioning trajectory of the proposed algorithm has a higher coincidence degree with the true trajectory,and the absolute position and pose root-mean-square error is lower,and the average time of the algorithm is 52.6ms,which meets the real-time requirements of unmanned vehicles.Then,the temporary road detection algorithm proposed in this paper based on Delaunay triangulation is tested in real vehicles.The experimental results show that the average time of the algorithm is 35.4ms,the absolute trajectory error of the detection of the drivable path is 0.2m,and the accuracy is 97%.The real-time requirement is met,the path detection error is reduced,and the accuracy of the path detection is improved.