Design And Implementation Of Large-scale Parking Lot Car Searching System With Inertial Navigation Based On DNN

2020 is the year when China will build a moderately prosperous society in an all-round way.With the development of economy and advancement of urbanization,China’s car ownership increases rapidly,which causes the difficulty of parking in cities.It often takes a lot of time for car owners to find a parking space in dense commercial area.Though the appendant large-scale parking lot has significantly alleviated this problem,it brings another new problem,that is the difficulty of finding car.Due to the large area of parking lot,the high density of vehicles,the high similarity of internal structure,and the unfamiliarity of car owners to the environment,it is hard to distinguish the direction.Starting from actual problems,and combining with internal environment characteristics of large-scale parking lot and current technology development trend,this thesis proposes largescale parking lot car searching system with inertial navigation based on deep neural network.Main works of this thesis are as follows:(1)This thesis conducts in-depth analysis of traditional inertial navigation methods,and innovatively proposes an improved window segmentation method based on deep neural network and unsupervised learning.In order to solve shortcomings of traditional zero-velocity detection methods,inspired by the Expectation-Maximization algorithm,this thesis proposes to use two sets of bidirectional GRU neural networks to complement each other.The former achieves dynamic adjustment of detection thresholds according to movement states and segments raw data into multiple weakly correlated windows step by step,while the latter estimates the polar vector over each window to track pedestrian trajectories.(2)In consideration of the complexity of WiFi signal propagation in indoor environment,this thesis proposes WiFi fingerprint positioning algorithm based on kernel density estimation,and extended Kalman filter is adopted to fuse WiFi positioning and inertial navigation to improve accuracy.(3)Detailed analysis and design of the proposed car searching system,realization as an App on Android,and several tests with analysis in the end.The experimental results show that the proposed system achieves dynamic adjustment of zero-velocity detection thresholds,thereby greatly improving detection accuracy,and is able to generate high-precision and long-lasting trajectories,successfully navigates car owner to place nearby the car,which confirms validity of the system.Compared with the reference model IONet,our proposed model performs better in pedestrian turning and new environments.The car searching system proposed in this thesis requires no additional changes to the environment,and is easy to use.It has certain value for the solution of practical problems and the intelligentization of modern parking lots.