Research On Automatic Parking Method Based On Panoramic Vision

The rapid development of automotive intelligence has brought great changes to the automotive industry,while providing unprecedented safety and comfort for people’s travel.Automatic parking is an important part of the intelligent car,it rescues the driver from the tedious and boring parking process,greatly reduces the burden of the driver,but also effectively reduces the incidence of parking accidents.At present,the automatic parking system equipped with most vehicle models is based on ultrasonic radar,which cannot detect the parking space in the scene without reference objects,which has great application limitations and is easy to misidentify.Vision-based detection method can effectively solve the above problems,so the research of automatic parking method based on panoramic vision is more important.This paper studies the automatic parking method based on panoramic vision.Based on panoramic vision,from perception to planning control,it analyzes and designs the parking space detection and positioning system,parking path planning and tracking control respectively.The main contents are as follows:(ⅰ).Design of panoramic top view acquisition system.The panoramic top view acquisition system designed in this paper is used for the image acquisition of parking space.Through the panoramic top view,the 360-degree environmental information around the car body can be obtained,which makes it easier to collect a relatively complete image of parking space.Firstly,the distortion model of the fish eye camera is constructed,and the internal and external parameters of the camera are obtained by Zhang Zhengyou’s calibration method.According to these parameters,the distortion of the original image collected by the fish eye camera is corrected.Secondly,a top view perpendicular to the ground is obtained through top view transformation,and then four images are stitched using marker points.Finally,the images are fused by a distance-based weighted average fusion algorithm to obtain a panoramic top view.(ⅱ).Design of parking space detection system.At present,most traditional image processing methods are used to recognize parking space lines to realize parking space detection,but this method is not good for image recognition with poor imaging conditions and complex information.In order to solve this problem,this paper designs a parking space detection model based on deep learning.The images collected by the panoramic top view acquisition system is used to make a data set,and the images are annotated by Labelimg software.Considering the two aspects of detection speed and accuracy,the parking space detection based on yolov4 tiny is selected.The parking space detection model is trained and tested by using the self-made data set,the detection effect of the model was evaluated by the mean of average accuracy.In order to determine the coordinate information of parking space,this paper designs a parking space location method based on parking space corner’s skeleton.determine the effective parking space according to the detection results of the parking space detection system and extract the parking space corner’s image.Perform preprocessing such as grayscale,median filter,binarization,and closed operation on the parking angle image.The skeleton of the parking space corner is obtained by skeleton extraction.Then,the effective straight line segment of the parking space corner is screened by Hough transform.The coordinates of the current parking space corner are calculated and transformed into the coordinate system of the panoramic top view.Finally,the corner coordinates of four parking spaces are calculated.(ⅲ).Parking path planning and tracking control.In order to design a parking path that is easy to track,this paper combines the characteristics of the fifth-order polynomial path and the Sigmoid function path,and proposes a superimposed curve with continuous curvature and zero endpoint curvature.In parallel parking and vertical parking scenarios,the parking paths based on superimposed curves are proposed respectively according to the coordinates of parking space and the constraints.At the same time,the path tracking model prediction controller is designed,the parking model of the vehicle running at low speed is established and linearized,and the error model is constructed and discretized.Based on the control effect,the objective function considering the tracking accuracy and control increment,as well as the constraint conditions considering the execution ability and dynamic characteristics of the actuator are designed.(ⅳ).Experimental verification.A real-vehicle experiment platform is built on the basis of the Apollo smart car to verify the parking space detection and positioning system.The results show that the system designed in this paper has a good detection and positioning effect.Using Matlab/Simulink and Car Sim co-simulation method to simulate and verify the proposed path planning and tracking controller.The results show that the curvature of the parking path is continuous and the parking path is easy to track,and the tracking effect of the model prediction controller has a good tracking effect,and the lateral and longitudinal errors can be controlled within centimeter level,and the vehicle can be straighten when the parking is completed.