Study On Localization And Semantic Mapping Technology Of Vehicle In The Indoor Parking Lot

With the increase in the number of cars,the demand for parking is increasing.The indoor parking lot is one of the main parking places in the city.In indoor parking lot,there are problems of ‘Parking difficultly’ and ‘Taking hard’.Real-time high-precision vehicle localization and visual parking map have become an effective way to solve this problem.In order to realize the accurate localization of vehicles,the vehicle needs to carry highprecision sensors or remodeling the parking scene greatly,which is of high cost and difficult.In the indoor parking lot,the existing vehicle localization method is easy to be disturbed by the environmental noise,and the overall localization accuracy of the system is not high,which seriously affects the accurate localization of the vehicle.At the same time,most of the maps used for vehicle localization only contain low-level geometric information,which is difficult to visualize for drivers.In order to overcome the shortcomings of the existing technology,this paper based on the visual simultaneous localization and mapping to realize the localization of vehicles,and combined with IMU,which is used for localization when visual failure.At the same time,a semantic map is constructed to facilitate drivers.In the process of research,the following research work has been completed:(1)Aiming at the problem of poor real-time semantic annotation of dense point clouds,a real-time semantic annotation method of RGB image-assisted spatial dense point clouds is constructed.The Yolact++ network is used to achieve real-time instance segmentation of RGB images.Based on the transformation relationship between the RGB camera and the depth camera,the semantic information of the RGB image is mapped to the depth image,so as to realize the semantic annotation of dense point clouds.Aiming at the problem of unstable segmentation of object instances and lack of depth information,a correction method based on the combination of Lucas-Kanade optical flow method and historical continuous image frame information is given,and the feasibility of this method is verified.(2)Because of moving objects reduce the accuracy of vehicle visual localization,a method for judging the motion state of instance objects between frames based on optical flow tracking is proposed.The removal of moving objects is achieved based on the results of the motion state of the instance objects,and the feasibility of the method is verified on the public data set.In order to solve the problem of visual failure,a combined localization mechanism of vision and IMU is established,and IMU is used to make up for the lack of robustness of visual localization.(3)In order to overcome the problems of traditional point cloud maps occupying large storage space and poor reusability,a semantic octree map of instance objects was constructed.The object instance database is established,and the relocation based on the instance object is preliminarily realized.