WO2023004595A1 - 一种停车数据修复方法、装置、计算机设备以及存储介质 - Google Patents

一种停车数据修复方法、装置、计算机设备以及存储介质 Download PDF

Info

Publication number
WO2023004595A1
WO2023004595A1 PCT/CN2021/108734 CN2021108734W WO2023004595A1 WO 2023004595 A1 WO2023004595 A1 WO 2023004595A1 CN 2021108734 W CN2021108734 W CN 2021108734W WO 2023004595 A1 WO2023004595 A1 WO 2023004595A1
Authority
WO
WIPO (PCT)
Prior art keywords
parking lot
parking
initial
space occupancy
occupancy information
Prior art date
Application number
PCT/CN2021/108734
Other languages
English (en)
French (fr)
Inventor
彭磊
吴伟伟
Original Assignee
中国科学院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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to PCT/CN2021/108734 priority Critical patent/WO2023004595A1/zh
Publication of WO2023004595A1 publication Critical patent/WO2023004595A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/14—Traffic control systems for road vehicles indicating individual free spaces in parking areas

Definitions

  • the present application relates to the technical field of the Internet, and in particular to a parking data restoration method, device, computer equipment and storage medium.
  • the city-wide parking guidance system (City-wide Parking Guidance System, CPGS) is an important means to alleviate the deteriorating parking problem in the urban area, and the parking space prediction is the key function of the system, and the training of its prediction model A large amount of historical parking data is required.
  • parking data restoration generation
  • the current data restoration method does not consider the impact of other related parking lot data in the geographic space on the target parking lot for the parking lot data that is geographically dependent, resulting in the similarity between the data generated by the existing data restoration method and the real data. Not enough, so that it affects the accuracy of the parking guidance system to predict the occupancy information of the parking space.
  • Embodiments of the present application provide a parking data restoration method, device, computer equipment, and storage medium, which can improve the accuracy of the parking guidance system in predicting parking space occupancy information.
  • Embodiments of the present application provide a method for repairing parking data on the one hand, which may include:
  • Determining a target parking lot determining a target area according to the target parking lot, and obtaining a parking lot topology map of a parking lot in the target area;
  • the historical parking space occupancy information of the target parking lot is generated by using the confrontational neural network; the historical parking space occupancy information is used to train a parking prediction system, and the parking prediction system is used to predict the parking space occupancy information of the parking lot.
  • the determining the target parking lot, determining the target area according to the target parking lot, and obtaining the parking lot topology map of the parking lot in the target area includes:
  • the target parking lot Determining the target parking lot, with the target parking lot as the center, and the area where the preset distance threshold is the radius is determined as the target area; the target area includes other parking lots except the target parking lot;
  • a parking lot topology map of parking lots in the target area is generated according to the adjacency matrix and the attribute matrix.
  • the statistics of the parking lot information in the target area, and determining the adjacency matrix and attribute matrix corresponding to the parking lot in the target area according to the parking lot information include:
  • the parking lot information including the location information and attribute information of the parking lot
  • the attribute information is converted into a standard value, and an attribute matrix corresponding to the parking lot is generated according to the standard value.
  • the acquiring the parking space occupancy information of the parking lot in the target area during the target time period includes:
  • the parking space occupancy information of the parking lot in the target time period is determined.
  • the constructing an adversarial neural network according to the parking lot topology map and the parking space occupancy information includes:
  • an initial confrontational neural network which includes an initial discriminator and an initial generator
  • a loss function of the initial adversarial neural network is generated by Wasserstein distance, and an adversarial neural network is generated by adjusting the initial adversarial neural network through the loss function.
  • both the initial discriminator and the initial generator include a graph convolution module, a space-time fusion module and a gated loop module;
  • the generating synthetic parking space occupancy information of the parking lot through the topology map of the parking lot and the initial generator includes:
  • the adjacency matrix and the attribute matrix in the parking lot topology map are input into the graph convolution module of the initial generator to generate the first graph feature matrix;
  • the first image feature matrix is fused with the initial noise to generate a first fusion matrix
  • Synthetic parking space occupancy information of the parking lot is generated through the first fusion matrix and the gated loop module of the initial generator.
  • the generation of the discrimination result of the training parking space occupancy information through the parking lot topology map, the training parking space occupancy information and the initial discriminator includes:
  • the second image feature matrix and the training parking space occupancy information are fused to generate a second fusion matrix
  • a discrimination result of the training parking space occupancy information is generated through the second fusion matrix and the gating cycle module of the initial discriminator.
  • the loss function of the initial confrontation neural network is generated through the Wasserstein distance, and the initial confrontation neural network is adjusted through the loss function Neural Network Generative Adversarial Neural Networks, including:
  • the model parameters of the initial confrontation neural network include the model parameters of the initial discriminator and the model parameters of the initial generator;
  • the initial adversarial neural network including the adjusted model parameters is determined as the adversarial neural network.
  • Embodiments of the present application provide a parking data restoration device on the one hand, which may include:
  • a topological map acquisition unit configured to determine a target parking lot, determine a target area according to the target parking lot, and obtain a parking lot topology map of a parking lot in the target area;
  • An information acquisition unit configured to acquire the parking space occupancy information of the parking lot in the target area during the target time period
  • a network construction unit configured to construct an adversarial neural network according to the parking lot topology map and the parking space occupancy information
  • An information generating unit configured to generate historical parking space occupancy information of the target parking lot by using the confrontational neural network; the historical parking space occupancy information is used to train a parking prediction system, and the parking prediction system is used to predict the parking space occupancy of the parking lot information.
  • the topology map acquisition unit includes:
  • the target area determination subunit is used to determine the target parking lot, and the target parking lot is the center, and the preset distance threshold is determined as the radius area as the target area; the target area includes other parking lots except the target parking lot ;
  • a matrix generating subunit configured to count the parking lot information in the target area, and determine the adjacency matrix and attribute matrix corresponding to the parking lot in the target area according to the parking lot information;
  • the topology map acquisition subunit is configured to generate a parking lot topology map of the parking lot in the target area according to the adjacency matrix and the attribute matrix.
  • the matrix generating subunit is specifically used for:
  • the parking lot information including the location information and attribute information of the parking lot
  • the attribute information is converted into a standard value, and an attribute matrix corresponding to the parking lot is generated according to the standard value.
  • the information acquisition unit is specifically configured to:
  • the parking space occupancy information of the parking lot in the target time period is determined.
  • the network construction unit includes:
  • the synthetic information generation subunit is used to obtain an initial confrontational neural network, the initial confrontational neural network includes an initial discriminator and an initial generator; through the parking lot topology and the initial generator, the synthetic parking space occupancy of the parking lot is generated information;
  • the discrimination result generating subunit is used to use the synthesized parking space occupancy information and the parking space occupancy information as training parking space occupancy information; through the parking lot topology map, the training parking space occupancy information and the initial discriminator, generate training Discrimination results of parking space occupancy information;
  • the network construction subunit is used to generate the loss function of the initial confrontation neural network through the Wasserstein distance based on the synthetic parking space occupancy information and the discrimination result, and adjust the initial confrontation neural network through the loss function to generate a confrontation Neural Networks.
  • both the initial discriminator and the initial generator include a graph convolution module, a space-time fusion module and a gated loop module;
  • the synthetic information generating subunit is specifically used for:
  • the adjacency matrix and the attribute matrix in the parking lot topology map are input into the graph convolution module of the initial generator to generate the first graph feature matrix;
  • the first image feature matrix is fused with the initial noise to generate a first fusion matrix
  • Synthetic parking space occupancy information of the parking lot is generated through the first fusion matrix and the gated loop module of the initial generator.
  • the discrimination result generating subunit is specifically used for:
  • the second image feature matrix and the training parking space occupancy information are fused to generate a second fusion matrix
  • a discrimination result of the training parking space occupancy information is generated through the second fusion matrix and the gating cycle module of the initial discriminator.
  • the network construction subunit is specifically used for:
  • the model parameters of the initial confrontation neural network include the model parameters of the initial discriminator and the model parameters of the initial generator;
  • the initial adversarial neural network including the adjusted model parameters is determined as the adversarial neural network.
  • Embodiments of the present application provide, on the one hand, a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and the computer program is adapted to be loaded by a processor to execute the above-mentioned method steps.
  • An embodiment of the present application provides a computer device, including: a processor, a memory, and a network interface; the processor is connected to the memory and the network interface, wherein the network interface is used to provide a network communication function , the memory is used to store program codes, and the processor is used to call the program codes to execute the above method steps.
  • An embodiment of the present application provides a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above-mentioned method steps.
  • the target area is determined according to the target parking lot, the parking lot topology map of the parking lot in the target area is obtained, and the parking lot in the target area is further obtained at the target time.
  • the information is used to train a parking prediction system for predicting occupancy information of parking lots.
  • the historical parking space occupancy information of the target parking lot generated by the adversarial neural network has a very high similarity with the real parking space occupancy information , which improves the accuracy of the parking guidance system in predicting the occupancy information of parking spaces.
  • Fig. 1 is a system architecture diagram of a kind of parking data restoration provided by the embodiment of the present application
  • Fig. 2 is a schematic flow chart of a method for repairing parking data provided by an embodiment of the present application
  • Fig. 3 is a schematic diagram of an example of determining a target area provided by an embodiment of the present application.
  • Fig. 4 is a schematic diagram of an example of a method for repairing parking data provided by an embodiment of the present application
  • Fig. 5 is a schematic flow chart of a method for repairing parking data provided by an embodiment of the present application
  • FIG. 6 is a schematic structural diagram of an adversarial neural network provided by an embodiment of the present application.
  • Fig. 7 is an exemplary schematic diagram of a cosine similarity comparison provided by an embodiment of the present application.
  • Fig. 8 is a schematic structural diagram of a parking data restoration device provided by an embodiment of the present application.
  • FIG. 9 is a schematic structural diagram of a computer device provided by an embodiment of the present application.
  • the network architecture diagram may include a service server 100 and a user terminal cluster
  • the user terminal cluster may include a user terminal 10a, a user terminal 10b, ..., a user terminal 10c, wherein there may be communication between the user terminal clusters connection, for example, there is a communication connection between the user terminal 10a and the user terminal 10b, there is a communication connection between the user terminal 10b and the user terminal 10c, and any user terminal in the user terminal cluster can have a communication connection with the service server 100, for example, the user There is a communication connection between the terminal 10 a and the service server 100 , and there is a communication connection between the user terminal 10 b and the service server 100 .
  • the above-mentioned user terminal cluster (also including the above-mentioned user terminal 10a, user terminal 10b, and user terminal 10c) may be integrated with target applications installed.
  • the target application may include an application with a function of displaying data information such as text, image, and video.
  • the user terminal determines the target parking lot, determines the target area according to the target parking lot, acquires the parking lot topology map of the parking lot in the target area, and further acquires the parking spaces of the parking lot in the target area in the target time period Occupancy information, and according to the parking lot topology map and the parking space occupancy information, construct an adversarial neural network, finally use the adversarial neural network to generate the historical parking space occupancy information of the target parking lot, and use the historical parking space occupancy information to train parking prediction
  • the system can further predict the parking space occupancy information of the parking lot according to the parking prediction system.
  • the above user terminal may be any user terminal selected from the user terminal cluster in the above embodiment corresponding to FIG. 1 , for example, the user terminal may be the above user terminal 10b.
  • the method provided in the embodiment of the present application can be executed by a computer device, which includes but not limited to a terminal or a server.
  • the service server 100 in the embodiment of the present application can be a computer device, a user terminal in a user terminal cluster It can also be computer equipment, which is not limited here.
  • the above-mentioned business server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, and can also provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication , middleware services, domain name services, security services, CDN, and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms.
  • the above-mentioned terminals may include: smart phones, tablet computers, notebook computers, desktop computers, smart TVs, smart speakers, desktop computers, smart watches and other smart terminals with image recognition functions, but are not limited thereto.
  • the user terminal and the service server may be connected directly or indirectly through wired or wireless communication, which is not limited in this application.
  • FIG. 2 is a schematic flowchart of a parking data restoration method provided by an embodiment of the present application.
  • the method may be executed by a user terminal (eg, the user terminal shown in FIG. 1 ), or jointly executed by the user terminal and a service server (such as the service server 100 in the embodiment corresponding to FIG. 1 ).
  • this embodiment takes the method executed by the above-mentioned user terminal as an example for description.
  • the parking data restoration method may at least include the following steps S101-step S104:
  • S101 Determine a target parking lot, determine a target area according to the target parking lot, and obtain a parking lot topology map of a parking lot in the target area;
  • the user terminal may determine a target parking lot, where the target parking lot is a parking lot whose parking space occupancy information needs to be repaired, and determine a target area according to the target parking lot.
  • the specific process of determining the target area is as follows.
  • the user terminal determines the target parking lot, and determines an area with the target parking lot as the center and a preset distance threshold as the radius as the target area, and the target area includes other parking lots except the target parking lot.
  • Fig. 3 is an example schematic diagram of determining the target area provided by the embodiment of the present application.
  • the parking lot is the center, and the area with the preset distance threshold as the radius is determined as the target area.
  • the distance threshold can be set to 500 meters. There are 6 parking lots in the target area in the figure. It should be noted that other methods may also be used to determine the target area. For example, after the target parking lot is determined, the area formed by the target parking lot and nearby parking lots with the shortest distance less than 500 meters may be determined as the target area.
  • the user terminal collects information on parking lots in the target area, and determines an adjacency matrix and an attribute matrix corresponding to the parking lots in the target area according to the parking lot information.
  • the user terminal collects parking lot information in the target area, the parking lot information includes location information and attribute information of the parking lot, and the location information is coordinate information of the parking lot, which represents the actual location of the parking lot.
  • the attribute information is the own attribute of the parking lot, which represents the actual characteristics of the parking lot.
  • the attribute information of the parking lot can include the total parking spaces of the parking lot, the parking rate, the main service type of the parking lot (public service, resident, business or office), etc.
  • the specific process of obtaining the adjacency matrix corresponding to the parking lot is as follows: the user terminal determines the shortest path between the parking lots according to the location information, and generates the adjacency matrix corresponding to the parking lot through the shortest path.
  • the shortest path is the real-world distance of the parking lot. Please refer to Fig. 3.
  • the target area includes 6 parking lots, which are respectively marked as P 1 , P 2 , ... P 6 .
  • Each parking lot is regarded as a node, and the shortest path between two parking lot nodes is calculated to generate The adjacency matrix E of the parking lot, where E is an n*n matrix, n is the number of parking lots, and the element E ij in E represents the shortest path between the parking lot P i and the parking lot P j .
  • the specific process of obtaining the attribute matrix corresponding to the parking lot is as follows: the user terminal converts the attribute information into a standard value, and generates the attribute matrix corresponding to the parking lot according to the standard value.
  • the standard value is a preset value range, for example, a number between 0-1.
  • Each parking lot in the target area may have a plurality of attribute information, but attributes related to factors of vehicle parking are: mainly service range, capacity, and price.
  • the service area indicates which vehicles are allowed to park there. For example, the parking lot of a large shopping center is open to all vehicles, while the residential parking lot only serves owners.
  • a factor can be used to represent the service scope of the parking lot. The larger the factor, the higher the service capacity.
  • the capacity corresponds to the parking spaces in the parking lot.
  • the user terminal generates a parking lot topology map of parking lots in the target area according to the adjacency matrix and the attribute matrix.
  • the adjacency matrix and attribute matrix of the parking lot are E and V respectively, then the topology map of the parking lot can be expressed as G(V,E).
  • the user terminal counts the number of vehicles in the parking lot in the target time period according to the vehicle counting rule.
  • the vehicle counting rule is to start counting from the time stamp when the first vehicle enters and exits the parking lot.
  • the initial value of the count is 0.
  • Subtract one add one to the count of each vehicle, and keep the count history of each time, or keep the average value at a fixed time interval, for example, take the average value every 5 minutes.
  • the number of vehicles is normalized to generate the vehicle occupancy of the parking lot, and the maximum value of the vehicle occupancy is taken as the full occupancy of the parking lot. Specifically, traverse all the vehicle data in the parking lot to obtain the historical minimum value of the vehicle data.
  • the historical minimum value can be positive, zero or negative, and normalize the number of vehicles according to the historical minimum value to generate the vehicle occupancy of the parking lot. quantity.
  • the historical minimum value is a negative number
  • add this negative number to each vehicle data in all counting histories, and if the historical minimum value is positive, then subtract this negative number from each vehicle data in all counting histories, that is, The historical minimum value is corrected to 0, and at the same time, the maximum value of the vehicle occupancy after normalization processing can also be obtained, and the maximum value of the vehicle occupancy is taken as the full occupancy of the parking lot.
  • the user terminal determines the parking space occupancy information of the parking lot in the target time period according to the full occupancy amount and the vehicle occupancy amount at each moment in the target time period. Specifically, the vehicle occupancy at each moment in the target time period is divided by the full occupancy to obtain the parking space occupancy information of the parking lot in the target time period.
  • the topological map and parking space occupancy information of the target area can be expressed as a space-time tensor.
  • STT space-time tensor
  • T is the time series data of the parking lot, expressed as parking space occupancy information.
  • T is a two-dimensional vector, the first dimension represents the ID of the time, and the second dimension represents the ID of the parking lot, that is, T ij represents the parking space occupancy information of the j-th parking lot at the i-th time, and T i represents the parking space at the i-th time Occupancy information for all car parks over time.
  • G(V,E) is the topological map of the parking lot in the target area
  • E is the adjacency matrix of the parking lot
  • V is the attribute matrix of the parking lot.
  • the user terminal obtains an initial adversarial neural network, and the initial adversarial neural network includes an initial discriminator and an initial generator. Generate synthetic parking space occupancy information of the parking lot through the topological map of the parking lot and the initial generator. Specifically, the user terminal inputs the adjacency matrix and attribute matrix in the topological map of the parking lot into the initial generator, and extracts the features and fusion to generate synthetic parking space occupancy information, which is untrue parking space occupancy information.
  • the user terminal uses the synthesized parking space occupancy information and the parking space occupancy information as the training parking space occupancy information, generating the discrimination of the training parking space occupancy information through the parking lot topology map, the training parking space occupancy information and the initial discriminator result.
  • the user terminal inputs the adjacency matrix and attribute matrix in the parking lot topology map into the initial discriminator, and after feature extraction and fusion, generates a discriminant result of training parking space occupancy information, and the discriminant result is scalar data, used It is used for judging whether the training parking space occupancy information is true or not, and whether the training parking space occupancy information is synthetic parking space occupancy information or parking space occupancy information.
  • a loss function of the initial adversarial neural network is generated by Wasserstein distance, and the initial adversarial neural network is adjusted by the loss function to generate an adversarial neural network.
  • Wasserstein distance is an index to measure the similarity of two distributions, the smaller the Wasserstein distance, the higher the similarity.
  • the user terminal can use the adversarial neural network to generate the historical parking space occupancy information of the target parking lot to repair the missing parking space occupancy information due to equipment failure or other reasons, and further use the historical parking space occupancy information to train Parking prediction system, the parking prediction system can predict the parking space occupancy information of the parking lot.
  • Fig. 4 is an example schematic diagram of the parking data restoration method provided by the embodiment of the present application. As shown in Fig. 4, first determine the target parking lot where the parking space occupancy information needs to be repaired, and then obtain the location of the target parking lot. Parking topology map of the target area.
  • the parking space occupancy information of the parking lot in the target area in the target time period is counted, and the initial confrontation neural network is trained according to the parking lot topology map and parking space occupancy information, and the parameters of the network are adjusted according to the loss function of the initial confrontation neural network until the initial Convergence of adversarial neural networks, generation of adversarial neural networks. Finally, the adversarial neural network is used to generate historical parking space occupancy information of the target parking lot.
  • the target area is determined according to the target parking lot, the parking lot topology map of the parking lot in the target area is obtained, and the parking lot in the target area is further obtained at the target time.
  • the information is used to train a parking prediction system for predicting occupancy information of parking lots.
  • the historical parking space occupancy information of the target parking lot generated by the adversarial neural network has a very high similarity with the real parking space occupancy information , which improves the accuracy of the parking guidance system in predicting the occupancy information of parking spaces.
  • FIG. 5 is a schematic flowchart of a parking data repair method provided in an embodiment of the present application
  • FIG. 6 is a schematic structural diagram of an adversarial neural network provided in an embodiment of the present application.
  • the method may be executed by a user terminal (eg, the user terminal shown in FIG. 1 ), or jointly executed by the user terminal and a service server (such as the service server 100 in the embodiment corresponding to FIG. 1 ).
  • a service server such as the service server 100 in the embodiment corresponding to FIG. 1
  • this embodiment takes the method executed by the above-mentioned user terminal as an example for description.
  • the parking data restoration method may at least include the following steps S201-step S206:
  • S201 Determine a target parking lot, determine a target area according to the target parking lot, and acquire a parking lot topology map of a parking lot in the target area;
  • step S201 in the embodiment of the present invention refer to the specific description of step S101 in the embodiment shown in FIG. 2 , and details are not repeated here.
  • step S202 in the embodiment of the present invention refer to the specific description of step S102 in the embodiment shown in FIG. 2 , and details are not repeated here.
  • S203 Acquire an initial confrontational neural network, and generate synthetic parking space occupancy information of the parking lot through the parking lot topology map and the initial generator; use the synthetic parking space occupancy information and the parking space occupancy information as training parking space occupancy information;
  • the initial adversarial neural network includes an initial discriminator D and an initial generator G, and both the initial discriminator and the initial generator include a graph convolution module, a space-time fusion module, and a gated loop module.
  • the user terminal inputs the adjacency matrix and the attribute matrix in the topological map of the parking lot into the graph convolution module of the initial generator to generate the first graph feature matrix, and the graph convolution module is a two-layer graph convolution network (Graph Convolutional Networks, GCN), a two-layer GCN model is expressed as follows:
  • V is the attribute matrix and E is the adjacency matrix.
  • E is the adjacency matrix.
  • it can be realized by adding the adjacency matrix E and the identity matrix I, namely further to to normalize, that is, in is the node degree matrix, ⁇ ( ) represents the activation function, and Relu is generally used as the activation function.
  • W 0 and W 1 are the weight matrix
  • W 0 ⁇ R N ⁇ H represents the weight matrix from the input layer to the hidden layer
  • N is the number of parking lots
  • H is the number of hidden units
  • W 1 ⁇ R H ⁇ T represents the weight matrix from the hidden layer to the weight matrix of the output layer.
  • PSN_SPE ⁇ R N ⁇ T is represented as a low-dimensional spatial feature whose length is the same as that of the parking space occupancy information of the target area.
  • the user terminal fuses the feature matrix of the first image with the initial noise to generate a first fusion matrix.
  • the initial noise is preset, for example, the initial noise may be random noise.
  • the dimension of the initial noise is the same as that of the parking space occupancy information.
  • the user terminal performs dot product fusion of the feature matrix of the first graph and the initial noise to generate the first fusion matrix.
  • the dimension of the feature matrix of the first graph is N*N
  • the dimension of the initial noise is T*N, where N is the number of graph nodes number, T is the length of the parking space occupancy information, and the dimension of the first fusion matrix is T*N.
  • the user terminal generates synthetic parking space occupancy information of the parking lot through the first fusion matrix and the gated loop module of the initial generator.
  • the first fusion matrix is input into the gated loop module of the initial generator to extract time features, and the fusion features are extracted through the two-layer fully connected layer network in the gated loop module, and finally the output length of the fully connected layer is
  • the T vector is used as synthetic parking space occupancy information, and the synthetic parking space occupancy information and the parking space occupancy information are used as training parking space occupancy information.
  • the user terminal inputs the adjacency matrix and attribute matrix in the topological graph of the parking lot into the graph convolution module of the initial discriminator to generate a second graph feature matrix.
  • the graph convolution module is a two-layer GCN, initially
  • the graph convolution module of the discriminator has the same structure as the initial generator, and the process of generating the feature matrix of the second graph is the same as that of the initial generator to generate the feature matrix of the first graph.
  • the user terminal fuses the second image feature matrix and the training parking space occupancy information to generate a second fusion matrix through the spatio-temporal fusion module of the initial discriminator. Specifically, the user terminal performs dot product fusion of the feature matrix of the second image and the occupancy information of the training parking space to generate the second fusion matrix.
  • the dimension of the feature matrix of the second image is N*N
  • the dimension of the occupancy information of the training parking space is T*N , where N is the number of graph nodes
  • T is the length of training parking space occupancy information
  • the dimension of the second fusion matrix is T*N.
  • the user terminal generates a discrimination result of the training parking space occupancy information through the second fusion matrix and the gating cycle module of the initial discriminator.
  • the second fusion matrix is input into the gated loop module of the initial discriminator to extract time features, and the fusion features are extracted through the two-layer fully connected layer network in the gated loop module, and finally the training pair is output through the fully connected layer.
  • the discrimination result of the parking space occupancy information, the judgment result is scalar data, which is used to judge whether the training parking space occupancy information is true or false, and whether the training parking space occupancy information is synthetic parking space occupancy information or parking space occupancy information.
  • the user terminal generates the first objective function of the initial generator G according to the synthetic parking space occupancy information and the Wasserstein distance, and the first objective function is Generate the second objective function of the initial discriminator D according to the discrimination result and Wasserstein distance, the second objective function is
  • the Wasserstein distance is an index to measure the similarity of two distributions. The smaller the Wasserstein distance, the higher the similarity.
  • w and ⁇ are the parameters of the initial discriminator D and the initial generator G
  • D w ( ) is the output of the discriminator
  • G w ( ) the output of the generator
  • P data and P z are the distribution space of parking space occupancy information and initial noise respectively
  • p psn_spe is the distribution space of spatial features, where x obeys P data distribution space, y obeys the distribution space of p psn_spe , and z obeys the distribution space of P z .
  • a loss function of the initial adversarial neural network is constructed based on the first objective function and the second objective function.
  • the loss function is to make the first objective function and the second objective function reach the maximum value at the same time, that is, the loss function STGAN obj can be expressed as follows:
  • the model parameters of the initial confrontation neural network include the model parameters of the initial discriminator and the model parameters of the initial generator; when the adjusted initial confrontation When the neural network meets the convergence condition, the initial adversarial neural network including the adjusted model parameters is determined as the adversarial neural network.
  • the generator and discriminator training in the existing GAN-based confrontational neural network uses the JS or KL divergence loss function, but when the JS or KL divergence loss function is used, there will be a problem of gradient disappearance when the data distribution does not cross. , leading to problems of slow training and possibly non-convergence of the model.
  • Using the loss function based on Wasserstein distance in this scheme can avoid the problem of gradient disappearance during training, and improve the training speed and convergence effect.
  • the training of the initial adversarial neural network includes the adjustment of hyperparameters, including the learning rate (LR) of the Adam optimizer, the exponential decay rate (Adam_Beta_1) of the 1st-moment estimates, and the final effective layer of the initial generator G and the initial discriminator D subsequent discard rate.
  • Hyperparameters are sensitive to specific tasks, and it is necessary to explore and try to find the appropriate hyperparameter settings.
  • LR is finally set to 2e-4
  • Adam_Beta_1 is set to 0.8 to ensure that the initial confrontation network can converge stably
  • the discard rate is set to 30% to avoid overfitting. In the experiment, a better convergence effect can be obtained after training about 2000-3000 rounds.
  • step S206 in the embodiment of the present invention refer to the specific description of step S104 in the embodiment shown in FIG. 2 , and details are not repeated here.
  • FIG. 7 is an example schematic diagram of a cosine similarity comparison provided by the embodiment of the present application.
  • the cosine similarity between the historical parking space occupancy information generated by the method in this program and the real parking space occupancy information curve 2 in the figure represents the cosine similarity between the historical parking space occupancy information generated by the method of the prior art RCGAN and the real parking space occupancy information . It can be seen from the figure that the cosine similarity between the historical parking space occupancy information generated by the method in this scheme and the real parking space occupancy information is 98%, which is significantly improved compared with 96% in the prior art.
  • the target area is determined according to the target parking lot, the parking lot topology map of the parking lot in the target area is obtained, and the parking lot in the target area is further obtained at the target time.
  • the information is used to train a parking prediction system for predicting occupancy information of parking lots.
  • the historical parking space occupancy information of the target parking lot generated by the adversarial neural network has a very high similarity with the real parking space occupancy information , which improves the accuracy of the parking guidance system in predicting the occupancy information of parking spaces.
  • FIG. 8 is a schematic structural diagram of a parking data restoration device provided by an embodiment of the present application.
  • the parking data restoration device may be a computer program (including program code) running in a computer device, for example, the parking data restoration device is an application software; step.
  • the parking data restoration device 1 of the embodiment of the present application may include: a topology map acquisition unit 11 , an information acquisition unit 12 , a network construction unit 13 , and an information generation unit 14 .
  • the topological map acquisition unit 11 is configured to determine a target parking lot, determine a target area according to the target parking lot, and obtain a parking lot topology map of a parking lot in the target area;
  • An information acquisition unit 12 configured to acquire the parking space occupancy information of the parking lot in the target area during the target time period
  • a network construction unit 13 configured to construct an adversarial neural network according to the parking lot topology map and the parking space occupancy information
  • the information generation unit 14 is used to generate the historical parking space occupancy information of the target parking lot by using the confrontation neural network; the historical parking space occupancy information is used to train a parking prediction system, and the parking prediction system is used to predict the parking spaces of the parking lot occupancy information.
  • the topological map acquisition unit 11 of the embodiment of the present application may include: a target area determination subunit 111, a matrix generation subunit 112, and a topological map acquisition subunit 113.
  • the target area determination subunit 111 is used to determine the target parking lot, and the target parking lot is the center, and the preset distance threshold is the radius area determined as the target area; the target area includes other parking areas except the target parking lot field;
  • the matrix generation subunit 112 is used to count the parking lot information in the target area, and determine the adjacency matrix and attribute matrix corresponding to the parking lot in the target area according to the parking lot information;
  • the topological map acquisition subunit 113 is configured to generate a parking lot topological map of a parking lot in the target area according to the adjacency matrix and the attribute matrix.
  • the matrix generating subunit 112 is specifically configured to:
  • the parking lot information including the location information and attribute information of the parking lot
  • the attribute information is converted into a standard value, and an attribute matrix corresponding to the parking lot is generated according to the standard value.
  • the information acquiring unit 12 is specifically configured to:
  • the parking space occupancy information of the parking lot in the target time period is determined.
  • the network construction unit 13 of the embodiment of the present application may include: a synthetic information generation subunit 131 , a discrimination result generation subunit 132 , and a network construction subunit 133 .
  • the synthetic information generation subunit 131 is used to obtain an initial confrontation neural network, which includes an initial discriminator and an initial generator; through the parking lot topology map and the initial generator, generate a synthetic parking space in the parking lot occupancy information;
  • the discrimination result generation subunit 132 is used to use the synthetic parking space occupancy information and the parking space occupancy information as training parking space occupancy information; through the parking lot topology map, the training parking space occupancy information and the initial discriminator, generate Discrimination results of training parking space occupancy information;
  • the network construction subunit 133 is used to generate the loss function of the initial confrontational neural network through the Wasserstein distance based on the synthetic parking space occupancy information and the discrimination result, and adjust the initial confrontational neural network through the loss function to generate against neural networks.
  • both the initial discriminator and the initial generator include a graph convolution module, a space-time fusion module and a gated loop module;
  • the combined information generating subunit 131 is specifically used for:
  • the adjacency matrix and the attribute matrix in the parking lot topology map are input into the graph convolution module of the initial generator to generate the first graph feature matrix;
  • the first image feature matrix is fused with the initial noise to generate a first fusion matrix
  • Synthetic parking space occupancy information of the parking lot is generated through the first fusion matrix and the gated loop module of the initial generator.
  • the discrimination result generation subunit 132 is specifically configured to:
  • the second image feature matrix and the training parking space occupancy information are fused to generate a second fusion matrix
  • a discrimination result of the training parking space occupancy information is generated through the second fusion matrix and the gating cycle module of the initial discriminator.
  • the network construction subunit 133 is specifically configured to:
  • the model parameters of the initial confrontation neural network include the model parameters of the initial discriminator and the model parameters of the initial generator;
  • the initial adversarial neural network including the adjusted model parameters is determined as the adversarial neural network.
  • the target area is determined according to the target parking lot, the parking lot topology map of the parking lot in the target area is obtained, and the parking lot in the target area is further obtained at the target time.
  • the information is used to train a parking prediction system for predicting occupancy information of parking lots.
  • the historical parking space occupancy information of the target parking lot generated by the adversarial neural network has a very high similarity with the real parking space occupancy information , which improves the accuracy of the parking guidance system in predicting the occupancy information of parking spaces.
  • the computer device 1000 may include: at least one processor 1001 , such as a CPU, at least one network interface 1004 , user interface 1003 , memory 1005 , and at least one communication bus 1002 .
  • the communication bus 1002 is used to realize connection and communication between these components.
  • the user interface 1003 may include a display screen (Display), and the optional user interface 1003 may also include a standard wired interface and a wireless interface.
  • the network interface 1004 may include a standard wired interface and a wireless interface (such as a WI-FI interface).
  • the memory 1005 may be a random access memory (Random Access Memory, RAM), or a non-volatile memory (non-volatile memory, NVM), such as at least one disk memory.
  • RAM Random Access Memory
  • NVM non-volatile memory
  • the memory 1005 may also be at least one storage device located away from the aforementioned processor 1001 .
  • the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a data processing application program.
  • the network interface 1004 can provide a network communication function
  • the user interface 1003 is mainly used to provide an input interface for the user
  • the processor 1001 can be used to call the data processing application program stored in the memory 1005 , so as to realize the description of the parking data restoration method in any one of the above-mentioned embodiments corresponding to FIG. 2-FIG. 7, which will not be repeated here.
  • the computer device 1000 described in the embodiment of the present application can execute the description of the parking data restoration method in any one of the embodiments corresponding to Figure 2- Figure 7 above, and can also execute the description in the embodiment corresponding to Figure 8 above.
  • the description of the parking data restoration device will not be repeated here.
  • the description of the beneficial effect of adopting the same method will not be repeated here.
  • the embodiment of the present application also provides a computer-readable storage medium, and the computer-readable storage medium stores the computer program executed by the aforementioned parking data recovery device, and the above-mentioned
  • the computer program includes program instructions, and when the processor executes the program instructions, it can execute the description of the parking data repair method in any one of the embodiments corresponding to Figure 2- Figure 7 above, so it will not be repeated here to repeat.
  • the description of the beneficial effect of adopting the same method will not be repeated here.
  • program instructions may be deployed to execute on one computing device, or on multiple computing devices located at one site, or, alternatively, on multiple computing devices distributed across multiple sites and interconnected by a communication network
  • program instructions may be deployed to execute on one computing device, or on multiple computing devices located at one site, or, alternatively, on multiple computing devices distributed across multiple sites and interconnected by a communication network
  • multiple computing devices distributed in multiple locations and interconnected by a communication network can form a blockchain system.
  • the above-mentioned computer-readable storage medium may be a parking data restoration device provided in any one of the foregoing embodiments or an internal storage unit of the above-mentioned device, such as a hard disk or a memory of an electronic device.
  • the computer-readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk equipped on the electronic device, a smart memory card (smart media card, SMC), a secure digital (secure digital, SD) card, Flash card (flash card), etc.
  • the above-mentioned computer-readable storage medium may also include a magnetic disk, an optical disk, a read-only memory (read-only memory, ROM) or a random access memory, and the like.
  • the computer-readable storage medium may also include both an internal storage unit of the electronic device and an external storage device.
  • the computer-readable storage medium is used to store the computer program and other programs and quantities required by the electronic device.
  • the computer-readable storage medium can also be used to temporarily store data that has been output or will be output.
  • Each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may physically exist separately, or two or more units may be integrated into one unit.
  • the above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Traffic Control Systems (AREA)

Abstract

一种停车数据修复方法、装置、计算机设备以及存储介质，其中方法包括如下步骤：确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图（S101）；获取所述目标区域中停车场在目标时间段的车位占用信息（S102）；根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络（S103）；采用所述对抗神经网络生成所述目标停车场的历史车位占用信息（S104）；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。采用该方法，可以提高停车诱导系统预测车位占用信息的准确性。

Description

一种停车数据修复方法、装置、计算机设备以及存储介质 技术领域
本申请涉及互联网技术领域，尤其涉及一种停车数据修复方法、装置、计算机设备以及存储介质。
背景技术
随着我国经济的飞速发展和现代化进程的加快，汽车持有量和出行次数持续增加。汽车保有量的持续增长，导致了道路交通拥堵，也带来了停车困难的问题。近年来在智能交通领域展开的深入研究和实践已经证实了其在缓解交通拥堵、提高道路通行能力和服务水平方面的优越性。交通预测是智能交通的重要组成部分，它可以帮助管理者提前了解交通变化，从而制定相应的管控策略。
目前，城市级停车诱导系统(City-wide Parking Guidance System，CPGS)是一种用于缓解市区日益恶化的停车问题的重要手段，而停车位预测作为该系统的关键功能，其预测模型的训练需要大量的历史停车数据。但是，由于设备故障或其他原因导致的停车数据丢失无法避免，不完整的数据会影响预测模型的训练精度。因此，停车数据修复(生成)对于构建高度可靠的停车诱导系统非常必要。但目前的数据修复方法对于具有地理空间依赖的停车场数据来说，没有考虑地理空间内其他相关停车场数据对目标停车场的影响，导致现有的数据修复方法生成的数据与真实数据相似度不够，以至于影响了停车诱导系统预测车位占用信息的准确性。
发明内容
本申请实施例提供一种停车数据修复方法、装置、计算机设备以及存储介质，可以提高停车诱导系统预测车位占用信息的准确性。
本申请实施例一方面提供了一种停车数据修复方法，可包括：
确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图；
获取所述目标区域中停车场在目标时间段的车位占用信息；
根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络；
采用所述对抗神经网络生成所述目标停车场的历史车位占用信息；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。
在一种可行的实施方式中，所述确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图，包括：
确定目标停车场，将所述目标停车场为中心，预设的距离阈值为半径的区域确定为目标区域；所述目标区域包括除目标停车场外的其他停车场；
统计所述目标区域的停车场信息，根据所述停车场信息确定所述目标区域中停车场对应的邻接矩阵和属性矩阵；
根据所述邻接矩阵和所述属性矩阵生成所述目标区域的中停车场的停车场拓扑图。
在一种可行的实施方式中，所述统计所述目标区域的停车场信息，根据所述停车场信息确定所述目标区域中停车场对应的邻接矩阵和属性矩阵，包括：
统计所述目标区域的停车场信息，所述停车场信息包括停车场的位置信息和属性信息；
根据所述位置信息确定停车场之间的最短路径，通过所述最短路径生成所述停车场对应的邻接矩阵；
将所述属性信息转换为标准数值，根据所述标准数值生成停车场对应的属性矩阵。
在一种可行的实施方式中，所述获取所述目标区域中停车场在目标时间段的车位占用信息，包括：
根据车辆计数规则，统计停车场在目标时间段的车辆数量；
对所述车辆数量进行归一化处理生成停车场的车辆占用量，将所述车辆占用量的最大值作为停车场的满占用量；
根据所述满占用量和目标时间段中各时刻的车辆占用量，确定停车场在目标时间段中的车位占用信息。
在一种可行的实施方式中，所述根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络，包括：
获取初始对抗神经网络，所述初始对抗神经网络包括初始判别器和初始生 成器；
通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息；
将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息；
通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果；
基于所述合成车位占用信息和所述判别结果，通过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络。
在一种可行的实施方式中，所述初始判别器和所述初始生成器均包括图卷积模块、时空融合模块和门控循环模块；
所述通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息，包括：
将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始生成器的图卷积模块，生成第一图特征矩阵；
基于所述初始生成器的时空融合模块，将所述第一图特征矩阵与初始噪声进行融合生成第一融合矩阵；
通过所述第一融合矩阵和初始生成器的门控循环模块，生成停车场的合成车位占用信息。
在一种可行的实施方式中，所述通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果，包括：
将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始判别器的图卷积模块，生成第二图特征矩阵；
通过所述初始判别器的时空融合模块，将所述第二图特征矩阵和所述训练车位占用信息进行融合生成第二融合矩阵；
通过所述第二融合矩阵和所述初始判别器的门控循环模块，生成所述训练车位占用信息的判别结果。
在一种可行的实施方式中，所述基于所述合成车位占用信息和所述判别结果，通过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络，包括：
根据所述合成车位占用信息和沃瑟斯坦距离生成所述初始生成器的第一目标函数，根据所述判别结果和沃瑟斯坦距离生成所述初始判别器的第二目标函数，
基于所述第一目标函数和第二目标函数构建所述初始对抗神经网络的损失函数；
根据所述损失函数调整所述初始对抗神经网络的模型参数；所述初始对抗神经网络的模型参数包括所述初始判别器的模型参数和所述初始生成器的模型参数；
当调整后的初始对抗神经网络满足收敛条件时，将包含调整后的模型参数的初始对抗神经网络确定为对抗神经网络。
本申请实施例一方面提供了一种停车数据修复装置，可包括：
拓扑图获取单元，用于确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图；
信息获取单元，用于获取所述目标区域中停车场在目标时间段的车位占用信息；
网络构建单元，用于根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络；
信息生成单元，用于采用所述对抗神经网络生成所述目标停车场的历史车位占用信息；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。
在一种可行的实施方式中，所述拓扑图获取单元，包括：
目标区域确定子单元，用于确定目标停车场，将所述目标停车场为中心，预设的距离阈值为半径的区域确定为目标区域；所述目标区域包括除目标停车场外的其他停车场；
矩阵生成子单元，用于统计所述目标区域的停车场信息，根据所述停车场信息确定所述目标区域中停车场对应的邻接矩阵和属性矩阵；
拓扑图获取子单元，用于根据所述邻接矩阵和所述属性矩阵生成所述目标区域的中停车场的停车场拓扑图。
在一种可行的实施方式中，所述矩阵生成子单元具体用于：
统计所述目标区域的停车场信息，所述停车场信息包括停车场的位置信息 和属性信息；
根据所述位置信息确定停车场之间的最短路径，通过所述最短路径生成所述停车场对应的邻接矩阵；
将所述属性信息转换为标准数值，根据所述标准数值生成停车场对应的属性矩阵。
在一种可行的实施方式中，所述信息获取单元具体用于：
根据车辆计数规则，统计停车场在目标时间段的车辆数量；
对所述车辆数量进行归一化处理生成停车场的车辆占用量，将所述车辆占用量的最大值作为停车场的满占用量；
根据所述满占用量和目标时间段中各时刻的车辆占用量，确定停车场在目标时间段中的车位占用信息。
在一种可行的实施方式中，所述网络构建单元，包括：
合成信息生成子单元，用于获取初始对抗神经网络，所述初始对抗神经网络包括初始判别器和初始生成器；通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息；
判别结果生成子单元，用于将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息；通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果；
网络构建子单元，用于基于所述合成车位占用信息和所述判别结果，通过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络。
在一种可行的实施方式中，所述初始判别器和所述初始生成器均包括图卷积模块、时空融合模块和门控循环模块；
所述合成信息生成子单元具体用于：
将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始生成器的图卷积模块，生成第一图特征矩阵；
基于所述初始生成器的时空融合模块，将所述第一图特征矩阵与初始噪声进行融合生成第一融合矩阵；
通过所述第一融合矩阵和初始生成器的门控循环模块，生成停车场的合成车位占用信息。
在一种可行的实施方式中，所述判别结果生成子单元具体用于：
将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始判别器的图卷积模块，生成第二图特征矩阵；
通过所述初始判别器的时空融合模块，将所述第二图特征矩阵和所述训练车位占用信息进行融合生成第二融合矩阵；
通过所述第二融合矩阵和所述初始判别器的门控循环模块，生成所述训练车位占用信息的判别结果。
在一种可行的实施方式中，所述网络构建子单元具体用于：
根据所述合成车位占用信息和沃瑟斯坦距离生成所述初始生成器的第一目标函数，根据所述判别结果和沃瑟斯坦距离生成所述初始判别器的第二目标函数，
基于所述第一目标函数和第二目标函数构建所述初始对抗神经网络的损失函数；
根据所述损失函数调整所述初始对抗神经网络的模型参数；所述初始对抗神经网络的模型参数包括所述初始判别器的模型参数和所述初始生成器的模型参数；
当调整后的初始对抗神经网络满足收敛条件时，将包含调整后的模型参数的初始对抗神经网络确定为对抗神经网络。
本申请实施例一方面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中存储有计算机程序，所述计算机程序适于由处理器加载并执行上述的方法步骤。
本申请实施例一方面提供了一种计算机设备，包括：处理器、存储器以及网络接口；所述处理器与所述存储器、所述网络接口相连，其中，所述网络接口用于提供网络通信功能，所述存储器用于存储程序代码，所述处理器用于调用所述程序代码执行上述的方法步骤。
本申请实施例一方面提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述的方法步骤。
在本申请实施例中，通过确定目标停车场，根据所述目标停车场确定目标 区域，获取所述目标区域的中停车场的停车场拓扑图，进一步获取所述目标区域中停车场在目标时间段的车位占用信息，并根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络，最后采用所述对抗神经网络生成所述目标停车场的历史车位占用信息，所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。采用上述方法，考虑了地理空间内其他相关停车场数据对目标停车场的影响，采用所述对抗神经网络生成所述目标停车场的历史车位占用信息与真实的车位占用信息具有极高的相似度，提高了停车诱导系统预测车位占用信息的准确性。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的一种停车数据修复的系统架构图；
图2是本申请实施例提供的一种停车数据修复方法的流程示意图；
图3是本申请实施例提供的一种确定目标区域的举例示意图；
图4是本申请实施例提供的一种停车数据修复方法的举例示意图；
图5是本申请实施例提供的一种停车数据修复方法的流程示意图；
图6是本申请实施例提供的一种对抗神经网络的结构示意图；
图7是本申请实施例提供的一种余弦相似度对比的举例示意图；
图8是本申请实施例提供的一种停车数据修复装置的结构示意图；
图9是本申请实施例提供的一种计算机设备的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
如图1所示，该网络架构图可以包括业务服务器100以及用户终端集群，该用户终端集群可以包括用户终端10a、用户终端10b、…、用户终端10c，其中，用户终端集群之间可以存在通信连接，例如用户终端10a与用户终端10b之间存在通信连接，用户终端10b与用户终端10c之间存在通信连接，且用户终端集群中的任一用户终端可以与业务服务器100存在通信连接，例如用户终端10a与业务服务器100之间存在通信连接，用户终端10b与业务服务器100之间存在通信连接。
其中，上述用户终端集群(也包括上述的用户终端10a、用户终端10b以及用户终端10c)均可以集成安装有目标应用。可选的，该目标应用可以包括具有展示文字、图像以及视频等数据信息功能的应用。具体的，用户终端确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图，进一步获取所述目标区域中停车场在目标时间段的车位占用信息，并根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络，最后采用所述对抗神经网络生成所述目标停车场的历史车位占用信息，采用历史车位占用信息训练停车预测系统，进一步可以根据停车预测系统预测停车场的车位占用信息。可选的，上述用户终端可以为在上述图1所对应实施例的用户终端集群中所选取的任意一个用户终端，比如，该用户终端可以为上述用户终端10b。
可以理解的是，本申请实施例所提供的方法可以由计算机设备执行，计算机设备包括但不限于终端或服务器，本申请实施例中的业务服务器100可以为计算机设备，用户终端集群中的用户终端也可以为计算机设备，此处不限定。上述业务服务器可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、CDN、以及大数据和人工智能平台等基础云计算服务的云服务器。上述终端可以包括：智能手机、平板电脑、笔记本电脑、桌上型电脑、智能电视、智能音箱、台式计算机、智能手表等携带图像识别功能的智能终端，但并不局限于此。其中，用户终端以及业务服务器可以通过有线或无线通信方式进行直接或间接地连接，本申请在此不做限制。
进一步地，为便于理解，请参见图2，图2是本申请实施例提供的停车数据修复方法的流程示意图。该方法可以由用户终端(例如，上述图1所示的用户终端)执行，也可以由用户终端和业务服务器(如上述图1所对应实施例中的业务服务器100)共同执行。为便于理解，本实施例以该方法由上述用户终端执行为例进行说明。其中，该停车数据修复方法至少可以包括以下步骤S101-步骤S104：
S101，确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图；
具体的，用户终端可以确定目标停车场，所述目标停车场为车位占用信息需要进行修复的停车场，并根据所述目标停车场确定目标区域。确定目标区域具体过程如下。
用户终端确定目标停车场，将所述目标停车场为中心，预设的距离阈值为半径的区域确定为目标区域，所述目标区域包括除目标停车场外的其他停车场。为便于理解，请参见图3，图3是本申请实施例提供的确定目标区域的举例示意图，如图3所示，首先将车位占用信息需要进行修复的停车场确定为目标停车场，以目标停车场为中心，预设的距离阈值为半径的区域确定为目标区域，距离阈值可以设定为500米，在图中目标区域内总共有6个停车场。需要说明的是，确定目标区域也可以采用其他方式，例如，确定目标停车场后，可以将目标停车场与附近最短距离小于500米的停车场共同形成的区域，确定为目标区域。
进一步的，用户终端统计所述目标区域的停车场信息，根据所述停车场信息确定所述目标区域中停车场对应的邻接矩阵和属性矩阵。具体的，用户终端统计所述目标区域的停车场信息，所述停车场信息包括停车场的位置信息和属性信息，所述位置信息是停车场的坐标信息，表征停车场的实际位置。所述属性信息是停车场的自有属性，表征停车场的实际特征，停车场的属性信息可以包括停车场的总停车位，停车费率，停车场主要服务类型(公共服务，居民，商业或者办公)等。
获取停车场对应的邻接矩阵的具体过程如下：用户终端根据所述位置信息确定停车场之间的最短路径，通过所述最短路径生成所述停车场对应的邻接矩阵。最短路径是停车场在真实世界中的距离。请参见图3，目标区域包括6个停 车场，分别记为P 1、P 2、...P 6，每个停车场视为一个节点，计算两个停车场节点之间的最短路径，生成停车场的邻接矩阵E，其中E为n*n矩阵，n为停车场的个数，E中的元素E ij表示停车场P i和停车场P j之间的最短路径。
获取停车场对应的属性矩阵的具体过程如下：用户终端将所述属性信息转换为标准数值，根据所述标准数值生成停车场对应的属性矩阵。标准数值为预设的取值范围，例如，0-1之间的数。目标区域中的每个停车场可以具有多个属性信息，但是与车辆停车的因素有关的属性是：主要是服务范围、容量和价格。服务范围表示可以允许将哪些车辆停放在此处。例如，大型购物中心的停车场对所有车辆开放，而住宅停车场仅为业主服务，可以采用因子来表示停车场的服务范围，因子越大，服务能力就越高。容量对应停车场的停车位，停车位越多，服务能力就越高。价格采用每小时的收费来表示。因此，可以将服务范围、容量和价格转换为标准数值，根据所述标准数值生成停车场对应的属性矩阵，则属性矩阵为三维向量，例如，属性矩阵V可以表示为V i＝[x i,y i,z i]，向量中的三维分别对应服务范围、容量和价格。
进一步的，用户终端根据所述邻接矩阵和所述属性矩阵生成所述目标区域的中停车场的停车场拓扑图。具体的，停车场的邻接矩阵和属性矩阵分别为E和V，则停车场拓扑图可以表示为G(V,E)。
S102，获取所述目标区域中停车场在目标时间段的车位占用信息；
具体的，用户终端根据车辆计数规则，统计停车场在目标时间段的车辆数量，车辆计数规则是从第一个车辆进出停车场的时间戳开始计数，计数的初始数值为0，出一辆计数减一，进一辆计数加一，并保留每一次的计数历史，也可以固定时间间距保留平均值，例如每5分钟取一次平均值。
进一步的，对所述车辆数量进行归一化处理生成停车场的车辆占用量，将所述车辆占用量的最大值作为停车场的满占用量。具体的，遍历停车场的所有车辆数据，得到车辆数据的历史最小值，历史最小值可以为正数、零或者是负数，根据历史最小值对车辆数量进行归一化处理生成停车场的车辆占用量。例如，若历史最小值为负数，则将所有计数历史的每个车辆数据都加上这个负数，若历史最小值为正数，则将所有计数历史的每个车辆数据都减去这个负数，即将历史最小值修正为0，同时，也可以得到归一化处理后的车辆占用量的最大值，将所述车辆占用量的最大值作为停车场的满占用量。
进一步的，用户终端根据所述满占用量和目标时间段中各时刻的车辆占用量，确定停车场在目标时间段中的车位占用信息。具体的，将目标时间段中各时刻的车辆占用量除以满占用量得到停车场在目标时间段中的车位占用信息。
为了能够更直接理解目标区域，可以将目标区域的拓扑图和车位占用信息表示为时空张量。假设目标区域具有n个停车场，则目标区域的时空张量(STT)定义如下：
STT＝<T,G(V,E)>
其中，T是停车场的时间序列数据，表示为车位占用信息。T是二维向量，第一个维度表示时间的ID，第二个维度表示停车场的ID，即T ij表示第i个时间的第j个停车场的车位占用信息，而T i表示在i时间所有停车场的车位占用信息。G(V,E)为目标区域的停车场拓扑图，E为停车场的邻接矩阵，V为停车场的属性矩阵。
S103，根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络；
具体的，用户终端获取初始对抗神经网络，所述初始对抗神经网络包括初始判别器和初始生成器。通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息，具体的，用户终端将停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始生成器，经过特征提取和融合，生成合成车位占用信息，合成车位占用信息是非真实的车位占用信息。
进一步的，将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息，通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果。具体的，用户终端将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始判别器，经过特征提取和融合，生成训练车位占用信息的判别结果，所述判断结果为标量数据，用于判断训练车位占用信息的真假，及判断训练车位占用信息是合成车位占用信息还是车位占用信息。
进一步的，基于所述合成车位占用信息和所述判别结果，通过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络。沃瑟斯坦距离是衡量两个分布的相似性的一个指标，沃瑟斯坦距离越小，相似度越高。
S104，采用所述对抗神经网络生成所述目标停车场的历史车位占用信息；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停 车场的车位占用信息。
具体的，用户终端可以采用所述对抗神经网络生成所述目标停车场的历史车位占用信息，用以修复因设备故障或其他原因导致缺失的车位占用信息，并进一步采用所述历史车位占用信息训练停车预测系统，停车预测系统可以预测停车场的车位占用信息。请参见图4，图4是本申请实施例提供的停车数据修复方法的举例示意图，如图4所示，首先确定车位占用信息需要进行修复的目标停车场，进而获取所述目标停车场所在的目标区域的停车场拓扑图。进一步的，统计目标区域中停车场在目标时间段的车位占用信息，并根据停车场拓扑图和车位占用信息，训练初始对抗神经网络，根据初始对抗神经网络的损失函数调整网络的参数，直到初始对抗神经网络收敛，生成对抗神经网络。最后采用所述对抗神经网络生成所述目标停车场的历史车位占用信息。
在本申请实施例中，通过确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图，进一步获取所述目标区域中停车场在目标时间段的车位占用信息，并根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络，最后采用所述对抗神经网络生成所述目标停车场的历史车位占用信息，所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。采用上述方法，考虑了地理空间内其他相关停车场数据对目标停车场的影响，采用所述对抗神经网络生成所述目标停车场的历史车位占用信息与真实的车位占用信息具有极高的相似度，提高了停车诱导系统预测车位占用信息的准确性。
请参见图5和图6，图5是本申请实施例提供的停车数据修复方法的流程示意图，图6是本申请实施例提供的对抗神经网络的结构示意图。该方法可以由用户终端(例如，上述图1所示的用户终端)执行，也可以由用户终端和业务服务器(如上述图1所对应实施例中的业务服务器100)共同执行。为便于理解，本实施例以该方法由上述用户终端执行为例进行说明。其中，该停车数据修复方法至少可以包括以下步骤S201-步骤S206：
S201，确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图；
其中，本发明实施例的步骤S201参见图2所示实施例的步骤S101的具体 描述，在此不进行赘述。
S202，获取所述目标区域中停车场在目标时间段的车位占用信息；
其中，本发明实施例的步骤S202参见图2所示实施例的步骤S102的具体描述，在此不进行赘述。
S203，获取初始对抗神经网络，通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息；将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息；
具体的，所述初始对抗神经网络包括初始判别器D和初始生成器G，所述初始判别器和所述初始生成器均包括图卷积模块、时空融合模块和门控循环模块。用户终端将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始生成器的图卷积模块，生成第一图特征矩阵，所述图卷积模块为两层图卷积网络(Graph Convolutional Networks，GCN)，一个两层GCN模型表示如下：
Figure PCTCN2021108734-appb-000001
其中，V是属性矩阵，E是邻接矩阵。为了在聚合节点特征的过程中保留自身信息，一般需要给每个节点添加自环。具体来说，可以通过邻接矩阵E和单位矩阵I相加来实现，即
Figure PCTCN2021108734-appb-000002
进一步对
Figure PCTCN2021108734-appb-000003
进行归一化处理，即
Figure PCTCN2021108734-appb-000004
其中
Figure PCTCN2021108734-appb-000005
为节点度矩阵，
Figure PCTCN2021108734-appb-000006
σ(·)代表激活函数，一般采用Relu作为激活函数。W 0和W 1是权重矩阵，W 0∈R N×H表示从输入层到隐藏层的权重矩阵，N为停车场数量，H为隐藏单元数，W 1∈R H×T表示从隐藏层到输出层的权重矩阵。最后PSN_SPE∈R N×T表示为低维空间特征，其长度与目标区域的车位占用信息的长度相同。
进一步的，用户终端基于所述初始生成器的时空融合模块，将所述第一图特征矩阵与初始噪声进行融合生成第一融合矩阵。具体的，所述初始噪声是预先设定的，例如，初始噪声可以是随机噪声。所述初始噪声的维度与车位占用信息的维度相同。用户终端将第一图特征矩阵与初始噪声进行点积融合生成第一融合矩阵，例如，第一图特征矩阵的维度为N*N，初始噪声的维度为T*N，其中N为图节点个数，T为车位占用信息的长度，第一融合矩阵的维度为T*N。
进一步的，用户终端通过所述第一融合矩阵和初始生成器的门控循环模块，生成停车场的合成车位占用信息。具体的，将所述第一融合矩阵输入初始生成 器的门控循环模块，提取时间特征，通过门控循环模块中的两层全连接层网络，提取融合特征，最后通过全连接层输出长度为T向量，作为合成车位占用信息，将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息。
S204，将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息；通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果；
具体的，用户终端将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始判别器的图卷积模块，生成第二图特征矩阵，所述图卷积模块为两层GCN，初始判别器的图卷积模块与初始生成器的结构相同，生成第二图特征矩阵的过程与初始生成器生成第一图特征矩阵的过程相同。
进一步的，用户终端通过所述初始判别器的时空融合模块，将所述第二图特征矩阵和所述训练车位占用信息进行融合生成第二融合矩阵。具体的，用户终端将第二图特征矩阵与训练车位占用信息进行点积融合生成第二融合矩阵，例如，第二图特征矩阵的维度为N*N，训练车位占用信息的维度为T*N，其中N为图节点个数，T为训练车位占用信息的长度，第二融合矩阵的维度为T*N。
进一步的，用户终端通过所述第二融合矩阵和所述初始判别器的门控循环模块，生成所述训练车位占用信息的判别结果。具体的，将所述第二融合矩阵输入初始判别器的门控循环模块，提取时间特征，通过门控循环模块中的两层全连接层网络，提取融合特征，最后通过全连接层输出对训练车位占用信息的判别结果，所述判断结果为标量数据，用于判断训练车位占用信息的真假，及判断训练车位占用信息是合成车位占用信息还是车位占用信息。
S205，基于所述合成车位占用信息和所述判别结果，通过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络。
具体的，用户终端根据所述合成车位占用信息和沃瑟斯坦距离生成所述初始生成器G的第一目标函数，第一目标函数为
Figure PCTCN2021108734-appb-000007
根据所述判别结果和沃瑟斯坦距离生成所述初始判别器D的第二目标函数，第二目标函数为
Figure PCTCN2021108734-appb-000008
沃瑟斯坦距离是衡量两个分布的相似性的一个指标，沃瑟斯坦距离越小，相似度越高，所述其中w和θ是初始判别器D和初始生成器G的参数，D w(·)是判别器的输出，G w(·)生成器的输出，P data和 P z分别是车位占用信息和初始噪声的分布空间，p psn_spe是空间特征的分布空间，其中，x服从P data的分布空间，y服从p psn_spe的分布空间，z服从P z的分布空间。
进一步的，基于所述第一目标函数和第二目标函数构建所述初始对抗神经网络的损失函数。损失函数是使第一目标函数和第二目标函数同时达到最大值，即损失函数STGAN obj可以表示为如下：
Figure PCTCN2021108734-appb-000009
根据所述损失函数调整所述初始对抗神经网络的模型参数；所述初始对抗神经网络的模型参数包括所述初始判别器的模型参数和所述初始生成器的模型参数；当调整后的初始对抗神经网络满足收敛条件时，将包含调整后的模型参数的初始对抗神经网络确定为对抗神经网络。现有的基于GAN的对抗神经网络中的生成器和判别器训练是采用JS或KL散度损失函数，但是采用JS或KL散度损失函数在数据分布无交叉的情况时会有梯度消失的问题，从而导致训练慢和模型可能不收敛的问题。采用本方案中的基于沃瑟斯坦距离的损失函数，可以避免训练时的梯度消失问题，提高了训练速度和收敛效果。
初始对抗神经网络的训练包括超参数的调整，超参数包括Adam优化器的学习率(LR)，1st-moment estimates的指数衰减率(Adam_Beta_1)以及初始生成器G和初始判别器D的最终有效层之后的丢弃率。超参数对于特定任务是敏感的，需要探索尝试出合适的超参数设置。经过实验对比，最终将LR设置为2e-4，将Adam_Beta_1设置为0.8，以确保初始对抗网络可以稳定收敛，将丢弃率设置为30％，以避免过拟合。在实验中大概在训练2000-3000轮之后可以得到比较好的收敛效果。
S206，采用所述对抗神经网络生成所述目标停车场的历史车位占用信息；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。
其中，本发明实施例的步骤S206参见图2所示实施例的步骤S104的具体描述，在此不进行赘述。
下面结合现有的技术对本方案中的方法进行说明，请参见图7，图7是本申请实施例提供的一种余弦相似度对比的举例示意图，如图7所示，图中的曲线1表示采用本方案中的方法生成的历史车位占用信息与真实车位占用信息的余弦 相似度，图中的曲线2表示采用现有技术RCGAN的方法生成的历史车位占用信息与真实车位占用信息的余弦相似度。从图中可以看出，采用本方案中的方法最终生成的历史车位占用信息与真实车位占用信息的余弦相似度为98％，相比于现有技术的96％来说有明显提升。
在本申请实施例中，通过确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图，进一步获取所述目标区域中停车场在目标时间段的车位占用信息，并根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络，最后采用所述对抗神经网络生成所述目标停车场的历史车位占用信息，所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。采用上述方法，考虑了地理空间内其他相关停车场数据对目标停车场的影响，采用所述对抗神经网络生成所述目标停车场的历史车位占用信息与真实的车位占用信息具有极高的相似度，提高了停车诱导系统预测车位占用信息的准确性。
请参见图8，图8是本申请实施例提供的一种停车数据修复装置的结构示意图。所述停车数据修复装置可以是运行于计算机设备中的一个计算机程序(包括程序代码)，例如该停车数据修复装置为一个应用软件；该装置可以用于执行本申请实施例提供的方法中的相应步骤。如图8所示，本申请实施例的所述停车数据修复装置1可以包括：拓扑图获取单元11、信息获取单元12、网络构建单元13、信息生成单元14。
拓扑图获取单元11，用于确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图；
信息获取单元12，用于获取所述目标区域中停车场在目标时间段的车位占用信息；
网络构建单元13，用于根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络；
信息生成单元14，用于采用所述对抗神经网络生成所述目标停车场的历史车位占用信息；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。
请参见图8，本申请实施例的所述拓扑图获取单元11可以包括：目标区域 确定子单元111、矩阵生成子单元112、拓扑图获取子单元113。
目标区域确定子单元111，用于确定目标停车场，将所述目标停车场为中心，预设的距离阈值为半径的区域确定为目标区域；所述目标区域包括除目标停车场外的其他停车场；
矩阵生成子单元112，用于统计所述目标区域的停车场信息，根据所述停车场信息确定所述目标区域中停车场对应的邻接矩阵和属性矩阵；
拓扑图获取子单元113，用于根据所述邻接矩阵和所述属性矩阵生成所述目标区域的中停车场的停车场拓扑图。
在一种可行的实施方式中，所述矩阵生成子单元112具体用于：
统计所述目标区域的停车场信息，所述停车场信息包括停车场的位置信息和属性信息；
根据所述位置信息确定停车场之间的最短路径，通过所述最短路径生成所述停车场对应的邻接矩阵；
将所述属性信息转换为标准数值，根据所述标准数值生成停车场对应的属性矩阵。
在一种可行的实施方式中，所述信息获取单元12具体用于：
根据车辆计数规则，统计停车场在目标时间段的车辆数量；
对所述车辆数量进行归一化处理生成停车场的车辆占用量，将所述车辆占用量的最大值作为停车场的满占用量；
根据所述满占用量和目标时间段中各时刻的车辆占用量，确定停车场在目标时间段中的车位占用信息。
请参见图8，本申请实施例的所述网络构建单元13可以包括：合成信息生成子单元131、判别结果生成子单元132、网络构建子单元133。
合成信息生成子单元131，用于获取初始对抗神经网络，所述初始对抗神经网络包括初始判别器和初始生成器；通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息；
判别结果生成子单元132，用于将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息；通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果；
网络构建子单元133，用于基于所述合成车位占用信息和所述判别结果，通 过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络。
在一种可行的实施方式中，所述初始判别器和所述初始生成器均包括图卷积模块、时空融合模块和门控循环模块；
所述合成信息生成子单元131具体用于：
将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始生成器的图卷积模块，生成第一图特征矩阵；
基于所述初始生成器的时空融合模块，将所述第一图特征矩阵与初始噪声进行融合生成第一融合矩阵；
通过所述第一融合矩阵和初始生成器的门控循环模块，生成停车场的合成车位占用信息。
在一种可行的实施方式中，所述判别结果生成子单元132具体用于：
将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始判别器的图卷积模块，生成第二图特征矩阵；
通过所述初始判别器的时空融合模块，将所述第二图特征矩阵和所述训练车位占用信息进行融合生成第二融合矩阵；
通过所述第二融合矩阵和所述初始判别器的门控循环模块，生成所述训练车位占用信息的判别结果。
在一种可行的实施方式中，所述网络构建子单元133具体用于：
根据所述合成车位占用信息和沃瑟斯坦距离生成所述初始生成器的第一目标函数，根据所述判别结果和沃瑟斯坦距离生成所述初始判别器的第二目标函数，
基于所述第一目标函数和第二目标函数构建所述初始对抗神经网络的损失函数；
根据所述损失函数调整所述初始对抗神经网络的模型参数；所述初始对抗神经网络的模型参数包括所述初始判别器的模型参数和所述初始生成器的模型参数；
当调整后的初始对抗神经网络满足收敛条件时，将包含调整后的模型参数的初始对抗神经网络确定为对抗神经网络。
在本申请实施例中，通过确定目标停车场，根据所述目标停车场确定目标 区域，获取所述目标区域的中停车场的停车场拓扑图，进一步获取所述目标区域中停车场在目标时间段的车位占用信息，并根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络，最后采用所述对抗神经网络生成所述目标停车场的历史车位占用信息，所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。采用上述方法，考虑了地理空间内其他相关停车场数据对目标停车场的影响，采用所述对抗神经网络生成所述目标停车场的历史车位占用信息与真实的车位占用信息具有极高的相似度，提高了停车诱导系统预测车位占用信息的准确性。
请参见图9，图9是本申请实施例提供的一种计算机设备的结构示意图。如图9所示，所述计算机设备1000可以包括：至少一个处理器1001，例如CPU，至少一个网络接口1004，用户接口1003，存储器1005，至少一个通信总线1002。其中，通信总线1002用于实现这些组件之间的连接通信。其中，用户接口1003可以包括显示屏(Display)，可选用户接口1003还可以包括标准的有线接口、无线接口。网络接口1004可选的可以包括标准的有线接口、无线接口(如WI-FI接口)。存储器1005可以是随机存取存储器(Random Access Memory，RAM)，也可以是非易失性存储器(non-volatile memory，NVM)，例如至少一个磁盘存储器。存储器1005可选的还可以是至少一个位于远离前述处理器1001的存储装置。如图9所示，作为一种计算机存储介质的存储器1005中可以包括操作系统、网络通信模块、用户接口模块以及数据处理应用程序。
在图9所示的计算机设备1000中，网络接口1004可提供网络通讯功能，用户接口1003主要用于为用户提供输入的接口；而处理器1001可以用于调用存储器1005中存储的数据处理应用程序，以实现上述图2-图7任一个所对应实施例中对所述停车数据修复方法的描述，在此不再赘述。
应当理解，本申请实施例中所描述的计算机设备1000可执行前文图2-图7任一个所对应实施例中对所述停车数据修复方法的描述，也可执行前文图8所对应实施例中对所述停车数据修复装置的描述，在此不再赘述。另外，对采用相同方法的有益效果描述，也不再进行赘述。
此外，这里需要指出的是：本申请实施例还提供了一种计算机可读存储介质，且所述计算机可读存储介质中存储有前文提及的停车数据修复装置所执行 的计算机程序，且所述计算机程序包括程序指令，当所述处理器执行所述程序指令时，能够执行前文图2-图7任一个所对应实施例中对所述停车数据修复方法的描述，因此，这里将不再进行赘述。另外，对采用相同方法的有益效果描述，也不再进行赘述。对于本申请所涉及的计算机可读存储介质实施例中未披露的技术细节，请参照本申请方法实施例的描述。作为示例，程序指令可被部署为在一个计算设备上执行，或者在位于一个地点的多个计算设备上执行，又或者，在分布在多个地点且通过通信网络互连的多个计算设备上执行，分布在多个地点且通过通信网络互连的多个计算设备可以组成区块链系统。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，上述的程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。其中，上述计算机可读存储介质可以是前述任一实施例提供的一种停车数据修复装置或者上述设备的内部存储单元，例如电子设备的硬盘或内存。该计算机可读存储介质也可以是该电子设备的外部存储设备，例如该电子设备上配备的插接式硬盘，智能存储卡(smart media card,SMC)，安全数字(secure digital,SD)卡，闪存卡(flash card)等。上述计算机可读存储介质还可以包括磁碟、光盘、只读存储记忆体(read-only memory，ROM)或随机存储记忆体等。进一步地，该计算机可读存储介质还可以既包括该电子设备的内部存储单元也包括外部存储设备。该计算机可读存储介质用于存储该计算机程序以及该电子设备所需的其它程序和数量。该计算机可读存储介质还可以用于暂时地存储已经输出或者将要输出的数据。
本发明的权利要求书和说明书及附图中的术语“第一”、“第二”等是用于区别不同对象，而不是用于描述特定顺序。此外，术语“包括”和“具有”以及它们任何变形，意图在于覆盖不排他的包含。例如包含了一系列步骤或单元的过程、方法、系统、产品或设备没有限定于已列出的步骤或单元，而是可选地还包括没有列出的步骤或单元，或可选地还包括对于这些过程、方法、产品或设备固有的其它步骤或单元。在本文中提及“实施例”意味着，结合实施例描述的特定特征、结构或特性可以包含在本发明的至少一个实施例中。在说明书中的各个位置展示该短语并不一定均是指相同的实施例，也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是， 本文所描述的实施例可以与其它实施例相结合。在本发明说明书和所附权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合，并且包括这些组合。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本发明的范围。
在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以是两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
以上所揭露的仅为本申请较佳实施例而已，当然不能以此来限定本申请之权利范围，因此依本申请权利要求所作的等同变化，仍属本申请所涵盖的范围。

Claims (11)

  1. 一种停车数据修复方法，其特征在于，包括：
    确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图；
    获取所述目标区域中停车场在目标时间段的车位占用信息；
    根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络；
    采用所述对抗神经网络生成所述目标停车场的历史车位占用信息；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。
  2. 根据权利要求1所述的方法，其特征在于，所述确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图，包括：
    确定目标停车场，将所述目标停车场为中心，预设的距离阈值为半径的区域确定为目标区域；所述目标区域包括除目标停车场外的其他停车场；
    统计所述目标区域的停车场信息，根据所述停车场信息确定所述目标区域中停车场对应的邻接矩阵和属性矩阵；
    根据所述邻接矩阵和所述属性矩阵生成所述目标区域的中停车场的停车场拓扑图。
  3. 根据权利要求2所述的方法，其特征在于，所述统计所述目标区域的停车场信息，根据所述停车场信息确定所述目标区域中停车场对应的邻接矩阵和 属性矩阵，包括：
    统计所述目标区域的停车场信息，所述停车场信息包括停车场的位置信息和属性信息；
    根据所述位置信息确定停车场之间的最短路径，通过所述最短路径生成所述停车场对应的邻接矩阵；
    将所述属性信息转换为标准数值，根据所述标准数值生成停车场对应的属性矩阵。
  4. 根据权利要求1所述的方法，其特征在于，所述获取所述目标区域中停车场在目标时间段的车位占用信息，包括：
    根据车辆计数规则，统计停车场在目标时间段的车辆数量；
    对所述车辆数量进行归一化处理生成停车场的车辆占用量，将所述车辆占用量的最大值作为停车场的满占用量；
    根据所述满占用量和目标时间段中各时刻的车辆占用量，确定停车场在目标时间段中的车位占用信息。
  5. 根据权利要求1所述的方法，其特征在于，所述根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络，包括：
    获取初始对抗神经网络，所述初始对抗神经网络包括初始判别器和初始生成器；
    通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息；
    将所述合成车位占用信息和所述车位占用信息作为训练车位占用信息；
    通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果；
    基于所述合成车位占用信息和所述判别结果，通过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络。
  6. 根据权利要求5所述的方法，其特征在于，所述初始判别器和所述初始生成器均包括图卷积模块、时空融合模块和门控循环模块；
    所述通过所述停车场拓扑图和所述初始生成器，生成停车场的合成车位占用信息，包括：
    将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始生成器的图卷积模块，生成第一图特征矩阵；
    基于所述初始生成器的时空融合模块，将所述第一图特征矩阵与初始噪声进行融合生成第一融合矩阵；
    通过所述第一融合矩阵和初始生成器的门控循环模块，生成停车场的合成车位占用信息。
  7. 根据权利要求5所述的方法，其特征在于，所述通过所述停车场拓扑图、所述训练车位占用信息和所述初始判别器，生成训练车位占用信息的判别结果，包括：
    将所述停车场拓扑图中的邻接矩阵和属性矩阵输入所述初始判别器的图卷积模块，生成第二图特征矩阵；
    通过所述初始判别器的时空融合模块，将所述第二图特征矩阵和所述训练 车位占用信息进行融合生成第二融合矩阵；
    通过所述第二融合矩阵和所述初始判别器的门控循环模块，生成所述训练车位占用信息的判别结果。
  8. 根据权利要求5所述的方法，其特征在于，所述基于所述合成车位占用信息和所述判别结果，通过沃瑟斯坦距离生成所述初始对抗神经网络的损失函数，通过所述损失函数调整所述初始对抗神经网络生成对抗神经网络，包括：
    根据所述合成车位占用信息和沃瑟斯坦距离生成所述初始生成器的第一目标函数，根据所述判别结果和沃瑟斯坦距离生成所述初始判别器的第二目标函数，
    基于所述第一目标函数和第二目标函数构建所述初始对抗神经网络的损失函数；
    根据所述损失函数调整所述初始对抗神经网络的模型参数；所述初始对抗神经网络的模型参数包括所述初始判别器的模型参数和所述初始生成器的模型参数；
    当调整后的初始对抗神经网络满足收敛条件时，将包含调整后的模型参数的初始对抗神经网络确定为对抗神经网络。
  9. 一种停车数据修复装置，其特征在于，包括：
    拓扑图获取单元，用于确定目标停车场，根据所述目标停车场确定目标区域，获取所述目标区域的中停车场的停车场拓扑图；
    信息获取单元，用于获取所述目标区域中停车场在目标时间段的车位占用信息；
    网络构建单元，用于根据所述停车场拓扑图和所述车位占用信息，构建对抗神经网络；
    信息生成单元，用于采用所述对抗神经网络生成所述目标停车场的历史车位占用信息；所述历史车位占用信息用于训练停车预测系统，所述停车预测系统用于预测停车场的车位占用信息。
  10. 一种计算机设备，其特征在于，包括：处理器、存储器以及网络接口；
    所述处理器与所述存储器、所述网络接口相连，其中，所述网络接口用于提供网络通信功能，所述存储器用于存储程序代码，所述处理器用于调用所述程序代码，以执行权利要求1-8任一项所述的方法。
  11. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质中存储有计算机程序，该计算机程序适于由处理器加载并执行权利要求1-8任一项所述的方法。
PCT/CN2021/108734 2021-07-27 2021-07-27 一种停车数据修复方法、装置、计算机设备以及存储介质 WO2023004595A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/108734 WO2023004595A1 (zh) 2021-07-27 2021-07-27 一种停车数据修复方法、装置、计算机设备以及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/108734 WO2023004595A1 (zh) 2021-07-27 2021-07-27 一种停车数据修复方法、装置、计算机设备以及存储介质

Publications (1)

Publication Number Publication Date
WO2023004595A1 true WO2023004595A1 (zh) 2023-02-02

Family

ID=85086215

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/108734 WO2023004595A1 (zh) 2021-07-27 2021-07-27 一种停车数据修复方法、装置、计算机设备以及存储介质

Country Status (1)

Country Link
WO (1) WO2023004595A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116612458A (zh) * 2023-05-30 2023-08-18 易飒(广州)智能科技有限公司 基于深度学习的泊车路径确定方法与系统
CN117877313A (zh) * 2024-03-12 2024-04-12 浙江宇泛精密科技有限公司 基于物联网感知的停车场管理方法及装置
CN117877313B (zh) * 2024-03-12 2024-05-31 浙江宇泛精密科技有限公司 基于物联网感知的停车场管理方法及装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109214422A (zh) * 2018-08-02 2019-01-15 深圳先进技术研究院 基于dcgan的停车数据修补方法、装置、设备及存储介质
CN111210656A (zh) * 2020-01-23 2020-05-29 北京百度网讯科技有限公司 停车场空闲车位预测方法、装置、电子设备及存储介质
CN111348034A (zh) * 2020-04-03 2020-06-30 苏州大学 基于生成对抗模仿学习的自动泊车方法及系统
WO2020254418A1 (en) * 2019-06-18 2020-12-24 Continental Automotive Gmbh System and method for populating a database with occupancy data of parking facilities
CN112419171A (zh) * 2020-10-28 2021-02-26 云南电网有限责任公司昆明供电局 一种多残差块条件生成对抗网络的图像复原方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109214422A (zh) * 2018-08-02 2019-01-15 深圳先进技术研究院 基于dcgan的停车数据修补方法、装置、设备及存储介质
WO2020254418A1 (en) * 2019-06-18 2020-12-24 Continental Automotive Gmbh System and method for populating a database with occupancy data of parking facilities
CN111210656A (zh) * 2020-01-23 2020-05-29 北京百度网讯科技有限公司 停车场空闲车位预测方法、装置、电子设备及存储介质
CN111348034A (zh) * 2020-04-03 2020-06-30 苏州大学 基于生成对抗模仿学习的自动泊车方法及系统
CN112419171A (zh) * 2020-10-28 2021-02-26 云南电网有限责任公司昆明供电局 一种多残差块条件生成对抗网络的图像复原方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116612458A (zh) * 2023-05-30 2023-08-18 易飒(广州)智能科技有限公司 基于深度学习的泊车路径确定方法与系统
CN117877313A (zh) * 2024-03-12 2024-04-12 浙江宇泛精密科技有限公司 基于物联网感知的停车场管理方法及装置
CN117877313B (zh) * 2024-03-12 2024-05-31 浙江宇泛精密科技有限公司 基于物联网感知的停车场管理方法及装置

Similar Documents

Publication Publication Date Title
Zhang et al. Deep spatio-temporal residual networks for citywide crowd flows prediction
Zhang et al. Predicting citywide crowd flows using deep spatio-temporal residual networks
US11586992B2 (en) Travel plan recommendation method, apparatus, device and computer readable storage medium
WO2022126970A1 (zh) 金融欺诈风险识别方法、装置、计算机设备及存储介质
US10575162B1 (en) Detecting and validating planned event information
CN110096634B (zh) 一种基于粒子群优化的房产数据向量对齐方法
WO2023123624A1 (zh) 城市流感发病趋势预测方法、系统、终端以及存储介质
Wang et al. BSV-PAGS: Blockchain-based special vehicles priority access guarantee scheme
WO2019237840A1 (zh) 数据集生成方法及装置
CN112861972A (zh) 一种展业区域的选址方法、装置、计算机设备和介质
WO2023004595A1 (zh) 一种停车数据修复方法、装置、计算机设备以及存储介质
CN111738474A (zh) 交通状态预测方法和装置
WO2023000261A1 (zh) 一种区域交通的预测方法以及设备
Luo et al. Spatiotemporal hashing multigraph convolutional network for service-level passenger flow forecasting in bus transit systems
Zhuang et al. Integrating a deep forest algorithm with vector‐based cellular automata for urban land change simulation
Xuqin Application of network protocol improvement and image content search in mathematical calculus 3D modeling video analysis
CN111861139A (zh) 商户推荐方法、装置及计算机设备
CN116186629A (zh) 基于个性化联邦学习的金融客户分类及预测方法、装置
Zhao et al. Learning region similarities via graph-based deep metric learning
CN113643564B (zh) 一种停车数据修复方法、装置、计算机设备以及存储介质
Bouillet et al. Fusing traffic sensor data for real-time road conditions
Zhou et al. Refined taxi demand prediction with ST-Vec
CN115222109A (zh) 信息处理方法、装置、介质及设备
CN114880296B (zh) 适用于园区大数据的储存方法及系统
US11861912B2 (en) Methods and internet of things systems for counting and regulating pedestrian volume in public places of smart cities

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: 21951211

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE