Research And Implementation Of SLAM For Unmanned Vehicle Based On Monocular Camera And IMU In Rejection Scenario

Localization is an important guarantee for safety navigation of unmanned vehicle.In the applications of unmanned vehicle,the localization method is mainly based on the expensive GNSS(Global Navigation Satellite System).However,due to the severe dependence on external signals,the GNSS rejection scene(the scene where signals cannot be received normally in mountains,tunnels,canyons,parking lots,etc.)cannot play a role in localization.The vision-based SLAM(Simultaneous Localization and Mapping)technology can be used in the rejection scene.The SLAM positioning technology of the monocular camera combined with the IMU(Inertial Measurement Unit)has the true scale that monocular vision lacks,and has computing advantages compared with Stereo Vision and Lidar,which has great significance to the localization system of unmanned vehicles.Therefore,relying on the resources of the Robotics Research Centre,this paper research on the application of the SLAM algorithm of the monocular camera and IMU on the unmanned vehicle platform,and proposes a method suitable for unmanned vehicles that provides stable positioning results in rejection scenarios.The specific research content of the localization algorithm is as follows:(1)According to the characteristics of unmanned vehicles with large number of plane and linear motion,this paper modifies the traditional fusion initialization and fusion positioning methods,and uses the Quaternion-based perturbation derivation method to calculate the relationship between the IMU and the camera position,the parameters and the scale recovery of the camera are iterative calculated,which solve the problem of difficulty in initializing the traditional SLAM system set for UAVs.At the same time,the state variables of the IMU are processed online,the problem of error increase caused by long-distance linear motion is avoided by online estimation and a translation constraints.Compared with the original SLAM system,the localization accuracy is improved by 28.77%.(2)According to the high speed of unmanned vehicles and the complex outdoor environment,the Object Detection results is introduced to localization method.The traditional visual inter-frame positioning method of feature points has low accuracy positioning in outdoor environments.Based on human’s attention to semantic cognitive information while driving,this article uses the results of deep learning Object Detection to help with localization.The feature points information and cognitive information are combined to establish an observation error equation to minimize the overall error each moment.Compared with traditional methods,the robustness of localization is significantly improved by 16.49%.(3)In the SLAM application of unmanned vehicles,the localization error will continue to increase due to the cumulative visual error and IMU error,the unmanned vehicle rarely passes through repeated positions,and the error cannot be eliminated by the traditional loop relocation method.Aiming at this characteristic,this paper uses the priori information of the map for SLAM relocation,we adopts the method of maximum posterior probability and expresses the position through probability distribution,we uses the location result output of SLAM as the location update incentive;the road width,lane angle,and road position information in the map will be used as a prior information,the prediction-update model is used to output the position distribution to obtain the positioning result.Compared with the SLAM system without map relocation,the accuracy has been greatly improved,and the error has been reduced by 74.11%.