| With the development of economy and the improvement of people’s living standards,the number of cars in the country continues to reach a new high.At the same time,it also brings the problem of parking.The collision and scratch that are easy to occur in the process of parking make driver headache,especially for novice drivers who have just obtained their driving licenses.In order to assist the driver to park better,the thesis has carried out a series of research on the auxiliary parking system based on UAV.The system uses the characteristics of small volume,flexible flight and low environmental requirements of UAV and applies them to the field of parking assistance.It transmits the surrounding conditions of the vehicle within the vision of UAV back to the driver in time and helps the driver to complete the process of parking.This system can make up for the disadvantages of limited observation angle and low flexibility of traditional auxiliary parking solutions,such as reversing radar system,auto panoramic image system and so on.The system needs to detect the 3D pose of the vehicle during parking in time by the monocular image of the UAV and control the UAV to track the vehicle according to the results in order to ensure that the parking vehicle is always in the vision of the camera of UAV.The thesis proposes the auxiliary parking system based on UAV with machine learning method as the kernel,which uses the deep learning model to detect the 3D pose of vehicles from the perspective of UAV.In order to improve the accuracy of model and verify the function of the system quickly,simulation environment is utilized to collect nearly 10000 training data which are used to train the model with the data from real-world scene.Data acquisition and annotation,model training and evaluation are built to be as a standard process.Finally,experiments are carried out in the real-world scene to verify the feasibility of the system.The thesis proposes the method of obtaining the 3D pose of the vehicle based on license plate location,which can improve the speed of the pose recognition of the vehicle,aiming at that the vehicle during parking is always in the perspective of the camera of UAV and the driver can get complete images of the vehicle to park better.This method uses characteristics of position and shape of the contour of plate to obtain the vehicle pose,which can greatly improve the speed of 3D pose detection and realize the real-time tracking of the vehicle during parking.The system needs to detect the vehicle posture by the method of machine learning and control the UAV to fly to the front or rear of the vehicle during the initialization in order to ensure that the license plate needs to appear in the field of vision of the UAV camera at the initial stage.Then,it uses the method based on the license plate location to track the vehicle during parking,aiming to monitor surrounding conditions of the vehicle in time and transmit the video back to the driver,assisting to complete the task of parking.The improved system combines the advantages of the full-angle detection of machine learning method and the fast detection of traditional method.The results in the real scene prove its superiority in real-time performance. |
