Research On Multi-sensor Fusion SLAM Applied To Automatic Driving Based On VINS-MONO

SLAM(simultaneous localization and mapping)technology is one of the key technologies for most applications in current world,including popular au-tomatic driving application.Automatic driving is comprehensive application of various technologies,including deep learning,SLAM,path planning,au-tomatic control and so on.SLAM mainly provides location information and description of surrounding scene map,so a robust and stable SLAM system is very important for autopilot application.Due to the complex and diverse driving environment of the car,it is difficult to meet the needs of accurate po-sitioning and mapping with independent visual SLAM or laser SLAM.In addi-tion,at present,the sensors equipped on the automatic driving vehicle are also abundant.It is very necessary to make full use of the existing resources to im-prove the safety of automatic driving.Lidar sensors are expensive.With the development of computer vision,it is a trend that laser is replaced by camera.Therefore,this thesis gives up lidar sensor and chooses to do research on fusion SLAM based on the open source system VINS-MONO with multiple cameras,IMU and wheel odometer sensor to provide reliable positioning and mapping services for automatic driving applications.The main work of this thesis is as follows:1.Dynamic objects have a great influence on SLAM,which often intro-duces wrong estimation in the visual part and makes the system unstable or even collapse.This paper proposes a dynamic object detection and elimination algorithm based on SegNet deep learning network.SegNet is responsible for semantic segmentation to get the mask,and then morphological processing is used to enlarge the mask edge and filter out the noise.At the same time,two methods based on depth information and epipolar distance are applied to detect the dynamic points,and then the dynamic points of the two ways are fused,and the fused dynamic points are used to repair the mask.Finally,the fixed mask is used to remove dynamic points,and more accurate static feature points are obtained to serve the SLAM system.2.In order to solve the problem that multiple cameras can’t be closed-loop when the car goes forward and backward through the same place,a cross closed-loop detection algorithm based on multiple cameras is realized.The observation of multiple cameras is put into the same container,and cross matching is carried out during matching,so that one camera can be closed-loop with any other camera,so as to improve the recall rate of closed-loop.3.The design of fusion SLAM based on multiple cameras,IMU and wheel odometer sensor which is applied to automatic driving.In this thesis,the ac-celerometer in IMU is abandoned on the basis of VINS-MONO open source system,and the translation information is provided by wheel odometer instead of accelerometer,and the online initialization strategy of the SLAM system is redesigned.Because the measurement of gyroscope and wheel odometer adopts the pre-integration theory,this paper deduces the pre-integration process,and then gives the objective residual function of system tight coupling optimization.