A vehicle-person correlation analysis method based on word vectors

Abstract

The application discloses a kind of vehicle-person association analysis methods based on word vector.The method is first by the data collected is preprocessed, the mobile phone IMSI code information and vehicle information of the resident population in each card are deleted.Then, data is divided according to time period, and the IMSI code and license plate number collected in each time period form a training sample.Afterwards, using word embedding technology CBOW model is trained, and the word vector of IMSI code and license plate number is obtained respectively.Finally, the similarity between the word vector corresponding to IMSI code and license plate number is calculated using cosine similarity method.Through the method of the application, the solving effect of vehicle-person association analysis can be effectively improved, the cost of manual analysis is reduced, and it has very important practical significance for intelligent traffic safety event processing.

Description

TECHNICAL FIELD

[0001] The application discloses a person-vehicle correlation analysis method, in particular to a person-vehicle correlation analysis method based on word vectors, and belongs to the technical field of computer application. BACKGROUND

[0002] With the rapid growth of China's economic strength and the sharp increase of people's travel demand, China's road transportation industry has achieved leap-forward development. The number of motor vehicles, drivers and road mileage in China is continuously increasing.

[0003] Therefore, the person-vehicle correlation analysis is of great significance. The current person-vehicle correlation analysis is mainly to establish a space-time trajectory vector, and then calculate the similarity between trajectories by using a similarity method to realize matching. However, this method usually does not consider the influence of the time factor in the trajectory, that is, similar trajectories appearing at different times may also be considered as similar trajectories, which may lead to matching errors. SUMMARY

[0004] The purpose of the present application is to solve the above problems, and provide a person-vehicle correlation analysis method based on word vectors. The method pairs the IMSI code of the mobile phone and the license plate number, then trains the word embedding model, respectively obtains the vector of the IMSI code and the license plate number, and then calculates the similarity of the IMSI code and the license plate number by using the similarity matching method to realize matching. This method can effectively solve the influence of the time factor and effectively improve the accuracy of the person-vehicle correlation.

[0005] To achieve the above purpose, the technical scheme of the present application is:

[0006] A person-vehicle correlation analysis method based on word vectors, comprising the following steps:

[0007] Step 1: a traffic portal collects the IMSI code of a mobile phone and the license plate number of a vehicle by using two different existing devices, respectively uses and to represent the collected data set, wherein and respectively represent the number of data related to the IMSI code and the license plate number collected by the portal. In the data set , each data can be represented as , wherein represents the number of the portal, represents the IMSI code of the mobile phone, represents the time when the IMSI code is collected. In , each data can be represented as , wherein represents the license plate number of the vehicle.

[0008] Step 2: Due to the influence of traffic congestion or traffic accidents, there may be a large amount of repeated data collected in the card slot, and the resident population and resident vehicles will affect the final association analysis, so it is necessary to preprocess the data and delete a large amount of repeated data. The specific implementation is as follows:

[0009] Step 2-1, resident population data removal:

[0010] If a certain IMSI code is captured at least times in a continuous specified period of time , all data related to the IMSI code are deleted from ; similarly, if a certain license plate number is captured at least times in a continuous specified period of time , all data related to the license plate number are deleted from .

[0011] Step 2-2, invalid data filtering:

[0012] If a certain IMSI code or license plate number is captured in only one card slot, all data related to the IMSI code or license plate number are deleted.

[0013] Step 2-3, the preprocessed data set is represented as and , and respectively represent the number of preprocessed IMSI code or license plate number related data.

[0014] Step 3: training sample data generation:

[0015] The data in and of the same card slot are divided by time period. Set the time interval with 00:00:00 as the starting time, and form a sample data with all the IMSI codes and license plate numbers collected in the time interval . The collection of sample data formed by all card slots is represented as , where is the total number of samples.

[0016] Step 4: each sample is regarded as a document, ; IMSI code or license plate number in the sample as a word in the document. All samples are used as a training set, using the word embedding technology in the continuous bag-of-words model (CBOW) for training. CBOW model is a neural network model, used to learn the vector representation of words from the unlabeled training set.

[0017] Step 5: initialize the parameter matrix in CBOW model , where is the number of words in the training set, that is, the total number of different IMSI codes and license plate numbers, represents the number of hidden layer neurons, that is, the dimension of the final word vector, and each word is represented as a one-hot encoding vector .

[0018] Step 6: define CBOW in a fixed window size The center word vector is predicted by the context vector of the center word. Assuming the center word vector is , its context vector is used as the input of the model, combined with the parameter matrix , CBOW can propagate the input vector to the hidden layer .

[0019] Step 7: vector representation of hidden layer According to the matrix value of another parameter matrix in CBOW, the information is further propagated, and the probability distribution of each word is obtained.

[0020] Step 8: the actual output of CBOW is quite different from the real target output, and CBOW model uses back propagation algorithm to continuously optimize the value of parameter matrix , the ultimate goal is to make the maximum probability of the actual output of the network close to the target output, and finally get the optimized , and further get the vector representation of the optimized word calculated by .

[0021] Step 9: for any two words and , the corresponding word vectors and , calculate the cosine similarity between .

[0022] ​Step 10: input a license plate number (or IMSI code), sort all IMSI codes (or license plate numbers) according to similarity from large to small, and obtain a final recommended result sequence.

[0023] Compared with the prior art, the present application has the beneficial effects that:

[0024] The person-vehicle correlation analysis method based on word vectors provided by the present application can effectively obtain the correlation between the person and the vehicle and improve the problem solving efficiency. In the present application, the mobile phone IMSI code and the license plate number are paired, and the word embedding method is used for training to obtain the word vectors of the IMSI code and the license plate number, which can effectively avoid the influence of time factors. In addition, the data scale is reduced, and the algorithm calculation cost is effectively reduced. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0026] Figure 1 The flowchart of the person-vehicle correlation analysis method based on word vector technology provided by the present application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0028] Data source: by deploying equipment at some traffic intersections in a city, traffic data within 5 days is collected. A total of 56 checkpoints are included, of which only 21 checkpoints include license plate number data. There is a large difference in the number of IMSI code and license plate number related data collected by each checkpoint. A total of 15,815,764 data are collected, of which 14,968,620 data are related to IMSI code, and 847,144 data are related to license plate number. A total of 158,496 different license plate numbers and 1,248,597 different mobile phone IMSI codes are involved.

[0029] Step 1: the traffic checkpoint collects the mobile phone IMSI code and the vehicle license plate number by two different devices, respectively, and uses and Represents the collected data set, where and These represent the number of IMSI codes and license plate numbers collected by the checkpoint, respectively. In this context, each piece of data can be represented as... ,in This indicates the checkpoint number. This indicates the phone's IMSI code. This indicates the time when the IMSI code was collected. In this context, each piece of data can be represented as... ,in This indicates the vehicle's license plate number.

[0030] Step 2: Due to traffic congestion or traffic accidents, the data collected at checkpoints may contain a large number of duplicate data. At the same time, the number of permanent residents and permanent vehicles will affect the final correlation analysis. Therefore, it is necessary to delete these data.

[0031] Step 2-1, Resident Population Data Removal: If a certain IMSI has been in a continuous period of time... ( Captured within 2 hours (at least) ( (Next, from) Delete all data related to that IMSI code; similarly, if a license plate number appears consecutively for a period of time... At least captured inside Next, from Delete all data related to this license plate number.

[0032] Step 2-2, Invalid data filtering: If a certain IMSI code or license plate number is captured only in one checkpoint, then delete all related data for that IMSI code or license plate number.

[0033] Steps 2-3: Represent the preprocessed dataset as follows and , and These represent the number of IMSI codes or license plate number-related data entries after preprocessing.

[0034] Step 3: Generate training sample data, including data from the same checkpoint. and The data is segmented into time periods. Starting at 00:00:00, the time intervals are set. ( (5 minutes), all IMSI codes and license plate numbers collected during this time period are combined into a single sample data set. Let the set of sample data formed by all the card slots be denoted as where is the total number of samples.

[0035] Step 4: For each sample When taken as a document, the IMSI code or license plate number in the sample is taken as a word in the document. All samples are taken as a training set, and the Continuous Bag-of-Words model (CBOW) in the word embedding technology is used for training.

[0036] Step 5: Initialize the parameter matrix in the CBOW model where is the number of words in the training set, i.e., the total number of different IMSI codes and license plate numbers, represents the number of hidden layer neurons, i.e., the dimension of the final word vector, and each word is represented as a one-hot encoding vector .

[0037] Step 6: Define the CBOW model using the context vector of the center word to predict the vector of the center word under a fixed window size Assuming the center word vector is , its context vector is used as the input of the model, combined with , the CBOW model can propagate the input vector to the hidden layer :

[0038]

[0039] Step 7: Vector representation of the hidden layer According to the matrix value of , the information is further propagated to obtain the probability distribution of each word :

[0040]

[0041] where is the actual output of the center word in the network.

[0042] Step 8: The actual output and the true output differ greatly, and the CBOW model uses the backpropagation algorithm to continuously optimize the value of , and the final goal is to make the maximum probability of the output of the network close to the target output:

[0043]

[0044] Finally, we get for the word The vector representation of the word "car" can be calculated as .

[0045] Step 9: For any two words and , the corresponding word vectors and , calculate the cosine similarity between and : :

[0046]

[0047] Step 10: Input a license plate number (or IMSI code), and sort all IMSI codes (or license plate numbers) according to the similarity from large to small to get the final recommended result sequence.

Claims

1.A vehicle-person association analysis method based on word vectors, characterized in that, The method comprises the following steps: Step 1: Traffic kiosk collects the IMSI code of the mobile phone and the license plate number of the vehicle through two different devices, respectively using S = (s1, s2, …, s n ) and R = (r1, r2, …, r m ) to represent the collected data sets; Step 2: Preprocessing the collected data to obtain data sets S' and R'; Step 3: Training sample data generation, cutting the preprocessed S' and R' data of the same card mouth according to time periods; Step 4: Each sample is regarded as a document, and the IMSI code or license plate number in the sample is regarded as a word in the document; all samples are used as a training set, and a continuous bag-of-words model in a word embedding technology is used; Step 5: Initialize the parameter matrix W in CBOW model Voc×N where Voc is the number of words in the training set, i.e., the total number of different IMSI codes and license plate numbers, N represents the number of hidden layer neurons, i.e., the dimension of the resulting word vector, and each word t is represented as a one-hot encoding vector V t ; Step 6: Define CBOW to predict the vector of the center word using its context vectors at a fixed window size z; assume the center word vector is V t , then its context vectors As input to the model, combine the parameter matrix W Voc×N , CBOW propagates the input vectors to the hidden layer h; Step 7: The vector representation h of the hidden layer is then used in conjunction with other parameter matrices W′ in CBOW. N×Voc The matrix values ​​continue to propagate information forward, yielding the probability distribution V for each word. 1×Voc ; Step 8: The actual output of CBOW is quite different from the real target output, the CBOW model uses the back propagation algorithm to continuously optimize the parameter matrix W Voc×N , the final goal is to make the actual output of the network close to the target output with the maximum probability, and finally get the optimized W Voc×N , and further get the vector representation of the optimized word t calculated by V t ·W Voc×N ; Step 8 Final goal L M is to make the output of the network close to the target output: Finally, we get W Voc×N The vector representation of a word t is given by V t ·W Voc×N is computed Step 9: For any two words t x and t y the corresponding word vectors and are computed and the cosine similarity between them is computed Step 10: Input a license plate number or IMSI code, and sort all IMSI codes or license plate numbers according to similarity from large to small to obtain a final recommended result sequence. 2.The vehicle-person association analysis method based on word vectors according to claim 1, characterized in that, In step 1, the mobile phone's IMSI code and the vehicle's license plate number are respectively represented by S = (s1, s2, ..., s n ) and R = (r1, r2, ..., r m The data set S is represented as follows: where n and m represent the number of IMSI codes and license plate numbers collected by the checkpoint, respectively; each data entry in the data set S is represented as s. i = (id, imsi, time), i = 1, 2, ..., n, where id represents the checkpoint number, imsi represents the IMSI code of the mobile phone, and time represents the time when the IMSI code was collected; in R, each data entry is represented as r. j = (id, carnum, time), j = 1, 2, ..., m, where carnum represents the vehicle's license plate number. 3.The vehicle-person association analysis method based on word vectors according to claim 2, characterized in that The preprocessing in step 2 is implemented as follows: Step 2-1: Resident population data removal; If an IMSI is captured at least num times in a continuous specified time period ConTime, all data related to the IMSI code are deleted from S; similarly, if a license plate number is captured at least num times in a continuous time period ConTime, all data related to the license plate number are deleted from R; Step 2-2: Invalid data filtering, if an IMSI code or license plate number is captured only in one card mouth, all related data of the IMSI code or license plate number are deleted; Step 2-3, the preprocessed data set is represented as S' = (s'1, s'2,..., s' n') and R' = (r'1, r'2,..., r' m'), n' and m' represent the number of preprocessed IMSI codes or license plate number related data respectively. n′ ) and R' = (r'1, r'2,..., r' m'), n' and m' represent the number of preprocessed IMSI codes or license plate number related data respectively. m′ ) and R' = (r'1, r'2,..., r' m'), n' and m' represent the number of preprocessed IMSI codes or license plate number related data respectively. 4.The vehicle-person association analysis method based on word vectors according to claim 2 or 3, characterized in that The training sample data generation in step 3 is implemented as follows: The S' and R' data of the same card slot are divided by time period, with 00:00:00 as the starting time, setting the time interval gap, forming a sample data d with all the IMSI codes and license plate numbers collected in the time period, and the collection of all sample data formed by all card slots is expressed as D=(d1, d2, …, dl), wherein l is the total number of samples. l ) 5.The vehicle-person association analysis method based on word vectors according to claim 2 or 3, characterized in that In step 6, the CBOW propagates the input vector to the hidden layer h: 6.The vehicle-person association analysis method based on word vectors according to claim 5, characterized in that Step 7 obtains the probability distribution V for each word 1×Voc : V 1×Voc = softmax(h · W' + b'). (1) N×Voc ). (2)

Images