Multi-scale based urban traffic fine-grained flow prediction method and system
By employing a multi-scale urban traffic flow prediction method, which utilizes multi-head self-attention mechanism and deformable convolution to learn spatial dependencies, the problems of model complexity and neglect of spatial dependencies in existing technologies are solved, thus achieving efficient urban traffic flow prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2023-10-18
- Publication Date
- 2026-06-30
AI Technical Summary
Existing urban traffic flow prediction methods are complex and fail to fully consider the spatiotemporal characteristics of traffic flow data, neglecting spatial dependence at different scales, resulting in low prediction accuracy.
We adopt a multi-scale fine-grained traffic flow prediction method for urban traffic. By constructing global and local spatial dependencies of the city, we obtain multi-scale spatial features. We use multi-head self-attention mechanism and deformable convolution to learn spatial dependencies at different scales, and improve prediction accuracy through feature discrimination loss and structural constraints.
It improves the accuracy of urban traffic flow prediction, simplifies the network structure, shortens the training cycle, and enhances prediction performance.
Smart Images
Figure CN117409581B_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of intelligent transportation technology, specifically to a method and system for fine-grained urban traffic flow prediction based on multiple scales. Background Technology
[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.
[0003] Smart cities represent an innovation that fully utilizes artificial intelligence technology in the field of traffic flow data. This area encompasses road flow monitoring, traffic signal control, traffic information collection, and intelligent transportation, with intelligent transportation playing a crucial role. Intelligent transportation requires granular monitoring and analysis of urban traffic flow to assist urban planners in optimizing road layouts and reducing the risk of traffic congestion.
[0004] Currently, there are two main approaches to urban traffic flow monitoring technology. One approach involves deploying a large number of traffic flow monitoring sensors throughout the city, but this method has high operation and maintenance costs. The other approach utilizes image super-resolution technology to intelligently process urban traffic flow data.
[0005] However, the inventors discovered that existing fine-grained flow prediction methods have some problems, such as complex models, failure to fully consider the spatiotemporal characteristics of flow data, and neglect of key factors such as spatial dependence at different scales. Summary of the Invention
[0006] To address the aforementioned issues, this disclosure proposes a multi-scale urban traffic fine-grained flow prediction method and system. Based on the concept of multi-scale, it constructs urban global spatial dependence and local spatial dependence, fully considering how they complement each other, and provides non-redundant and mutually complementary urban spatial information, thereby improving the accuracy of fine-grained prediction.
[0007] According to some embodiments, the present disclosure adopts the following technical solutions:
[0008] Multi-scale urban traffic fine-grained flow prediction methods include:
[0009] Obtain multiple traffic flow distribution maps of the city to be predicted;
[0010] A fine-grained flow distribution map is constructed based on the flow distribution map, and a coarse-grained flow distribution map is obtained by scaling the fine-grained flow distribution map according to a set scaling factor.
[0011] High-level multi-scale spatial feature extraction is performed on the coarse-grained flow distribution map. The spatial dependency between a certain area of the city and geographically close and distant areas is learned to obtain coarse-grained flow distribution feature maps at different scales. The coarse-grained flow distribution feature maps at different scales are decoupled, and then the decoupled coarse-grained flow distribution feature maps at different scales are fused with complementary multi-scale spatial features to obtain the final regional features.
[0012] The final regional features are used to predict fine-grained urban traffic flow.
[0013] According to some embodiments, the present disclosure adopts the following technical solutions:
[0014] A multi-scale urban traffic fine-grained flow prediction system includes:
[0015] The data acquisition module is used to acquire multiple traffic flow distribution maps of the city to be predicted.
[0016] The data processing module is used to construct a fine-grained flow distribution map based on the flow distribution map, and to scale the fine-grained flow distribution map according to a set scaling factor to obtain a coarse-grained flow distribution map.
[0017] The feature extraction module is used to extract high-level multi-scale spatial features from the coarse-grained flow distribution map, learn the spatial dependency between a certain area of the city and geographically close and distant areas, obtain coarse-grained flow distribution feature maps at different scales, decouple the coarse-grained flow distribution feature maps at different scales, and then perform complementary fusion of multi-scale spatial features on the decoupled coarse-grained flow distribution feature maps at different scales to obtain the final regional features.
[0018] The prediction module is used to predict fine-grained urban traffic flow using the final regional features.
[0019] According to some embodiments, the present disclosure adopts the following technical solutions:
[0020] A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the multi-scale urban traffic fine-grained flow prediction method.
[0021] According to some embodiments, the present disclosure adopts the following technical solutions:
[0022] An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the multi-scale urban traffic fine-grained flow prediction method.
[0023] Compared with the prior art, the beneficial effects of this disclosure are as follows:
[0024] One embodiment of this disclosure provides a multi-scale urban traffic fine-grained flow prediction method. Based on the characteristics of flow data, it learns spatial dependencies at different scales, fully considers the redundancy and complementarity of information between spatial dependencies at different scales, and uses additional factors such as time and weather to make the urban fine-grained flow prediction model have rich spatiotemporal data characteristics, thereby improving the prediction accuracy.
[0025] In the training process of the final fine-grained traffic flow prediction model by combining multi-scale spatial information, structural constraints are set to ensure that the sum of traffic flows in fine-grained regions is strictly equal to the corresponding coarse-grained traffic flow, further improving the prediction performance. Compared with previous urban traffic flow prediction models, it has the advantages of simpler network, shorter training cycle, and better prediction results. Attached Figure Description
[0026] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.
[0027] Figure 1 This is a flowchart of the urban fine-grained traffic prediction method based on multi-scale representation in the embodiments of this disclosure;
[0028] Figure 2 This is a diagram showing the coarse-grained and fine-grained flow distribution and structural constraints in the embodiments of this disclosure;
[0029] Figure 3 This is a flowchart of the region-scale embedding training process in an embodiment of this disclosure;
[0030] Figure 4 This is a flowchart of the city-scale embedding training process in an embodiment of this disclosure;
[0031] Figure 5 This is a schematic diagram of the overall architecture of the urban fine-grained traffic prediction model based on multi-scale representation in the embodiments of this disclosure. Detailed Implementation
[0032] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0033] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.
[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0035] Example 1
[0036] One embodiment of this disclosure provides a multi-scale urban traffic fine-grained flow prediction method, including:
[0037] Step 1: A multi-scale, fine-grained urban traffic flow prediction method.
[0038] Step 2: Construct a fine-grained flow distribution map based on the flow distribution map, and scale the fine-grained flow distribution map according to the set scaling factor to obtain a coarse-grained flow distribution map;
[0039] Step 3: Perform high-level multi-scale spatial feature extraction on the coarse-grained flow distribution map, learn the spatial dependency between a certain area of the city and geographically close and distant areas, and obtain coarse-grained flow distribution feature maps at different scales.
[0040] Step 4: Decouple the coarse-grained flow distribution feature maps at different scales, and then perform complementary fusion of multi-scale spatial features on the decoupled coarse-grained flow distribution feature maps at different scales to obtain the final regional features;
[0041] Step 5: Utilize the final regional features to predict fine-grained urban traffic flow.
[0042] As one embodiment, the multi-scale urban traffic fine-grained flow prediction method disclosed herein aims to infer fine-grained urban flow based on observed coarse-grained flow (pedestrian flow, bicycle flow, vehicle flow). This embodiment discloses a multi-scale representation-based urban fine-grained flow prediction method. It employs a multi-scale approach to predict fine-grained urban flow, designing a complementary network of two scales to enable a simple network structure to obtain powerful representational capabilities from the data. Then, it combines the spatiotemporal characteristics of the flow data to perform fine-grained urban flow prediction. Specifically, flow data has characteristics at different scales. From the city-scale perspective, within a city, there are many urban areas with similar functions (e.g., office areas, entertainment areas, and residential areas), which are geographically different but have similar flow distribution patterns. A global feature learning structure is designed for the city-scale to allow features of urban areas with similar functions to learn from each other, while keeping features of areas with different functions separate. From the regional scale perspective, traffic flow data exhibits proximity. The traffic flow of one region is strongly influenced by nearby regions. A dynamic regional feature learning structure is designed for the regional scale to dynamically integrate features of geographically proximate regions. By extracting features at two different scales, the two encoders learn the multi-scale characteristics of the data. Finally, the two multi-scale fused features are used to perform fine-grained traffic prediction for the city. The prediction process uses a fine-grained traffic prediction model, which includes a dependency learning layer, a decoupling layer, a fusion layer, and an upsampling layer for multi-scale spatial features.
[0043] Furthermore, the decoder is a convolutional layer with a ReLU activation function, which obtains the unique characteristics and mutual complementarity of geographical features at different scales.
[0044] Furthermore, during the training process of the fine-grained flow prediction model, structural constraints are also set, requiring that the sum of the regional flows in the downsampled coarse-grained flow distribution map be equal to the flow magnitude of the corresponding region in the coarse-grained flow distribution map.
[0045] like Figure 1 and Figure 5 As shown, the specific implementation of the multi-scale urban traffic fine-grained flow prediction method of this disclosure is as follows:
[0046] Step 1: Obtain multiple traffic flow distribution maps of the city to be predicted over a period of time, construct a fine-grained traffic flow distribution map based on each traffic flow distribution map, and then construct a coarse-grained traffic flow distribution map.
[0047] The traffic distribution map was obtained from publicly available data websites, including pedestrian data, bicycle data, and motor vehicle data. It was preprocessed to obtain a coarse-grained traffic distribution map from the fine-grained traffic distribution map.
[0048] Further, step 2: construct a fine-grained flow distribution map based on the flow distribution map, and scale the fine-grained flow distribution map according to the set scaling factor to obtain a coarse-grained flow distribution map;
[0049] Based on different latitude and longitude scales, coarse-grained and fine-grained flow distribution maps can be obtained, and the scaling factor N∈Z between the coarse and fine flow distribution maps can be determined. + ,For example Figure 2 The scaling factor N = 2. The final coarse-grained flow distribution map X = [x] is obtained. 1 , ..., x t , ..., x T ] and fine-grained flow distribution map Y=[y 1 , ..., y t , ..., y T X and Y represent coarse-grained and fine-grained traffic flow distribution maps, respectively; and then a fine-grained traffic flow prediction task is carried out for urban traffic.
[0050] Specifically, a fine-grained flow distribution map is obtained, and then the fine-grained flow distribution map is processed according to a set scaling factor to obtain a coarse-grained flow distribution map.
[0051] Step 3: Input the coarse-grained flow distribution map into the fine-grained flow prediction model, and use the multi-scale spatial feature dependency learning layer to learn multi-scale features from the coarse-grained flow distribution map;
[0052] High-level multi-scale spatial feature extraction is performed on the coarse-grained flow distribution map. The spatial dependencies between a certain area of the city and geographically proximate and distant areas are learned to obtain coarse-grained flow distribution feature maps at different scales. Specifically:
[0053] The spatial dependence of a certain area of a city with geographically proximate and distant areas refers to the spatial dependence at both the regional and urban scales. For a given area, the spatial dependence of geographically proximate areas is learned by combining it with surrounding areas, while the spatial dependence of similar functional areas in the entire city is learned by combining it with distant areas with high relevance.
[0054] First, for region-scale features, previous work used standard convolutions to learn the features of a region, where features are obtained by considering information about the region itself and its neighboring regions. However, a limitation of standard convolutions is their inability to effectively handle neighboring regions with irregular shapes. To address this shortcoming, this disclosure introduces an additional two-dimensional offset into the receptive field using deformable convolutions, allowing for more flexible and adaptive sampling of the input feature map. Deformable convolutions are chosen as a component of the encoder. The representation of each region (i, j) at each time step is as follows: Z represents the region, t represents the time step, and we have:
[0055]
[0056] Where R is defined as R = {(-1, -1), (-1, 0), ..., (0, 1), (1, 1)}, (i n j n In R, (Δi) n , Δj n ) is the offset.
[0057] Further processing uses the ReLU activation function to obtain a neighborhood-level representation.
[0058] h b =ReLu(z) t ),
[0059] E B (x t ) = h b .
[0060] Among them, E B () indicates a region encoder, which obtains a region-level scale representation.
[0061] For city-scale features, previous work on modeling spatial dependencies has not considered establishing spatial relationships at the level of distant cities; however, this is crucial for functionally similar but geographically dispersed regions. To capture global urban spatial dependencies and inspired by the advantages of multi-head self-attention mechanisms, this disclosure introduces a multi-head transformer spatial encoder to learn the relationships between any two spatial intervals across spatial scales. The multi-head self-attention (MHA) function can be defined as:
[0062]
[0063] Among them, w q w k w v It is the weight of multi-head self-attention, x p t It is coarse-grained traffic flow with location coding.
[0064] The multi-head converter spatial encoder consists of a multi-head self-attention layer, layer normalization (LN) in each block, and residual connections.
[0065] h c =LN(x t +z c ),
[0066] E c (x t ) = hc .
[0067] Here, Ec() represents the city encoder, which obtains a city-scale representation.
[0068] Step 4: In the decoupling layer, the coarse-grained flow distribution feature maps at different scales are decoupled to remove redundant information from features at different scales. A loss function is designed to remove duplicate information, and duplicates are removed through features at different scales, including:
[0069] Representations at different scales may inadvertently learn redundant geographic information. Both neighborhood-level and city-level representations may contain spatial details surrounding a specific area. However, the intent of this disclosure is for neighborhood-level representations to capture finer regional nuances, while city-level representations should present a broader context of the area within the urban landscape. The goal is to avoid representations of the same area being too similar at different scales. To achieve this, this disclosure introduces a feature discrimination loss to reduce redundancy.
[0070] The formula for feature discrimination loss is as follows:
[0071]
[0072] Among them, h b i，j h represents the regional scale. c i,j This indicates a city-level scale representation.
[0073] Finally, mean squared error and weighted feature discrimination loss are used as the training loss for the model.
[0074] Step 5: In the fusion layer, the coarse-grained flow distribution feature maps of different scales after decoupling are fused with complementary spatial features at multiple scales to obtain the final regional features for the prediction of fine-grained urban traffic flow.
[0075] To address the challenge of integrating geographic information at different scales, a private decoder is used to acquire specific information at each scale, while a shared decoder is used to ensure the interaction between information at different scales.
[0076] The architecture of private decoders can be designed based on neighborhood-level and city-level encoders. They employ similar network architectures to ensure consistency and interoperability. Specifically, the private neighborhood-level decoder uses a deformable convolutional structure, while the private city-level decoder uses a Transformer decoder structure. In summary, private and shared decoders work together to maintain the specificity of information at each scale while promoting collaborative integration to improve the performance of multi-scale geographic information representation and inference.
[0077] ob =D B (h n ),
[0078] o c =D C (h c ).
[0079] Among them, the regional private decoder uses D B () indicates that the city's private decoder uses D C () indicates that h n h c These are the two latent variables learned at different scales.
[0080] The shared decoder blends representations from the neighborhood level and the city level through simple convolutional operations.
[0081] o bc =D BC (concat(h b h c )).
[0082] Among them, the interactive decoder uses D BC ()express.
[0083] Private and shared decoders work together to maintain the specificity of information at each scale while promoting collaborative integration to improve the performance of multi-scale geographic information representation and inference.
[0084] After passing a coarse-grained traffic map with additional information through the encoder and decoder, the upsampling layer employs a PixelShuffle (PS) layer to increase their size using a magnification factor. Inspired by the distributed upsampling technique used in UrbanFM, an M... 2 The normalization method ensures that the sum of the subregions equals the sum of their corresponding superregions. This normalization scheme can be defined as:
[0085]
[0086] To satisfy structural constraints, for x t Perform nearest neighbor interpolation upsampling, then with W t Element-wise multiplication to obtain fine-grained images
[0087]
[0088] The traditional mean squared error (MSE) loss is combined with the weighted feature discrimination loss as the overall loss of the model, which is used to optimize the model:
[0089]
[0090] L = L MSE +λL d ,
[0091] Among them, y t This represents a fine-grained flow plot at time t.
[0092] Example 2
[0093] One embodiment of this disclosure provides a multi-scale urban traffic fine-grained flow prediction system, including:
[0094] The data acquisition module is used to acquire multiple traffic flow distribution maps of the city to be predicted.
[0095] The data processing module is used to construct a fine-grained flow distribution map based on the flow distribution map, and to scale the fine-grained flow distribution map according to a set scaling factor to obtain a coarse-grained flow distribution map.
[0096] The feature extraction module is used to extract high-level multi-scale spatial features from the coarse-grained flow distribution map, learn the spatial dependency between a certain area of the city and geographically close and distant areas, obtain coarse-grained flow distribution feature maps at different scales, decouple the coarse-grained flow distribution feature maps at different scales, and then perform complementary fusion of multi-scale spatial features on the decoupled coarse-grained flow distribution feature maps at different scales to obtain the final regional features.
[0097] The prediction module is used to predict fine-grained urban traffic flow using the final regional features.
[0098] Example 3
[0099] One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the multi-scale urban traffic fine-grained flow prediction method.
[0100] Example 4
[0101] One embodiment of this disclosure provides an electronic device, including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the multi-scale urban traffic fine-grained flow prediction method.
[0102] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0104] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.
Claims
1. A multi-scale urban traffic fine-grained flow prediction method, characterized in that, include: Obtain multiple traffic flow distribution maps of the city to be predicted; A fine-grained flow distribution map is constructed based on the flow distribution map, and a coarse-grained flow distribution map is obtained by scaling the fine-grained flow distribution map according to a set scaling factor. High-level multi-scale spatial feature extraction is performed on the coarse-grained flow distribution map. The spatial dependency between a certain area of the city and geographically close and distant areas is learned to obtain coarse-grained flow distribution feature maps at different scales. The coarse-grained flow distribution feature maps at different scales are decoupled, and then the decoupled coarse-grained flow distribution feature maps at different scales are fused with complementary multi-scale spatial features to obtain the final regional features. The method for obtaining the coarse-grained flow distribution feature maps at different scales is as follows: High-level multi-scale spatial feature extraction is performed on the coarse-grained flow distribution map. For the regional scale features of a certain region and geographically similar regions, an additional two-dimensional offset is introduced into the receptive field in the variable convolution to adaptively sample the input coarse-grained flow distribution map. The deformable convolution is used as a component of the encoder to obtain the representation of each region in each time step. For the urban-scale characteristics of a certain region and a remote region, a multi-head converter spatial encoder is introduced to learn the relationship between any two spatial intervals across spatial scales. The multi-head converter spatial encoder consists of a multi-head self-attention layer, layer normalization in each block, and residual connections. Decoupling coarse-grained flow distribution feature maps at different scales includes: enabling representations at the level of geographically proximate regions to capture finer regional differences, while representing remote regions at the city level to present a broader context of the city and region; and introducing feature discrimination loss to decouple coarse-grained flow distribution feature maps at different scales. The decoupled coarse-grained flow distribution feature maps at different scales are subjected to complementary fusion of multi-scale spatial features. This includes: using a private decoder to obtain specific information of the feature map at each scale, while using a shared decoder to ensure the interaction between feature map information at different scales. Then, the graph is fused with scale-specific information and interaction information through weights. The structure of the private decoder is designed based on neighborhood-level and city-level encoders. The private neighborhood-level decoder adopts a deformable convolutional structure, while the private city-level decoder adopts a Transformer decoder structure. The final regional features are used to predict fine-grained urban traffic flow. During the training of the fine-grained traffic flow prediction model, structural constraints are also set, requiring that the sum of regional flows in the downsampled coarse-grained traffic flow distribution map is equal to the corresponding regional flow in the coarse-grained traffic flow distribution map.
2. The multi-scale urban traffic fine-grained flow prediction method as described in claim 1, characterized in that, The traffic distribution map is obtained based on the traffic distribution data of the city to be predicted, which includes pedestrian data, bicycle data, and motor vehicle data.
3. The multi-scale urban traffic fine-grained flow prediction method as described in claim 1, characterized in that, We further use the ReLU activation function to obtain representations at the level of geographically proximate regions.
4. A multi-scale urban traffic fine-grained flow prediction system, employing the multi-scale urban traffic fine-grained flow prediction method as described in any one of claims 1-3, characterized in that, include: The data acquisition module is used to acquire multiple traffic flow distribution maps of the city to be predicted. The data processing module is used to construct a fine-grained flow distribution map based on the flow distribution map, and to scale the fine-grained flow distribution map according to a set scaling factor to obtain a coarse-grained flow distribution map. The feature extraction module is used to extract high-level multi-scale spatial features from the coarse-grained flow distribution map, learn the spatial dependency between a certain area of the city and geographically close and distant areas, obtain coarse-grained flow distribution feature maps at different scales, decouple the coarse-grained flow distribution feature maps at different scales, and then perform complementary fusion of multi-scale spatial features on the decoupled coarse-grained flow distribution feature maps at different scales to obtain the final regional features. The prediction module is used to predict fine-grained urban traffic flow using the final regional features.
5. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the multi-scale fine-grained urban traffic flow prediction method as described in any one of claims 1-3.
6. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the multi-scale urban traffic fine-grained flow prediction method as described in any one of claims 1-3.