Research On Parking Space Status Detection Algorithm Based On Re-parameterization Network And Improved YOLOv5

With the rapid growth of the number of cars in China,the problem of parking difficulty in daily travel is becoming more and more serious.By building intelligent parking lots,timely and accurate access to parking status information and integrated release is of great significance to alleviate the problem of parking difficulty.For parking space status detection,the traditional technology is solved by installing sensors on parking spaces.Although the use of sensors can accurately detect the parking space status,the installation of sensors is vulnerable to variable physical environment and the subsequent maintenance is complicated,which increases the cost.In the continuous advancement of computer vision technology,a variety of image technologies are used to detect the status of parking spaces.Not only is the cost of detection low,but the cameras used to acquire the images can also be used for parking lot security.Therefore,this paper proposes two deep learning-based algorithms for car parking status detection,as follows:(1)In order to improve the accuracy and speed of parking space status detection,a parking space status detection algorithm based on convolutional neural network structure reparameterization is proposed.This method considers parking space status detection as a classification problem of parking space images,and requires segmentation of each parking space during training and testing.During training,multiple branches consisting of small convolution kernels of different scales are used to extract local detail information in parking space images,enabling the model to achieve higher detection accuracy.After training,the model parameters obtained during training are transformed into equivalent forms using the convolutional network structure re-parameterization method,and the corresponding inference structure is also transformed.The resulting model has the same detection ability as the original model,but with faster detection speed and fewer parameters.The detection results on two datasets show that this method has significant advantages in both detection accuracy and speed.(2)To improve the vehicle detection capability,reduce misjudgment of obscured parking spaces,and increase the accuracy of parking space status detection,a parking space status detection algorithm based on improved YOLOv5 is proposed.This method is based on vehicle detection to achieve parking space status detection,therefore,three parts of YOLOv5 are improved to enhance the model’s detection performance for vehicle targets.Firstly,Coordinate Attention Module(CAM)is introduced in the backbone network of the model to guide the model to improve the attention to vehicle location information and channel features under the occlusion condition;then the feature fusion network is improved by using bidirectional scale connection and weighted feature fusion to replace the feature fusion module in the original model to enhance the multi-scale feature complement each other to reduce information loss;finally decoupling the prediction head of YOLOv5 and leaving the regression task and classification task to two branches to improve the expression capability of the detection head.When performing the parking space status detection,the improved YOLOv5 is first used to accurately detect the bounding box of all vehicles,and then it is judged with the parking space box based on Intersection over Union(IOU)to infer the parking space status.This paper addresses the problem of parking space status detection,and proposes two parking space status detection methods based on two approaches,image recognition and object detection,respectively,based on the summary of existing research.Tests on the experimental dataset and in the parking lot show that the proposed method has excellent detection results.