WO2020024206A1 - 基于dcgan的停车数据修补方法、装置、设备及存储介质 - Google Patents

基于dcgan的停车数据修补方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2020024206A1
WO2020024206A1 PCT/CN2018/098261 CN2018098261W WO2020024206A1 WO 2020024206 A1 WO2020024206 A1 WO 2020024206A1 CN 2018098261 W CN2018098261 W CN 2018098261W WO 2020024206 A1 WO2020024206 A1 WO 2020024206A1
Authority
WO
WIPO (PCT)
Prior art keywords
parking lot
parking
data
sample
empty vehicle
Prior art date
Application number
PCT/CN2018/098261
Other languages
English (en)
French (fr)
Inventor
彭磊
邹万
李慧云
Original Assignee
深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Priority to PCT/CN2018/098261 priority Critical patent/WO2020024206A1/zh
Publication of WO2020024206A1 publication Critical patent/WO2020024206A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the invention belongs to the technical field of data repair, and particularly relates to a DCGAN-based parking data repair method, device, equipment and storage medium.
  • Any machine learning algorithm requires a large amount of sample data as support, and the development of intelligent parking systems also requires a large amount of parking data in a certain area.
  • Traditional methods for obtaining parking data include finding public data sets, or collecting through the purchase and installation of sensors. These traditional methods have high economic and time costs, making it difficult to obtain parking data for all parking lots in cities, resulting in most of the cities Parking data is missing. In addition, construction damage, line failures, and processing errors also cause incomplete parking data in some parking lots.
  • the above problems can be expressed as model distortion caused by missing data.
  • data augmentation techniques are generally used to deal with data missing problems, that is, data generation (or patching) techniques in a popular sense.
  • the latest application of this technology is mainly reflected in the style transfer and repair of images, and the research on traffic data repair is relatively scarce.
  • researches in the field of traffic data repair are mainly interpolation methods, such as data repair method based on Lagrange interpolation method, data repair method based on Newton interpolation method, and data repair method based on piecewise linear interpolation method.
  • interpolation methods such as data repair method based on Lagrange interpolation method, data repair method based on Newton interpolation method, and data repair method based on piecewise linear interpolation method.
  • the characteristics of these methods It requires certain prior knowledge and is often used in the field of traffic data repair with certain historical data.
  • due to economic costs, installation and construction, and economic property rights historical parking data for most parking lots is difficult to obtain.
  • the purpose of the present invention is to provide a DCGAN-based parking lot data repairing method, device, equipment and storage medium, which aims to solve the high cost of collecting parking lot data in all parking lots in the city in the prior art, and repair existing data.
  • the method has a large dependence on the historical parking data of the parking lot.
  • the present invention provides a DCGAN-based parking lot data repair method, which includes the following steps:
  • the present invention provides a parking lot data repairing device based on DCGAN, the device includes:
  • a collection data acquisition unit configured to obtain a pre-collected collection of urban parking lots and a set of urban geographic points of interest, where the parking lots in the collection of urban parking lots include a sample parking lot and a parking lot to be repaired;
  • a data processing unit configured to cluster the parking lot into a plurality of parking lot clusters according to the set of urban geographic points of interest, and generate an empty vehicle rate curve image of the sample parking lot according to the parking data of the sample parking lot ;
  • a network construction training unit is configured to construct a deep convolution-based generative adversarial network, and perform the deep convolution-based generative adversarial network based on an image of an empty vehicle rate curve of the sample parking lot in the parking lot cluster. Training to generate a simulation image of the empty vehicle rate curve corresponding to each parking lot cluster;
  • the parking data repair unit is configured to process and map the vacancy rate curve simulation image corresponding to each parking lot cluster to corresponding one-dimensional vacancy rate data, and according to the one-dimensional vacancy corresponding to each parking lot cluster
  • the vehicle rate data is used to repair the parking data in the parking lot to be repaired in each parking lot cluster.
  • the present invention also provides a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor.
  • the processor is implemented when the processor executes the computer program. The steps described in the above DCGAN-based parking lot data repair method.
  • the present invention also provides a computer-readable storage medium storing a computer program, and the computer program is executed by a processor to implement a method for repairing parking lot data based on DCGAN as described above. The steps described.
  • the present invention collects a collection of urban parking lots and a set of urban geographic points of interest in advance.
  • the parking lots in the urban parking lot set include a sample parking lot and a parking lot to be patched, and the parking lots are clustered into parking lot clusters according to the set of urban geographic points of interest.
  • an empty car rate curve image of the sample parking lot is generated.
  • a deep convolution-based generative adversarial network is trained to pass the training.
  • the deep convolution-based generative adversarial network generates empty car rate curve simulation images corresponding to each parking lot cluster, processes these empty car rate curve simulation images, and maps them into corresponding one-dimensional empty car rate data, so that according to these One-dimensional space rate data is used to repair the parking data in each parking lot in the parking lot cluster, so that the parking lot to be repaired can be accurately repaired without relying on the prior knowledge of the parking lot to be repaired. Parking data patching, thereby saving data collection costs for parking lots.
  • FIG. 1 is a flowchart of a DCGAN-based parking lot data repair method provided by Embodiment 1 of the present invention
  • FIG. 2 is a schematic structural diagram of a DCGAN-based parking lot data repair device provided by Embodiment 2 of the present invention.
  • FIG. 3 is a schematic diagram of a preferred structure of a parking lot data repairing device based on DCGAN provided by Embodiment 2 of the present invention.
  • FIG. 4 is a schematic structural diagram of a computing device according to a third embodiment of the present invention.
  • FIG. 1 shows the implementation process of a DCGAN-based parking lot data repair method provided by Embodiment 1 of the present invention. For convenience of explanation, only the parts related to the embodiment of the present invention are shown, and the details are as follows:
  • step S101 a pre-collected urban parking lot set and a geographic geographic point of interest set are acquired.
  • the parking lots in the urban parking lot set include a sample parking lot and a parking lot to be repaired.
  • the embodiments of the present invention are applicable to a data repair platform, system, and equipment.
  • the pre-collected collection of urban parking lots contains information about multiple parking lots in the city and each parking lot. These parking lots include parking lots that lack parking data, and parking lots that do not lack parking data.
  • parking lots lacking parking data are referred to as parking lots to be repaired, and parking lots not lacking parking data are referred to as sample parking lots.
  • the relevant information of the parking lot to be repaired may include the geographic location of the parking lot to be repaired in the city and the size of the parking lot to be repaired, and the relevant information of the sample parking lot may include the geographical location of the parking lot in the city and the size of the sample parking lot. And parking data for the sample parking lot.
  • the pre-collected collection of urban geographic points of interest includes the geographic locations of geographic points of interest (POIs, such as parking lots, shopping malls, stations, and schools) in the city. This information is available on the website, and the geographic location can be longitude and latitude information.
  • POIs geographic points of interest
  • step S102 the parking lot is clustered into a plurality of parking lot clusters according to the set of urban geographic points of interest, and the empty parking rate curve image of the sample parking lot is generated according to the parking data of the sample parking lot.
  • a tolerance radius value is set in advance, and a circle is formed with the parking lot as the center and the tolerance radius as the circle radius.
  • These POIs are the POIs around the parking lot, and the corresponding spatial characteristics of each parking lot are constructed based on the POIs around each parking lot.
  • DBSCAN Density-Based Spatial Clustering Applications with Noise
  • the core point in DBSCAN is determined according to the data size of the data set formed by the spatial characteristics of all parking lots.
  • Threshold value MinPts based on the mean value of the Euclidean distance between parking lots randomly selected from the urban parking lot collection, determines the density field radius ⁇ in DBSCAN, thereby effectively improving the classification effect of parking lots.
  • 10 parking lots are randomly selected from 310 parking lots, and 7 (the number of types of POIs) is calculated to obtain a 7-dimensional space.
  • the average European-style distance of these parking lots is 22, and 22 is set as the DBSCAN medium density.
  • the field radius ⁇ is taken as 49, which is the core point threshold MinPts in DBSCAN.
  • the parking data of the sample parking lot is the accurate time information of vehicles entering and leaving the sample parking lot, and according to the vehicle entering and exiting the sample parking lot.
  • the accurate time information can obtain the number of free parking spaces in the sample parking lot at different times. Considering that the size of different parking lots may be different (that is, the total number of parking spaces may be different), the number of free parking spaces in the sample parking lot at different times can be obtained.
  • the normalization process is performed to obtain the free parking space rate at different times of the sample parking lot.
  • the free parking space rate is referred to herein as the empty parking rate.
  • two-dimensional coordinates corresponding to the vacancy rate of the sample parking lot at different times are calculated, and these two-dimensional coordinates constitute a vacancy rate curve corresponding to the sample parking lot. Then, the vacancy rate curve image of the sample parking lot is obtained.
  • the vacant rate of the sample parking lot at time t p is expressed as Is the number of free parking spaces in the sample parking lot at time t p , Total is the total number of parking spaces in the sample parking lot, L is the length of the curve in the curve parameters, H is the height of the curve in the curve parameters, and J is the number of sampling points.
  • the POIs of different parking lots in the same parking lot cluster after clustering must be similar, but the parking data of different parking lots in the same parking lot cluster after clustering are not necessarily similar, so they pass Pearson after the clustering
  • the correlation coefficient calculates the degree of similarity of the empty vehicle rate curves of different sample parking lots in the same parking lot cluster to determine whether the empty vehicle rate curves of different parking lots in the same parking lot cluster are similar. For example, when the Pearson correlation coefficient of the vacancy rate curves of two sample parking lots is greater than 0.5, the parking data of the two sample parking lots is considered to have a strong correlation. If more than 70% of the number of permutations and combinations of all parking lots in the parking lot cluster, the parking data of different parking lots in the parking lot cluster are considered similar.
  • step S103 a generative adversarial network based on deep convolution is constructed, and the generative adversarial network based on deep convolution is trained according to the empty car rate curve image of the sample parking lot in the parking lot cluster to generate each parking lot. Simulation image of empty vehicle rate curve corresponding to the cluster.
  • a preset generative adversarial network (GAN) is extended to obtain a deep volume-based
  • the productive generative adversarial network (Deep Convolutional, Adversarial Networks, DCGAN for short) here does not limit the number of network layers of supervised convolutional neural networks and generative adversarial networks, and the number of units in each layer.
  • the main changes in the model structure of the constructed DCGAN are: the pooling layer is replaced with a convolution layer, and the pooling layer in the discriminator is replaced with a convolution layer with a step size.
  • the pooling layer is replaced by a micro-step convolutional layer; batch normalization is used on both the discriminator and the generator; the fully connected layer is removed.
  • a filter for normality check is connected to the output layer of the generator in the DCGAN, and only the simulation image of the empty rate curve that meets the normality check is output to ensure the empty rate generated by the generator in the DCGAN. Curve simulation effect.
  • the filter for normality verification uses a D'Agostino-Pearson test to improve the effect of normality verification on the simulation image of the empty vehicle rate curve.
  • the empty vehicle rate curve image of the sample parking lot in the parking lot cluster is input as training data to DCGAN, the DCGAN is trained, and the trained DCGAN generates the empty vehicle rate curve corresponding to the parking lot cluster.
  • the simulation image that is, the simulation image of the empty vehicle rate curve corresponding to the parking lot to be repaired in the parking lot cluster.
  • the empty vehicle rate curve image generated by DCGAN is called the empty vehicle rate curve simulation image), so that it can be generated. Simulation image of empty car rate curve corresponding to each parking lot cluster.
  • step S104 the vacancy rate curve simulation image corresponding to each parking lot cluster is processed and mapped into corresponding one-dimensional vacancy rate data. According to the one-dimensional vacancy rate data corresponding to each parking lot cluster, Parking data in each parking lot cluster to be repaired.
  • the empty vehicle rate curve image is mapped to the corresponding one-dimensional empty vehicle rate data (time series of empty vehicle rate).
  • the one-dimensional vacancy rate data corresponding to the parking lot cluster and the size of the parking lot to be repaired in the parking lot cluster the number of free parking spaces at different times in the parking lot to be repaired in the parking lot cluster can be calculated, and then each The parking data of each parking lot in the parking lot cluster is repaired.
  • the formula for graying the empty vehicle rate curve simulation image is:
  • a threshold value is set in advance, each pixel point of the vacancy rate curve simulation image is traversed, and the pixel point whose pixel value exceeds the threshold value is set. It is white, otherwise it is set to black, so as to improve the processing effect of the empty vehicle rate curve simulation image.
  • the noise reduction processing of the empty vehicle rate curve simulation image includes the removal of burr points and the outlier points.
  • the curve burr points in the empty vehicle rate curve simulation image are caused by too dense pixels, the pixel points need to be adjusted. Density is reduced, so the average filter is used to remove the burr points to improve the processing effect of the burr points on the simulation image of the empty vehicle rate curve.
  • the formula for removing the glitch points by using the average filter is:
  • M represents the preset filter window size
  • f (w, e) is a simulation image of the empty vehicle rate curve before removing the burr point
  • f '(x, y) is a simulation image of the empty vehicle rate curve after removing the burr point.
  • a distance-based anomaly detection algorithm is used to detect the outliers on the simulation image of the empty vehicle rate curve to improve the detection effect of the outliers.
  • the distance-based anomaly detection algorithm is a local anomaly factor LOF algorithm (Local Outlier Factor) to improve the outlier detection effect.
  • the parking lot is clustered into a parking lot cluster, and the empty parking rate curve image of the sample parking lot is generated based on the parking data of the sample parking lot.
  • the deep convolution-based generative adversarial network is trained to generate the simulated empty car rate curve simulation images corresponding to each parking lot cluster through the trained deep convolutional generative adversarial network.
  • FIG. 2 shows a structure of a parking lot data repairing device based on DCGAN provided in Embodiment 3 of the present invention. For convenience of explanation, only parts related to the embodiment of the present invention are shown, including:
  • the collection data acquisition unit 21 is configured to obtain a pre-collected collection of urban parking lots and a set of urban geographic points of interest.
  • the parking lots in the urban parking lot collection include a sample parking lot and a parking lot to be repaired.
  • the pre-collected collection of urban parking lots includes multiple parking lots in the city and related information of each parking lot, and these parking lots include a parking lot to be repaired and a sample parking lot.
  • the relevant information of the parking lot to be repaired may include the geographic location of the parking lot to be repaired in the city and the size of the parking lot to be repaired
  • the relevant information of the sample parking lot may include the geographical location of the parking lot in the city and the size of the sample parking lot.
  • parking data for the sample parking lot may be used to be repaired.
  • the pre-collected urban geographic point of interest collection includes the geographic position of the geographic point of interest (POI) in the city, which can be obtained on a map website through a crawler program, and the geographic position may be longitude and latitude information.
  • POI geographic point of interest
  • the data processing unit 22 is configured to cluster a parking lot into a plurality of parking lot clusters according to a set of urban geographic points of interest, and generate an empty vehicle rate curve image of the sample parking lot according to the parking data of the sample parking lot.
  • a tolerance radius value is set in advance, and a circle is formed with the parking lot as the center and the tolerance radius as the circle radius.
  • These POIs are the POIs around the parking lot, and the corresponding spatial characteristics of each parking lot are constructed based on the POIs around each parking lot.
  • a parking lot cluster According to the space features corresponding to each parking lot and the preset density-based spatial clustering algorithm (DBSCAN), All parking lots in the collection are clustered to obtain multiple clusters, which effectively improves the classification effect of parking lots. For ease of description, the cluster obtained by the clustering is referred to as a parking lot cluster here.
  • the core point in DBSCAN is determined according to the data size of the data set formed by the spatial characteristics of all parking lots.
  • Threshold value MinPts based on the mean value of the Euclidean distance between parking lots randomly selected from the urban parking lot collection, determines the density field radius ⁇ in DBSCAN, thereby effectively improving the classification effect of parking lots.
  • the parking data of the sample parking lot is the accurate time information of vehicles entering and leaving the sample parking lot, and according to the vehicle entering and exiting the sample parking lot.
  • the accurate time information can obtain the number of free parking spaces in the sample parking lot at different times. Considering that the size of different parking lots may be different, the number of free parking spaces in the sample parking lot at different times can be normalized to obtain the sample parking lot. Empty vacancy rate at different times of the game.
  • two-dimensional coordinates corresponding to the vacancy rate of the sample parking lot at different times are calculated, and these two-dimensional coordinates constitute a vacancy rate curve corresponding to the sample parking lot. Then, the vacancy rate curve image of the sample parking lot is obtained.
  • the vacant rate of the sample parking lot at time t p is expressed as Is the number of free parking spaces in the sample parking lot at time t p , Total is the total number of parking spaces in the sample parking lot, L is the length of the curve in the curve parameters, H is the height of the curve in the curve parameters, and J is the number of sampling points.
  • the POIs of different parking lots in the same parking lot cluster after clustering must be similar, but the parking data of different parking lots in the same parking lot cluster after clustering are not necessarily similar, so they pass Pearson after the clustering
  • the correlation coefficient calculates the degree of similarity of the empty vehicle rate curves of different sample parking lots in the same parking lot cluster to determine whether the empty vehicle rate curves of different parking lots in the same parking lot cluster are similar.
  • the network construction training unit 23 is configured to construct a generative adversarial network based on deep convolution, and train the generative adversarial network based on deep convolution according to the empty car rate curve image of the sample parking lot in the parking lot cluster to generate each Simulation image of empty car rate curve corresponding to each parking lot cluster.
  • the preset generative adversarial network is extended according to a preset supervised convolutional neural network (CNN) to obtain a deep convolution-based generative adversarial network (DCGAN) in This does not limit the number of network layers of the supervised convolutional neural network and the generative adversarial network, and the number of units in each layer.
  • the main changes in the model structure of the constructed DCGAN are: the pooling layer is replaced with a convolution layer, and the pooling layer in the discriminator is replaced with a convolution layer with a step size.
  • the pooling layer is replaced by a micro-step convolutional layer; batch normalization is used on both the discriminator and the generator; the fully connected layer is removed.
  • a filter for normality check is connected to the output layer of the generator in the DCGAN, and only the simulation image of the empty rate curve that meets the normality check is output to ensure the empty rate generated by the generator in the DCGAN. Curve simulation effect.
  • the filter for normality verification uses a D'Agostino-Pearson test to improve the effect of normality verification on the simulation image of the empty vehicle rate curve.
  • the empty vehicle rate curve image of the sample parking lot in the parking lot cluster is input as training data to DCGAN, the DCGAN is trained, and the trained DCGAN generates the empty vehicle rate curve corresponding to the parking lot cluster.
  • the simulation image can generate the vacancy rate curve simulation image corresponding to each parking lot cluster.
  • the parking data repair unit 24 is configured to process and map the vacancy rate curve simulation image corresponding to each parking lot cluster to corresponding one-dimensional vacancy rate data, and according to the one-dimensional vacancy rate data corresponding to each parking lot cluster , To repair parking data in each parking lot cluster to be repaired.
  • the empty vehicle rate curve image is mapped to the corresponding one-dimensional empty vehicle rate data.
  • the one-dimensional vacancy rate data corresponding to the parking lot cluster and the size of the parking lot to be repaired in the parking lot cluster the number of free parking spaces at different times in the parking lot to be repaired in the parking lot cluster can be calculated, and then each The parking data of each parking lot in the parking lot cluster is repaired.
  • the formula for graying the empty vehicle rate curve simulation image is:
  • a threshold value is set in advance, each pixel point of the vacancy rate curve simulation image is traversed, and the pixel point whose pixel value exceeds the threshold value is set. It is white, otherwise it is set to black, so as to improve the processing effect of the empty vehicle rate curve simulation image.
  • the noise reduction processing of the empty vehicle rate curve simulation image includes the removal of burr points and the outlier points.
  • the curve burr points in the empty vehicle rate curve simulation image are caused by too dense pixels, the pixel points need to be adjusted. Density is reduced, so the average filter is used to remove the burr points to improve the processing effect of the burr points on the simulation image of the empty vehicle rate curve.
  • the formula for removing the glitch points by using the average filter is:
  • M represents the preset filter window size
  • f (w, e) is a simulation image of the empty vehicle rate curve before removing the burr point
  • f '(x, y) is a simulation image of the empty vehicle rate curve after removing the burr point.
  • a distance-based anomaly detection algorithm is used to detect the outliers on the simulation image of the empty vehicle rate curve to improve the detection effect of the outliers.
  • the distance-based anomaly detection algorithm is a local anomaly factor LOF algorithm to improve the outlier point detection effect.
  • the data processing unit 22 includes:
  • a geographic point of interest determination unit 321, configured to determine, according to a preset tolerance radius, a geographic point of interest around each parking lot in a set of urban parking lots;
  • the parking lot clustering unit 322 is configured to construct the spatial features corresponding to each parking lot according to the geographic interest points around each parking lot, and according to the spatial features corresponding to each parking lot and a preset density-based spatial clustering algorithm, Perform high-dimensional clustering on parking lots.
  • the network construction training unit 23 includes:
  • a network construction unit 331 is configured to expand a preset generative adversarial network according to a preset supervised convolutional neural network to construct a generative adversarial network based on deep convolution and a generative adversarial network based on deep convolution.
  • the output layer in is connected to a filter for normality check.
  • the parking lot is clustered into a parking lot cluster, and the empty parking rate curve image of the sample parking lot is generated based on the parking data of the sample parking lot.
  • the deep convolution-based generative adversarial network is trained to generate the simulated empty car rate curve simulation images corresponding to each parking lot cluster through the trained deep convolutional generative adversarial network.
  • each unit of the DCGAN-based parking lot data repair device may be implemented by corresponding hardware or software units.
  • Each unit may be an independent software and hardware unit, or may be integrated into one software and hardware unit.
  • FIG. 4 shows the structure of a computing device provided in Embodiment 3 of the present invention. For ease of description, only parts related to the embodiment of the present invention are shown.
  • the computing device 4 includes a processor 40, a memory 41, and a computer program 42 stored in the memory 41 and executable on the processor 40.
  • the processor 40 executes the computer program 42, the steps in the foregoing method embodiment are implemented, for example, steps S101 to S104 shown in FIG. 1.
  • the processor 40 executes the computer program 42, the functions of the units in the above-mentioned device embodiment are realized, for example, the functions of the units 21 to 24 shown in FIG. 2.
  • the parking lot is clustered into a parking lot cluster, and the empty parking rate curve image of the sample parking lot is generated based on the parking data of the sample parking lot.
  • the deep convolution-based generative adversarial network is trained to generate the simulated empty car rate curve simulation images corresponding to each parking lot cluster through the trained deep convolutional generative adversarial network.
  • Embodiment 4 is a diagrammatic representation of Embodiment 4:
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the foregoing method embodiment are implemented, for example, as shown in FIG. 1.
  • the functions of the units in the foregoing device embodiments are implemented, for example, the functions of units 21 to 24 shown in FIG. 2.
  • the parking lot is clustered into a parking lot cluster, and the empty parking rate curve image of the sample parking lot is generated based on the parking data of the sample parking lot.
  • the deep convolution-based generative adversarial network is trained to generate the simulated empty car rate curve simulation images corresponding to each parking lot cluster through the trained deep convolutional generative adversarial network.
  • the computer-readable storage medium of the embodiment of the present invention may include any entity or device capable of carrying computer program code, a recording medium, for example, a memory such as a ROM / RAM, a magnetic disk, an optical disk, a flash memory, or the like.
  • a recording medium for example, a memory such as a ROM / RAM, a magnetic disk, an optical disk, a flash memory, or the like.

Abstract

本发明适用数据修补技术领域，提供了一种基于DCGAN的停车数据修补方法、装置、设备及存储介质，该方法包括：将城市停车场集合中的停车场聚类为停车场簇，停车场包括样本停车场和待修补停车场，根据样本停车场的停车数据生成对应的空车率曲线图像，根据停车场簇中样本停车场的空车率曲线图像，训练基于深度卷积的生成式对抗网络，以生成每个停车场簇的空车率曲线仿真图像，对这些空车率曲线仿真图像进行处理并映射为对应的一维空车率数据，以根据这些一维空车率数据对每个停车场簇中的待修补停车场进行数据修补，从而在不依赖待修补停车场的先验知识时对待修补停车场进行较为精准的数据修补，实现城市级的停车数据修补,节约了停车场的数据采集成本。

Description

基于DCGAN的停车数据修补方法、装置、设备及存储介质 技术领域
本发明属于数据修补技术领域，尤其涉及一种基于DCGAN的停车数据修补方法、装置、设备及存储介质。
背景技术
交通运输作为现代经济的重要部分，但伴随着社会经济的发展，日益增长的交通需求和城市交通负载量之间的矛盾日益凸显。其中，停车难问题比较突出，智能停车系统是解决该问题、降低停车时间成本的最有效方法之一。
任何机器学习算法都需要大量样本数据作为支撑，智能停车系统的研发同样需要一定区域内大量停车场的停车数据。获取停车数据的传统方法包括查找公开数据集、或者通过购买和安装传感器进行采集等，这些传统方法经济成本和时间成本较高，很难获得城市中所有停车场的停车数据，导致城市中大部分停车场缺少停车数据。此外，施工损坏、线路故障和处理错误等，也使得部分停车场存在停车数据不完备的情况。因此，面对大范围停车数据缺失的情况，如何在经济成本可控的前提下，提高城市范围内停车数据的完整性、有效性和可预测性，为相关的机器学习算法和交通分析模型提供有效支持，是当前智能停车算法设计和训练所需考虑的问题。
上述问题可表述为因数据缺失引起的模型失真，在机器学习中一般采用数据增强技术来处理数据缺失问题，即通俗意义上的数据生成(或修补)技术。这一技术的最新应用成果主要体现在图像的风格迁移和修复上，交通数据修补上的研究相对匮乏。目前交通数据修补领域的研究主要为一些插值法，例如基于拉格朗日插值法的数据修补方法、基于牛顿插值法的数据修补法和基于分段线性插值法的数据修补法，这些方法的特点是需要一定的先验知识，常常被用 于有一定历史数据的交通数据修补领域。然而，由于经济成本、安装施工以及经济产权等原因，多数停车场的历史停车数据难以获得。
发明内容
本发明的目的在于提供一种基于DCGAN的停车场数据修补方法、装置、设备及存储介质，旨在解决现有技术中采集城市范围内所有停车场停车数据的成本较高、且现有数据修补方法对停车场历史停车数据依赖性较大的问题。
一方面，本发明提供了一种基于DCGAN的停车场数据修补方法，所述方法包括下述步骤：
获取预先采集的城市停车场集合和城市地理兴趣点集合，所述城市停车场集合中的停车场包括样本停车场和待修补停车场；
根据所述城市地理兴趣点集合将所述停车场聚类为多个停车场簇，并根据所述样本停车场的停车数据生成所述样本停车场的空车率曲线图像；
构建基于深度卷积的生成式对抗网络，根据所述停车场簇中所述样本停车场的空车率曲线图像，对所述基于深度卷积的生成式对抗网络进行训练，以生成所述每个停车场簇对应的空车率曲线仿真图像；
对所述每个停车场簇对应的空车率曲线仿真图像进行处理并映射为相应的一维空车率数据，根据所述每个停车场簇对应的一维空车率数据，对所述每个停车场簇中的所述待修补停车场进行停车数据修补。
另一方面，本发明提供了一种基于DCGAN的停车场数据修补装置，所述装置包括：
采集数据获取单元，用于获取预先采集的城市停车场集合和城市地理兴趣点集合，所述城市停车场集合中的停车场包括样本停车场和待修补停车场；
数据处理单元，用于根据所述城市地理兴趣点集合将所述停车场聚类为多个停车场簇，并根据所述样本停车场的停车数据生成所述样本停车场的空车率曲线图像；
网络构建训练单元，用于构建基于深度卷积的生成式对抗网络，根据所述停车场簇中所述样本停车场的空车率曲线图像，对所述基于深度卷积的生成式对抗网络进行训练，以生成所述每个停车场簇对应的空车率曲线仿真图像；以及
停车数据修补单元，用于对所述每个停车场簇对应的空车率曲线仿真图像进行处理并映射为相应的一维空车率数据，根据所述每个停车场簇对应的一维空车率数据，对所述每个停车场簇中的所述待修补停车场进行停车数据修补。
另一方面，本发明还提供了一种计算设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上述基于DCGAN的停车场数据修补方法所述的步骤。
另一方面，本发明还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现如上述基于DCGAN的停车场数据修补方法所述的步骤。
本发明预先采集城市停车场集合和城市地理兴趣点集合，城市停车场集合中的停车场包括样本停车场和待修补停车场，根据城市地理兴趣点集合将这些停车场聚类为停车场簇，并根据样本停车场的停车数据生成样本停车场的空车率曲线图像，根据停车场簇中样本停车场的空车率曲线图像对基于深度卷积的生成式对抗网络进行训练，以通过训练好的基于深度卷积的生成式对抗网络生成每个停车场簇对应的空车率曲线仿真图像，对这些空车率曲线仿真图像进行处理并映射为对应的一维空车率数据，以根据这些一维空间率数据对每个停车场簇中的待修补停车场进行停车数据修补，从而在不依赖待修补停车场的先验知识时对待修补停车场进行较为精准的数据修补，实现城市级的停车数据修补，进而节约了停车场的数据采集成本。
附图说明
图1是本发明实施例一提供的基于DCGAN的停车场数据修补方法的实现 流程图；
图2是本发明实施例二提供的基于DCGAN的停车场数据修补装置的结构示意图；
图3是本发明实施例二提供的基于DCGAN的停车场数据修补装置的优选结构示意图；以及
图4是本发明实施例三提供的计算设备的结构示意图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
以下结合具体实施例对本发明的具体实现进行详细描述：
实施例一：
图1示出了本发明实施例一提供的基于DCGAN的停车场数据修补方法的实现流程，为了便于说明，仅示出了与本发明实施例相关的部分，详述如下：
在步骤S101中，获取预先采集的城市停车场集合和城市地理兴趣点集合，城市停车场集合中的停车场包括样本停车场和待修补停车场。
本发明实施例适用于数据修补平台、系统及设备。预先采集的城市停车场集合中包含城市中的多个停车场和每个停车场的相关信息，这些停车场中有缺乏停车数据的停车场，也有不缺乏停车数据的停车场，为了支持智能停车系统的研发需要对缺乏停车数据的停车场进行停车数据修补，因此为了便于描述，将缺乏停车数据的停车场称为待修补停车场，将不缺乏停车数据的停车场称为样本停车场。待修补停车场的相关信息可包括待修补停车场在城市中的地理位置和待修补停车场的规模，样本停车场的相关信息可包括样本停车场在城市中的地理位置、样本停车场的规模以及样本停车场的停车数据。
在本发明实施例中，预先采集的城市地理兴趣点集合中包括城市中地理兴 趣点(Point of Interest，简称POI，例如停车场、商场、车站、学校)的地理位置，可通过爬虫程序在地图网站上获取这些信息，地理位置可为经纬度信息。
在步骤S102中，根据城市地理兴趣点集合将停车场聚类为多个停车场簇，并根据样本停车场的停车数据生成样本停车场的空车率曲线图像。
在本发明实施例中，在生活中人们往往会寻找离目的地最近的停车场，如果最近的停车场没有停车位才会考虑距离远一点的停车场，但最终选择的停车场和目的地的距离不会无限远，即存在一个人们可接受的最远距离，在此将该最远距离称为容忍度半径。因此，以停车场为圆心、且以容忍度半径为圆半径形成一个圆，在该圆内的POI都将对停车场的停车数据产生影响，圆外的POI对停车场停车数据产生的影响可忽略不计，即可认为周围POI类型数量相似的两个停车场停车数据也很可能相似。
优选地，在根据城市地理兴趣点集合将停车场聚类为多个停车场簇时，预先设置容忍度半径的值，以停车场为圆心、且以容忍度半径为圆半径形成一个圆，从城市地理兴趣点集合中获取地理位置出现在该圆内的POI，这些POI即停车场周围的POI，根据每个停车场周围的POI构建每个停车场对应的空间特征，若当前城市中主要的POI有n种(或者城市地理兴趣点集合中采集了n种POI)，则每个停车场对应的空间特征为一个n维的特征向量，根据每个停车场对应的空间特征和预设的基于密度的空间聚类算法(Density-Based Spatial Clustering of Applications with Noise，简称DBSCAN)，对城市停车场集合中所有停车场进行聚类，得到多个簇，从而有效地提高了停车场的分类效果。为了便于描述，在此将聚类得到的簇称为停车场簇。
进一步优选地，在根据每个停车场对应的空间特征和DBSCAN对城市停车场集合中所有停车场进行聚类时，根据所有停车场的空间特征所形成数据集的数据规模确定DBSCAN中的核心点阈值MinPts，根据从城市停车场集合中随机选取的停车场之间欧式距离的均值，确定DBSCAN中密度的领域半径ε，从而有效提高停车场的分类效果。
作为示例地，在具体实验过程中，从310个停车场随机抽取10个停车场，计算7得到(POI的类型数量)维空间这些停车场欧式距离的均值为22，将22设置为DBSCAN中密度的领域半径ε，取49为DBSCAN中的核心点阈值MinPts，由DBSCAN将这310个停车场聚类为6类，即6个停车场簇。
在本发明实施例中，在根据样本停车场的停车数据生成样本停车场的空车率曲线图像时，样本停车场的停车数据为车辆进出样本停车场的准确时间信息，依据车辆进出样本停车场的准确时间信息可得到样本停车场在不同时刻空闲的停车位数量，考虑到不同停车场的规模可能不同(即停车位总数量可能不同)，可对样本停车场在不同时刻空闲的停车位数量进行归一化处理，得到样本停车场不同时刻的空闲车位率，为了简洁描述，在此将空闲车位率称为空车率。
在本发明实施例中，根据预先设置的曲线参数和采样点数目，计算样本停车场不同时刻的空车率对应的二维坐标，由这些二维坐标构成样本停车场对应的空车率曲线，进而得到样本停车场的空车率曲线图像。
优选地，样本停车场时刻t p的空车率为
Figure PCTCN2018098261-appb-000001
样本停车场时刻t p的空车率对应的二维坐标为
Figure PCTCN2018098261-appb-000002
Figure PCTCN2018098261-appb-000003
Figure PCTCN2018098261-appb-000004
其中，样本停车场的时间序列表示为
Figure PCTCN2018098261-appb-000005
Figure PCTCN2018098261-appb-000006
为样本停车场在时刻t p的空闲车位数量，Total为样本停车场中停车位总数，L为曲线参数中的曲线长度，H为曲线参数中的曲线高度，J为采样点数目。
优选地，由于聚类后同一停车场簇内不同停车场的POI一定是相似的，但聚类后同一停车场簇内不同停车场的停车数据不一定相似，因此在聚类结束后通过皮尔逊相关系数计算同一停车场簇中不同样本停车场空车率曲线的相似程度，以判断同一停车场簇中不同停车场空车率曲线是否相似。例如，当两个样本停车场空车率曲线的皮尔逊相关系数大于0.5，认为两个样本停车场的停车数据具有强相关性，若一个停车场簇中具有强相关性的停车场排列组合数量超过该停车场簇中所有停车场排列组合数的70％，则认为该停车场簇中不同停车场 的停车数据相似。
在步骤S103中，构建基于深度卷积的生成式对抗网络，根据停车场簇中样本停车场的空车率曲线图像，对基于深度卷积的生成式对抗网络进行训练，以生成每个停车场簇对应的空车率曲线仿真图像。
在本发明实施例中，根据预设的有监督的卷积神经网络(Convolutional Neural Networks，简称CNN)对预设的生成式对抗网络(Generative Adversarial Networks，简称GAN)进行扩展，以得到基于深度卷积的生成式对抗网络(Deep Convolutional Generative Adversarial Networks，简称DCGAN)在此对有监督的卷积神经网络和生成式对抗网络的网络层数、以及每层的单元数目不进行限定。构建的DCGAN相较于扩展前的GAN，模型结构上的主要变化有：池化层用卷积层替代，其中，鉴别器中的池化层用带步长的卷积层替代，生成器中的池化层用微步幅的卷积层替代；在鉴别器和生成器上都采用批标准化；移除全连接层。优选地，在DCGAN中生成器的输出层连接着用于正态性校验的过滤器，仅输出符合正态性校验的空车率曲线仿真图像，以保证DCGAN中生成器生成的空车率曲线仿真图像的效果。进一步优选地，用于正态性校验的过滤器采用达戈斯提诺算法(D'Agostino-Pearson test)，以提高对空车率曲线仿真图像进行正态性校验的效果。
在本发明实施例中，将停车场簇中样本停车场的空车率曲线图像作为训练数据输入DCGAN中，对DCGAN进行训练，并由训练好的DCGAN生成该停车场簇对应的空车率曲线仿真图像(即该停车场簇中待修补停车场对应的空车率曲线仿真图像，此处为了便于区分，将DCGAN生成的空车率曲线图像称为空车率曲线仿真图像)，从而可生成每个停车场簇对应的空车率曲线仿真图像。
在步骤S104中，对每个停车场簇对应的空车率曲线仿真图像进行处理并映射为相应的一维空车率数据，根据每个停车场簇对应的一维空车率数据，对每个停车场簇中的待修补停车场进行停车数据修补。
在本发明实施例中，在获得每个停车场簇对应的空车率曲线仿真图像后， 对空车率曲线图像进行灰度化、二值化以及降噪处理等图像处理操作，将处理后的空车率曲线图像映射为相应的一维空车率数据(空车率的时间序列)。根据停车场簇对应的一维空车率数据和该停车场簇中待修补停车场的规模，可计算得到该停车场簇中待修补停车场不同时刻的空闲停车位数量，进而实现对每个停车场簇中每个待修补停车场的停车数据修补。在将处理后的空车率曲线图像映射为相应的一维空车率数据时，可采用上述样本停车场不同时刻的空车率到空车率曲线图像的映射关系的逆关系实现。
优选地，对空车率曲线仿真图像进行灰度化的公式为：
R 2＝G 2＝B 2＝R 1*a 1+G 1*a 2+B 1*a 3，其中，R 2、G 2和B 2为空车率曲线仿真图像灰度化之后的RGB值，R 1、G 1和B 1为空车率曲线仿真图像灰度化之前的RGB值，a 1、a 2和a 3为预设的灰度化参数。
优选地，在对灰度化后的空车率曲线仿真图像进行二值化时，预先设置一个阈值，遍历空车率曲线仿真图像的每个像素点，将像素值超过该阈值的像素点设置为白色，否则设为黑色，从而提高空车率曲线仿真图像的处理效果。
优选地，对空车率曲线仿真图像的降噪处理包括毛刺点去除和离群点去除，考虑到空车率曲线仿真图像中曲线的毛刺点是由像素点过密造成的，需要将像素点密度降低，因此使用均值滤波来去除毛刺点，以提高提高空车率曲线仿真图像上毛刺点的处理效果。进一步优选地，使用均值滤波去除毛刺点的公式为：
Figure PCTCN2018098261-appb-000007
其中，M表示预设的滤波器窗口大小，f(w,e)为去除毛刺点之前的空车率曲线仿真图像，f‘(x,y)为去除毛刺点之后的空车率曲线仿真图像。
优选地，在对空车率曲线仿真图像进行离群点去除时，采用基于距离的异常检测算法检测空车率曲线仿真图像上的离群点，以提高离群点检测效果。进一步优选地，基于距离的异常检测算法为局部异常因子LOF算法(Local Outlier Factor)，以提高离群点检测效果。
在本发明实施例中，将停车场聚类为停车场簇，根据样本停车场的停车数据生成样本停车场的空车率曲线图像，根据停车场簇中样本停车场的空车率曲线图像对基于深度卷积的生成式对抗网络进行训练，以通过训练好的基于深度卷积的生成式对抗网络生成每个停车场簇对应的空车率曲线仿真图像，对这些空车率曲线仿真图像进行处理并映射为对应的一维空车率数据，以根据这些一维空间率数据对每个停车场簇中的待修补停车场进行停车数据修补，从而在不依赖待修补停车场的先验知识时对待修补停车场进行较为精准的数据修补，实现城市级的停车数据修补，进而节约了停车场的数据采集成本。
实施例二：
图2示出了本发明实施例三提供的基于DCGAN的停车场数据修补装置的结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
采集数据获取单元21，用于获取预先采集的城市停车场集合和城市地理兴趣点集合，城市停车场集合中的停车场包括样本停车场和待修补停车场。
在本发明实施例中，预先采集的城市停车场集合中包含城市中的多个停车场和每个停车场的相关信息，这些停车场包括待修补停车场和样本停车场。待修补停车场的相关信息可包括待修补停车场在城市中的地理位置和待修补停车场的规模，样本停车场的相关信息可包括样本停车场在城市中的地理位置、样本停车场的规模以及样本停车场的停车数据。
在本发明实施例中，预先采集的城市地理兴趣点集合中包括城市中地理兴趣点(POI)的地理位置，可通过爬虫程序在地图网站上获取这些信息，地理位置可为经纬度信息。
数据处理单元22，用于根据城市地理兴趣点集合将停车场聚类为多个停车场簇，并根据样本停车场的停车数据生成样本停车场的空车率曲线图像。
在本发明实施例中，在生活中人们往往会寻找离目的地最近的停车场，如果最近的停车场没有停车位才会考虑距离远一点的停车场，但最终选择的停车场和目的地的距离不会无限远，即存在一个人们可接受的最远距离，在此将该 最远距离称为容忍度半径。因此，以停车场为圆心、且以容忍度半径为圆半径形成一个圆，在该圆内的POI都将对停车场的停车数据产生影响，圆外的POI对停车场停车数据产生的影响可忽略不计，即可认为周围POI类型数量相似的两个停车场停车数据也很可能相似。
优选地，在根据城市地理兴趣点集合将停车场聚类为多个停车场簇时，预先设置容忍度半径的值，以停车场为圆心、且以容忍度半径为圆半径形成一个圆，从城市地理兴趣点集合中获取地理位置出现在该圆内的POI，这些POI即停车场周围的POI，根据每个停车场周围的POI构建每个停车场对应的空间特征，若当前城市中主要的POI有n种，则每个停车场对应的空间特征为一个n维的特征向量，根据每个停车场对应的空间特征和预设的基于密度的空间聚类算法(DBSCAN)，对城市停车场集合中所有停车场进行聚类，得到多个簇，从而有效地提高了停车场的分类效果。为了便于描述，在此将聚类得到的簇称为停车场簇。
进一步优选地，在根据每个停车场对应的空间特征和DBSCAN对城市停车场集合中所有停车场进行聚类时，根据所有停车场的空间特征所形成数据集的数据规模确定DBSCAN中的核心点阈值MinPts，根据从城市停车场集合中随机选取的停车场之间欧式距离的均值，确定DBSCAN中密度的领域半径ε，从而有效提高停车场的分类效果。
在本发明实施例中，在根据样本停车场的停车数据生成样本停车场的空车率曲线图像时，样本停车场的停车数据为车辆进出样本停车场的准确时间信息，依据车辆进出样本停车场的准确时间信息可得到样本停车场在不同时刻空闲的停车位数量，考虑到不同停车场的规模可能不同，可对样本停车场在不同时刻空闲的停车位数量进行归一化处理，得到样本停车场不同时刻的空车率。
在本发明实施例中，根据预先设置的曲线参数和采样点数目，计算样本停车场不同时刻的空车率对应的二维坐标，由这些二维坐标构成样本停车场对应的空车率曲线，进而得到样本停车场的空车率曲线图像。
优选地，样本停车场时刻t p的空车率为
Figure PCTCN2018098261-appb-000008
样本停车场时刻t p的空车率对应的二维坐标为
Figure PCTCN2018098261-appb-000009
Figure PCTCN2018098261-appb-000010
Figure PCTCN2018098261-appb-000011
其中，样本停车场的时间序列表示为
Figure PCTCN2018098261-appb-000012
Figure PCTCN2018098261-appb-000013
为样本停车场在时刻t p的空闲车位数量，Total为样本停车场中停车位总数，L为曲线参数中的曲线长度，H为曲线参数中的曲线高度，J为采样点数目。
优选地，由于聚类后同一停车场簇内不同停车场的POI一定是相似的，但聚类后同一停车场簇内不同停车场的停车数据不一定相似，因此在聚类结束后通过皮尔逊相关系数计算同一停车场簇中不同样本停车场空车率曲线的相似程度，以判断同一停车场簇中不同停车场空车率曲线是否相似。
网络构建训练单元23，用于构建基于深度卷积的生成式对抗网络，根据停车场簇中样本停车场的空车率曲线图像，对基于深度卷积的生成式对抗网络进行训练，以生成每个停车场簇对应的空车率曲线仿真图像。
在本发明实施例中，根据预设的有监督的卷积神经网络(CNN)对预设的生成式对抗网络(GAN)进行扩展，以得到基于深度卷积的生成式对抗网络(DCGAN)在此对有监督的卷积神经网络和生成式对抗网络的网络层数、以及每层的单元数目不进行限定。构建的DCGAN相较于扩展前的GAN，模型结构上的主要变化有：池化层用卷积层替代，其中，鉴别器中的池化层用带步长的卷积层替代，生成器中的池化层用微步幅的卷积层替代；在鉴别器和生成器上都采用批标准化；移除全连接层。优选地，在DCGAN中生成器的输出层连接着用于正态性校验的过滤器，仅输出符合正态性校验的空车率曲线仿真图像，以保证DCGAN中生成器生成的空车率曲线仿真图像的效果。进一步优选地，用于正态性校验的过滤器采用达戈斯提诺算法(D'Agostino-Pearson test)，以提高对空车率曲线仿真图像进行正态性校验的效果。
在本发明实施例中，将停车场簇中样本停车场的空车率曲线图像作为训练数据输入DCGAN中，对DCGAN进行训练，并由训练好的DCGAN生成该停 车场簇对应的空车率曲线仿真图像，从而可生成每个停车场簇对应的空车率曲线仿真图像。
停车数据修补单元24，用于对每个停车场簇对应的空车率曲线仿真图像进行处理并映射为相应的一维空车率数据，根据每个停车场簇对应的一维空车率数据，对每个停车场簇中的待修补停车场进行停车数据修补。
在本发明实施例中，在获得每个停车场簇对应的空车率曲线仿真图像后，对空车率曲线图像进行灰度化、二值化以及降噪处理等图像处理操作，将处理后的空车率曲线图像映射为相应的一维空车率数据。根据停车场簇对应的一维空车率数据和该停车场簇中待修补停车场的规模，可计算得到该停车场簇中待修补停车场不同时刻的空闲停车位数量，进而实现对每个停车场簇中每个待修补停车场的停车数据修补。在将处理后的空车率曲线图像映射为相应的一维空车率数据时，可采用上述样本停车场不同时刻的空车率到空车率曲线图像的映射关系的逆关系实现。
优选地，对空车率曲线仿真图像进行灰度化的公式为：
R 2＝G 2＝B 2＝R 1*a 1+G 1*a 2+B 1*a 3，其中，R 2、G 2和B 2为空车率曲线仿真图像灰度化之后的RGB值，R 1、G 1和B 1为空车率曲线仿真图像灰度化之前的RGB值，a 1、a 2和a 3为预设的灰度化参数。
优选地，在对灰度化后的空车率曲线仿真图像进行二值化时，预先设置一个阈值，遍历空车率曲线仿真图像的每个像素点，将像素值超过该阈值的像素点设置为白色，否则设为黑色，从而提高空车率曲线仿真图像的处理效果。
优选地，对空车率曲线仿真图像的降噪处理包括毛刺点去除和离群点去除，考虑到空车率曲线仿真图像中曲线的毛刺点是由像素点过密造成的，需要将像素点密度降低，因此使用均值滤波来去除毛刺点，以提高提高空车率曲线仿真图像上毛刺点的处理效果。进一步优选地，使用均值滤波去除毛刺点的公式为：
Figure PCTCN2018098261-appb-000014
其中，M表示预设的滤波器窗口大小，f(w,e)为去除 毛刺点之前的空车率曲线仿真图像，f‘(x,y)为去除毛刺点之后的空车率曲线仿真图像。
优选地，在对空车率曲线仿真图像进行离群点去除时，采用基于距离的异常检测算法检测空车率曲线仿真图像上的离群点，以提高离群点检测效果。进一步优选地，基于距离的异常检测算法为局部异常因子LOF算法，以提高离群点检测效果。
优选地，数据处理单元22包括：
地理兴趣点确定单元321，用于根据预设的容忍度半径在城市地理兴趣点集合中确定城市停车场集合中每个停车场周围的地理兴趣点；以及
停车场聚类单元322，用于根据每个停车场周围的地理兴趣点构建每个停车场对应的空间特征，根据每个停车场对应的空间特征和预设的基于密度的空间聚类算法，对停车场进行高维聚类。
优选地，网络构建训练单元23包括：
网络构建单元331，用于根据预设的有监督的卷积神经网络对预设的生成式对抗网络进行扩展，以构建基于深度卷积的生成式对抗网络，基于深度卷积的生成式对抗网络中的输出层连接着用于正态性校验的过滤器。
在本发明实施例中，将停车场聚类为停车场簇，根据样本停车场的停车数据生成样本停车场的空车率曲线图像，根据停车场簇中样本停车场的空车率曲线图像对基于深度卷积的生成式对抗网络进行训练，以通过训练好的基于深度卷积的生成式对抗网络生成每个停车场簇对应的空车率曲线仿真图像，对这些空车率曲线仿真图像进行处理并映射为对应的一维空车率数据，以根据这些一维空间率数据对每个停车场簇中的待修补停车场进行停车数据修补，从而在不依赖待修补停车场的先验知识时对待修补停车场进行较为精准的数据修补，实现城市级的停车数据修补，进而节约了停车场的数据采集成本。
在本发明实施例中，基于DCGAN的停车场数据修补装置的各单元可由相应的硬件或软件单元实现，各单元可以为独立的软、硬件单元，也可以集成为 一个软、硬件单元，在此不用以限制本发明。
实施例三：
图4示出了本发明实施例三提供的计算设备的结构，为了便于说明，仅示出了与本发明实施例相关的部分。
本发明实施例的计算设备4包括处理器40、存储器41以及存储在存储器41中并可在处理器40上运行的计算机程序42。该处理器40执行计算机程序42时实现上述方法实施例中的步骤，例如图1所示的步骤S101至S104。或者，处理器40执行计算机程序42时实现上述装置实施例中各单元的功能，例如图2所示单元21至24的功能。
在本发明实施例中，将停车场聚类为停车场簇，根据样本停车场的停车数据生成样本停车场的空车率曲线图像，根据停车场簇中样本停车场的空车率曲线图像对基于深度卷积的生成式对抗网络进行训练，以通过训练好的基于深度卷积的生成式对抗网络生成每个停车场簇对应的空车率曲线仿真图像，对这些空车率曲线仿真图像进行处理并映射为对应的一维空车率数据，以根据这些一维空间率数据对每个停车场簇中的待修补停车场进行停车数据修补，从而在不依赖待修补停车场的先验知识时对待修补停车场进行较为精准的数据修补，实现城市级的停车数据修补，进而节约了停车场的数据采集成本。
实施例四：
在本发明实施例中，提供了一种计算机可读存储介质，该计算机可读存储介质存储有计算机程序，该计算机程序被处理器执行时实现上述方法实施例中的步骤，例如，图1所示的步骤S101至S104。或者，该计算机程序被处理器执行时实现上述装置实施例中各单元的功能，例如图2所示单元21至24的功能。
在本发明实施例中，将停车场聚类为停车场簇，根据样本停车场的停车数据生成样本停车场的空车率曲线图像，根据停车场簇中样本停车场的空车率曲线图像对基于深度卷积的生成式对抗网络进行训练，以通过训练好的基于深度 卷积的生成式对抗网络生成每个停车场簇对应的空车率曲线仿真图像，对这些空车率曲线仿真图像进行处理并映射为对应的一维空车率数据，以根据这些一维空间率数据对每个停车场簇中的待修补停车场进行停车数据修补，从而在不依赖待修补停车场的先验知识时对待修补停车场进行较为精准的数据修补，实现城市级的停车数据修补，进而节约了停车场的数据采集成本。
本发明实施例的计算机可读存储介质可以包括能够携带计算机程序代码的任何实体或装置、记录介质，例如，ROM/RAM、磁盘、光盘、闪存等存储器。
以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种基于DCGAN的停车数据修补方法，其特征在于，所述方法包括下述步骤：
    获取预先采集的城市停车场集合和城市地理兴趣点集合，所述城市停车场集合中的停车场包括样本停车场和待修补停车场；
    根据所述城市地理兴趣点集合将所述停车场聚类为多个停车场簇，并根据所述样本停车场的停车数据生成所述样本停车场的空车率曲线图像；
    构建基于深度卷积的生成式对抗网络，根据所述停车场簇中所述样本停车场的空车率曲线图像，对所述基于深度卷积的生成式对抗网络进行训练，以生成所述每个停车场簇对应的空车率曲线仿真图像；
    对所述每个停车场簇对应的空车率曲线仿真图像进行处理并映射为相应的一维空车率数据，根据所述每个停车场簇对应的一维空车率数据，对所述每个停车场簇中的所述待修补停车场进行停车数据修补。
  2. 如权利要求1所述的方法，其特征在于，根据所述城市地理兴趣点集合将所述停车场聚类为多个停车场簇的步骤，包括：
    根据预设的容忍度半径在所述城市地理兴趣点集合中确定所述城市停车场集合中每个停车场周围的地理兴趣点；
    根据所述每个停车场周围的地理兴趣点构建所述每个停车场对应的空间特征，根据所述每个停车场对应的空间特征和预设的基于密度的空间聚类算法，对所述停车场进行高维聚类。
  3. 如权利要求1所述的方法，其特征在于，根据所述样本停车场的停车数据生成所述样本停车场的空车率曲线图像的步骤，包括：
    根据所述样本停车场的停车数据计算所述样本停车场不同时刻的空车率；
    根据预设的曲线参数和采样点数目，计算所述样本停车场不同时刻的空车率所对应的二维坐标。
  4. 如权利要求1所述的方法，其特征在于，构建基于深度卷积的生成式对 抗网络的步骤，包括：
    根据预设的有监督的卷积神经网络对预设的生成式对抗网络进行扩展，以构建所述基于深度卷积的生成式对抗网络，所述基于深度卷积的生成式对抗网络中生成器的输出层连接着用于正态性校验的过滤器。
  5. 如权利要求1所述的方法，其特征在于，对所述每个停车场簇对应的空车率曲线仿真图像进行处理并映射为相应的一维空车率数据的步骤，包括：
    对所述空车率曲线仿真图像进行灰度化、二值化以及降噪处理，所述降噪处理包括毛刺点去除和离群点去除；
    将处理后的所述空车率曲线仿真图像映射为相应的所述一维空车率数据。
  6. 一种基于DCGAN的停车数据修补装置，其特征在于，所述装置包括：
    采集数据获取单元，用于获取预先采集的城市停车场集合和城市地理兴趣点集合，所述城市停车场集合中的停车场包括样本停车场和待修补停车场；
    数据处理单元，用于根据所述城市地理兴趣点集合将所述停车场聚类为多个停车场簇，并根据所述样本停车场的停车数据生成所述样本停车场的空车率曲线图像；
    网络构建训练单元，用于构建基于深度卷积的生成式对抗网络，根据所述停车场簇中所述样本停车场的空车率曲线图像，对所述基于深度卷积的生成式对抗网络进行训练，以生成所述每个停车场簇对应的空车率曲线仿真图像；以及
    停车数据修补单元，用于对所述每个停车场簇对应的空车率曲线仿真图像进行处理并映射为相应的一维空车率数据，根据所述每个停车场簇对应的一维空车率数据，对所述每个停车场簇中的所述待修补停车场进行停车数据修补。
  7. 如权利要求6所述的装置，其特征在于，所述数据处理单元包括：
    地理兴趣点确定单元，用于根据预设的容忍度半径在所述城市地理兴趣点集合中确定所述城市停车场集合中每个停车场周围的地理兴趣点；以及
    停车场聚类单元，用于根据所述每个停车场周围的地理兴趣点构建所述每 个停车场对应的空间特征，根据所述每个停车场对应的空间特征和预设的基于密度的空间聚类算法，对所述停车场进行高维聚类。
  8. 如权利要求6所述的装置，其特征在于，所述网络构建训练单元包括：
    网络构建单元，用于根据预设的有监督的卷积神经网络对预设的生成式对抗网络进行扩展，以构建所述基于深度卷积的生成式对抗网络，所述基于深度卷积的生成式对抗网络中生成器的输出层连接着用于正态性校验的过滤器。
  9. 一种计算设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如权利要求1至5任一项所述方法的步骤。
  10. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1至5任一项所述方法的步骤。
PCT/CN2018/098261 2018-08-02 2018-08-02 基于dcgan的停车数据修补方法、装置、设备及存储介质 WO2020024206A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/098261 WO2020024206A1 (zh) 2018-08-02 2018-08-02 基于dcgan的停车数据修补方法、装置、设备及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/098261 WO2020024206A1 (zh) 2018-08-02 2018-08-02 基于dcgan的停车数据修补方法、装置、设备及存储介质

Publications (1)

Publication Number Publication Date
WO2020024206A1 true WO2020024206A1 (zh) 2020-02-06

Family

ID=69231176

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/098261 WO2020024206A1 (zh) 2018-08-02 2018-08-02 基于dcgan的停车数据修补方法、装置、设备及存储介质

Country Status (1)

Country Link
WO (1) WO2020024206A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103235933A (zh) * 2013-04-15 2013-08-07 东南大学 一种基于隐马尔科夫模型的车辆异常行为检测方法
AU2017101166A4 (en) * 2017-08-25 2017-11-02 Lai, Haodong MR A Method For Real-Time Image Style Transfer Based On Conditional Generative Adversarial Networks
CN108053454A (zh) * 2017-12-04 2018-05-18 华中科技大学 一种基于深度卷积生成对抗网络的图结构数据生成方法
CN108334941A (zh) * 2018-03-06 2018-07-27 陕西师范大学 一种基于生成式对抗网络的桥梁裂缝图像生成模型

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103235933A (zh) * 2013-04-15 2013-08-07 东南大学 一种基于隐马尔科夫模型的车辆异常行为检测方法
AU2017101166A4 (en) * 2017-08-25 2017-11-02 Lai, Haodong MR A Method For Real-Time Image Style Transfer Based On Conditional Generative Adversarial Networks
CN108053454A (zh) * 2017-12-04 2018-05-18 华中科技大学 一种基于深度卷积生成对抗网络的图结构数据生成方法
CN108334941A (zh) * 2018-03-06 2018-07-27 陕西师范大学 一种基于生成式对抗网络的桥梁裂缝图像生成模型

Similar Documents

Publication Publication Date Title
CN109214422B (zh) 基于dcgan的停车数据修补方法、装置、设备及存储介质
WO2018068653A1 (zh) 点云数据处理方法、装置及存储介质
JP2018523875A (ja) 車線認識のモデリング方法、装置、記憶媒体及び機器、並びに車線の認識方法、装置、記憶媒体及び機器
CN104700099A (zh) 识别交通标志的方法和装置
CN113947766B (zh) 一种基于卷积神经网络的实时车牌检测方法
CN111428558A (zh) 一种基于改进YOLOv3方法的车辆检测方法
CN109522831A (zh) 一种基于微卷积神经网络的车辆实时检测方法
WO2021236006A1 (en) Route deviation quantification and vehicular route learning based thereon
CN116597270A (zh) 基于注意力机制集成学习网络的道路损毁目标检测方法
CN114519819B (zh) 一种基于全局上下文感知的遥感图像目标检测方法
Camilleri et al. Detecting road potholes using computer vision techniques
CN116778146A (zh) 基于多模态数据的道路信息提取方法及装置
CN110636248B (zh) 目标跟踪方法与装置
CN112507867B (zh) 一种基于EDLines线特征的车道线检测方法
CN112329886A (zh) 双车牌识别方法、模型训练方法、装置、设备及存储介质
WO2020024206A1 (zh) 基于dcgan的停车数据修补方法、装置、设备及存储介质
CN117058069A (zh) 一种全景影像中路面表观病害自动检测方法
CN116628531A (zh) 众包地图道路对象要素聚类方法、系统及存储介质
CN112733782B (zh) 基于道路网的城市功能区识别方法、存储介质和电子设备
CN112016534B (zh) 车辆违停检测的神经网络的训练方法、检测方法和装置
CN112926482A (zh) 一种基于多尺度残差卷积神经网络的道路提取方法
Zou et al. Application of DCGAN in Data Repair of Parking Lots
Liu et al. Extracting campus’ road network from walking gps trajectories
CN116824277B (zh) 用于道路病害检测的视觉目标检测模型、构建方法及应用
CN112883840B (zh) 基于关键点检测的输电线路提取方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18928487

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS (EPO FORM 1205A DATED 12.07.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18928487

Country of ref document: EP

Kind code of ref document: A1