Method for predicting spatial distribution of pedestrian flow and method for early warning of regional public safety

Abstract

The present invention discloses a method for predicting the spatial distribution of pedestrian flow and a method for regional public safety warning, including obtaining existing pedestrian flow observation sample data and geographical environment factor data and preprocessing them to construct a training data set; constructing an initial model for predicting the spatial distribution of pedestrian flow based on a multi-layer perceptron, an attention mechanism, and a deep neural network and training to obtain a model for predicting the spatial distribution of pedestrian flow; using the obtained model for predicting the spatial distribution of pedestrian flow to predict the spatial distribution of pedestrian flow in a target area. By learning the characteristics of regional pedestrian flow to extract a pattern matrix and combining geographical environment factor data for weight mining, the present invention not only realizes the prediction of the spatial distribution of pedestrian flow and the corresponding regional public safety warning, but also has higher reliability and better accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of urban intelligent computing, and particularly relates to a method for predicting the spatial distribution of pedestrian flow and a method for early warning of regional public safety. Background Art

[0002] The prediction of the spatial distribution of pedestrian flow can provide data support for the formulation of urban planning, the control of traffic flow, the early warning of regional public safety, etc. Therefore, the prediction of the spatial distribution of pedestrian flow is of great significance.

[0003] At present, the commonly used schemes for predicting the spatial distribution of pedestrian flow are mainly divided into mechanism-driven methods and data-driven methods. The mechanism-driven methods are based on the physical laws of human movement, and a prediction framework is established through parameters such as regional attraction and distance decay effect. A typical scheme is the improved gravity model, which improves the prediction accuracy by quantifying the dynamic association between the activity space and urban functions; however, such schemes rely extremely on the construction and input of prior knowledge, and the subjective deviation is relatively large. The data-driven methods are to use machine learning models to mine the objective non-linear relationships of spatio-temporal data, such as multi-scale spatio-temporal network schemes and deep geographically weighted regression model schemes; however, such schemes have problems of insufficient understanding of spatial heterogeneity and insufficient utilization of spatial context information, so the accuracy of such schemes is not high. Summary of the Invention

[0004] One of the purposes of the present invention is to provide a method for predicting the spatial distribution of pedestrian flow with high reliability and good accuracy.

[0005] Another purpose of the present invention is to provide a method for early warning of regional public safety.

[0006] The method for predicting the spatial distribution of pedestrian flow provided by the present invention includes the following steps:

[0007] S1. Obtain existing pedestrian flow observation sample data and geographical environment factor data;

[0008] S2. Preprocess the data information obtained in step S1 to construct a training data set;

[0009] S3. Based on a multi-layer perceptron, an attention mechanism, and a deep neural network, construct an initial model for predicting the spatial distribution of pedestrian flow including a feature pattern extraction module, a spatial feature distance weight extraction module, and a feature fusion regression module;

[0010] Among them, the feature pattern extraction module is used to learn the time series features of the pedestrian flow data and obtain the corresponding pattern matrix; the spatial feature distance weight extraction module is used to mine the weights of the geographical environment factor information and obtain the corresponding spatial feature distance weights; the feature fusion regression module is used to fuse the pattern matrix and the spatial feature distance weights and calculate the final predicted data of the spatial distribution of pedestrian flow.

[0011] S4. Use the training dataset constructed in step S2 to train the initial model for predicting the spatial distribution of pedestrian flow constructed in step S3 to obtain a model for predicting the spatial distribution of pedestrian flow.

[0012] S5. Use the model for predicting the spatial distribution of pedestrian flow obtained in step S4 to predict the spatial distribution of pedestrian flow in the target area.

[0013] The specific steps of step S1 are as follows:

[0014] Obtain the existing pedestrian flow observation data.

[0015] Obtain the existing geographical environment factor data; the geographical environment factor data includes POI data, road network data, and urban environment data.

[0016] The preprocessing described in step S2 specifically includes the following steps:

[0017] The preprocessing includes averaging processing and standardization processing.

[0018] The averaging processing specifically includes the following steps:

[0019] Align the data information obtained in step S1 in terms of spatial resolution according to the set resolution grid.

[0020] Average the aligned data and project it into each grid element.

[0021] The standardization processing specifically includes the following steps:

[0022] Perform standardization processing using the following formula: In the formula is the attribute value of the r1-th data information corresponding to the i1-th grid after standardization processing; is the attribute value of the r1-th data information corresponding to the i1-th grid before standardization processing; is the maximum value of the r1-th data information; is the minimum value of the r1-th data information.

[0023] The specific steps of step S3 are as follows:

[0024] Based on a multi-layer perceptron neural network, a feature pattern extraction module is constructed;

[0025] Based on the attention mechanism and the deep neural network, a spatial feature distance weight extraction module is constructed;

[0026] Based on the dimension conversion scheme, a feature fusion regression module is constructed.

[0027] The processing process of the feature pattern extraction module specifically includes the following steps:

[0028] The following formula is used to calculate the pattern matrix: In the formula is the k-th column in the pattern matrix P; MLP() is the processing function of the multi-layer perceptron neural network; X is the input grid data; is the longitude of the input grid i; is the latitude of the input grid i.

[0029] The processing process of the spatial feature distance weight extraction module specifically includes the following steps:

[0030] The following formula is used to calculate the spatial distance between sample point i and sample point j in the sample point data: In the formula is the spatial distance between grid i and grid j; is the longitude of the input grid j; is the latitude of the input grid j;

[0031] The following formula is used to calculate the feature dimension distance vector between grid i and grid j in the grid data: In the formula is the feature dimension distance vector between grid i and grid j; is the feature vector of grid i; is the feature vector of grid j;

[0032] Based on the attention mechanism, the following formula is used to fuse the obtained spatial distance and feature dimension distance vector to obtain a unified spatial feature distance matrix: In the formula is the unified spatial feature distance of grid i to other spatial points; a is the self-attention mechanism coefficient; b is the bias term; Q is the query vector in the self-attention mechanism, and , is the linear weight for query vector transformation; K is the key vector in the self-attention mechanism, and , is the linear weight for key vector transformation; V is the value vector in the self-attention mechanism, and , is the linear weight for value vector transformation; F is the concatenation matrix, and , is the i2-th element in F; is the feature dimension;

[0033] Based on the deep neural network, the spatial feature distance weight matrix is calculated using the following formula: In the formula is the spatial feature distance weight of the grid i with coordinates in the spatial feature distance weight matrix; is the unified spatial feature distance between grid i and grid n; SFWNN() is the processing function of the spatial feature distance weight extraction network constructed based on the deep neural network.

[0034] The processing process of the spatial feature distance weight extraction network specifically includes the following steps:

[0035] The spatial feature distance weight extraction network includes several hidden layers and an output layer; several hidden layers and the output layer are connected in series in turn;

[0036] For the i-th hidden layer, the processing process is expressed as: In the formula is the output of the i-th hidden layer; is the hyperbolic tangent activation function; is the layer normalization operation; is the weight matrix of the i-th linear layer; is the bias vector of the i-th linear layer; is the set annihilation probability; is the annihilation operation, and the specific operation is: The value of is annihilated to 0 with the probability corresponding to the value of and remains unchanged with the probability corresponding to the value of

[0037] For the output layer, the processing process is expressed as: In the formula is the output of the output layer; is the weight matrix of the output linear layer; x is the input of the output layer; is the bias vector of the output linear layer; is the set annihilation probability.

[0038] The processing process of the feature fusion regression module specifically includes the following steps:

[0039] Based on the dimension conversion scheme, the predicted pedestrian flow value of the i-th grid is calculated using the following formula: In the formula is the predicted pedestrian flow value of the i-th grid; is the h-th column in the spatial feature distance weight matrix; is the h-th column in the pattern matrix P; is the set coefficient vector; is the feature vector of the i-th grid of the input; is the bias term of the i-th grid of the input.

[0040] The present invention also provides a method for regional public safety early warning, including the following steps:

[0041] A. Obtain the data information of the target area;

[0042] B. According to the data information obtained in step A, use the described method for predicting the spatial distribution of pedestrian flow to predict the spatial distribution of pedestrian flow in the target area;

[0043] C. According to the prediction result of the spatial distribution of pedestrian flow in the target area obtained in step B, use the following rules for regional public safety early warning:

[0044] If the predicted result of the pedestrian flow in a certain grid in the target area is greater than the set pedestrian flow threshold, conduct public safety early warning for this grid area.

[0045] This method for predicting the spatial distribution of pedestrian flow and the method for regional public safety early warning provided by the present invention extract the pattern matrix by learning the regional pedestrian flow characteristics, and conduct weight mining in combination with the geographical environment factor data, which not only realizes the prediction of the spatial distribution of pedestrian flow and the corresponding regional public safety early warning, but also has higher reliability and better accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the schematic diagram of the method flow of the prediction method of the present invention.

[0047] Figure 2 is the schematic diagram of the research area of the embodiment of the prediction method of the present invention.

[0048] Figure 3 is the schematic diagram of the method flow of the early warning method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] As Figure 1 shown is the schematic diagram of the method flow of the prediction method of the present invention: This method for predicting the spatial distribution of pedestrian flow disclosed by the present invention includes the following steps:

[0050] S1. Obtain the existing pedestrian flow observation sample data and geographical environment factor data; specifically including the following steps:

[0051] Obtain the existing pedestrian flow observation data;

[0052] Obtain existing geographical environment factor data; the geographical environment factor data includes POI data, road network data, and urban environment data.

[0053] S2. Preprocess the data information obtained in step S1 to construct a training data set;

[0054] Among them, the preprocessing process specifically includes the following steps:

[0055] The preprocessing includes averaging processing and standardization processing;

[0056] The averaging processing includes the following steps:

[0057] Align the data information obtained in step S1 in terms of spatial resolution according to the set resolution grid;

[0058] Average the aligned data and project it into each grid element;

[0059] The standardization processing specifically includes the following steps:

[0060] Perform standardization processing using the following formula: In the formula is the attribute value of the r1th data information corresponding to the i1th grid after standardization processing; is the attribute value of the r1th data information corresponding to the i1th grid before standardization processing; is the maximum value of the r1th data information; is the minimum value of the r1th data information.

[0061] S3. Based on a multi-layer perceptron, attention mechanism, and deep neural network, construct an initial model for predicting the spatial distribution of pedestrian flow, including a feature pattern extraction module, a spatial feature distance weight extraction module, and a feature fusion regression module;

[0062] Among them, the feature pattern extraction module is used to learn the time series features of pedestrian flow data and obtain the corresponding pattern matrix; the spatial feature distance weight extraction module is used to mine the weights of geographical environment factor information and obtain the corresponding spatial feature distance weights; the feature fusion regression module is used to fuse the pattern matrix and spatial feature distance weights and calculate the final predicted data of the spatial distribution of pedestrian flow;

[0063] Among them, adaptive spatial weight mining is carried out through the feature pattern extraction module, and then the local heterogeneous driving mechanism of regional pedestrian flow is mined and perceived through the spatial feature distance weight extraction module;

[0064] During specific implementation: Based on the multi-layer perceptron neural network, a feature pattern extraction module is constructed; based on the attention mechanism and the deep neural network, a spatial feature distance weight extraction module is constructed; based on the dimension conversion scheme, a feature fusion regression module is constructed.

[0065] The processing process of the feature pattern extraction module specifically includes the following steps:

[0066] The pattern matrix is calculated using the following formula: In the formula is the k-th column in the pattern matrix P; MLP() is the processing function of the multi-layer perceptron neural network; X is the input grid data; is the longitude of the input grid i; is the latitude of the input grid i.

[0067] The processing process of the spatial feature distance weight extraction module specifically includes the following steps:

[0068] The following formula is used to calculate the spatial distance between sample point i and sample point j in the sample point data: In the formula is the spatial distance between grid i and grid j; is the longitude of the input grid j; is the latitude of the input grid j;

[0069] The following formula is used to calculate the feature dimension distance vector between grid i and grid j in the grid data: In the formula is the feature dimension distance vector between grid i and grid j; is the feature vector of grid i; is the feature vector of grid j;

[0070] Based on the attention mechanism, the following formula is used to fuse the obtained spatial distance and feature dimension distance vector to obtain the unified spatial feature distance matrix: In the formula is the unified spatial feature distance of grid i to other spatial points; a is the self-attention mechanism coefficient; b is the bias term; Q is the query vector in the self-attention mechanism, and , is the query vector transformation linear weight; K is the key vector in the self-attention mechanism, and , is the key vector transformation linear weight; V is the value vector in the self-attention mechanism, and , is the value vector transformation linear weight; F is the concatenation matrix, and , is the i2-th element in F; is the feature dimension;

[0071] Based on a deep neural network, the spatial feature distance weight matrix is calculated using the following formula: In the formula is the spatial feature distance weight of grid i with coordinates in the spatial feature distance weight matrix; is the unified spatial feature distance between grid i and grid n; SFWNN() is the processing function of the spatial feature distance weight extraction network constructed based on a deep neural network.

[0072] Among them, the processing process of the spatial feature distance weight extraction network specifically includes the following steps:

[0073] The spatial feature distance weight extraction network includes several hidden layers and an output layer; several hidden layers and the output layer are connected in series in sequence;

[0074] For the i-th hidden layer, the processing process is expressed as: In the formula is the output of the i-th hidden layer; is the hyperbolic tangent activation function; is the layer normalization operation; is the weight matrix of the i-th linear layer; is the bias vector of the i-th linear layer; is the set annihilation probability; is the annihilation operation, and the specific operation is: The value of is annihilated to 0 with the probability corresponding to the value of and remains unchanged with the probability corresponding to the value of

[0075] For the output layer, the processing process is expressed as: In the formula is the output of the output layer; is the weight matrix of the output linear layer; x is the input of the output layer; is the bias vector of the output linear layer; is the set annihilation probability.

[0076] The processing process of the feature fusion regression module specifically includes the following steps:

[0077] Based on the dimension conversion scheme, the predicted pedestrian flow value of the i-th grid is calculated using the following formula: In the formula is the predicted pedestrian flow value of the i-th grid; is the h-th column in the spatial feature distance weight matrix; is the h-th column in the pattern matrix P; is the set coefficient vector; is the feature vector of the i-th grid of the input; is the bias term of the i-th grid of the input.

[0078] S4. Use the training data set constructed in step S2 to train the initial model for predicting the spatial distribution of pedestrian flow constructed in step S3 to obtain a model for predicting the spatial distribution of pedestrian flow;

[0079] During the training process, add the regularization error of the pattern matrix, expressed as: In the formula is the regularization error value; is the weight parameter; P is the pattern matrix; I is the identity matrix; is the Frobenius norm.

[0080] S5. Use the model for predicting the spatial distribution of pedestrian flow obtained in step S4 to predict the spatial distribution of pedestrian flow in the target area.

[0081] The prediction method of the present invention innovatively introduces a mining scheme for the pedestrian flow spatial feature pattern matrix. By identifying the potential pattern distribution of the pedestrian flow in different regions, a pattern matrix is constructed, thereby effectively capturing the similarity of the pedestrian flow patterns between regions, enhancing the model's understanding of the regional pedestrian flow characteristics, and making the operation process of the pedestrian flow spatial prediction model more systematic. The prediction method of the present invention constructs a model to extract the spatial weight matrix in the data, mines the spatial local differences of the regional pedestrian flow, and carefully reflects the local details of the regional pedestrian flow driving mechanism.

[0082] The following further illustrates the prediction method of the present invention in conjunction with an embodiment:

[0083] Based on the geographical and population information data of the main urban area of XM City in 2022 (as Figure 2 shown), the specific implementation process of the present invention will be described.

[0084] In this embodiment, taxi trajectory data is used to simulate the inflow and outflow of the population in the observation area; collect taxi trajectory data, map POI data, OSM road network data, CLCD land use data, building height data, and NDVI data, perform preprocessing, construct an index system for regional pedestrian flow driving factors, and use it as the input feature sample set of the model; construct a grid with a resolution of 500m, align the spatial resolution of the data, and crop the data according to the research area.

[0085] Taking taxi trajectory data as material, calculate the population inflow and outflow of regional grids, which are used as the observed data of regional pedestrian flow. Count the departure trips and arrival trips of taxi trajectories within the grid. A total of 7007 grids are collected and the data is standardized. Construct and calculate and standardize the index system of regional pedestrian flow geographical environment factors according to three dimensions: infrastructure, built environment, and traffic accessibility, and construct a model training data set.

[0086] The constructed index system of geographical environment influencing factors includes: infrastructure, built environment, and traffic accessibility; among them, the infrastructure includes 6 indicators, including the density of hotel accommodation facilities, the density of business companies, the science and education culture index, the density of medical facilities, the density of business residences, and the density of tourist attractions. The built environment includes 3 indicators, including the greening intensity of the built environment, the proportion of water area, and the building height; the traffic accessibility includes 5 indicators, including the cost of traffic travel, airport accessibility, railway station accessibility, bus stop density, and subway station density.

[0087] Construct a model through the prediction method of the present invention. According to the division of training, validation, and test sets, train the regional pedestrian flow prediction model for the main urban area of SM City. Among them, the training set is used for fitting the model parameters, the validation set is used to test the fitting effect of the model parameters and judge whether to stop training, and the test set is used to test the accuracy of the model's pedestrian flow prediction effect.

[0088] Finally, compare the simulation results of the present invention with the goodness of fit of the OLS least squares method and the GNNWR method; the comparison index used is the goodness of fit Root Mean Square Error RMSE and Mean Absolute Error MAE; among them, the OLS least squares method was proposed by Adrien-Marie Legendre in the paper "Nouvelles méthodes pour la détermination des orbites des comètes" in 1805; the GNNWR method was proposed by Du Zhenhong in the paper "Geographically neural network weighted regression for the accurate estimation of spatial non-stationarity" in 2020;

[0089] The comparison of the finally obtained evaluation indexes is shown in Tables 1 and 2:

[0090] Table 1 Comparison of evaluation indexes of regional population outflow

[0091]

[0092] Table 2 Comparative schematic table of evaluation indicators for population inflow in the region

[0093]

[0094] As can be seen from Table 1 and Table 2, the method of the present invention performs optimally in each indicator, thus proving the reliability and effectiveness of the method of the present invention.

[0095] Such as Figure 3 shown is the schematic flowchart of the method of the early warning method of the present invention: The regional public safety early warning method disclosed by the present invention includes the following steps:

[0096] A. Obtain the data information of the target area;

[0097] B. According to the data information obtained in step A, use the above-mentioned method for predicting the spatial distribution of the number of people flow to predict the spatial distribution of the number of people flow in the target area;

[0098] C. According to the prediction result of the spatial distribution of the number of people flow in the target area obtained in step B, use the following rules for regional public safety early warning:

[0099] If the predicted result of the number of people flow in a certain grid in the target area is greater than the set threshold of the number of people flow, public safety early warning is carried out for the grid area.

Claims

1. A method for predicting the spatial distribution of pedestrian flow, characterized in that It includes the following steps: S1. Obtain existing human flow observation sample data and geographical environment factor data; S2. Preprocess the data information obtained in step S1 to construct a training data set; S3. Based on a multi-layer perceptron, an attention mechanism, and a deep neural network, construct an initial model for predicting the spatial distribution of human flow, including a feature pattern extraction module, a spatial feature distance weight extraction module, and a feature fusion regression module; Among them, the feature pattern extraction module is used to learn the time series features of human flow data and obtain the corresponding pattern matrix; the spatial feature distance weight extraction module is used to mine the weights of geographical environment factor information and obtain the corresponding spatial feature distance weights; the feature fusion regression module is used to fuse the pattern matrix and the spatial feature distance weights and calculate the final predicted data of the spatial distribution of human flow; The processing process of the feature pattern extraction module specifically includes the following steps: The pattern matrix is calculated using the following formula: In the formula is the k-th column in the pattern matrix P; MLP() is the processing function of the multi-layer perceptron neural network; X is the input grid data; is the longitude of the input grid i; is the latitude of the input grid i; The processing process of the spatial feature distance weight extraction module specifically includes the following steps: The spatial distance between sample point i and sample point j in the sample point data is calculated using the following formula: In the formula is the spatial distance between grid i and grid j; is the longitude of the input grid j; is the latitude of the input grid j; The characteristic dimension distance vector between grid i and grid j in the grid data is calculated using the following formula: In the formula is the characteristic dimension distance vector between grid i and grid j; is the characteristic vector of grid i; is the characteristic vector of grid j; Based on the attention mechanism, the following formula is used to fuse the obtained spatial distance and feature dimension distance vectors to obtain a unified spatial feature distance matrix: In the formula is the unified spatial feature distance from grid i to other points in space; a is the self-attention mechanism coefficient; b is the bias term; Q is the query vector in the self-attention mechanism, and , is the linear weight for query vector transformation; K is the key vector in the self-attention mechanism, and , is the linear weight for key vector transformation; V is the value vector in the self-attention mechanism, and , is the linear weight for value vector transformation; F is the concatenation matrix, and , is the i2-th element in F; is the feature dimension; Based on a deep neural network, the spatial feature distance weight matrix is calculated using the following formula: In the formula is the spatial feature distance weight of grid i in the spatial feature distance weight matrix, representing the coordinate ; is the unified spatial feature distance between grid i and grid n; SFWNN() is the processing function of the spatial feature distance weight extraction network constructed based on the deep neural network; The processing process of the feature fusion regression module specifically includes the following steps: Based on the dimension transformation scheme, the predicted value of the pedestrian flow of the i-th grid is calculated using the following formula: Where is the predicted value of the pedestrian flow of the i-th grid; is the h-th column in the spatial feature distance weight matrix; is the h-th column in the pattern matrix P; is the set coefficient vector; is the feature vector of the input i-th grid; is the bias term of the input i-th grid; S4. Use the training data set constructed in step S2 to train the initial model for predicting the spatial distribution of human flow constructed in step S3 to obtain a model for predicting the spatial distribution of human flow; S5. Use the model for predicting the spatial distribution of human flow obtained in step S4 to predict the spatial distribution of human flow in the target area.

2. The method for predicting the spatial distribution of pedestrian flow according to claim 1, wherein The step S1 specifically includes the following steps: Obtain existing human flow observation data; Obtain existing geographical environment factor data; the geographical environment factor data includes POI data, road network data, and urban environment data.

3. The method for predicting the spatial distribution of pedestrian flow according to claim 1, characterized in that The preprocessing described in step S2 specifically includes the following steps: The preprocessing includes averaging processing and standardization processing; The averaging processing specifically includes the following steps: Align the spatial resolution of the data information obtained in step S1 according to the set resolution grid; Average the aligned data and project it into each grid element; The standardization processing specifically includes the following steps: The standardization process is carried out using the following formula: In the formula is the attribute value of the r1-th data information corresponding to the i1-th grid after standardization; is the attribute value of the r1-th data information corresponding to the i1-th grid before standardization; is the maximum value of the r1-th data information; is the minimum value of the r1-th data information.

4. The method for predicting the spatial distribution of the number of people according to claim 3, wherein The step S3 includes the following steps: Based on a multi-layer perceptron neural network, construct a feature pattern extraction module; Based on an attention mechanism and a deep neural network, construct a spatial feature distance weight extraction module; Based on a dimension conversion scheme, construct a feature fusion regression module.

5. The method for predicting the spatial distribution of pedestrian flow according to claim 4, wherein The processing process of the spatial feature distance weight extraction network specifically includes the following steps: The spatial feature distance weight extraction network includes a number of hidden layers and an output layer; the number of hidden layers and the output layer are connected in series in sequence; For the i-th hidden layer, the processing is represented as: where is the output of the i-th hidden layer; is the hyperbolic tangent activation function; is the layer normalization operation; is the weight matrix of the i-th linear layer; is the bias vector of the i-th linear layer; is the set annihilation probability; is the annihilation operation, and the specific operation is: The value of is annihilated to 0 with the probability corresponding to the value of and remains unchanged with the probability corresponding to the value of For the output layer, the processing is expressed as: where is the output of the output layer; is the weight matrix of the output linear layer; x is the input of the output layer; is the bias vector of the output linear layer; is the set annihilation probability.

6. A regional public security early warning method, characterized in that It includes the following steps: A. Obtain the data information of the target area; B. According to the data information obtained in step A, use the method for predicting the spatial distribution of human flow according to any one of claims 1 to 5 to predict the spatial distribution of human flow in the target area; C. According to the prediction result of the spatial distribution of human flow in the target area obtained in step B, use the following rules for regional public safety early warning: If the predicted result of the human flow in a certain grid in the target area is greater than the set human flow threshold, conduct a public safety early warning for the grid area.

Images