The Research On Parking Slot Detection Methods Based On Convolutional Neural Network

In recent years,as an important part of the Advanced Driving Assistance System,the Automatic Parking System is gradually becoming a research hotspot in the field of automatic driving.Equipping the vehicle with an automatic parking system can effectively avoid scratching accidents that occur when the driver is parking,and reduce the time required for parking.Its implementation mainly includes three steps: parking slots detection,path planning and vehicle control.Obviously,as the first step to achieve accurate automatic parking,the accuracy of parking slots detection is very important.At present,parking slots detection mainly uses a vision-based detection method,and then uses a distance sensor to detect static objects to determine the parking slot occupancy.This kind of methods need to merge the results of parking slots detection and obstacle detection,moreover the detection accuracy is greatly affected by the environment and weather.In view of the above problems,this thesis proposes a semantic segmentation detection method,which can quickly and effectively detect parking slots on panoramic surround view images,and the accuracy will not be affected by environmental conditions.The main research contents of this thesis are as follows:Firstly,this thesis produces dataset and augments it by data augmentation.Currently there are few public datasets with parking slots labels for semantic segmentation.In this thesis,based on the dataset ps2.0,a dataset ps2.0-SS with parking slots labels for semantic segmentation is made,which contains 3000 panoramic surround view images in different environments such as normal outdoor normal daylight,outdoor rainy,outdoor shadow,indoor parking-lot,etc.data augmentation is used to expand the dataset to 45000 images.Secondly,this thesis applys the semantic segmentation method based on fully convolutional neural network to parking slots detection.Based on the research of multiple semantic segmentation models,the technical problems of semantic segmentation and their corresponding optimization directions are summarized.A comparative experiment is designed,in which U-net,HFCN,VH-HFCN and the network proposed in this thesis are trained by learning the training set of ps2.0-SS to obtain the semantic segmentation models,and hyperparameters are adjusted to obtain the optimal weights.Then the images in the testing set of ps2.0-SS are tested to obtain the result that will be analyzed.Thirdly,this thesis proposes a new semantic segmentation model DVH-U-net based on fully convolutional neural network.The network structure,the loss function and the model optimization methods are designed.The model includes three parts: a down-sampling part for extracting feature information,an up-sampling part for restoring spatial resolution,and a multiple prediction part for combining feature maps of different scales.For the first time,a linear convolution module with dilated convolution is proposed,which can effectively extract linear features.By combining feature maps of different scales,the multiple prediction part merges deep features and shallow features better.The experimental results prove that the DVH-U-net proposed in this thesis achieves a pixel accuracy of 88.69% on the ps2.0-SS dataset,which is 2.33%-2.56% higher than other semantic segmentation models.Therefore,DVH-U-net has better accuracy and stronger robustness.