Road network safety early warning method and device based on hologram and storage medium

Abstract

The invention provides a road network safety early warning method and device based on a hologram and a storage medium. The method comprises the following steps: collecting original traffic data flow of traffic participants through multi-element perception equipment; and processing a data stream by using edge calculation and a multi-source data fusion technology, and generating a digital twinning hologram of a road network through a deep learning algorithm and a Leiye fitting technology. And combining the hologram with the road network analysis and prediction model to carry out space-time analysis and situation prediction on the original data to generate a road network digital portrait. According to the road network digital portrait, a traffic management scheme and an emergency plan set are retrieved and matched from a preset traffic control strategy library, and safety early warning is provided in the digital twinning hologram. According to the method, multi-source data fusion and situation prediction are carried out on key nodes of a target road network, a digital twin hologram and a road network digital portrait are generated, the road network state is analyzed in real time, a corresponding traffic management scheme and a corresponding emergency plan are matched, and thus a more accurate safety early warning service is provided.

Description

Technical Field

[0001] The present invention relates to the field of urban management technology, and in particular to a road network safety early warning method, device and storage medium based on a hologram. Background Art

[0002] With the acceleration of urbanization and the surge in the number of motor vehicles, traffic safety issues have become increasingly prominent, especially in complex road network structures and during peak hours, where traffic accidents occur frequently, posing huge challenges to urban traffic management. Traditional traffic safety warning systems mainly rely on a single data source or manual monitoring methods, and are unable to perceive the behavior of traffic participants and the dynamic changes of the road network in real time and comprehensively. In addition, existing systems are generally difficult to effectively process and analyze large-scale, diversified traffic data, resulting in the inability to accurately predict potential traffic accidents or safety hazards. Especially in complex road network structures, the warning mechanism of traditional systems lags behind and cannot respond quickly to emergencies, reducing the efficiency and safety of traffic management. Summary of the invention

[0003] The main purpose of the present invention is to solve the technical problem that the existing road network safety early warning method cannot perceive the behavior of traffic participants and the dynamic changes of the road network in real time and comprehensively; A first aspect of the present invention provides a hologram-based road network safety early warning method, the hologram-based road network safety early warning method comprising: Deploy multi-sensing devices at key nodes of the target road network, collect multi-dimensional data of traffic participants in the target road network through the multi-sensing devices, and obtain original traffic data streams; The original traffic data stream is processed by edge computing and multi-source data fusion, and the processed original traffic data stream is modeled and analyzed by deep learning algorithm and radar fitting technology to obtain a digital twin hologram of the target road network; Perform multi-dimensional spatiotemporal analysis and situation prediction on the original traffic data stream through a preset road network analysis and prediction model to obtain analysis and prediction results, and generate a road network digital portrait of the target road network by combining the analysis and prediction results with the original traffic data stream; Based on the digital portrait of the road network, the preset traffic control strategy library is searched and matched to obtain a traffic management plan and an emergency plan set, and a safety warning is issued for the target road network in the digital twin hologram based on the digital portrait of the road network, the traffic management plan and the emergency plan set.

[0004] Optionally, in a first implementation of the first aspect of the present invention, the multi-sensing devices are deployed on key nodes of the target road network, and multi-dimensional data collection is performed on traffic participants in the target road network through the multi-sensing devices to obtain the original traffic data stream including: The topological structure of the target road network is analyzed by graph theory algorithm and traffic flow theory to obtain the key node weight matrix of the target road network, and the multi-sensing devices are optimized and configured according to the key node weight matrix to obtain the sensing device deployment strategy; Based on the sensing device deployment strategy, multiple sensing devices are deployed on the target road network, and data collection parameters of the deployed multiple sensing devices are configured through edge computing units and 5G networks to obtain a distributed sensing network; The various multi-sensing devices in the distributed sensing network are coordinated and scheduled to obtain differentiated original data collection plans, and the various multi-sensing devices perform multi-dimensional data collection on traffic participants in the target road network according to the original data collection plan to obtain an original traffic data stream.

[0005] Optionally, in a second implementation of the first aspect of the present invention, the holographic processing of the road network state representation by a preset radar vision fitting model to obtain a holographic road network model includes: Extracting and preprocessing radar data and visual data in the original traffic data stream to obtain a radar and visual data set; Applying a preset radar vision fitting model to perform feature fusion on the road network state representation and the radar vision data set to obtain a multimodal fusion feature representation; Classify and identify static targets and dynamic targets in a road network environment based on the multimodal fusion feature representation through a target detection algorithm and a trajectory prediction model, and generate a time-varying road network scene description based on the classification and identification results; The time-varying road network scene description is semantically segmented and structured through a graph convolutional network and a temporal attention mechanism to obtain a holographic road network model.

[0006] Optionally, in a third implementation of the first aspect of the present invention, the holographic processing of the road network state representation by a preset radar vision fitting model to obtain a holographic road network model includes: Extracting and preprocessing radar data and visual data in the original traffic data stream to obtain a radar and visual data set; Applying a preset radar vision fitting model to perform feature fusion on the road network state representation and the radar vision data set to obtain a multimodal fusion feature representation; Classify and identify static targets and dynamic targets in a road network environment based on the multimodal fusion feature representation through a target detection algorithm and a trajectory prediction model, and generate a time-varying road network scene description based on the classification and identification results; The time-varying road network scene description is semantically segmented and structured through a graph convolutional network and a temporal attention mechanism to obtain a holographic road network model.

[0007] Optionally, in a fourth implementation of the first aspect of the present invention, the multi-dimensional spatiotemporal analysis and situation prediction of the original traffic data stream is performed by a preset road network analysis and prediction model, and the analysis and prediction results obtained include: Performing multi-scale time decomposition on the original traffic data stream by using preset average pooling kernels of different sizes to obtain separated periodic components and trend components; Generate a feature matrix representing local time features and global time features according to the periodic components through a gated one-dimensional local-global time convolution module, and perform weighted aggregation on features of different time scales in the feature matrix through an attention mechanism to obtain a multi-scale time feature vector; The long-term evolution characteristics of the traffic state in the trend component are extracted and enhanced by multi-layer linear regression and residual connection to obtain a trend feature vector; The multi-scale time feature vector and the trend feature vector are integrated through an adaptive fusion network to obtain a comprehensive time feature representation; Based on the comprehensive temporal feature representation, the spatial dependency of the road network is dynamically modeled to generate a time-varying graph structure, and the road connection relationships and traffic status propagation at different levels in the target road network are captured and aggregated based on the time-varying graph structure through a preset multi-granularity graph convolutional network to obtain analysis and prediction results.

[0008] Optionally, in a fifth implementation of the first aspect of the present invention, based on the comprehensive time feature representation, the spatial dependency of the road network is dynamically modeled to generate a time-varying graph structure, and the road connection relationships and traffic state propagation at different levels in the target road network are captured and aggregated based on the time-varying graph structure through a preset multi-granularity graph convolutional network, and the analysis and prediction results obtained include: Dynamically modeling the spatial dependency of the road network based on the comprehensive time feature representation through an adaptive adjacency matrix generation algorithm, calculating the time-varying correlation strength between key nodes in the target road network, and obtaining a time-varying graph structure that represents the dynamic topological structure of the road network; By using a multi-level graph convolutional network, feature extraction is performed on node relationships at different levels in the road network according to the time-varying graph structure, thereby obtaining multi-scale spatial features containing local and global spatial dependency information; The multi-scale spatial features and the comprehensive temporal feature representation are fused through the spatiotemporal attention mechanism to obtain the spatiotemporal features of each key node, and the spatiotemporal features are weighted to obtain the spatiotemporal feature representation of the key nodes; The future state of each key node is evaluated and classified based on the temporal and spatial characteristics of the key nodes through a preset multi-dimensional safety threshold to obtain an analysis and prediction result.

[0009] Optionally, in a sixth implementation of the first aspect of the present invention, based on the digital portrait of the road network, searching and matching a preset traffic control strategy library to obtain a traffic management plan and an emergency plan set, and performing a safety warning on the target road network in the digital twin hologram according to the digital portrait of the road network, the traffic management plan and the emergency plan set includes: Generate a road network status description of the target road network according to the road network digital portrait, and match the road network status description with a preset traffic control strategy library to obtain a traffic management plan and an emergency plan set; Mapping the traffic management plan and emergency plan set to the digital twin hologram for virtual simulation and evaluation, identifying potential high-risk areas and sections in the digital twin hologram, and obtaining the road network safety risk distribution; According to the safety risk distribution of the road network, areas with different risk levels are marked on the digital twin hologram, and safety warnings are issued to the target road network in combination with the traffic management plan, emergency plan and digital portrait of the road network.

[0010] A second aspect of the present invention provides a hologram-based road network safety warning device, the hologram-based road network safety warning device comprising: A deployment collection module is used to deploy multi-sensing devices on key nodes of the target road network, and to collect multi-dimensional data of traffic participants in the target road network through the multi-sensing devices to obtain original traffic data streams; A hologram generation module is used to perform edge computing processing and multi-source data fusion processing on the original traffic data stream, and to model and analyze the processed original traffic data stream through a deep learning algorithm and a radar fitting technique to obtain a digital twin hologram of the target road network; An analysis and prediction module is used to perform multi-dimensional spatiotemporal analysis and situation prediction on the original traffic data stream through a preset road network analysis and prediction model to obtain analysis and prediction results, and generate a road network digital portrait of the target road network by combining the analysis and prediction results with the original traffic data stream; The safety warning module is used to search and match the preset traffic control strategy library based on the digital portrait of the road network, obtain the traffic management plan and the emergency plan set, and issue a safety warning to the target road network in the digital twin hologram according to the digital portrait of the road network, the traffic management plan and the emergency plan set.

[0011] A third aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the steps of the above-mentioned hologram-based road network safety warning method.

[0012] The above-mentioned hologram-based road network safety warning method, device and storage medium collect the original traffic data stream of traffic participants through multi-sensing devices. The data stream is processed by edge computing and multi-source data fusion technology, and a digital twin hologram of the road network is generated by deep learning algorithm and radar fitting technology. The hologram is combined with the road network analysis and prediction model to perform spatiotemporal analysis and situation prediction on the original data to generate a digital portrait of the road network. According to the digital portrait of the road network, the matching traffic management plan and emergency plan set are retrieved from the preset traffic control strategy library, and a safety warning is provided in the digital twin hologram. The present invention performs multi-source data fusion and situation prediction on the key nodes of the target road network, generates a digital twin hologram and a digital portrait of the road network, analyzes the road network status in real time and matches the corresponding traffic management plan and emergency plan, and then provides a more accurate safety warning service.

[0013] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0014] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of a first embodiment of a road network safety early warning method based on a hologram in an embodiment of the present invention; Figure 2 It is a schematic diagram of an embodiment of a road network safety warning device based on a hologram in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0018] To facilitate understanding of this embodiment, a hologram-based road network safety warning method disclosed in an embodiment of the present invention is first introduced in detail. The hologram-based road network safety warning method is applied to a PCB processing system, and the PCB processing system includes a PCB processing workpiece and a PCB processing device. Figure 1 As shown, the method comprises the following steps: 101. Deploy multi-sensing devices at key nodes of the target road network, collect multi-dimensional data of traffic participants in the target road network through the multi-sensing devices, and obtain the original traffic data stream; In one embodiment of the present invention, the deployment of multi-sensing devices on key nodes of the target road network, and the multi-dimensional data collection of traffic participants in the target road network through the multi-sensing devices to obtain the original traffic data flow includes: performing a topological structure analysis on the target road network through graph theory algorithms and traffic flow theory to obtain a key node weight matrix of the target road network, and optimizing the multi-sensing devices according to the key node weight matrix to obtain a sensing device deployment strategy; based on the sensing device deployment strategy, deploying multi-sensing devices on the target road network, and configuring data collection parameters of the deployed multi-sensing devices through edge computing units and 5G networks to obtain a distributed sensing network; collaboratively scheduling each multi-sensing device in the distributed sensing network to obtain a differentiated original data collection scheme, and using each multi-sensing device to collect multi-dimensional data on traffic participants in the target road network according to the original data collection scheme to obtain an original traffic data flow.

[0019] Specifically, the topological structure of the target road network is analyzed by graph theory and traffic flow theory, and the key node weight matrix of the target road network is obtained by modeling the road network as a weighted directed graph G(V,E), where V represents the intersection set and E represents the road set. The Dijkstra algorithm is used to calculate the shortest path between all node pairs to obtain the accessibility matrix of the road network. Then, the PageRank algorithm is applied to calculate the importance score of each node, which takes into account the connectivity of the node and the importance of the connecting node. At the same time, the Betweenness Centrality indicator is used to calculate the frequency of each node as an intermediate node of the shortest path in the network. In terms of traffic flow theory, the LWR (Lighthill-Whitham-Richards) model is used to describe the dynamics of traffic flow on the road. Using this model, combined with historical traffic data, the average flow and density of each road are calculated. These traffic flow characteristics are combined with the static topological characteristics obtained by graph theory, and a multi-attribute decision-making method (such as TOPSIS) is used for comprehensive scoring, and finally the weight of each node is obtained to form a key node weight matrix.

[0020] Specifically, the multi-sensing devices are optimized and configured according to the key node weight matrix to obtain the sensing device deployment strategy. It is necessary to define the objective function f(x), where x represents the device deployment plan. The objective function contains three sub-goals: coverage, cost, and data quality. The coverage is quantified by calculating the sum of the weights of the key nodes covered by the sensing devices; the cost considers the number of devices and the installation and maintenance costs; and the data quality is based on the device performance parameters and node importance evaluation. The deployment plan is optimized using a genetic algorithm. The specific steps include: (1) initializing the population, where each individual represents a deployment plan; (2) evaluating the fitness, that is, calculating the objective function value of each plan; (3) selecting an excellent individual using the tournament selection method; (4) crossover operation, using the uniform crossover method to generate a new deployment plan; (5) mutation operation, randomly changing the deployment location of some devices; (6) repeating steps 2-5 until the termination condition is reached. Finally, the individual with the highest fitness is selected as the optimal sensing device deployment strategy.

[0021] Specifically, based on the perception device deployment strategy, multiple perception devices are deployed on the target road network, and the data collection parameters of the deployed multiple perception devices are configured through edge computing units and 5G networks. The distributed perception network needs to install multiple perception devices at selected locations according to the deployment strategy, such as video surveillance cameras, microwave radars, acoustic sensors, etc. Each device is equipped with an edge computing unit, such as a Raspberry Pi or an industrial-grade embedded system. The edge computing unit is pre-installed with a lightweight deep learning framework (such as TensorFlow Lite) and a data preprocessing algorithm. Then, a communication system based on the 5G network is established, including small base station deployment and network slice configuration. Using the network function virtualization (NFV) and software defined network (SDN) characteristics of the 5G network, dedicated network slices are configured for different types of perception data to ensure low latency and high reliability of data transmission. Then, each edge computing unit is initialized and configured through a remote configuration tool (such as Ansible) to set data collection parameters, including sampling frequency, resolution, trigger conditions, etc. Finally, a distributed data management system (such as Apache Kafka) is deployed for real-time processing and distribution of data streams to form a complete distributed perception network.

[0022] Specifically, in order to coordinate and schedule the various sensing devices in the distributed sensing network and obtain differentiated raw data collection solutions, a central coordination system needs to be built and deployed on the cloud platform using containerization technology (such as Docker) to ensure the scalability and flexibility of the system. The system includes the following core modules: (1) Data access module: Receive the status information of each sensing device and external data sources (such as traffic flow data and weather data) in real time through the MQTT protocol; (2) Status evaluation module: Use the sliding time window method to analyze real-time data and evaluate the current road network status; (3) Resource allocation module: Implement a dynamic resource allocation algorithm based on reinforcement learning, such as Deep Q-Network (DQN), take the road network status as input, and output the optimal resource allocation strategy; (4) Task scheduling module: Use the improved Johnson algorithm to schedule multiple device tasks to minimize the completion time of data collection; (5) Communication module: Responsible for safely sending scheduling instructions to each device. When the system is running, it continuously monitors the network status and triggers the rescheduling process when significant changes are detected (such as sudden changes in traffic flow and sudden changes in weather). The scheduling results form a dynamically updated raw data collection plan to guide the specific working mode of each sensing device.

[0023] Specifically, the original traffic data stream is obtained by collecting multi-dimensional data of traffic participants in the target road network through various multi-sensing devices according to the original data collection scheme. The video surveillance camera needs to use the YOLO v5 algorithm for real-time target detection and tracking to identify traffic participants such as vehicles and pedestrians; the microwave radar uses frequency shift keying (FSK) modulation technology to measure vehicle speed and position; the acoustic sensor uses voiceprint analysis technology to identify abnormal traffic events. The edge computing unit of each device performs data preprocessing, including noise reduction, outlier detection and data compression. The processed data is transmitted to the central data processing platform through the 5G network. The platform uses Apache Flink for stream processing to achieve real-time aggregation, correlation analysis and spatiotemporal alignment of data. The final generated original traffic data stream contains multi-dimensional information, such as traffic flow, speed, vehicle trajectory, pedestrian density, etc., which is stored in a distributed database (such as Apache HBase) to provide basic data support for subsequent in-depth analysis and modeling.

[0024] 102. Perform edge computing and multi-source data fusion processing on the original traffic data stream, and model and analyze the processed original traffic data stream through deep learning algorithms and radar fitting technology to obtain a digital twin hologram of the target road network; In one embodiment of the present invention, the edge computing processing and multi-source data fusion processing are performed on the original traffic data stream, and the processed original traffic data stream is modeled and analyzed through a deep learning algorithm and a radar fitting technology to obtain a digital twin hologram of the target road network, including: performing distributed edge computing processing on the original traffic data stream to obtain an edge feature set, and globally integrating the edge feature set through a preset spatiotemporal alignment algorithm and a semantic annotation model to obtain a fused feature matrix; applying a preset deep learning algorithm to perform feature mapping and pattern recognition on the fused feature matrix to obtain a high-dimensional road network state representation; holographically processing the road network state representation through a preset radar fitting model to obtain a holographic road network model; applying a preset point cloud registration and three-dimensional reconstruction algorithm to virtually reproduce the road network environment of the target road network according to the holographic road network model to obtain a digital twin hologram of the target road network.

[0025] Specifically, the original traffic data stream is first processed by distributed edge computing to obtain an edge feature set, and the edge feature set is globally integrated through a preset spatiotemporal alignment algorithm and a semantic annotation model to obtain a fused feature matrix. In the specific implementation process, a lightweight deep learning model, such as MobileNetV3 or EfficientNet-Lite, is deployed at each edge computing node to process the raw data from multiple sensing devices in real time. These models are pre-trained and fine-tuned to quickly extract key features such as traffic flow, vehicle speed, and pedestrian density. For video data, an improved YOLOv5-s model is used for target detection and tracking; for radar data, 1D-CNN is used to extract speed and position features; for acoustic data, an LSTM network is used for event classification. Each edge node processes data in parallel to generate a local feature vector. Subsequently, a distributed time synchronization protocol such as PTP (Precision Time Protocol) is used to synchronize the clocks of all edge nodes to ensure the time consistency of the data. An improved dynamic time warping (DTW) algorithm is used for spatiotemporal alignment to handle sampling rate differences and time delays of different sensor data. The aligned data is semantically annotated using the pre-trained BERT model to add structured descriptions to traffic events and states. Finally, a federated learning framework such as the FedAvg algorithm is used to globally integrate the features of each node while protecting privacy to form a fused feature matrix.

[0026] Specifically, a preset deep learning algorithm is applied to perform feature mapping and pattern recognition on the fused feature matrix to obtain a high-dimensional representation of the road network state. In this step, the fused feature matrix is ​​first denoised and feature extracted using an autoencoder network. The autoencoder consists of a multi-layer fully connected neural network that learns the hidden representation of the data by minimizing the reconstruction error. Then, the processed features are input into a deep spatiotemporal graph convolutional network (DSTGCN). The network combines graph convolutional layers and temporal convolutional layers to simultaneously capture the spatial topological structure and temporal evolution characteristics of the road network. The graph convolutional layer uses the ChebNet algorithm to achieve efficient spatial information propagation, while the temporal convolutional layer uses dilated convolution to expand the receptive field and capture long-term dependencies. The network also introduces an attention mechanism to dynamically adjust the importance weights of different nodes and time steps. After multi-layer processing, the network outputs a high-dimensional feature representation that describes the overall state of the road network. In order to further improve the generalization ability of the model, a contrastive learning strategy is adopted to learn more robust features by maximizing the mutual information of different perspective representations of the same scene. Finally, the t-SNE algorithm is used to visualize high-dimensional features and generate two-dimensional or three-dimensional road network status representations for easy intuitive understanding and subsequent analysis.

[0027] Specifically, the road network state representation is holographically processed by a preset radar-vision fitting model to obtain a holographic road network model. This step first constructs a multimodal fusion network to process radar and visual data at the same time. For radar data, the PointNet++ network is used to extract point cloud features; for visual data, ResNet-50 is used as the backbone network to extract image features. The two features are adaptively fused through the attention mechanism to generate a unified representation. Next, a generative model based on a variational autoencoder (VAE) is designed to map the fused features to the latent space. Both the encoder and decoder of VAE adopt a multi-layer perceptron structure, and noise is added in the middle to enhance the generalization ability of the model. By minimizing the reconstruction error and KL divergence, the model learns the probability distribution of the road network state. In the inference stage, sampling is performed from the learned distribution to generate multiple possible road network states. In order to improve the authenticity of the generated results, an adversarial training mechanism is introduced, and WGAN-GP is used as a discriminator to encourage the generative model to produce a more realistic road network representation. Finally, the time series is modeled through recurrent neural network (RNN) to predict the evolution of road network status in the future, thus obtaining a dynamic holographic road network model.

[0028] Specifically, the preset point cloud registration and three-dimensional reconstruction algorithms are applied to virtually reproduce the road network environment of the target road network according to the holographic road network model, so as to obtain a digital twin hologram of the target road network. First, the point cloud data from different sensors are registered using the improved ICP (Iterative Closest Point) algorithm to align them to the same coordinate system. In order to process large-scale point clouds, the octree structure is used for spatial division, and the KD tree is used to accelerate the nearest neighbor search. The registered point cloud is fused with the holographic road network model generated in the previous step to form a preliminary three-dimensional scene representation. Next, a deep learning-based three-dimensional reconstruction algorithm, such as Mesh R-CNN, is applied to reconstruct a detailed three-dimensional mesh model from a sparse point cloud. The algorithm first uses PointNet++ to extract point cloud features, and then generates a voxel grid through a series of 3D convolution and upsampling operations. Finally, the geometric details of the grid are refined through a graph convolutional network. In order to improve the reconstruction quality, multi-view is introduced. Figure 1 Consistency constraints are imposed, and image information from different angles is used to guide the reconstruction process. After the reconstruction is completed, physically based rendering technology is used to add textures and materials to the 3D model to improve visual realism. Finally, the reconstructed static environment is combined with dynamic traffic flow simulation, and the SUMO (Simulation of Urban MObility) tool is used to generate the movement trajectories of vehicles and pedestrians. Real-time rendering and interaction are achieved in the browser through WebGL technology, and finally a highly realistic digital twin hologram of the target road network is obtained.

[0029] Furthermore, the holographic processing of the road network state representation by a preset radar vision fitting model to obtain a holographic road network model includes: extracting and preprocessing the radar data and visual data in the original traffic data stream to obtain a radar vision data set; applying the preset radar vision fitting model to perform feature fusion on the road network state representation and the radar vision data set to obtain a multimodal fusion feature representation; classifying and identifying static targets and dynamic targets in the road network environment based on the multimodal fusion feature representation through a target detection algorithm and a trajectory prediction model, and generating a time-varying road network scene description based on the results of the classification and identification; performing semantic segmentation and structured representation on the time-varying road network scene description through a graph convolutional network and a temporal attention mechanism to obtain a holographic road network model.

[0030] Specifically, the radar data and visual data in the original traffic data stream are first extracted and preprocessed to obtain a radar vision data set. In the specific process, the distance amplitude compensation algorithm is applied to the radar data to eliminate the influence of distance attenuation, and then the two-dimensional discrete Fourier transform is used to convert the time domain signal into a frequency domain representation. Then, the constant false alarm rate detection (CFAR) algorithm is used for target detection to screen out valid target echoes. For visual data, image defogging is first performed, and the atmospheric illumination and transmittance are estimated using the dark channel prior algorithm, and then image restoration is performed. Subsequently, adaptive histogram equalization is applied to enhance the image contrast, and bilateral filtering is used for edge-preserving denoising. Finally, the image is geometrically corrected to eliminate lens distortion and perspective deviation. The two types of data are time synchronized and spatially registered to form a unified radar vision data set. The network time protocol (NTP) is used for time synchronization to ensure the accurate timestamp of data acquisition, and the external calibration method is used for spatial registration to obtain the rigid body transformation relationship between the radar and the camera.

[0031] Specifically, the preset radar-vision fitting model is applied to perform feature fusion on the road network state representation and the radar-vision dataset to obtain a multimodal fusion feature representation. This step first uses two parallel feature extraction networks to process radar and visual data respectively. The radar data is processed by an improved PointNet++ network, which adds a dynamic point cloud clustering layer to better capture the local structural information of the point cloud. The visual data is extracted through a pre-trained ResNeXt-101 network to extract deep features, and a feature pyramid network (FPN) is used to fuse multi-scale features. The road network state representation is processed by a fully connected layer for dimensionality reduction. Then, a multimodal attention fusion module is designed, which contains three branches: self-attention, cross-attention, and adaptive fusion. The self-attention branch performs internal information interaction on radar, vision, and road network state features respectively; the cross-attention branch calculates the correlation between different modalities; and the adaptive fusion branch dynamically adjusts the fusion weight according to the reliability of different modalities. Finally, all features are integrated into a unified multimodal fusion feature representation through residual connections and layer normalization.

[0032] Specifically, static and dynamic targets in the road network environment are classified and identified based on multimodal fusion feature representation through target detection algorithm and trajectory prediction model, and time-varying road network scene description is generated based on the classification and recognition results. The target detection adopts the improved CenterNet algorithm, which adds a multi-scale feature aggregation module and deformable convolution on the original basis to improve the detection ability of small targets and dense targets. The detection network outputs the category, position, size and direction information of the target. For dynamic targets, the Deep Sorting Network (Deep SORT) algorithm is used for multi-target tracking. This algorithm combines appearance features and motion information to effectively solve the problems of occlusion and ID switching. The trajectory prediction adopts the social LSTM model, considering the interaction between targets and predicting the movement trajectory in the next few seconds. Static targets such as traffic signs and street lights are accurately located and classified through the semantic segmentation network (DeepLabv3+). Finally, all recognition and prediction results are integrated into a structured scene description, including the type, position, speed, trajectory and other time-varying information of each target.

[0033] Specifically, the time-varying road network scene description is semantically segmented and structured through the graph convolution network and the temporal attention mechanism to obtain a holographic road network model. First, a dynamic graph structure is constructed, in which nodes represent the entities in the scene (such as vehicles, pedestrians, and road sections), and edges represent the spatial relationships and interactions between entities. The spatial-temporal graph convolution network (ST-GCN) is used to process this dynamic graph, which is composed of multiple layers of graph convolution and temporal convolution stacked alternately. The graph convolution uses the ChebNet algorithm, which can efficiently process large-scale graph structures; the temporal convolution uses causal convolution to capture temporal dependencies. The gating mechanism is introduced to dynamically adjust the weights of different edges to adapt to complex traffic scenarios. At the top layer of the network, the temporal attention mechanism is applied to calculate the correlation between different time steps and capture long-term dependencies. Finally, the high-dimensional features are mapped to the low-dimensional space through the graph embedding layer to form a compact scene representation. In order to enhance the interpretability of the model, the knowledge distillation technology is introduced to integrate traffic rules and empirical knowledge into the model training process. The final holographic road network model not only contains rich semantic information, but also retains the spatiotemporal dynamic characteristics of the scene, providing a comprehensive data basis for subsequent traffic analysis and decision-making.

[0034] 103. Perform multi-dimensional spatiotemporal analysis and situation prediction on the original traffic data stream through the preset road network analysis and prediction model to obtain the analysis and prediction results, and generate a road network digital portrait of the target road network by combining the analysis and prediction results with the original traffic data stream; In one embodiment of the present invention, the original traffic data stream is subjected to multi-dimensional spatiotemporal analysis and situation prediction by a preset road network analysis and prediction model, and the analysis and prediction results obtained include: performing multi-scale time decomposition on the original traffic data stream by preset average pooling kernels of different sizes to obtain separated periodic components and trend components; generating a feature matrix representing local time features and global time features according to the periodic components by a gated one-dimensional local-global time convolution module, and weightedly aggregating features of different time scales in the feature matrix by an attention mechanism to obtain a multi-scale time feature vector; extracting and enhancing the long-term evolution features of the traffic state in the trend component by multi-layer linear regression and residual connection to obtain a trend feature vector; integrating the multi-scale time feature vector and the trend feature vector by an adaptive fusion network to obtain a comprehensive time feature representation; based on the comprehensive time feature representation, dynamically modeling the spatial dependency of the road network to generate a time-varying graph structure, and capturing and aggregating the road connection relationships and traffic state propagation at different levels in the target road network based on the time-varying graph structure by a preset multi-granularity graph convolution network to obtain the analysis and prediction results.

[0035] Specifically, the original traffic data stream is first subjected to multi-scale temporal decomposition by using preset average pooling kernels of different sizes to obtain separated periodic components and trend components. In the specific implementation process, a multi-scale pooling layer is designed, which contains multiple parallel one-dimensional average pooling operations, and the pooling kernel sizes are set to 1 hour, 4 hours, 12 hours and 24 hours respectively to capture periodic patterns at different time scales. These pooling operations are applied to the original traffic data stream to obtain smoothed sequences of multiple scales. Then, the original sequence is subtracted from the sequence obtained by the maximum pooling kernel (24 hours) to extract the high-frequency periodic components. For the remaining low-frequency part, the exponentially weighted moving average (EWMA) algorithm is applied to further separate the medium-frequency periodic components and the long-term trend components. The smoothing parameters of EWMA are optimized by grid search to balance trend capture and noise suppression. Finally, the high-frequency and medium-frequency components are merged into the periodic component, and the remaining part is used as the trend component, thus completing the multi-scale temporal decomposition of the original data.

[0036] Specifically, a gated one-dimensional local-global temporal convolution module is used to generate a feature matrix representing local and global temporal features according to the periodic components, and the features of different time scales in the feature matrix are weighted and aggregated through the attention mechanism to obtain a multi-scale temporal feature vector. The module first uses multiple parallel one-dimensional convolutional layers with convolution kernel sizes set to 3, 5, and 7 to capture local temporal features. At the same time, global average pooling and fully connected layers are used to extract global temporal features. Local and global features are fused through a gating mechanism. The gating unit is implemented by a sigmoid-activated fully connected layer to dynamically adjust the importance of local and global information. The fused features pass through a multi-head self-attention layer, where each attention head is responsible for capturing dependencies at different time scales. The attention weights are normalized by a softmax function to ensure that the model focuses on the most relevant time steps. Finally, the outputs of different attention heads are merged using residual connections and layer normalization to form a multi-scale temporal feature vector.

[0037] Specifically, the long-term evolution characteristics of the traffic status in the trend component are extracted and enhanced through multi-layer linear regression and residual connection to obtain the trend feature vector. A deep residual network is constructed, which contains multiple residual blocks, each of which consists of two layers of linear transformation and ReLU activation function. A batch normalization layer is added in each residual block to improve the training stability and generalization ability of the model. The input of the network is the trend component, and the deep features are extracted through layer-by-layer transformation. The residual connection allows the gradient to flow directly, alleviating the gradient vanishing problem of the deep network. The attention mechanism is introduced in the middle layer of the network, so that the model can dynamically pay attention to the importance of different time steps. The last layer uses linear transformation to map the features to the required dimension to obtain the trend feature vector. In order to enhance the robustness of the model, the dropout technique is applied during the training process, and L2 regularization is used to prevent overfitting.

[0038] Specifically, the multi-scale time feature vector and trend feature vector are integrated through an adaptive fusion network to obtain a comprehensive time feature representation. A transformer-based fusion network is designed, which contains a multi-layer encoder-decoder structure. The encoder part processes multi-scale time features and trend features respectively, and uses a multi-head self-attention mechanism to capture the internal correlation of the features. The decoder part uses a cross-attention mechanism to interactively fuse the two features. In the fusion process, learnable feature importance weights are introduced to dynamically adjust the contribution of different features. At the same time, the gated linear unit (GLU) is used as the activation function to enhance the nonlinear expression ability of the network. In order to process input sequences of different lengths, the position encoding technology is used to inject time information into the feature representation. Finally, the final comprehensive time feature representation is generated through a feedforward neural network and a residual connection.

[0039] Specifically, based on the comprehensive temporal feature representation, the spatial dependency of the road network is dynamically modeled to generate a time-varying graph structure. The preset multi-granularity graph convolution network is used to capture and aggregate the road connection relationship and traffic state propagation at different levels in the target road network based on the time-varying graph structure to obtain the analysis and prediction results. First, a dynamic graph structure is constructed, in which the nodes of the graph represent the key positions in the road network and the edges represent the connection relationship between the nodes. The weights of the edges are dynamically adjusted according to the comprehensive temporal features and historical traffic patterns using an adaptive adjacency matrix generation algorithm. Then, a multi-granularity graph convolution network is designed, which contains multiple parallel graph convolution branches, each of which is responsible for capturing dependencies at different spatial scales. In each branch, an improved ChebNet is used to implement efficient graph convolution operations. An attention mechanism is introduced to dynamically adjust the importance of different nodes and edges. In order to capture long-term spatial dependencies, gated jump connections are inserted between graph convolution layers. Finally, the graph convolution results of different granularities are fused through a multi-layer perceptron to obtain the final analysis and prediction results. The result includes the traffic state prediction of each node in the future period of time and the situation assessment of the overall road network.

[0040] Furthermore, based on the comprehensive temporal feature representation, the spatial dependency of the road network is dynamically modeled to generate a time-varying graph structure, and the road connection relationship and traffic state propagation at different levels in the target road network are captured and aggregated based on the time-varying graph structure through a preset multi-granularity graph convolutional network, and the analysis and prediction results obtained include: based on the comprehensive temporal feature representation, the spatial dependency of the road network is dynamically modeled through an adaptive adjacency matrix generation algorithm, and the time-varying correlation strength between key nodes in the target road network is calculated to obtain a time-varying graph structure that characterizes the dynamic topological structure of the road network; based on the time-varying graph structure, feature extraction is performed on the node relationship at different levels in the road network through a multi-level graph convolutional network to obtain multi-scale spatial features containing local and global spatial dependency information; the multi-scale spatial features and the comprehensive temporal feature representation are fused through a spatiotemporal attention mechanism to obtain the spatiotemporal features of each key node, and the spatiotemporal features are weighted to obtain the spatiotemporal feature representation of the key nodes; based on the spatiotemporal feature representation of the key nodes through a preset multi-dimensional safety threshold, the future state of each key node is evaluated and classified to obtain the analysis and prediction results.

[0041] Specifically, the spatial dependency of the road network is dynamically modeled based on the comprehensive temporal feature representation through an adaptive adjacency matrix generation algorithm, and the time-varying correlation strength between key nodes in the target road network is calculated to obtain a time-varying graph structure that represents the dynamic topological structure of the road network. In the specific implementation process, the road network is modeled as a graph G(V,E), where V is a node set and E is an edge set. For each pair of nodes (i,j), its time-varying correlation strength A_ij(t) is calculated. First, the comprehensive temporal features of nodes i and j are mapped to the same latent space using a multi-layer perceptron (MLP). Then, the cosine similarity of the two mapped features is calculated to obtain the initial correlation strength. Next, the historical traffic pattern information is introduced, and the long short-term memory network (LSTM) is used to model the historical correlation strength and output the predicted correlation strength. The initial correlation strength and the predicted correlation strength are fused through a gating mechanism, and the gating weight is dynamically generated by another MLP according to the current temporal features. Finally, the softmax function is applied to normalize the fused correlation strength to ensure the numerical stability of the adjacency matrix. In this way, the obtained time-varying adjacency matrix A(t) not only takes into account the current traffic status, but also includes the historical evolution pattern, and can accurately characterize the dynamic topological structure of the road network.

[0042] Specifically, a multi-level graph convolutional network is used to extract features from node relationships at different levels in the road network according to the time-varying graph structure, and a multi-scale spatial feature containing local and global spatial dependency information is obtained. The network adopts a hierarchical design and contains multiple parallel graph convolution branches, each of which is responsible for capturing the dependency relationship of a specific spatial scale. In each branch, the graph convolution operation is implemented using an improved ChebNet, and the feature propagation process is simplified through Chebyshev polynomial approximation to improve computational efficiency. In order to capture spatial dependencies in different ranges, multiple convolution kernels are set, corresponding to the 1st, 2nd and 3rd order neighborhoods. The attention mechanism is introduced to dynamically adjust the importance weights of neighborhoods of different orders. Residual connections are inserted between graph convolution layers to alleviate the gradient vanishing problem of deep networks. In order to enhance the adaptability of the model to node degree, a degree normalization operation is added after each graph convolution layer. At the top layer of the network, a global pooling operation is used to aggregate global information, and it is fused with local features of each layer through jump connections to form a multi-scale spatial feature representation. Finally, through the adaptive feature aggregation module, the weights of features of different scales are dynamically adjusted according to their importance to obtain the final multi-scale spatial feature.

[0043] Specifically, the multi-scale spatial features and comprehensive temporal feature representations are fused through the spatiotemporal attention mechanism to obtain the spatiotemporal features of each key node, and the spatiotemporal features are weighted to obtain the spatiotemporal feature representation of the key nodes. In the implementation process, a two-stream attention module is first designed to process spatial features and temporal features respectively. In the spatial stream, the graph attention network (GAT) is used to calculate the spatial dependency strength between nodes, and the multi-head attention mechanism is used to capture different types of spatial relationships. In the temporal stream, the Self-Attention mechanism is used to model long-term temporal dependencies, and relative position encoding is introduced to enhance the model's perception of temporal patterns. Next, a cross-attention layer is designed to achieve interactive fusion of spatial and temporal features. This layer calculates the attention weights of spatial features on temporal features, and the attention weights of temporal features on spatial features, thereby capturing complex spatiotemporal correlations. The fused features are nonlinearly transformed through a feedforward neural network to obtain preliminary node spatiotemporal features. Finally, a meta-learning module is introduced to dynamically adjust the feature weights of different nodes according to their historical importance. This module uses the MAML (Model-Agnostic Meta-Learning) algorithm to quickly adapt to the feature distribution of different nodes through a small number of samples, realize personalized weight allocation, and finally obtain the spatiotemporal feature representation of key nodes after weight optimization.

[0044] Specifically, the future states of key nodes are evaluated and classified based on the spatiotemporal features of key nodes through preset multi-dimensional safety thresholds to obtain analytical prediction results. First, a multi-task learning framework is designed to simultaneously predict multiple safety-related indicators, such as traffic flow, average speed, accident risk, etc. For each task, a dedicated prediction head is constructed, which contains a multi-layer fully connected network and task-specific activation functions. A hard parameter sharing strategy is used to allow different tasks to share the underlying feature extraction network while maintaining the task-specific output layer. In order to deal with the correlation between indicators, a task association matrix is ​​introduced to explicitly model the dependencies between different prediction tasks. Then, multi-dimensional safety thresholds are set, which are pre-defined based on historical data statistics and expert knowledge. The prediction results are compared with the safety thresholds to perform a multi-dimensional evaluation of the future state of each node. A fuzzy logic system is introduced to map continuous prediction values ​​to discrete safety levels to improve the robustness of classification. Finally, an ensemble learning method, such as random forest or gradient boosting tree, is used to comprehensively consider the evaluation results of multiple dimensions to obtain the final safety state classification of each node. The prediction results include the status prediction of each key node at different time scales (such as 15 minutes, 1 hour, and 4 hours), as well as the safety situation assessment of the overall road network.

[0045] 104. Based on the digital portrait of the road network, the preset traffic control strategy library is searched and matched to obtain the traffic management plan and emergency plan set, and a safety warning is issued for the target road network based on the digital portrait of the road network, the traffic management plan and the emergency plan set in the digital twin hologram.

[0046] In one embodiment of the present invention, based on the digital portrait of the road network, a preset traffic control strategy library is retrieved and matched to obtain a traffic management plan and an emergency plan set, and a safety warning is issued for the target road network in the digital twin hologram according to the digital portrait of the road network, the traffic management plan and the emergency plan set, including: generating a road network status description of the target road network according to the digital portrait of the road network, and matching the road network status description with a preset traffic control strategy library to obtain a traffic management plan and an emergency plan set; mapping the traffic management plan and the emergency plan set to the digital twin hologram for virtual simulation and evaluation, identifying potential high-risk areas and sections in the digital twin hologram, and obtaining a road network safety risk distribution; marking areas of different risk levels on the digital twin hologram according to the road network safety risk distribution, and issuing a safety warning for the target road network in combination with the traffic management plan, emergency plan and digital portrait of the road network.

[0047] Specifically, firstly, the road network state description of the target road network is generated according to the road network digital portrait, and the road network state description is matched with the preset traffic control strategy library to obtain the traffic management plan and emergency plan set. In the specific implementation process, a deep neural network is used to extract features from the road network digital portrait, which includes a multi-layer convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN is used to extract spatial features, such as road topology, traffic density distribution, etc.; LSTM is used to capture time series features, such as traffic flow change trends, periodic patterns, etc. The extracted features are weighted fused through the attention mechanism to generate a compact road network state vector. Then, the state vector is input into a pre-trained natural language generation model, such as GPT-3, to generate a structured road network state description text. Then, a semantic similarity algorithm, such as BERT-based cosine similarity, is used to match the generated state description with the strategies in the preset traffic control strategy library. The KNN algorithm is used in the matching process to select the K strategies with the highest similarity. In order to improve the accuracy of matching, a context-aware re-ranking mechanism is introduced to consider external factors such as the current time and weather. Finally, a multi-objective optimization algorithm, such as NSGA-II, is used to select the optimal traffic management scheme and emergency plan set from the high-similarity strategies, while considering multiple objectives such as implementation cost and expected effect.

[0048] Specifically, the traffic management scheme and emergency plan set are mapped to the digital twin hologram for virtual simulation and evaluation, and the potential high-risk areas and sections in the digital twin hologram are identified to obtain the road network safety risk distribution. In the implementation process, a traffic flow simulation model based on graph neural network is first constructed. The model represents the road network as a dynamic graph structure, with nodes representing intersections and edges representing road sections. The graph convolutional network (GCN) is used to encode the attributes of nodes and edges to capture the local traffic state. Then, the graph attention network (GAT) is applied to simulate the propagation of vehicles in the road network and calculate the interaction between nodes. In order to improve the authenticity of the simulation, a vehicle agent model is introduced, and each agent has an adaptive behavior strategy that is continuously optimized through reinforcement learning. The selected traffic management scheme and emergency plan are used as input, and multiple rounds of simulation are performed in the digital twin hologram. The Monte Carlo Tree Search (MCTS) algorithm is used to explore different implementation orders and parameter combinations to evaluate the effect of each scheme. The evaluation indicators include average travel time, congestion level, accident risk, etc. Through anomaly detection algorithms, such as Isolation Forest, abnormal states that occur during the simulation process are identified to locate potential high-risk areas and sections. Finally, the kernel density estimation (KDE) method is used to generate a continuous road network safety risk heat map based on the cumulative distribution of risk events.

[0049] Specifically, according to the distribution of road network safety risks, areas of different risk levels are marked on the digital twin hologram, and safety warnings are issued to the target road network in combination with traffic management plans, emergency plans and road network digital portraits. In the implementation process, the generated safety risk heat map is first discretized, and the risk areas are divided into different levels using clustering algorithms such as DBSCAN. For each risk level, the corresponding visualization style is defined, such as color, transparency, texture, etc. These risk areas are rendered in the three-dimensional scene of the digital twin hologram using WebGL technology to ensure real-time interactive performance. Then, a multi-level warning system is designed, including global warning, regional warning and node warning. Global warning is based on the overall risk distribution statistics, and time series anomaly detection algorithms such as Prophet are used to identify system-level risk trends. Regional warning focuses on medium-scale risk aggregation, and spatial autocorrelation analysis such as Getis-Ord Gi* statistics are used to identify hot spots. Node warning targets a single key node and uses a sliding time window and threshold trigger mechanism to monitor state changes in real time. The warning information is associated with the traffic management plan and emergency plan, and corresponding disposal suggestions are specified for each warning scenario. Finally, a visual cockpit interface is developed to integrate digital images of the road network, real-time traffic status, risk distribution, and warning information. Data-driven document (D3.js) technology is used to achieve interactive data visualization, supporting multi-dimensional data exploration and decision support. Real-time data push is achieved through the WebSocket protocol to ensure the timeliness of warning information.

[0050] In this embodiment, the original traffic data stream of traffic participants is collected by multi-sensing devices. The data stream is processed by edge computing and multi-source data fusion technology, and a digital twin hologram of the road network is generated by deep learning algorithm and radar fitting technology. The hologram combines the road network analysis and prediction model to perform spatiotemporal analysis and situation prediction on the original data to generate a digital portrait of the road network. According to the digital portrait of the road network, the matching traffic management plan and emergency plan set are retrieved from the preset traffic control strategy library, and a safety warning is provided in the digital twin hologram. The present invention performs multi-source data fusion and situation prediction on the key nodes of the target road network, generates a digital twin hologram and a digital portrait of the road network, analyzes the road network status in real time and matches the corresponding traffic management plan and emergency plan, and then provides a more accurate safety warning service.

[0051] The above describes the road network safety warning method based on hologram in the embodiment of the present invention. The following describes the road network safety warning device based on hologram in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a road network safety early warning device based on a hologram includes: A deployment collection module 201 is used to deploy multi-sensing devices on key nodes of the target road network, and to collect multi-dimensional data of traffic participants in the target road network through the multi-sensing devices to obtain original traffic data streams; The hologram generation module 202 is used to perform edge computing processing and multi-source data fusion processing on the original traffic data stream, and to model and analyze the processed original traffic data stream through deep learning algorithms and radar fitting technology to obtain a digital twin hologram of the target road network; The analysis and prediction module 203 is used to perform multi-dimensional spatiotemporal analysis and situation prediction on the original traffic data stream through a preset road network analysis and prediction model to obtain analysis and prediction results, and generate a road network digital portrait of the target road network by combining the analysis and prediction results with the original traffic data stream; The safety warning module 204 is used to search and match the preset traffic control strategy library based on the digital portrait of the road network, obtain the traffic management plan and the emergency plan set, and issue a safety warning to the target road network in the digital twin hologram based on the digital portrait of the road network, the traffic management plan and the emergency plan set.

[0052] In an embodiment of the present invention, the hologram-based road network safety warning device runs the above-mentioned hologram-based road network safety warning method, and the hologram-based road network safety warning device collects the original traffic data stream of traffic participants through a multi-sensing device. The data stream is processed by edge computing and multi-source data fusion technology, and a digital twin hologram of the road network is generated by a deep learning algorithm and a radar fitting technology. The hologram is combined with a road network analysis and prediction model to perform spatiotemporal analysis and situation prediction on the original data to generate a digital portrait of the road network. According to the digital portrait of the road network, a matching traffic management plan and emergency plan set are retrieved from a preset traffic control strategy library, and a safety warning is provided in the digital twin hologram. The present invention performs multi-source data fusion and situation prediction on the key nodes of the target road network, generates a digital twin hologram and a digital portrait of the road network, analyzes the road network status in real time and matches the corresponding traffic management plan and emergency plan, and then provides a more accurate safety warning service.

[0053] above Figure 2 The hologram-based road network safety warning device in the embodiment of the present invention is described in detail from the perspective of modular functional entities.

[0054] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the hologram-based road network safety warning method.

[0055] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device, or unit can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0056] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0057] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A road network safety early warning method based on hologram, characterized in that: The road network safety early warning method based on hologram includes: Deploy multi-sensing devices at key nodes of the target road network, collect multi-dimensional data of traffic participants in the target road network through the multi-sensing devices, and obtain original traffic data streams; The original traffic data stream is processed by edge computing and multi-source data fusion, and the processed original traffic data stream is modeled and analyzed through deep learning algorithms and radar fitting technology to obtain a digital twin hologram of the target road network. Perform multi-dimensional spatiotemporal analysis and situation prediction on the original traffic data stream through a preset road network analysis and prediction model to obtain analysis and prediction results, and generate a road network digital portrait of the target road network by combining the analysis and prediction results with the original traffic data stream; Based on the digital portrait of the road network, the preset traffic control strategy library is searched and matched to obtain a traffic management plan and an emergency plan set, and a safety warning is issued for the target road network in the digital twin hologram based on the digital portrait of the road network, the traffic management plan and the emergency plan set.

2. The hologram-based road network safety early warning method according to claim 1 is characterized in that: The multi-sensing devices are deployed on the key nodes of the target road network, and the multi-dimensional data of the traffic participants in the target road network are collected by the multi-sensing devices to obtain the original traffic data stream including: The topological structure of the target road network is analyzed by graph theory algorithm and traffic flow theory to obtain the key node weight matrix of the target road network, and the multi-sensing devices are optimized and configured according to the key node weight matrix to obtain the sensing device deployment strategy; Based on the sensing device deployment strategy, multiple sensing devices are deployed on the target road network, and data collection parameters of the deployed multiple sensing devices are configured through edge computing units and 5G networks to obtain a distributed sensing network; The various multi-sensing devices in the distributed sensing network are coordinated and scheduled to obtain differentiated original data collection plans, and the various multi-sensing devices perform multi-dimensional data collection on traffic participants in the target road network according to the original data collection plan to obtain an original traffic data stream.

3. The hologram-based road network safety early warning method according to claim 1 is characterized in that: The edge computing and multi-source data fusion processing are performed on the original traffic data stream, and the processed original traffic data stream is modeled and analyzed by deep learning algorithm and radar fitting technology to obtain the digital twin hologram of the target road network, including: Perform distributed edge computing processing on the original traffic data stream to obtain an edge feature set, and globally integrate the edge feature set through a preset spatiotemporal alignment algorithm and semantic annotation model to obtain a fusion feature matrix; Applying a preset deep learning algorithm to perform feature mapping and pattern recognition on the fused feature matrix to obtain a high-dimensional representation of the road network state; Holographically processing the road network state representation through a preset radar vision fitting model to obtain a holographic road network model; The preset point cloud registration and three-dimensional reconstruction algorithms are applied to virtually reproduce the road network environment of the target road network according to the holographic road network model to obtain a digital twin hologram of the target road network.

4. The hologram-based road network safety early warning method according to claim 3 is characterized in that: The holographic processing of the road network state representation by a preset radar vision fitting model to obtain a holographic road network model includes: Extracting and preprocessing radar data and visual data in the original traffic data stream to obtain a radar and visual data set; Applying a preset radar vision fitting model to perform feature fusion on the road network state representation and the radar vision data set to obtain a multimodal fusion feature representation; Classify and identify static targets and dynamic targets in a road network environment based on the multimodal fusion feature representation through a target detection algorithm and a trajectory prediction model, and generate a time-varying road network scene description based on the classification and identification results; The time-varying road network scene description is semantically segmented and structured through a graph convolutional network and a temporal attention mechanism to obtain a holographic road network model.

5. The hologram-based road network safety early warning method according to claim 1, characterized in that: The original traffic data stream is subjected to multi-dimensional spatiotemporal analysis and situation prediction by using a preset road network analysis and prediction model, and the analysis and prediction results obtained include: Performing multi-scale time decomposition on the original traffic data stream by using preset average pooling kernels of different sizes to obtain separated periodic components and trend components; A feature matrix representing local time features and global time features is generated according to the periodic components through a gated one-dimensional local-global time convolution module, and features of different time scales in the feature matrix are weighted and aggregated through an attention mechanism to obtain a multi-scale time feature vector; The long-term evolution characteristics of the traffic state in the trend component are extracted and enhanced by multi-layer linear regression and residual connection to obtain a trend feature vector; The multi-scale time feature vector and the trend feature vector are integrated through an adaptive fusion network to obtain a comprehensive time feature representation; Based on the comprehensive temporal feature representation, the spatial dependency of the road network is dynamically modeled to generate a time-varying graph structure, and the road connection relationships and traffic status propagation at different levels in the target road network are captured and aggregated based on the time-varying graph structure through a preset multi-granularity graph convolutional network to obtain analysis and prediction results.

6. The hologram-based road network safety early warning method according to claim 5, characterized in that: Based on the comprehensive time feature representation, the spatial dependency of the road network is dynamically modeled to generate a time-varying graph structure, and the road connection relationships and traffic state propagation at different levels in the target road network are captured and aggregated based on the time-varying graph structure through a preset multi-granularity graph convolution network, and the analysis and prediction results obtained include: Dynamically modeling the spatial dependency of the road network based on the comprehensive time feature representation through an adaptive adjacency matrix generation algorithm, calculating the time-varying correlation strength between key nodes in the target road network, and obtaining a time-varying graph structure that represents the dynamic topological structure of the road network; By using a multi-level graph convolutional network, feature extraction is performed on node relationships at different levels in the road network according to the time-varying graph structure, thereby obtaining multi-scale spatial features containing local and global spatial dependency information; The multi-scale spatial features and the comprehensive temporal feature representation are fused through the spatiotemporal attention mechanism to obtain the spatiotemporal features of each key node, and the spatiotemporal features are weighted to obtain the spatiotemporal feature representation of the key nodes; The future state of each key node is evaluated and classified based on the temporal and spatial characteristics of the key nodes through a preset multi-dimensional safety threshold to obtain an analysis and prediction result.

7. The hologram-based road network safety early warning method according to claim 1, characterized in that: Based on the digital portrait of the road network, searching and matching a preset traffic control strategy library to obtain a traffic management plan and an emergency plan set, and providing a safety warning for the target road network in the digital twin hologram according to the digital portrait of the road network, the traffic management plan and the emergency plan set include: Generate a road network status description of the target road network according to the road network digital portrait, and match the road network status description with a preset traffic control strategy library to obtain a traffic management plan and an emergency plan set; Mapping the traffic management plan and emergency plan set to the digital twin hologram for virtual simulation and evaluation, identifying potential high-risk areas and sections in the digital twin hologram, and obtaining the road network safety risk distribution; According to the safety risk distribution of the road network, areas with different risk levels are marked on the digital twin hologram, and safety warnings are issued to the target road network in combination with the traffic management plan, emergency plan and digital portrait of the road network.

8. A road network safety warning device based on hologram, characterized in that: The road network safety early warning device based on hologram includes: A deployment collection module is used to deploy multi-sensing devices on key nodes of the target road network, and to collect multi-dimensional data of traffic participants in the target road network through the multi-sensing devices to obtain original traffic data streams; A hologram generation module is used to perform edge computing processing and multi-source data fusion processing on the original traffic data stream, and to model and analyze the processed original traffic data stream through a deep learning algorithm and a radar fitting technique to obtain a digital twin hologram of the target road network; An analysis and prediction module is used to perform multi-dimensional spatiotemporal analysis and situation prediction on the original traffic data stream through a preset road network analysis and prediction model to obtain analysis and prediction results, and generate a road network digital portrait of the target road network by combining the analysis and prediction results with the original traffic data stream; The safety warning module is used to search and match the preset traffic control strategy library based on the digital portrait of the road network, obtain the traffic management plan and the emergency plan set, and issue a safety warning to the target road network in the digital twin hologram according to the digital portrait of the road network, the traffic management plan and the emergency plan set.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the steps of the hologram-based road network safety warning method as described in any one of claims 1-7 are implemented.

Images