A deep learning passenger flow prediction method fusing dynamic influence of built environment

Abstract

The application discloses a kind of deep learning crowd flow prediction methods of fusion built environment dynamic influence, comprising the following steps: S1, data acquisition and processing;S2, built environment dynamic influence modeling, the extraction of built environment dynamic influence effect is carried out;S3, construct spatiotemporal graph neural network fusion model;S4, the evaluation of model performance and the diagnosis of error are carried out;S5, output multi-time scale people flow prediction by spatiotemporal graph neural network fusion model.According to the present application, higher precision, stronger interpretability, crowd activity prediction can be adapted to multiple regions and multiple time periods.

Description

Technical Field

[0001] This invention relates to the field of quantity processing technology, and in particular to a deep learning method for predicting pedestrian traffic that integrates the dynamic influence of the built environment. Background Technology

[0002] Predicting and characterizing pedestrian flow is a crucial foundation for achieving efficient and orderly urban management. As my country's urbanization continues, urban travel behavior is becoming increasingly complex, and traditional static scheduling and experience-driven management models are no longer sufficient to meet the demands of intelligent transportation, emergency response, and precise allocation of public resources. This places higher demands on the ability to model and predict pedestrian flow with high spatiotemporal accuracy.

[0003] In recent years, deep learning technology has made significant progress in urban pedestrian flow prediction. By introducing structures such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and graph neural networks (GNNs), researchers have been able to capture the temporal characteristics and spatial dependencies of pedestrian flow to a certain extent, thereby improving the accuracy and stability of predictions.

[0004] Existing similar technologies, such as "A Method for Predicting Crowd Activity Based on a Dynamic Graph-Assisted Neural Differential Model" (patent application number CN202310227491), use a dynamic graph structure to assist a neural differential model to simulate crowd activity. This type of method breaks through the traditional point-in-time prediction approach, characterizing crowd activity trajectories through continuously changing differential modeling, and can, to some extent, perceive the disturbances to crowd distribution caused by dynamic factors such as holidays and emergencies, thereby improving the model's adaptability to complex real-world situations. However, this method does not consider the significant impact of the built environment on pedestrian flow, and lacks modeling of its dynamic mechanisms. In fact, the built environment, as the spatial carrier of crowd travel and aggregation, is a key external factor driving crowd activity. More importantly, the type, function, and scale of the built environment have significantly different effects on pedestrian flow in different time periods and spatial areas. For example, business districts experience significant pedestrian peaks during morning and evening rush hours, while commercial areas attract crowds during holidays or after get off work. Therefore, to improve prediction accuracy, it is necessary to introduce a dynamic impact modeling mechanism for the built environment. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a deep learning-based pedestrian flow prediction method that integrates the dynamic influence of the built environment, achieving higher accuracy, stronger interpretability, and adaptability to multi-regional and multi-time period pedestrian activity prediction. To achieve the above-mentioned objectives and other advantages of this invention, a deep learning-based pedestrian flow prediction method integrating the dynamic influence of the built environment is provided, comprising: S1. Perform data collection and processing; S2. Model the dynamic impact of the built environment and extract the dynamic impact effects of the built environment; S3. Construct a spatiotemporal graph neural network fusion model; S4. Evaluate model performance and diagnose errors; S5. Output multi-timescale pedestrian flow predictions through a spatiotemporal graph neural network fusion model.

[0006] Preferably, in step S1, the pedestrian flow data comes from China Unicom mobile signaling, and the population access volume of each street unit is recorded with an hourly time resolution; after the data is corrected by coordinate projection, it is spatially matched with the street boundary and constructed into a continuous pedestrian flow time series tensor in chronological order, providing a basis for capturing periodic and sudden changes in travel.

[0007] Preferably, in step S1, street units are used as a unified spatial modeling granularity, and all built environment features, pedestrian flow observations, and control variables are mapped and aligned according to consistent street numbers.

[0008] Preferably, the time attribute information in step S1 is generated based on the timestamp, including semantic features of weekdays, weekends, and morning and evening rush hours, and is embedded in the model using one-hot encoding.

[0009] Preferably, in step S1, the spatial adjacency structure is constructed using the Queen adjacency method to create a static adjacency matrix, treating street units sharing boundaries or corners as adjacent nodes, in order to more completely express the interdependencies of real urban space.

[0010] Preferably, step S2 specifically includes the following steps: Several key variables were selected from the built environment data obtained in step S1 as regression explanatory factors. Using hours as the time granularity, a crowd flow prediction model is constructed hourly. GTWR is used to independently estimate the marginal impact of each built environment variable on crowd flow in each time slice. GTWR's spatial weighting matrix introduces a variable adaptive bandwidth mechanism; The regression coefficients of each variable estimated by the GTWR model at each time step and in each spatial grid are used as feature outputs, and finally a [T×N×D]-dimensional dynamic built environment influence feature tensor is constructed.

[0011] Compared with the prior art, the advantages and positive effects of the present invention are: 1) A modeling method for the dynamic impact of the built environment is introduced. Based on the multidimensional built environment characteristics at the spatial grid scale, GTWR technology is introduced to generate built environment impact factors that change over time, so as to characterize the dynamic driving effect of the built environment on pedestrian flow.

[0012] 2) A deep learning model with ST-GNN as the core and incorporating recurrent neural networks was constructed to achieve collaborative modeling of the spatial dependence, time series and periodic changes of crowd flow, thereby improving prediction accuracy and generalization ability.

[0013] 3) A multi-source heterogeneous feature fusion mechanism was designed, which combines the dynamic effects of the built environment, historical flow of the population and time-related information into the network model for dynamic fusion in the feature space, effectively enhancing the model's adaptability to complex scenarios with multiple factors. Attached Figure Description

[0014] Figure 1 The flowchart is for a deep learning-based pedestrian flow prediction method that incorporates the dynamic influence of the built environment, according to the present invention.

[0015] Figure 2 This is a comparison chart of the prediction results of the deep learning-based pedestrian flow prediction method based on the dynamic influence of the built environment according to the present invention and the traditional deep learning method. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Reference Figure 1 A deep learning-based method for predicting pedestrian traffic that integrates the dynamic influences of the built environment includes: Step 1. Data Acquisition and Preprocessing 1) Using street units as the unified spatial modeling granularity, all built environment features, pedestrian flow observations and control variables are mapped and aligned according to consistent street numbers, thereby ensuring spatial consistency in subsequent model construction.

[0018] 2) The pedestrian flow data comes from China Unicom mobile signaling, recording the population visits of each street unit with an hourly time resolution. After coordinate projection correction, the data is spatially matched with the street boundaries and constructed into a continuous pedestrian flow time series tensor in chronological order, providing a basis for capturing periodic and sudden changes in travel patterns.

[0019] 3) Built environment characteristics consist of multi-source data, including land use distribution extracted from Landsat-8 high-resolution remote sensing imagery, POI (Point of Interest) data from Baidu Maps, and road network topology from OpenStreetMap. All indicators were aggregated at the street scale and then spatiotemporally aligned with pedestrian flow sequences.

[0020] 4) Time attribute information is generated based on timestamps, including semantic features such as weekdays, weekends, and morning and evening rush hours, and is embedded in the model using one-hot encoding to enhance the model's ability to identify periodic time patterns.

[0021] 5) Spatial adjacency structure: A static adjacency matrix is ​​constructed using the Queen adjacency method, treating street units with shared boundaries or corners as adjacent nodes, in order to more completely express the interdependencies of real urban space and provide support for subsequent graph convolution to capture spatial interaction effects.

[0022] Step 2. Establish environmental dynamic impact modeling 1) Several key variables were selected from the built environment data obtained in step 1 as regression explanatory factors, including: residential POI density, commercial POI density, green space coverage, road density, average building height, and land use mix. These indicators were aggregated at the street scale and strictly aligned with the pedestrian flow time series to ensure consistency between the spatial and temporal benchmarks of the explanatory factors.

[0023] 2) Using hours as the time granularity, a crowd flow prediction model is constructed hourly. GTWR is used to independently estimate the marginal impact of each built environment variable on crowd flow in each time slice, thereby preserving spatial heterogeneity and temporal dynamics.

[0024] 3) To enhance the model's adaptability to the spatial behavior characteristics of different variables, the spatial weighting matrix of GTWR introduces a variable adaptive bandwidth mechanism. Different built environment factors have varying spatial influence ranges. The adaptive bandwidth automatically adjusts the weight decay rate according to the variable's own spatial scale, thereby avoiding oversmoothing or exaggeration of local fluctuations caused by a uniform bandwidth, and improving the stability and interpretability of the regression results.

[0025] 4) The regression coefficients of each variable estimated by the GTWR model at each time step and in each spatial grid are used as feature outputs. Finally, a dynamic built environment impact feature tensor of [T × N × D] dimensions is constructed (T is the time step, N is the number of street units, and D is the dimension of built environment variables), which serves as the structured input for subsequent deep prediction models.

[0026] Step 3. Spatiotemporal graph neural network fusion modeling and prediction 1) The aforementioned uniformly divided street units are used as nodes in a graph neural network, with each street unit considered as a node in the graph; adjacent grids are connected by edges according to the Queen adjacency rule, thereby constructing an adjacency matrix. , used for subsequent graph convolution operations.

[0027] 2) Construct a multi-dimensional input feature sequence for each node over time, including historical pedestrian traffic, the dynamic effects of the built environment output by GTWR, and temporal attributes, and integrate them into a three-dimensional tensor. ,in Indicates the number of historical time steps. For the number of nodes, For node feature dimensions.

[0028] 3) For each time step Input features Graph convolution operations are applied to propagate neighborhood information using an adjacency matrix, extracting crowd interaction features between local spatial grids. Chebyshev GCN is used for graph convolution, and the output is a spatial encoding tensor. .

[0029] 4) After graph convolution, 1D temporal convolution (TCN) is applied to perform sliding window convolution on the time series of each node. The convolution kernel is set to 3, and the stride is 1 to extract local features and improve the model's ability to model short-term travel (such as peak periods). The output is a temporal convolution tensor. .

[0030] 5) The features after temporal convolution are input into a gated recurrent unit (GRU) to capture travel evolution trends under non-stationary disturbances such as holidays and peak hours. The number of GRU units is set to 64, and the number of layers is 1 to 2. The final output is the state representation vector of each node. .

[0031] 6) This model supports both single-step prediction and multi-step rolling prediction tasks. In multi-step prediction, the output state of the GRU is input into the decoder, and future predictions are continuously generated through fully connected layers. Predict the sequence step by step. The output tensor dimension is... , where 2 represents the predicted pedestrian flow.

[0032] 7) The loss function uses the weighted average absolute error (MAE), in the form of: , used for supervised prediction. For the weight hyperparameters, an L2 regularization term and Dropout (set to 0.2) are introduced during training to prevent overfitting.

[0033] 8) The Adam optimizer was used for training, with an initial learning rate of 0.001, a batch size of 64, a maximum number of training epochs of 200, and an early stopping strategy to avoid overfitting.

[0034] For example, in an empirical analysis in Shenzhen, the deep learning-based pedestrian flow prediction method integrating the dynamic influence of the built environment, as described in this invention, was applied for pedestrian flow prediction analysis, and the classic deep learning method LSTM was used as a comparative model for experimental verification. (Reference) Figure 2 As can be seen, the prediction error of the proposed method is significantly smaller than that of the LSTM model in both peak and off-peak periods, indicating that the proposed method has higher prediction accuracy and better stability in traffic flow prediction tasks at different time periods.

[0035] This application employs three core modules working in synergy: a built environment dynamic impact modeling module, a spatiotemporal graph neural network fusion modeling module, and a multi-source feature fusion and prediction module. The first is the built environment dynamic impact modeling module. This invention introduces a spatiotemporal geographic weighted regression (GTWR) method to dynamically quantify the actual impact of built environment features on pedestrian flow in each time period, obtaining high-resolution, interpretable dynamic driving features of the built environment. The role of this module is to explicitly model this key external variable, the built environment, in a spatiotemporally sensitive manner, providing structural interpretive support for subsequent predictions.

[0036] Second, the spatiotemporal graph neural network fusion modeling module. This invention constructs a spatiotemporal graph neural network (ST-GNN) as its core, fused with a recurrent neural network from the temporal modeling module, to jointly extract the spatial structural dependence, temporal dynamic features, and periodic patterns of population flow. This structure can model the adjacency relationships and dynamic interaction patterns between regions, and effectively improve the expressive power and prediction accuracy of travel behavior across multiple time scales.

[0037] Thirdly, there is a multi-source feature fusion and prediction module. This invention inputs multi-source heterogeneous features, such as dynamic environmental influences, historical population flow, and time-slice attributes (e.g., peak hours, non-working days), into a deep learning network. Dynamic weight fusion is then performed in the feature space to output a predicted population flow value for each area in the future time period. The purpose of this module is to achieve collaborative modeling of multi-dimensional influencing factors, thereby improving the overall prediction accuracy.

[0038] In this application, the adjacency matrix currently uses Queen adjacency to construct a static spatial graph. Alternative solutions include: using K-nearest neighbors or distance thresholding to construct the graph structure, which is suitable for large-scale urban areas; and using dynamic adjacency matrix generation strategies, such as dynamically adjusting edge weights based on temporal residuals or prediction errors to achieve an adaptive graph structure.

[0039] This application currently adopts a fusion structure of ST-GNN and GRU. Alternative solutions may be to use GRU modules such as LSTM, TCN, or Transformer for time series modeling.

[0040] The number of devices and processing scale described herein are for simplification purposes. Applications, modifications, and variations of this invention will be readily apparent to those skilled in the art. Although embodiments of the invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for this invention, and further modifications can be readily implemented by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, this invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A deep learning-based method for predicting pedestrian traffic that integrates the dynamic influences of the built environment, characterized in that, Includes the following steps: S1. Perform data collection and processing; S2. Model the dynamic impact of the built environment and extract the dynamic impact effects of the built environment; S3. Construct a spatiotemporal graph neural network fusion model; S4. Evaluate model performance and diagnose errors; S5. Output multi-timescale pedestrian flow predictions through a spatiotemporal graph neural network fusion model.

2. The deep learning-based pedestrian flow prediction method that integrates the dynamic influence of the built environment as described in claim 1, characterized in that, In step S1, the pedestrian flow data comes from China Unicom mobile signaling and records the population visits of each street unit with an hourly time resolution. After the data is corrected by coordinate projection, it is spatially matched with the street boundary and constructed into a continuous pedestrian flow time series tensor in chronological order, providing a basis for capturing periodic and sudden changes in travel.

3. The deep learning-based pedestrian flow prediction method that integrates the dynamic influence of the built environment as described in claim 1, characterized in that, In step S1, street units are used as the unified spatial modeling granularity, and all built environment features, pedestrian flow observations, and control variables are mapped and aligned according to consistent street numbers.

4. The deep learning-based pedestrian flow prediction method that integrates the dynamic influence of the built environment as described in claim 2, characterized in that, In step S1, the time attribute information is generated based on the timestamp, including semantic features of weekdays, weekends, and morning and evening rush hours, and is embedded in the model using one-hot encoding.

5. The deep learning-based pedestrian flow prediction method integrating the dynamic influence of the built environment as described in claim 3, characterized in that, In step S1, the spatial adjacency structure constructs a static adjacency matrix using the Queen adjacency method, treating street units sharing boundaries or corners as adjacent nodes to more completely express the interdependencies of real urban space.

6. The deep learning-based pedestrian flow prediction method that integrates the dynamic influence of the built environment as described in claim 1, characterized in that, Step S2 specifically includes the following steps: Several key variables were selected from the built environment data obtained in step S1 as regression explanatory factors. Using hours as the time granularity, a crowd flow prediction model is constructed hourly. GTWR is used to independently estimate the marginal impact of each built environment variable on crowd flow in each time slice. GTWR's spatial weighting matrix introduces a variable adaptive bandwidth mechanism; The regression coefficients of each variable estimated by the GTWR model at each time step and in each spatial grid are used as feature outputs, and finally a [T × N × D]-dimensional dynamic built environment influence feature tensor is constructed.

7. The deep learning-based pedestrian flow prediction method integrating the dynamic influence of the built environment as described in claim 1, characterized in that, Step S3 specifically includes the following steps: The uniformly divided street units are used as nodes in the graph neural network, with each street unit considered as a node in the graph; adjacent grids are connected by edges according to the Queen adjacency rule, thereby constructing an adjacency matrix; Construct a multidimensional input feature sequence for each node over time; For each time step Input features By applying graph convolution operations and using the adjacency matrix to propagate neighborhood information, we can extract the interaction features between people in local spatial grids. After graph convolution, 1D temporal convolution (TCN) is used to perform sliding window convolution on the time series of each node; The features after temporal convolution are input into a gated recurrent unit (GRU) to capture travel evolution trends under non-stationary disturbances during holidays and morning and evening rush hours. Perform single-step prediction or multi-step rolling prediction tasks, outputting tensor dimensions as follows: , where 2 represents the predicted pedestrian flow.

Images