Research On Parking Lot Semantic Map Based On Around View Monitor System And Visual Inertial

With the development of autonomous driving,autonomous valet parking,as an important branch of autonomous driving technology,has gradually attracted the attention of academic and industrial circles.Mapping and positioning are important modules of the autonomous valet parking.Semantic SLAM is a novel technology to implement this module.However,the existing semantic SLAM is difficult to extract and track the semantic features of the ground.Considering that the parking spaces in the indoor parking lot are arranged regularly and the colors are bright,which is a stable and significant semantic feature,this paper proposes a semantic SLAM algorithm for the indoor parking lot.The algorithm generates a parking lot semantic map by fusing visual inertia and parking space information in the around view.The main research contents are as follows:1.The multi-sensor joint calibration is completed,and the data of the frontviewing monocular camera,the fish-eye camera and the IMU are unified into the ground coordinate system.First,the fish-eye camera was calibrated in the study,the internal distortion coefficient of the fish-eye camera was obtained,the conversion matrix from the corrected image to the ground was calculated,the original image to the top view mapping table was established,and the top view was spliced and fused to generate the around view image;Secondly,this paper calibrated the monocular camera,obtained the internal distortion coefficient of the monocular camera,and calculated the conversion matrix of the monocular camera to the ground;Finally,this paper calibrates the IMU,obtains the noise and random walk of the IMU,and calculates the conversion matrix from the IMU to the camera.The data of each sensor is converted to the same coordinate system,which is convenient for calculation and fusion.2.In the study,the visual-inertial SLAM algorithm with relatively better performance is selected for vehicle positioning.The visual inertial SLAM is used to fuse the image data of the monocular camera and the motion measurement data of the IMU to estimate the vehicle pose.Four mature visual SLAM algorithms based on different sensors are selected for comparison.Through experimental verification,the ORB-SLAM3 with the smallest absolute trajectory error is selected as the basic algorithm for follow-up research.3.This research proposes a parking space association model.The correlation model is based on parking space detection,and carries out data correlation according to the confidence and geometric information of the parking space,establishes a parking space library,stores the information of each parking space,and optimizes the parking space coordinates.The accuracy of parking space detection directly affects the accuracy of the parking space association model.Three parking space detection algorithms based on the around view image are selected,and the indoor parking lot data set is used for verification,and the best performing VPS-Net is selected for parking space detection.4.This research proposes PS-SLAM,a semantic SLAM algorithm for indoor parking lots.Visual inertia SLAM technology ORB-SLAM3 combined with panoramic parking space detection technology VPS-Net,fusion and optimization of visual inertia and the around view image information,build a parking space association model,can construct parking in real time during the movement of the vehicle Field semantic map.The verification experiment of the algorithm was completed,and the wire-controlled chassis equipped with fish-eye camera,monocular camera and IMU was used to test in indoor parking lots.The results show that the algorithm can generate real-time maps during vehicle movement with good positioning accuracy.After research,the parking lot semantic SLAM system designed in this paper solves the problem of multi-sensor data fusion,and proposes a parking space data association model,which is suitable for indoor parking lot positioning and semantic map construction.