| In the context of the development of automotive intelligence,the application of automatic parking can effectively improve travel comfort and reduce the incidence of parking accidents.In the automatic parking technology,fast and accurate detection and location of empty parking spaces is a key part.Currently,the main method used is to achieve visionbased detection and localization of empty parking spaces using a surround view image composed of four fisheye cameras as input.However,the existing methods are easily affected by external factors such as illumination,shadows,and limited field of view of the surround view image,resulting in false detection,missed detection,and unstable positioning results.To address these problems,this paper proposes a vision-based empty parking space detection and localization system based on convolutional neural network,which improves the detection performance under the interference of external factors,and has a faster detection speed to meet the real-time demand of parking scenarios.The main research work is as follows:The circumferential image established by using the four-way fisheye camera to capture the image of the parking space around the vehicle can better reflect the geometric features of the parking space line.However,the fisheye camera has a large amount of radial distortion,so it is necessary to use the tessellation calibration method to calibrate the fisheye camera for orthorectification,convert the orthorectified image into a top view perspective through perspective transformation,and then stitch the four-way fisheye camera based on the marker points around the body to obtain the surround view image.For the task of empty parking space detection,the collected surround view images are used with the public dataset to produce the empty parking space dataset of this paper,which is labeled using labelimg and mainly divided into three kinds of empty parking spaces,horizontal,vertical and inclined,and three kinds of parking space angles,T-shaped,L-shaped and inclined.In order to meet the requirements for the accuracy and speed of empty parking space detection in the automatic parking scenario,the study is based on the YOLOv5-s algorithm,and considering the small scale variation of empty parking spaces in the surround view image,only the two detection branches(13,13)and(26,26)are used,and the Ghost Module is introduced to reduce the model parameters and improve the inference speed.Decoupled Head is used to decouple the classification and regression tasks to improve the detection accuracy.The experimental results show that the improved model m AP reaches94.34% and the detection speed reaches 47.7 frames/s,which has better detection effect relative to YOLOv5-s under the interference of environmental factors such as light and shadow.For the empty parking position localization task,according to the detection results of the empty parking position detection algorithm,the parking position angle screening and the tail parking position angle center point coordinate inference are carried out to realize the parking position line reconstruction and complete the positioning of empty parking position under the pixel coordinate system,and a pixel coordinate prediction method of empty parking position based on Kalman filtering is proposed to improve the stability of system positioning.Then,the actual position of the empty parking space is determined by studying the conversion relationship of pixel coordinates in the surround view image under the vehicle coordinate system to realize the positioning of the empty parking space under the vehicle coordinate system.Finally,the three parts are fused to propose a vision-based empty parking space detection and localization system,which is implemented and tested on a laboratory mobile platform.The experimental results prove the feasibility of the system,which can accurately,stably and in real time complete the task of detecting and localizing empty parking spaces. |
