Research On Parking Slot Detection Technology Based On Key Point Features In Panoramic Video

With the development of intelligent urban traffic,automatic driving technology has become a hot research topic,automatic parking is an important part of automatic driving,accurate and efficient access to the location and direction of available parking slot is an important task of automatic parking.Parking slot detection methods are mainly divided into three categories.The first one uses ultrasonic radar sensors to detect available parking slots,using the principle of ultrasonic reflection.This method requires that there are obstacles next to parking slots to reflect sound waves.The second is based on infrared or magnetic field detection technology,which usually requires the transformation of parking slots,it is costly and requires a lot of work.The third kind is the detection method based on image,through the camera around the vehicle image acquisition,image processing and testing according to the characteristics of the parking slot,compared with the other two methods,the method based on image can adapt to various types of parking slot,state requirements for automatic parking slot is low,and is now the mainstream of the method.Current image-based parking slot detection methods can better complete the detection task,however,there is still much room for improvement in accurately detecting the location and direction of parking slots.To solve these problems,this paper designed a parking slot detection solution,and the specific research content is as follows:(1)Stitching of panoramic images.In this paper,image-based methods are used to detect parking slots,so obtaining images is a necessary condition for detecting parking slots,panoramic images can provide the distance and angle information of any position in the image relative to the vehicle.Firstly,four fisheye cameras installed around the vehicle are used to collect images,and the collected images are corrected to eliminate the distortion caused by the fisheye lens.Then,four images are transformed into the same plane through registration and perspective transformation,and the overlapping parts of adjacent images are weighted and fused to eliminate the joint.(2)Detect parking slots in panoramic images.In this paper,the method of deep learning is used to detect parking slots.The detection task of parking slots is simplified to the detection of key points of parking slots,and the gaussian heatmap is used to locate key points.A dual-path multi-resolution detection network is designed to predict the position and angle of key points respectively,and the feature extraction ability of the network is strengthened through the multiresolution network.Through the relative position and angle information between the detected key points,fitting the parking slot information.(3)Deploy neural network to embedded devices.Embedded devices are limited in size and power consumption and usually have low computing power,so the designed neural network needs to be optimized appropriately to reduce the amount of computation and improve the operating efficiency.In this paper,TensorRT is used as a deployment tool to quantify and deploy the designed network.Enable it to run in real time in Jetson AGX Xavier.The experimental results show that this algorithm can effectively reduce the positioning and angle errors in parking slot detection,and provides a reference for the application of parking slot detection in automatic parking.