Research On Automatic Parking Scene Segmentation And Parking Space Detection Based On Deep Learning

In recent years,with the development of the economy and the continuous improvement of people’s living standards,the number of cars has increased year by year.As of the end of 2020,the number of cars in my country has reached 281 million,which has further aggravated parking difficulties.Although some vehicles have been equipped with automatic parking systems,their parking success rate is low and practicality is poor.The main problem of the current automatic parking system is that its parking space detection technology cannot be applied to both structured parking scenes with parking space marking lines and unstructured parking scenes without parking space marking lines.This topic originated from a research and development project of an intelligent logistics vehicle.In response to the above problems,the research and application of parking space detection algorithms suitable for multiple scenarios were launched.This paper proposes a new parking space detection method that does not rely on multi-sensor fusion technology.It takes the image of the on-board 4-channel fisheye camera as input.Through the semantic segmentation algorithm,the automatic parking scene is segmented,and parking space detection is performed according to the segmentation result.The specific research content is as follows.(1)Research on the internal parameter calibration,image distortion correction and external parameter calibration algorithms of vehicle-mounted fisheye cameras based on machine vision theory.At the same time,study the plane ranging algorithm based on monocular fisheye cameras,and conduct vehicle-mounted fisheye camera calibration experiments.(2)For the automatic parking scene segmentation task,construct an automatic parking scene segmentation data set,including three categories of free driving area,parking space marking line and vehicle.(3)Based on the Bi Se Net network and Deep Lab V3＿plus network design,2 sets of 8automatic parking scene segmentation models in total are designed and trained on the constructed data set.Perform segmentation experiments on the designed model,and select the model with the best performance according to the evaluation index,and use it as the semantic segmentation model of the parking space detection algorithm.(4)Based on the results of automatic parking scene segmentation,research the parking space detection method based on vertical grid search and the parking space detection method based on Radon transform,and select the best algorithm for real car experiment through parking space detection experiment.(5)Research the video stream data compression algorithm,explore the construction of a terminal-cloud collaborative computing cluster based on the Ray distributed architecture,build a hardware and software real vehicle experiment platform,and conduct real vehicle experiments to verify the feasibility of the algorithm studied in this paper.Experimental results show that the algorithm has a strong ability to generalize parking scenes,and can realize both structured parking scenes and unstructured parking scenes without relying on sensor fusion technology.