Dynamic traffic distribution-oriented traffic demand prediction method and system

Abstract

The invention relates to the field of dynamic traffic distribution, and particularly discloses a traffic demand prediction method and system for dynamic traffic distribution, and the method comprises the steps: firstly obtaining historical traffic flow data and traffic demand external driving data collected by a database, and real-time traffic video data collected by a camera, and then carrying out the deep learning technology, the method comprises the following steps: carrying out feature extraction and correlation analysis on the three, and finally generating a dynamic traffic distribution suggestion through a generator, thereby carrying out timely adjustment and optimization according to a real-time condition, alleviating traffic congestion, improving travel efficiency, and further improving the real-time performance and flexibility of traffic distribution.

Description

Technical Field

[0001] The present application relates to the field of dynamic traffic allocation, and more specifically, to a traffic demand prediction method and system for dynamic traffic allocation. Background Art

[0002] With the rapid development of urban transportation, people's transportation needs are increasing day by day. Among them, by predicting the amount of transportation demand in each area of ​​the city, transportation resources can be reasonably allocated in advance, thereby alleviating traffic congestion, saving travelers' travel time and avoiding waste of public resources.

[0003] At present, most traffic demand forecasting models rely on building static spatial networks, which have a fixed structure and do not change over time. Although this method can reflect the operating status of the transportation system to a certain extent, it lacks the ability to adjust dynamically and cannot fully capture the complex spatiotemporal dependencies between regions or stations. In addition, existing methods often lack support from deeper data mining techniques and advanced forecasting models, which further limits the improvement of forecasting accuracy.

[0004] Therefore, a traffic demand prediction method and system for dynamic traffic allocation is desired. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed. The embodiment of this application provides a traffic demand prediction method and system for dynamic traffic distribution, which first obtains historical traffic flow data collected by the database, traffic demand external driving data and real-time traffic video data collected by the camera, and then uses deep learning technology to perform feature extraction and correlation analysis on the three, and finally generates dynamic traffic distribution suggestions through a generator, so as to make timely adjustments and optimizations according to real-time conditions, alleviate traffic congestion, improve travel efficiency, and thus improve the real-time and flexibility of traffic distribution.

[0006] According to one aspect of the present application, a traffic demand forecasting method for dynamic traffic allocation is provided, which includes: Obtain historical traffic flow data collected by the database, external driving data of traffic demand, and real-time traffic video data collected by cameras; Extracting a traffic flow analysis multimodal correlation feature vector and a real-time traffic global feature vector from the historical traffic flow data collected by the database, the traffic demand external driving data, and the real-time traffic video data collected by the camera; Based on the traffic flow analysis multimodal association feature vector and the real-time traffic global feature vector, a dynamic traffic allocation suggestion is generated.

[0007] According to another aspect of the present application, a traffic demand forecasting system for dynamic traffic allocation is provided, comprising: Dynamic traffic demand forecasting data acquisition module, used to acquire historical traffic flow data collected by the database, traffic demand external driving data and real-time traffic video data collected by the camera; A dynamic traffic demand prediction data extraction module, used to extract traffic flow analysis multimodal correlation feature vectors and real-time traffic global feature vectors from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera; The dynamic traffic allocation suggestion generating module is used to generate dynamic traffic allocation suggestions based on the traffic flow analysis multimodal correlation feature vector and the real-time traffic global feature vector.

[0008] Compared with the prior art, the present application provides a traffic demand prediction method and system for dynamic traffic distribution, which first obtains historical traffic flow data collected by a database, traffic demand external driving data and real-time traffic video data collected by a camera, and then uses deep learning technology to perform feature extraction and correlation analysis on the three, and finally generates dynamic traffic distribution recommendations through a generator, so as to make timely adjustments and optimizations according to real-time conditions, alleviate traffic congestion, improve travel efficiency, and thereby improve the real-time and flexibility of traffic distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 The present invention is a flowchart of a traffic demand prediction method for dynamic traffic allocation according to an embodiment of the present application.

[0011] Figure 2 The present invention is a flow chart of extracting traffic flow analysis multimodal correlation feature vectors and real-time traffic global feature vectors from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera in the traffic demand prediction method for dynamic traffic distribution according to an embodiment of the present application.

[0012] Figure 3The present invention is a flow chart of extracting features from the historical traffic flow data collected by the database to obtain a semantically associated feature vector of the historical traffic flow in the traffic demand prediction method for dynamic traffic allocation according to an embodiment of the present application.

[0013] Figure 4 The present invention is a flow chart of extracting features from the traffic demand external driving data to obtain a traffic demand external driving semantic feature vector in the traffic demand prediction method for dynamic traffic allocation according to an embodiment of the present application.

[0014] Figure 5 It is a block diagram of a traffic demand prediction system for dynamic traffic allocation according to an embodiment of the present application. DETAILED DESCRIPTION

[0015] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0016] Figure 1 FIG. 1 is a block diagram of a traffic demand prediction method for dynamic traffic allocation according to an embodiment of the present application. Figure 1 As shown, the traffic demand prediction method for dynamic traffic distribution according to the embodiment of the present application includes: S110, acquiring historical traffic flow data collected by a database, traffic demand external driving data and real-time traffic video data collected by a camera; S120, extracting traffic flow analysis multimodal correlation feature vectors and real-time traffic global feature vectors from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera; S130, generating dynamic traffic distribution recommendations based on the traffic flow analysis multimodal correlation feature vectors and the real-time traffic global feature vectors.

[0017] In the above-mentioned traffic demand prediction method for dynamic traffic allocation, the step S110 obtains historical traffic flow data collected by the database, traffic demand external driving data and real-time traffic video data collected by the camera. It should be understood that the historical traffic flow data records the traffic conditions of each road section in the past period of time, including information such as vehicle flow and speed, and provides a basis for long-term trend analysis for the model; the traffic demand external driving data covers factors such as weather, holidays, special events, etc., reflecting the external conditions that may affect traffic demand; and the real-time traffic video data captures the current road conditions through the camera and provides instant traffic flow information. Specifically, historical traffic flow data is usually stored in the database of the traffic management department or relevant institutions, which has accumulated and maintained a wealth of traffic history records for a long time. These data can be imported into the prediction system for processing by interface calling or batch export. Secondly, the traffic demand external driving data comes from multiple channels, such as weather forecast data provided by the Meteorological Bureau, holiday arrangements issued by government departments, and emergency information on social media. Such data needs to be cleaned, formatted and standardized before it can be used for analysis. Finally, the real-time traffic video data is collected by cameras deployed on urban main roads, intersections, etc. These cameras work around the clock, generating a large amount of video stream data. By integrating the above three types of data, we can not only capture the static characteristics of the traffic system, but also dynamically monitor its changing patterns. This method not only improves the accuracy of traffic demand forecasts, but also enhances the real-time and flexibility of traffic allocation recommendations, thereby effectively alleviating traffic congestion and improving overall travel efficiency.

[0018] Specifically, with the rapid development of urban transportation, residents' traffic demand continues to increase. In this case, predicting the traffic demand in each area can reasonably allocate traffic resources in advance, thereby alleviating traffic congestion, saving travel time, and avoiding waste of public resources. At present, most traffic demand prediction models rely on establishing a static spatial network, which is a fixed network structure and does not change over time. Although this method can reflect the operating status of the traffic system to a certain extent, it cannot fully capture the complex spatiotemporal dependencies between regions or sites due to the lack of dynamic adjustment capabilities. In addition, existing prediction methods usually lack in-depth data mining technology and advanced prediction model support, which makes it difficult to further improve the prediction accuracy. In the technical solution of the present application, by obtaining historical traffic flow data collected by the database, traffic demand external driving data and real-time traffic video data collected by the camera, and combining deep learning technology, a dynamic traffic resource allocation suggestion is generated to make timely adjustments and optimizations according to real-time traffic conditions, thereby alleviating traffic congestion, improving travel efficiency, and enhancing the real-time and flexibility of traffic distribution.

[0019] In the above-mentioned traffic demand prediction method for dynamic traffic distribution, the step S120 extracts the traffic flow analysis multimodal correlation feature vector and the real-time traffic global feature vector from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera. It should be understood that by extracting the traffic flow analysis multimodal correlation feature vector and the real-time traffic global feature vector, the system can fully integrate historical data, external driving factors and real-time monitoring information to build a more accurate and dynamic traffic demand prediction model, so as to convert data from different sources and different forms into a unified feature representation, thereby providing a basis for generating accurate dynamic traffic distribution recommendations. This multimodal feature extraction method significantly captures the complex spatiotemporal dependencies of the traffic system, and integrates multi-source information to enhance the expressiveness of the prediction model, thereby improving the real-time and flexibility of traffic distribution recommendations, and providing strong technical support for alleviating traffic congestion and optimizing travel efficiency.

[0020] Figure 2 The present invention is a flow chart of extracting a traffic flow analysis multimodal correlation feature vector and a real-time traffic global feature vector from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera in the traffic demand prediction method for dynamic traffic allocation according to an embodiment of the present application. Figure 2 As shown, in a specific embodiment of the present application, the step S120, extracting a traffic flow analysis multimodal association feature vector and a real-time traffic global feature vector from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera, includes: S121, performing feature extraction on the historical traffic flow data collected by the database to obtain a historical traffic flow semantic association feature vector; S122, performing feature extraction on the traffic demand external driving data to obtain a traffic demand external driving semantic feature vector; S123, associating the historical traffic flow semantic association feature vector with the traffic demand external driving semantic feature vector to obtain the traffic flow analysis multimodal association feature vector; S124, performing feature extraction on the real-time traffic video data collected by the camera to obtain the real-time traffic global feature vector.

[0021] In step S121, feature extraction is performed on the historical traffic flow data collected by the database to obtain a historical traffic flow semantic association feature vector. It should be understood that the original historical traffic flow data usually contains a large amount of time series information, such as vehicle volume, average speed, congestion level, etc. These data are difficult to be used directly for modeling without processing. By extracting features from these data, complex multidimensional data can be reduced to representative feature vectors while retaining their core information. This feature vector is not only easy to calculate and analyze, but also can better reflect the spatiotemporal characteristics of traffic flow and its changing laws.

[0022] In step S122, feature extraction is performed on the traffic demand external driving data to obtain a traffic demand external driving semantic feature vector. It should be understood that traditional prediction methods often rely only on historical traffic flow data, ignoring the significant impact that external driving factors may have. For example, bad weather may cause a sharp drop in traffic volume on certain sections of road, while holidays may trigger traffic peaks in specific areas. By extracting features from traffic demand external driving data, the impact of these external factors can be quantified and incorporated into the prediction model, thereby improving the comprehensiveness and adaptability of the prediction.

[0023] In step S123, the historical traffic flow semantic association feature vector and the traffic demand external driving semantic feature vector are associated to obtain the traffic flow analysis multimodal association feature vector. It should be understood that although the historical traffic flow data can reflect the long-term trend of the traffic system, it cannot explain the short-term fluctuations caused by external factors (such as weather, holidays, etc.) alone. The traffic demand external driving data provides a quantitative description of these external conditions, but lacks a direct connection with the historical traffic flow. By combining the two, the correlation between historical data and external driving factors can be established, thereby achieving a multi-dimensional understanding of traffic demand changes. Specifically, by designing an association algorithm, the two feature vectors are merged into a unified representation. This association algorithm may use weighted summation, splicing or attention mechanism to ensure that the information of the two feature vectors can fully interact and complement each other. Among them, the weighted summation method can assign different weights according to the importance of historical data and external driving factors; the splicing method simply combines the two feature vectors into a higher-dimensional vector; the attention mechanism allows the model to dynamically adjust the degree of attention to different features, so as to better capture complex traffic demand patterns. By clarifying the relationship between historical traffic flow data and external driving factors, the model can clearly show which external factors have a significant impact on traffic demand and how these impacts change over time. Through multimodal feature fusion, the system can comprehensively utilize multiple sources of information and reduce the risk of anomalies caused by a single data source.

[0024] In step S124, feature extraction is performed on the real-time traffic video data collected by the camera to obtain the real-time traffic global feature vector. It should be understood that real-time traffic video data usually contains a large amount of redundant information (such as background noise, irrelevant objects, etc.). Directly using the original video data for analysis not only has high computational costs, but may also lead to overfitting of the model or unstable prediction results. By extracting features from video data, these complex multidimensional data can be reduced to representative feature vectors while retaining their core information. This feature vector is not only easy to calculate and analyze, but also can better reflect the spatial distribution and temporal variation of the current traffic status.

[0025] Figure 3 The present invention is a flow chart of extracting features from the historical traffic flow data collected from the database to obtain a historical traffic flow semantic association feature vector in the traffic demand prediction method for dynamic traffic allocation according to an embodiment of the present application. Figure 3 As shown, in a specific embodiment of the present application, the step S121, performing feature extraction on the historical traffic flow data collected by the database to obtain a historical traffic flow semantic association feature vector, includes: S1211, passing the historical traffic flow data collected by the database through a historical traffic flow semantic encoder including an embedding layer to obtain a plurality of historical traffic flow feature vectors; S1212, cascading the plurality of historical traffic flow feature vectors into the historical traffic flow semantic association feature vector.

[0026] In step S1211, the historical traffic flow data collected by the database is passed through a historical traffic flow semantic encoder including an embedding layer to obtain multiple historical traffic flow feature vectors. It should be understood that historical traffic flow data is usually stored in the form of time series, including multi-dimensional variables such as vehicle flow and speed. Directly using these raw data will increase the complexity of model training and may introduce noise and redundant information. The role of the embedding layer is to map these discrete or continuous values ​​to a low-dimensional space, in which similar data points are geometrically closer, so that the model can learn the inherent laws of the data more efficiently. For example, for the traffic flow data of a certain section of road, the embedding layer can convert it from the original time-flow two-dimensional coordinates to a fixed-length vector, which not only retains the main features of the original data, but also enhances its computability. Semantic encoders are usually implemented based on deep learning technologies, such as long short-term memory networks (LSTM), gated recurrent units (GRU) or Transformers. These models are good at capturing long-term and short-term dependencies in time series data and can mine hidden patterns and trends from historical traffic flow data. Specifically, the semantic encoder analyzes the low-dimensional vectors output by the embedding layer one by one to generate a set of new feature vectors, each of which corresponds to the traffic status in a specific time period. Specifically, first, the historical traffic flow data collected by the database is input into the embedding layer. The embedding layer performs a dimensionality reduction operation on each data point to generate a vector representation of a fixed length. Next, the embedded vector sequence is sent to the historical traffic flow semantic encoder. The encoder processes each vector in the sequence layer by layer to extract time dependency and semantic information. Finally, the encoder outputs multiple historical traffic flow feature vectors, each of which represents the status of traffic flow in a certain time period and its relationship with other time periods. These feature vectors not only contain local traffic information, but also integrate global time dependency characteristics. In the technical solution of the present application, the historical traffic flow data collected from the database is passed through a historical traffic flow semantic encoder including an embedding layer to obtain a plurality of historical traffic flow feature vectors, including: performing word segmentation processing on the historical traffic flow data collected from the database to obtain a historical traffic flow word sequence; using the embedding layer of the historical traffic flow semantic encoder including the embedding layer to respectively map each historical traffic flow word in the historical traffic flow word sequence into a word embedding vector to obtain a sequence of historical traffic flow word embedding vectors; using the converter-based Bert model of the historical traffic flow semantic encoder including the embedding layer to perform global context semantic encoding on the sequence of historical traffic flow word embedding vectors to obtain a plurality of historical traffic flow feature vectors.

[0027] In step S1212, the multiple historical traffic flow feature vectors are cascaded into the historical traffic flow semantic association feature vector. It should be understood that each historical traffic flow feature vector usually only reflects the traffic status within a specific time period, and the operation of the traffic system is a dynamic and continuous process. A single time period feature is difficult to fully describe its complexity. By cascading multiple feature vectors into a larger vector, the traffic status information of different time periods can be integrated together, thereby capturing the changing patterns and correlations of traffic flow in the time dimension. The cascaded semantic association feature vector can simultaneously contain information from these different time periods, allowing the model to better understand the global characteristics of the traffic system.

[0028] Figure 4 The present invention is a flow chart of extracting features from the external driving data of traffic demand to obtain a semantic feature vector of the external driving data of traffic demand in the traffic demand prediction method for dynamic traffic allocation according to an embodiment of the present application. Figure 4 As shown, in a specific embodiment of the present application, the step S122, performing feature extraction on the traffic demand external driving data to obtain a traffic demand external driving semantic feature vector, includes: S1221, passing the traffic demand external driving data through a traffic demand external driving data analyzer to obtain a sequence of traffic demand external driving word vectors; S1222, passing the sequence of traffic demand external driving word vectors through a traffic demand external driving semantic context encoder to obtain the traffic demand external driving semantic feature vector.

[0029] In step S1221, the traffic demand external driving data is passed through a traffic demand external driving data analyzer to obtain a sequence of traffic demand external driving word vectors. It should be understood that traffic demand external driving data usually exists in text or other unstructured forms. These data themselves are not suitable for direct input into a deep learning model for analysis. By generating a word vector sequence through a traffic demand external driving data analyzer, these text information can be converted into a numerical representation, so that the model can better understand and utilize the influence of these external factors. Specifically, the process of generating a traffic demand external driving word vector sequence is divided into the following steps: The first step is to preprocess the original external driving data. This includes operations such as word segmentation, removal of stop words, and standardization, the purpose of which is to decompose text data into smaller units (such as words or phrases) for subsequent processing. The second step is to generate word vectors through a traffic demand external driving data analyzer. Data analyzers are usually implemented based on natural language processing technology, such as word embedding methods (such as Word2Vec, GloVe, or BERT). These methods can map each word or phrase to a fixed-length vector space in which similar words are geometrically closer. The third step is to arrange all word vectors in order to form a word vector sequence, which fully represents the content and semantic information of the original external driving data. By converting text data into a word vector sequence, the system can not only capture the meaning of individual words, but also retain the context of the entire sentence or paragraph. In this way, it can adapt to the complexity and diversity of external driving data and improve the model's ability to understand these factors.

[0030] In the step S1222, the sequence of the traffic demand external driving word vectors is passed through the traffic demand external driving semantic context encoder to obtain the traffic demand external driving semantic feature vector. It should be understood that although the sequence of traffic demand external driving word vectors can reflect the meaning of a single word or phrase, it lacks the overall understanding of the entire text context. Through the traffic demand external driving semantic context encoder, these word vector sequences can be converted into higher-level semantic feature vectors, thereby integrating context information and revealing the complex relationship between external driving factors. This semantic feature vector not only contains the independent meaning of each word, but also reflects their combined effect in the entire text, so that the model can better understand the impact of external driving factors on traffic demand. In the technical solution of the present application, the sequence of the traffic demand external driving word vectors is passed through the traffic demand external driving semantic context encoder to obtain the traffic demand external driving semantic feature vector, including: using the converter-based Bert model of the traffic demand external driving semantic context encoder to perform global context semantic encoding on the sequence of word embedding vectors to obtain multiple demand external driving feature vectors; and, cascading the multiple demand external driving feature vectors to obtain the traffic demand external driving semantic feature vector.

[0031] In a specific embodiment of the present application, the step S124, performing feature extraction on the real-time traffic video data collected by the camera to obtain the real-time traffic global feature vector, includes: S1241, performing local feature extraction on the real-time traffic video data collected by the camera to obtain multiple real-time traffic key frame local feature maps; S1242, performing global feature extraction on the multiple real-time traffic key frame local feature maps to obtain the real-time traffic global feature vector.

[0032] In step S1241, local feature extraction is performed on the real-time traffic video data collected by the camera to obtain multiple local feature maps of real-time traffic key frames. It should be understood that the significance of local feature extraction is to focus on key areas and important information in the video. Real-time traffic video data usually contains a large amount of background noise and irrelevant details (such as static buildings, sky, etc.), which are not important for traffic status analysis and may even interfere with model learning. Through local feature extraction, the system can focus on dynamic objects in the video (such as vehicles, pedestrians, etc.) and their motion characteristics, so as to more efficiently capture the changing patterns of traffic flow.

[0033] In a specific embodiment of the present application, the step S1241 performs local feature extraction on the real-time traffic video data collected by the camera to obtain multiple real-time traffic key frame local feature maps, including: extracting key frames from the real-time traffic video data collected by the camera to obtain multiple real-time traffic video key frames; passing the multiple real-time traffic video key frames through a real-time traffic key frame local feature extractor based on a deep residual network to obtain the multiple real-time traffic key frame local feature maps.

[0034] In the technical solution of the present application, key frames are extracted from the real-time traffic video data collected by the camera to obtain multiple real-time traffic video key frames. It should be understood that since real-time traffic video data usually contains a large number of continuous frames, and there is a high temporal correlation between these frames (that is, the contents of adjacent frames are similar), directly processing all frames will result in unnecessary computational overhead and storage burden. By extracting key frames, the most representative frames can be screened out, the core information in the traffic scene can be retained, and repeated or irrelevant parts can be ignored. Specifically, first, the selection criteria for key frames need to be defined. In traffic scenes, key frames usually refer to those frames that can reflect significant changes in traffic conditions, such as the moment when a vehicle enters or leaves a certain area, the state of a signal light switches, and pedestrians cross the road. In order to identify these key frames, a motion detection-based method can be used to analyze the differences between adjacent frames to determine whether there are significant dynamic changes. For example, pixel-level motion information can be detected by calculating the frame difference (Frame Difference) or the optical flow field (OpticalFlow). If a frame has a large change compared to the previous frame, the frame may be a key frame. Secondly, the keyframe extraction process can be further optimized by combining background modeling technology. Background modeling is a commonly used computer vision method to distinguish between background areas and foreground objects in videos. By building a stable background model, moving targets (such as vehicles, pedestrians, etc.) in the video can be separated. Only when there is a significant foreground change in the video frame is it marked as a keyframe. This method not only improves the accuracy of keyframe extraction, but also reduces the possibility of misjudgment. Finally, the extracted multiple real-time traffic video keyframes will become the basic input for subsequent local feature extraction.

[0035] In the technical solution of the present application, the multiple real-time traffic video key frames are respectively passed through the real-time traffic key frame local feature extractor based on the deep residual network to obtain the multiple real-time traffic key frame local feature maps. It should be understood that although the key frames have screened out the most representative traffic scene information, they are still original image data and contain a large amount of redundant information (such as background noise, irrelevant details, etc.), which is not important for traffic state analysis and may even interfere with the learning of the model. Through the local feature extractor based on the deep residual network, the spatial features in the key frames can be analyzed layer by layer, and local information related to traffic, such as low-level features such as vehicle shape, color, edge and road signs, can be extracted, and further integrated into a higher-level semantic representation. This feature extraction process not only reduces the data dimension, but also enhances the ability to focus on core information. Specifically, the working principle of the local feature extractor based on the deep residual network is as follows: The first step is to input each real-time traffic video key frame into the ResNet model. ResNet is a classic deep convolutional neural network. Its core idea is to alleviate the gradient vanishing problem in deep network training by introducing residual connections (Skip Connection), thereby allowing a deeper network design. The second step is to gradually extract the spatial features of the image through the multi-layer convolution operation of ResNet. For example, in the shallow part of the network, the convolution layer can identify low-level features in the key frame, such as texture, edge, and color; while in the deep part, more complex high-level features can be captured, such as the overall shape of the vehicle, the state of the traffic light, or the movement of pedestrians. The third step is to output the local feature map corresponding to each key frame. The local feature map is a multi-channel matrix, each channel represents a specific feature distribution. For example, a channel may reflect the distribution of densely populated areas of vehicles, while another channel may reflect the location of road signs. In this way, a high-level feature representation that can reflect the core information of the traffic scene can be extracted from each key frame to adapt to the problem that the original image data is high-dimensional, complex, and difficult to directly analyze. In the technical solution of the present application, the multiple real-time traffic video key frames are respectively passed through a real-time traffic key frame local feature extractor based on a deep residual network to obtain the multiple real-time traffic key frame local feature maps, including: using each layer of the real-time traffic key frame local feature extractor based on a deep residual network to perform convolution processing, mean pooling processing based on a local feature matrix and nonlinear activation processing on the input data in the forward transmission of the layer to output the multiple real-time traffic key frame local feature maps from the last layer of the real-time traffic key frame local feature extractor based on the deep residual network, wherein the input of the real-time traffic key frame local feature extractor based on the deep residual network is the multiple real-time traffic video key frames.

[0036] In step S1242, global feature extraction is performed on the local feature maps of the multiple real-time traffic key frames to obtain the real-time traffic global feature vector. It should be understood that although the local feature maps can capture important information of specific areas in each key frame, this information is usually scattered and isolated and cannot be directly used to describe the overall state of the entire traffic scene. For example, a single local feature map may reflect the vehicle density or traffic light status of a certain section of road, but it lacks sufficient expression ability for the global traffic conditions of the entire intersection or road network. Therefore, it is necessary to fuse the information of multiple local feature maps through global feature extraction to form a comprehensive global feature vector. This global feature vector not only contains the core information of each local feature map, but also can reveal the correlation between them, thereby more accurately reflecting the overall operating status of the traffic system.

[0037] In a specific embodiment of the present application, the step S1242, performing global feature extraction on the multiple real-time traffic key frame local feature maps to obtain the real-time traffic global feature vector, includes: splicing the multiple real-time traffic key frame local feature maps into a real-time traffic splicing feature map; passing the real-time traffic splicing feature map through a real-time traffic global feature extractor based on a deep residual network to obtain a real-time traffic global feature map; performing global mean pooling on the real-time traffic global feature map to obtain the real-time traffic global feature vector.

[0038] In the technical solution of the present application, the multiple local feature maps of real-time traffic key frames are spliced ​​into a real-time traffic spliced ​​feature map. It should be understood that each local feature map of a real-time traffic key frame is a high-level feature representation extracted from a single key frame, and these feature maps may contain local information such as vehicle distribution, road conditions, and pedestrian activities. However, a single key frame local feature map can usually only reflect the traffic status at a certain moment or within a small range, and cannot fully describe the dynamic changes of the entire traffic scene. Therefore, by splicing multiple local feature maps together, a comprehensive analysis of multi-frame data can be achieved, thereby capturing traffic characteristics over a longer time span and a larger spatial range. Common splicing methods include channel-wise concatenation and spatial concatenation. Channel-wise concatenation is to directly superimpose the number of channels of multiple feature maps to generate a new high-dimensional feature map; spatial-dimensional concatenation is to expand the feature map along the width or height direction to form a larger two-dimensional feature map. In this method, considering the spatial distribution characteristics of the traffic scene and the requirements of computational efficiency, the channel-wise concatenation method is adopted. In this way, the information from multiple key frames can be effectively fused to form a higher-level feature representation. This spliced ​​feature map not only retains the detailed information in the original local feature map, but also captures the time series change law of the traffic scene through the correlation between adjacent frames.

[0039] In the technical solution of the present application, the real-time traffic splicing feature map is passed through a real-time traffic global feature extractor based on a deep residual network to obtain a real-time traffic global feature map. It should be understood that the real-time traffic splicing feature map integrates local feature information from multiple key frames, but it still remains at a lower level of feature representation. These features may contain detailed information such as vehicle distribution and road conditions, but have not yet formed an overall description of the entire traffic scene. In order to capture deeper spatial and temporal dependencies, it is necessary to further process with the powerful expression ability of deep learning models. The real-time traffic global feature extractor based on the deep residual network (Deep Residual Network, ResNet) has become an ideal choice for achieving this goal due to its excellent performance and wide range of application scenarios. The core idea of ​​the deep residual network is to alleviate the gradient vanishing or gradient explosion problems that may occur during the training of the deep neural network by introducing residual connections, so that the network can learn complex nonlinear mapping relationships more effectively. In the task of extracting global features of real-time traffic, the deep residual network can gradually extract high-level features in the splicing feature map through multi-layer convolution operations. For example, low-level convolutional layers are mainly responsible for capturing local texture and edge information, while high-level convolutional layers focus on identifying structured patterns in a larger range, such as vehicle clusters, road layouts, etc. In addition, the design of residual connections can also promote the flow of information in the network, ensuring that the underlying features will not be over-smoothed or lost due to the increase in the number of layers. Specifically, the real-time traffic splicing feature map is input into the global feature extractor based on the deep residual network, which usually includes the following steps: first, the splicing feature map is passed as input to the first convolution layer of the network, and preliminary local features are extracted through the convolution kernel; second, these local features are processed layer by layer using a series of residual blocks. In each residual block, the input feature map is output after two convolution operations and one residual connection; finally, the final real-time traffic global feature map is generated through a pooling layer or other dimensionality reduction operations. This feature map not only retains the important information in the original splicing feature map, but also enhances its semantic expression ability through a multi-level feature extraction process. In this way, deep spatial and temporal dependencies can be further explored from the concatenated feature map to generate a higher-level, more semantically meaningful global feature representation to enhance the model's ability to understand complex traffic scenarios.

[0040] In the technical solution of the present application, the real-time traffic global feature map is subjected to global mean pooling to obtain the real-time traffic global feature vector. It should be understood that the real-time traffic global feature map is usually a three-dimensional tensor (height × width × number of channels), which contains rich spatial and semantic information. However, this high-dimensional data form may bring two major problems in practical applications: one is that the computational cost is too high, especially when large-scale traffic scenes need to be processed; the other is that it is difficult to be directly used for subsequent classification or regression tasks, because these tasks usually require the input to be in the form of a vector of fixed length. Therefore, it is necessary to convert the global feature map into a feature vector of fixed length through a dimensionality reduction operation. The principle of global mean pooling is to independently calculate the average value of all spatial positions for each channel of the feature map, and finally output a one-dimensional vector with a length equal to the number of channels. The specific steps are as follows: First, for each channel in the real-time traffic global feature map, it is regarded as a two-dimensional matrix, in which each element represents the eigenvalue of the channel at a specific spatial position; second, all elements of the two-dimensional matrix are averaged to obtain the global mean of the channel; finally, the global means of all channels are arranged in order to form the final real-time traffic global feature vector. This process can be seen as a process of extracting global statistical information from local details, which not only retains the main features of each channel but also removes redundant spatial information. Global mean pooling has natural spatial invariance. Even if the information in some areas of the input feature map changes slightly, the output feature vector will not be significantly affected as long as the overall distribution remains consistent. This makes the model more robust in the face of complex traffic scenarios. Secondly, global mean pooling does not introduce additional parameters, avoids the risk of overfitting, and also reduces computational overhead. Compared with maximum pooling, global mean pooling can better reflect the overall distribution characteristics of the global feature map, rather than just highlighting local extreme values.

[0041] In the above-mentioned traffic demand prediction method for dynamic traffic distribution, the step S130 generates a dynamic traffic distribution suggestion based on the traffic flow analysis multimodal association feature vector and the real-time traffic global feature vector, including: S131, fusing the traffic flow analysis multimodal association feature vector and the real-time traffic global feature vector to obtain a traffic demand distribution feature vector; S132, performing feature dynamic mask mixing adjustment based on gated synthesis on the traffic demand distribution feature vector to obtain an optimized traffic demand distribution feature vector; S133, passing the optimized traffic demand distribution feature vector through a generator to generate a dynamic traffic distribution suggestion.

[0042] In the technical solution of the present application, the step S131 is to fuse the multimodal correlation feature vector of traffic flow analysis and the real-time traffic global feature vector to obtain a traffic demand allocation feature vector. It should be understood that the multimodal correlation feature vector of traffic flow analysis is mainly derived from historical traffic flow data and traffic demand external driving data. It reflects the traffic mode under long-term trends and background conditions, captures the spatiotemporal dependencies between regions and the impact of external environment (such as weather, holidays, etc.) on traffic demand. However, such features usually lack sensitivity to the current real-time conditions and may not be able to respond to sudden traffic events or local congestion in a timely manner. In contrast, the real-time traffic global feature vector is based on the real-time traffic video data collected by the camera, generated by local feature extraction and global feature fusion, and focuses on reflecting the instantaneous changes in the current traffic state. Although the feature vector has a high timeliness, it may be difficult to fully consider historical laws and macro backgrounds because it only relies on real-time data. Through the fusion operation, the stability provided by historical data and the flexibility brought by real-time data can be used at the same time to form a comprehensive feature representation that includes both time series information and spatial distribution characteristics. This feature vector can not only better describe the overall state of the current traffic demand, but also support the subsequent generator module to generate more accurate and adaptable dynamic traffic allocation suggestions. In the technical solution of this application, the two types of feature vectors are combined by weighted summation and concatenation. In the weighted summation method, different weights can be assigned to each feature according to actual needs to emphasize a specific type of feature; in the concatenation method, the two feature vectors are directly connected into a new high-dimensional vector to retain all the original information.

[0043] In the technical solution of the present application, the step S132 performs a feature dynamic mask hybrid adjustment based on gated synthesis on the traffic demand allocation feature vector to obtain an optimized traffic demand allocation feature vector. It should be understood that, considering that in the technical solution of the present application, the feature sources involved include historical traffic flow data, traffic demand external driving data and real-time traffic video data. Among them, the historical traffic flow data is converted into multiple feature vectors through a semantic encoder and associated with the semantic feature vector of the external driving data. The external driving data itself may contain information related to the historical traffic data, and some traffic demand patterns may be repeatedly expressed in two different data sources. In addition, considering that historical traffic flow features are usually numerical data, traffic demand driving data is represented by a sequence of word vectors, has a higher dimension and semantic information, and focuses on capturing static patterns, while the feature vector of traffic video data comes from a deep residual network, contains rich visual information, and focuses more on the changes in dynamic and temporal information. These features vary greatly in dimension and representation. When fusing these features, there may be a dimensionality mismatch problem. This dimensionality difference may cause some features to have too much or too little influence during the fusion process, thereby causing the model training process to be unstable and even affecting the final traffic allocation recommendation. Therefore, in the technical solution of the present application, the traffic demand allocation feature vector is adjusted by a gated synthesis-based feature dynamic mask mixing to obtain an optimized traffic demand allocation feature vector.

[0044] The traffic demand allocation feature vector is subjected to feature dynamic mask hybrid adjustment based on gated synthesis to obtain an optimized traffic demand allocation feature vector, including: First, extract the traffic demand feature manifold projection basis matrix. It should be understood that extracting the traffic demand feature manifold projection basis matrix is ​​not just about obtaining information, but is actually a key operation to structure and computate prior knowledge. The principle is to condense domain expertise, model design concepts, or macro laws inverted from data into a mathematical form of the traffic demand feature manifold projection basis matrix for effective use of subsequent algorithms.

[0045] Then, the core prior information distillation is performed on the traffic demand feature manifold projection basis matrix to obtain a set of core prior information feature primitive embedding vectors of the traffic demand model, which is expressed as follows using the core prior information distillation formula: ;in, Represents the set of embedding vectors of the core prior information feature primitives of the traffic demand model, Respectively represent the first, second, and The core prior information feature primitives of the traffic demand model are embedded in the vector, represents the transpose operation, represents the core prior information distillation, represents the projection basis matrix of the traffic demand characteristic manifold, represents a diagonal matrix, Respectively represent the first and the second diagonal of the diagonal matrix The value of each position. It should be understood that the projection matrix of the traffic demand feature manifold may contain redundant or high-dimensional information, and direct application may lead to excessive computational burden and interference of non-critical information on the optimization process. Therefore, the principle of core prior information distillation is to reduce noise and refine prior knowledge, just like the filtering process in signal processing, retaining the main components and filtering out redundancy.

[0046] Next, a dynamic coupling mapping matrix of the core prior information of the traffic demand model is constructed between the traffic demand allocation feature vector and each of the core prior information feature primitive embedding vectors of the traffic demand model in the set of the core prior information feature primitive embedding vectors of the traffic demand model to obtain a set of dynamic coupling mapping matrices of the core prior information of the traffic demand model, which is expressed as a dynamic coupling mapping formula: ;in, represents the traffic demand allocation feature vector, Express After linear transformation, the feature vector has the same feature scale as the corresponding core prior information feature primitive embedding vector of the traffic demand model. Indicates The core prior information feature primitives of the traffic demand model are embedded in the vector, represents matrix multiplication, represents the length of the embedding vector of the core prior information feature primitive of the traffic demand model, Indicates The core prior information dynamic coupling mapping matrix of the traffic demand model is a dynamic coupling mapping matrix. It should be understood that the core of this step is to construct the interactive relationship between the traffic demand allocation feature vector and the refined prior knowledge. Specifically, through latent space mapping, the nonlinear response pattern of the traffic demand allocation feature vector to different aspects of prior knowledge is learned. In essence, the core prior information dynamic coupling mapping matrix of the traffic demand model is a re-encoding of the traffic demand allocation feature vector from the perspective of prior knowledge, which incorporates the interpretation and processing of knowledge. Its function exceeds information association and realizes the directional enhancement of feature representation and the extraction of multi-perspective feature information.

[0047] Next, the traffic demand prior information response characteristic coupling strength symbol of each traffic demand model core prior information dynamic coupling mapping matrix in the set of the traffic demand model core prior information dynamic coupling mapping matrix is ​​calculated to obtain a set of traffic demand prior information response characteristic coupling strength symbols, which is expressed as the characteristic coupling strength formula: ;in, represents the F-norm of the matrix, Indicates Traffic demand prior information response feature coupling strength symbol. It should be understood that the principle of this step is to extract key information and simplify the feature representation of the dynamic coupling mapping matrix of the core prior information of each traffic demand model, just like information summary generation, and extract the summary or feature vector that best represents the core information. The traffic demand prior information response feature coupling strength symbol must be representative and discriminative. It contains the ideas of feature selection and feature aggregation, selects the most informative parts to retain, and aggregates them into concise feature coupling strength symbols to achieve deeper compression and refinement. The function of the traffic demand prior information response feature coupling strength symbol is not only to compress information, but also to improve the efficiency and robustness of the subsequent fusion process, reduce data dimensions, and reduce computational burden, especially in high-dimensional matrices.

[0048] Subsequently, based on the set of the traffic demand prior information response feature coupling strength symbols, the set of the traffic demand model core prior information dynamic coupling mapping matrices is gated feature synthesized to obtain the traffic demand model prior information response projection coding matrix, which is expressed as a gated feature synthesis formula: ;in, represents the normalized exponential function, Represents the projection coding matrix of the prior information response of the traffic demand model. It should be understood that the core principle of the gated feature synthesis is to emphasize the adaptive and selective prior information integration strategy. Sparsity reflects that not all prior information responses are equally important. The fusion should be selective, focusing on more important responses, weakening or ignoring unimportant responses, improving feature selection and generalization capabilities, and avoiding overfitting. Dynamicity means that the weights or methods of gated feature synthesis are not fixed, but are adaptively adjusted according to input data or model status to improve the flexibility and adaptability of the model. The essence of gated feature synthesis of the set of dynamic coupling mapping matrices of the core prior information of the traffic demand model is to optimally combine the response information from different prior knowledge perspectives to form a traffic demand model prior information response projection coding matrix that comprehensively reflects the model's prior response.

[0049] Finally, the traffic demand allocation feature vector is mapped to the feature space of the traffic demand model prior information response projection coding matrix to obtain the optimized traffic demand allocation feature vector, which is expressed as: ;in, Represents the optimized traffic demand allocation feature vector. It should be understood that the entire adaptation process is finally completed by mapping the traffic demand allocation feature vector to the feature space of the traffic demand model prior information response projection coding matrix to obtain the optimized traffic demand allocation feature vector. The principle of this step is to use the feature space defined by the traffic demand model prior information response projection coding matrix to project the traffic demand allocation feature vector into the space. In essence, the traffic demand model prior information response projection coding matrix acts as a transformation matrix, which performs a linear or nonlinear mapping transformation on the traffic demand allocation feature vector so that it is embedded in the feature space of the fusion model prior information. The optimized traffic demand allocation feature vector better fits the model prior constraints and guidance, and realizes boundary adaptation on the feature manifold. Compared with the traffic demand allocation feature vector, the optimized traffic demand allocation feature vector is usually significantly improved in terms of expressiveness, distinguishability, and generalization, providing a better feature representation for subsequent machine learning tasks.

[0050] In the technical solution of the present application, the step S133 passes the optimized traffic demand allocation feature vector through the generator to generate dynamic traffic allocation suggestions. It should be understood that the traffic demand allocation feature vector integrates the multi-dimensional characteristics of historical data, external driving factors and real-time video information, and can fully reflect the overall state of current traffic demand. However, these feature vectors are essentially high-dimensional representations in mathematics, and it is difficult to directly use them to guide actual traffic management operations. Therefore, it is necessary to use the deep learning module of the generator to map the abstract feature vector into a specific dynamic traffic allocation suggestion. The generator is usually designed based on a neural network, has a strong nonlinear mapping ability and generalization performance, and can generate output results that meet actual needs based on the input feature vector. Specifically, first, the optimized traffic demand allocation feature vector is passed to the generator as input. The feature vector contains comprehensive information on historical trends, external conditions, and real-time status, providing a comprehensive data basis for the generator. Secondly, the generator processes the input feature vector through a series of fully connected layers or convolutional layers to gradually extract higher-level semantic information. In this process, the generator will learn the complex mapping relationship between traffic demand and allocation strategy, such as how to adjust the duration of traffic lights according to traffic flow forecasts, how to re-plan lane usage during peak hours, etc. Finally, the generator outputs dynamic traffic allocation suggestions, which may include but are not limited to the following: capacity adjustment of specific road sections, optimization of traffic light timing schemes, diversion measures in congested areas, etc. Among them, the generator can automatically learn complex nonlinear relationships without manually defining clear rules or formulas, which makes the system more flexible and adaptable, thereby helping traffic management departments to quickly respond to changes in current traffic conditions and formulate more scientific and reasonable long-term plans by predicting future needs in advance.

[0051] In summary, the embodiment of the present application first obtains historical traffic flow data collected by the database, traffic demand external driving data and real-time traffic video data collected by the camera, and then uses deep learning technology to perform feature extraction and correlation analysis on the three, and finally generates dynamic traffic distribution suggestions through a generator, so as to make timely adjustments and optimizations according to real-time conditions, alleviate traffic congestion, improve travel efficiency, and thereby improve the real-time and flexibility of traffic distribution.

[0052] Figure 5 FIG. 1 is a block diagram of a traffic demand prediction system for dynamic traffic allocation according to an embodiment of the present application. Figure 5 As shown, according to the traffic demand prediction system 100 for dynamic traffic distribution according to the embodiment of the present application, it includes: a dynamic traffic demand prediction data acquisition module 110, which is used to obtain historical traffic flow data collected by the database, traffic demand external driving data and real-time traffic video data collected by the camera; a dynamic traffic demand prediction data extraction module 120, which is used to extract traffic flow analysis multimodal correlation feature vectors and real-time traffic global feature vectors from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera; a dynamic traffic distribution suggestion generation module 130, which is used to generate dynamic traffic distribution suggestions based on the traffic flow analysis multimodal correlation feature vectors and the real-time traffic global feature vectors.

[0053] Here, those skilled in the art can understand that the specific operations of each step in the above-mentioned traffic demand prediction system for dynamic traffic allocation have been described in the above reference. Figures 1 to 4 The description of the traffic demand prediction method for dynamic traffic assignment has been introduced in detail, and therefore, its repeated description will be omitted.

[0054] As described above, the traffic demand forecasting system 100 for dynamic traffic allocation according to the embodiment of the present application can be implemented in various terminal devices. In one example, the traffic demand forecasting system 100 for dynamic traffic allocation can be integrated into the terminal device as a software module and / or a hardware module. For example, the traffic demand forecasting system 100 for dynamic traffic allocation can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the traffic demand forecasting system 100 for dynamic traffic allocation can also be one of the many hardware modules of the terminal device.

[0055] Alternatively, in another example, the traffic demand forecasting system 100 for dynamic traffic distribution and the terminal device may also be separate devices, and the traffic demand forecasting system 100 for dynamic traffic distribution may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

Claims

1. A traffic demand forecasting method for dynamic traffic distribution, characterized in that: include: Obtain historical traffic flow data collected by the database, external driving data of traffic demand, and real-time traffic video data collected by cameras; Extracting a traffic flow analysis multimodal correlation feature vector and a real-time traffic global feature vector from the historical traffic flow data collected by the database, the traffic demand external driving data, and the real-time traffic video data collected by the camera; Based on the traffic flow analysis multimodal association feature vector and the real-time traffic global feature vector, a dynamic traffic allocation suggestion is generated.

2. The traffic demand prediction method for dynamic traffic distribution according to claim 1 is characterized in that: Extracting a traffic flow analysis multimodal correlation feature vector and a real-time traffic global feature vector from the historical traffic flow data collected by the database, the traffic demand external driving data, and the real-time traffic video data collected by the camera, including: Performing feature extraction on the historical traffic flow data collected from the database to obtain a historical traffic flow semantic association feature vector; Extracting features from the traffic demand external driving data to obtain a traffic demand external driving semantic feature vector; Associating the historical traffic flow semantic association feature vector with the traffic demand external driving semantic feature vector to obtain the traffic flow analysis multimodal association feature vector; Feature extraction is performed on the real-time traffic video data collected by the camera to obtain the real-time traffic global feature vector.

3. The traffic demand forecasting method for dynamic traffic distribution according to claim 2 is characterized in that: Performing feature extraction on the historical traffic flow data collected from the database to obtain a historical traffic flow semantic association feature vector includes: Passing the historical traffic flow data collected from the database through a historical traffic flow semantic encoder including an embedding layer to obtain a plurality of historical traffic flow feature vectors; The multiple historical traffic flow feature vectors are cascaded into the historical traffic flow semantic association feature vector.

4. The traffic demand forecasting method for dynamic traffic distribution according to claim 3 is characterized in that: Extracting features from the traffic demand external driving data to obtain a traffic demand external driving semantic feature vector includes: Passing the traffic demand external driving data through a traffic demand external driving data analyzer to obtain a sequence of traffic demand external driving word vectors; The sequence of traffic demand external driving word vectors is passed through a traffic demand external driving semantic context encoder to obtain the traffic demand external driving semantic feature vector.

5. The traffic demand prediction method for dynamic traffic distribution according to claim 4 is characterized in that: Performing feature extraction on the real-time traffic video data collected by the camera to obtain the real-time traffic global feature vector includes: Extracting local features from the real-time traffic video data collected by the camera to obtain a plurality of local feature maps of real-time traffic key frames; Perform global feature extraction on the multiple local feature maps of real-time traffic key frames to obtain the real-time traffic global feature vector.

6. The traffic demand prediction method for dynamic traffic distribution according to claim 5 is characterized in that: Performing local feature extraction on the real-time traffic video data collected by the camera to obtain a plurality of real-time traffic key frame local feature maps, including: Extracting key frames from the real-time traffic video data collected by the camera to obtain a plurality of real-time traffic video key frames; The multiple real-time traffic video key frames are respectively passed through a real-time traffic key frame local feature extractor based on a deep residual network to obtain the multiple real-time traffic key frame local feature maps.

7. The traffic demand prediction method for dynamic traffic distribution according to claim 6 is characterized in that: Performing global feature extraction on the multiple local feature maps of real-time traffic key frames to obtain the real-time traffic global feature vector includes: splicing the multiple real-time traffic key frame local feature maps into a real-time traffic splicing feature map; The real-time traffic splicing feature map is passed through a real-time traffic global feature extractor based on a deep residual network to obtain a real-time traffic global feature map; Performing global mean pooling on the real-time traffic global feature map to obtain the real-time traffic global feature vector.

8. The traffic demand prediction method for dynamic traffic distribution according to claim 7, characterized in that: Generating a dynamic traffic allocation suggestion based on the traffic flow analysis multimodal association feature vector and the real-time traffic global feature vector, including: Fusion of the traffic flow analysis multimodal correlation feature vector and the real-time traffic global feature vector to obtain a traffic demand allocation feature vector; Performing gated synthesis-based feature dynamic mask hybrid adjustment on the traffic demand allocation feature vector to obtain an optimized traffic demand allocation feature vector; The optimized traffic demand allocation feature vector is passed through a generator to generate a dynamic traffic allocation suggestion.

9. The traffic demand prediction method for dynamic traffic distribution according to claim 8, characterized in that: The traffic demand allocation feature vector is subjected to a feature dynamic mask hybrid adjustment based on gated synthesis to obtain an optimized traffic demand allocation feature vector, including: Extract the projection basis matrix of the traffic demand feature manifold; Performing core prior information distillation on the traffic demand feature manifold projection basis matrix to obtain a set of core prior information feature primitive embedding vectors of the traffic demand model; Constructing a traffic demand model core prior information dynamic coupling mapping matrix between the traffic demand allocation feature vector and each traffic demand model core prior information feature primitive embedding vector in the set of traffic demand model core prior information feature primitive embedding vectors to obtain a set of traffic demand model core prior information dynamic coupling mapping matrices; Calculating the traffic demand prior information response characteristic coupling strength symbol of each traffic demand model core prior information dynamic coupling mapping matrix in the set of traffic demand model core prior information dynamic coupling mapping matrices to obtain a set of traffic demand prior information response characteristic coupling strength symbols; Based on the set of traffic demand prior information response feature coupling strength symbols, gated feature synthesis is performed on the set of traffic demand model core prior information dynamic coupling mapping matrices to obtain a traffic demand model prior information response projection coding matrix; The traffic demand allocation feature vector is mapped to the feature space of the traffic demand model prior information response projection coding matrix to obtain the optimized traffic demand allocation feature vector.

10. A traffic demand forecasting system for dynamic traffic distribution, characterized in that: include: Dynamic traffic demand forecasting data acquisition module, used to acquire historical traffic flow data collected by the database, traffic demand external driving data and real-time traffic video data collected by the camera; A dynamic traffic demand prediction data extraction module, used to extract traffic flow analysis multimodal correlation feature vectors and real-time traffic global feature vectors from the historical traffic flow data collected by the database, the traffic demand external driving data and the real-time traffic video data collected by the camera; The dynamic traffic allocation suggestion generating module is used to generate dynamic traffic allocation suggestions based on the traffic flow analysis multimodal correlation feature vector and the real-time traffic global feature vector.

Images