Sensing intelligent driving complex traffic scene dynamic risk prediction method

Abstract

The invention discloses a perception intelligent driving complex traffic scene dynamic risk prediction method, and relates to the technical field of risk prediction, and the method comprises the steps: obtaining the multi-source traffic dynamic data of a target region in real time; performing space-time alignment processing on the multi-source traffic dynamic data; inputting the space-time coupling feature matrix into a pre-trained depth space-time prediction model to generate a risk thermodynamic map; calculating a dynamic risk index of each traffic sub-region, and generating a risk level distribution sequence; and triggering a self-adaptive early warning response mechanism according to the risk level distribution sequence, and dynamically adjusting operation parameters of the variable information sign and the traffic signal controller. The technical problems of frequent traffic congestion and high accident risk caused by inaccurate traffic risk prediction and difficulty in dynamic adjustment according to the real-time traffic condition in the prior art are solved, and the technical effects of accurately predicting the traffic risk and improving the safety and traffic efficiency in a complex traffic scene are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of risk prediction, and particularly to a dynamic risk prediction method for complex traffic scenarios driven by perceptual intelligence. Background Art

[0002] In the existing traffic system, traditional traffic management methods mainly rely on manual experience and simple traffic monitoring devices, making it difficult to comprehensively and in real-time grasp multi-source traffic dynamic data. This results in difficulties in accurately predicting traffic risks in a timely manner. For example, it is impossible to know in advance which sections may experience congestion or high accident rates during specific time periods. At the same time, it is difficult to dynamically adjust the operating parameters of traffic facilities according to real-time traffic conditions. For instance, the signal timing plan is usually fixed and cannot flexibly respond to changes in traffic flow, leading to frequent traffic jams and high accident risks, seriously affecting safety and traffic efficiency in complex traffic scenarios.

[0003] The existing technology has technical problems such as inaccurate traffic risk prediction, difficulty in dynamically adjusting according to real-time traffic conditions, resulting in frequent traffic jams and high accident risks. Summary of the Invention

[0004] This application provides a dynamic risk prediction method for complex traffic scenarios driven by perceptual intelligence, aiming to solve the technical problems in the existing technology, including inaccurate traffic risk prediction, difficulty in dynamically adjusting according to real-time traffic conditions, resulting in frequent traffic jams and high accident risks.

[0005] In view of the above problems, this application provides a dynamic risk prediction method for complex traffic scenarios driven by perceptual intelligence. The method includes:

[0006] Obtain multi-source traffic dynamic data of the target area in real time; perform spatio-temporal alignment processing on the multi-source traffic dynamic data, extract a multi-dimensional risk feature vector including traffic flow mutation features, road topology constraint features, and environmental interference factors to construct a spatio-temporal coupling feature matrix; input the spatio-temporal coupling feature matrix into a pre-trained deep spatio-temporal prediction model to generate a risk heat map; perform real-time analysis on the risk heat map based on an edge computing node cluster, calculate the dynamic risk index of each traffic sub-area, and generate a risk level distribution sequence; trigger an adaptive early warning response mechanism according to the risk level distribution sequence, and dynamically adjust the operating parameters of variable message signs and traffic signal controllers in the target area.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] Obtain multi-source traffic dynamic data of the target area in real time; perform spatio-temporal alignment processing on the multi-source traffic dynamic data, extract a multi-dimensional risk feature vector including traffic flow mutation features, road topology constraint features, and environmental interference factors to construct a spatio-temporal coupling feature matrix; input the spatio-temporal coupling feature matrix into a pre-trained deep spatio-temporal prediction model to generate a risk heat map; perform real-time analysis on the risk heat map based on an edge computing node cluster, calculate the dynamic risk index of each traffic sub-region, and generate a risk level distribution sequence; trigger an adaptive early warning response mechanism according to the risk level distribution sequence, and dynamically adjust the operating parameters of variable message signs and traffic signal controllers in the target area. It achieves the technical effects of accurately predicting traffic risks and improving safety and traffic efficiency in complex traffic scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0010] Figure 1 It is a schematic flowchart of a method for predicting dynamic risks in a complex traffic scenario driven by perception intelligence provided by an embodiment of the present application;

[0011] Figure 2 It is a schematic flowchart of constructing a spatio-temporal coupling feature matrix in a method for predicting dynamic risks in a complex traffic scenario driven by perception intelligence provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] The present application provides a method for predicting dynamic risks in a complex traffic scenario driven by perception intelligence, which is used to solve the technical problems in the prior art that traffic risk prediction is inaccurate and it is difficult to dynamically adjust according to real-time traffic conditions, resulting in frequent traffic jams and high accident risks.

[0013] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0014] Embodiment, as Figure 1 shown, the present application provides a method for predicting dynamic risks in a complex traffic scenario driven by perception intelligence, and the method includes:

[0015] Step S100: Obtain multi-source traffic dynamic data of the target area in real time.

[0016] Specifically, various sensors, monitoring devices, and data acquisition systems are used to obtain multi-source traffic dynamic data of the target area in real time. These data include video image data, traffic flow monitoring data, environmental data, etc., comprehensively reflecting the immediate state of the traffic scene.

[0017] Step S200: Perform spatio-temporal alignment processing on the multi-source traffic dynamic data, extract a multi-dimensional risk feature vector containing traffic flow mutation features, road topology constraint features, and environmental interference factors, and construct a spatio-temporal coupling feature matrix.

[0018] Specifically, alignment processing is carried out on the obtained multi-source traffic dynamic data. Through a multi-modal feature alignment network, different types of data are transformed into a set of standardized feature vectors. Then, according to the spatio-temporal correlation degree distribution, the weights of traffic flow mutation features, road topology constraint features, and environmental interference factors are dynamically adjusted to generate a fused feature vector. A redundancy check mechanism is used to evaluate the confidence of the fused feature vector, eliminate abnormal features, and finally construct a spatio-temporal coupling feature matrix containing the mapping relationship between timestamp indexes and spatial coordinates, integrating multi-dimensional risk feature vectors, laying a foundation for subsequent analysis.

[0019] Step S300: Input the spatio-temporal coupling feature matrix into a pre-trained deep spatio-temporal prediction model to generate a risk heat map.

[0020] Specifically, first, the pre-trained deep spatio-temporal prediction model that has been completed training before will be called. This model is trained based on a large amount of historical traffic data and has learned the complex relationship between spatio-temporal laws and risk features in traffic data. At this time, the constructed spatio-temporal coupling feature matrix is provided as input to this pre-trained deep spatio-temporal prediction model. The multi-dimensional risk feature vectors such as traffic flow mutation features, road topology constraint features, and environmental interference factors contained in the matrix, as well as the mapping relationship between timestamp indexes and spatial coordinates, provide rich information for the model's analysis. After receiving the input, the deep spatio-temporal prediction model uses its internal complex neural network structure to perform in-depth analysis and processing on these features. After a series of calculations and inferences inside the model, a risk heat map is finally generated. This map visually presents the risk levels at different positions in the target area, for example, using different shades of color or heat to represent the high and low risks, enabling traffic managers and relevant personnel to quickly understand which areas have higher traffic risks and which areas have relatively lower risks, providing an important basis for subsequent targeted traffic management measures.

[0021] Step S400: Based on the edge computing node cluster, perform real-time analysis on the risk heat map, calculate the dynamic risk index of each traffic sub-region, and generate a risk level distribution sequence.

[0022] Specifically, use the edge computing cluster to perform real-time analysis on the risk heat map. According to the preset traffic sub-region division rules, divide the map into multiple sub-region heat maps. Each edge computing node is responsible for the analysis of the corresponding sub-region map, uses local computing resources to statistically analyze the risk data, and calculates the dynamic risk index of each sub-region according to the preset risk assessment model. Arrange these indices in chronological order to generate a risk level distribution sequence, presenting the changing trend of risks over time.

[0023] Step S500: Trigger an adaptive early warning response mechanism according to the risk level distribution sequence, and dynamically adjust the operating parameters of variable message signs and traffic signal controllers in the target area.

[0024] Specifically, trigger an adaptive early warning response mechanism according to the generated risk level distribution sequence. Set the early warning thresholds and methods corresponding to different risk levels in advance. When the risk level reaches or exceeds the threshold, push early warning information in a timely manner. At the same time, use the reinforcement learning algorithm to optimize the display content of variable message signs according to the changing trend of the risk level distribution sequence, and transmit accurate traffic information to road users; use the traffic signal optimization algorithm to dynamically adjust the signal timing plan of the traffic signal controller in combination with the risk level and traffic flow conditions, effectively relieve traffic congestion, and ensure traffic safety.

[0025] In a possible implementation manner, as Figure 2 shown, step S200 further includes:

[0026] Step S210: Perform spatio-temporal alignment processing on the multi-source traffic dynamic data through a multi-modal feature alignment network to generate a standardized feature vector set.

[0027] Step S220: Based on the spatio-temporal correlation degree distribution of the standardized feature vector set, dynamically adjust the weight distribution coefficients of the traffic flow mutation feature, road topology constraint feature, and environmental interference factor to generate a fused feature vector.

[0028] Step S230: Use a redundancy check mechanism to evaluate the confidence level of the fused feature vector, and eliminate abnormal features with a confidence level lower than a predetermined threshold to generate the multi-dimensional risk feature vector.

[0029] Step S240: According to the spatio-temporal correlation of the multi-dimensional risk feature vector, reconstruct a spatio-temporal coupling feature matrix including the mapping relationship between the time stamp index and the spatial coordinates.

[0030] Specifically, a multi-modal feature alignment network is used to perform spatio-temporal alignment processing on the obtained multi-source traffic dynamic data. For different types of data such as video image data, traffic flow monitoring data, and environmental data, this network respectively uses convolutional neural networks, wavelet transform decomposition and reconstruction algorithms, normalization processing, etc. to convert them into image feature sub-vectors, traffic flow feature sub-vectors, and environmental feature sub-vectors. Then, with the help of the attention mechanism, weighted fusion is carried out according to the importance of each sub-vector in different spatio-temporal dimensions, so as to generate a standardized feature vector set, making various types of data consistent and comparable in the spatio-temporal dimension.

[0031] Further process based on the spatio-temporal correlation degree distribution of the standardized feature vector set. By calculating the time series correlation and spatial distance correlation between vectors, a spatio-temporal correlation matrix is constructed. Then, the hierarchical clustering algorithm is used to perform clustering analysis on the standardized feature vectors to clarify the correlation degree between different types of feature vectors and traffic flow mutation features, road topology constraint features, and environmental interference factors. According to these correlation degrees, the adaptive weight adjustment algorithm is used to dynamically calculate the weight distribution coefficients of each feature and factor, and then weighted summation is performed on them to generate a fused feature vector, fully integrating various types of feature information.

[0032] Use the redundancy check mechanism to control the quality of the fused feature vector. The fused feature vector is randomly sampled multiple times to obtain multiple sampled subsets. The principal component analysis algorithm is used to reduce the dimension of the sampled subsets, extract the main feature components and calculate the similarity between them to construct a similarity matrix. Based on this matrix, an anomaly detection algorithm based on a threshold is used to evaluate the confidence of each feature in the fused feature vector, and the abnormal features with a confidence lower than the predetermined threshold are removed to generate a more reliable multi-dimensional risk feature vector, improving the data quality.

[0033] Based on the multi-dimensional risk feature vectors obtained in the early stage, the spatio-temporal coupling feature matrix begins to be constructed. First, deeply explore the internal connections of the multi-dimensional risk feature vectors in the time and space dimensions, such as analyzing the change trends of each feature vector at different times and their mutual influences in spatial positions. Then, incorporate the time information in the form of timestamp indexes, accurately recording the time points corresponding to each feature vector, which enables the data to be arranged in an orderly manner according to the time sequence, facilitating the tracking of the evolution of traffic risks over time. At the same time, establish a spatial coordinate mapping relationship, associating the feature vectors with specific geographical locations within the target area, ensuring that each feature vector can clearly correspond to the actual traffic scene location. In this way, the timestamp indexes and the spatial coordinate mapping relationship are organically integrated with the multi-dimensional risk feature vectors to reconstruct the spatio-temporal coupling feature matrix. This matrix completely retains the information of the multi-dimensional risk feature vectors and clearly shows the distribution and variation laws of these features in the spatio-temporal dimensions, providing a key data basis for subsequent dynamic risk prediction of complex traffic scenes using deep spatio-temporal prediction models, helping to more accurately grasp the spatio-temporal dynamic changes of traffic risks, and thus realizing efficient traffic management and risk warning.

[0034] In a possible implementation manner, step S210 further includes:

[0035] Step S211: For the video image data in the multi-source traffic dynamic data, use a convolutional neural network for feature extraction to obtain image feature sub-vectors.

[0036] Step S212: For the traffic flow monitoring data in the multi-source traffic dynamic data, use the wavelet transform decomposition and reconstruction algorithm to extract the flow change features in different frequency bands and generate flow feature sub-vectors.

[0037] Step S213: Normalize the environmental data in the multi-source traffic dynamic data to form environmental feature sub-vectors.

[0038] Step S214: Input the image feature sub-vectors, flow feature sub-vectors, and environmental feature sub-vectors into the multi-modal feature alignment network. Through the attention mechanism in the network, perform weighted fusion according to the importance of each sub-vector in different spatio-temporal dimensions to generate a standardized feature vector set.

[0039] Specifically, for the video image data in multi-source traffic dynamic data, a convolutional neural network is used for processing. With its powerful image feature extraction ability, the convolutional neural network can automatically learn and identify the key information in video images. Through a series of operations of convolutional layers, pooling layers, and activation functions, features such as the number of vehicles, driving directions, speed changes, and traffic congestion areas are extracted from the video images, converting these complex image information into concise image feature sub-vectors, which contain the key information closely related to the traffic conditions in the video images.

[0040] For traffic flow monitoring data, a wavelet transform decomposition and reconstruction algorithm is used. This algorithm can decompose the traffic flow monitoring data in different frequency bands, respectively extracting the long-term trend information reflected in the low-frequency band and the short-term fluctuation details reflected in the high-frequency band. For example, the low-frequency band can present the overall change trend of traffic flow at different times of the day, and the high-frequency band captures the instantaneous changes in traffic flow. Integrating the flow change characteristics of these different frequency bands to generate a flow feature sub-vector, comprehensively reflecting the dynamic changes in traffic flow.

[0041] When processing environmental data, since environmental data may have different dimensions and value ranges, in order to facilitate subsequent fusion calculations, normalization processing is required. Through a specific normalization formula, environmental data such as temperature, humidity, and light intensity are uniformly mapped to a standard interval, eliminating the influence of dimensional differences, forming environmental feature sub-vectors, and ensuring that environmental data can be compared and fused with other types of data on the same scale.

[0042] Input the previously generated image feature sub-vectors, flow feature sub-vectors, and environmental feature sub-vectors into the multi-modal feature alignment network. The attention mechanism in this network plays a key role. It will automatically evaluate the importance of each sub-vector in different spatio-temporal dimensions. For example, during traffic congestion periods, the flow feature sub-vector may be more important; while when bad weather affects the traffic vision, the importance of the image feature sub-vector and the environmental feature sub-vector will increase. According to these evaluation results, the attention mechanism performs weighted fusion on each sub-vector, highlighting the role of important sub-vectors and weakening the influence of relatively unimportant sub-vectors, and finally generating a set of standardized feature vectors. These sets of standardized feature vectors integrate the key information of multi-source traffic dynamic data, laying a solid data foundation for subsequent construction of spatio-temporal coupling feature matrices and dynamic risk prediction of complex traffic scenarios.

[0043] In a possible implementation manner, step S220 further includes:

[0044] Step S221: Calculate the time series correlation and spatial distance correlation between the vectors in the set of standardized feature vectors, and construct a spatio-temporal correlation matrix.

[0045] Step S222: According to the spatio-temporal correlation matrix, use the hierarchical clustering algorithm to perform clustering analysis on the standardized feature vectors, and determine the correlation degrees of different types of feature vectors with the traffic flow mutation features, road topology constraint features, and environmental interference factors.

[0046] Step S223: According to the correlation degrees, use the adaptive weight adjustment algorithm to dynamically calculate the weight distribution coefficients of the traffic flow mutation features, road topology constraint features, and environmental interference factors.

[0047] Step S224: According to the weight distribution coefficients, perform weighted summation on the traffic flow mutation features, road topology constraint features, and environmental interference factors to generate a fused feature vector.

[0048] Specifically, for the obtained standardized feature vector set, in order to deeply explore its internal connections in the time and space dimensions, it is necessary to calculate the time series correlation and spatial distance correlation between each vector respectively. For the time series correlation, use the Pearson correlation coefficient to analyze the standardized feature vectors at different times, and quantify the similarity degree and the consistency of the change trends between them. For example, through calculation, it can be clarified whether the change of the traffic flow feature vector in a certain time period is closely related to the feature vectors at other times, and the strength of this relationship. In terms of calculating the spatial distance correlation, methods such as Euclidean distance and cosine similarity are used to consider the differences and similarities between the standardized feature vectors at different spatial positions. For example, compare the feature vectors of adjacent road sections or different regions to judge their spatial correlations, and analyze the influence range and degree of the change of the traffic conditions in a certain region on the surrounding regions. Integrate the calculated time series correlation and spatial distance correlation results, arrange them according to certain rules, and finally construct a spatio-temporal correlation matrix. This matrix intuitively shows the mutual correlation degrees of the vectors in the standardized feature vector set in time and space, provides an important data basis for subsequent clustering analysis and feature fusion operations, and helps to more comprehensively understand and analyze the spatio-temporal characteristics of multi-source traffic dynamic data.

[0049] Based on the constructed spatio-temporal correlation matrix, use the hierarchical clustering algorithm to perform clustering analysis on the standardized feature vectors. This algorithm will divide the feature vectors into different categories according to the similarity degree between the vectors. After clustering is completed, further explore the connections between different types of feature vectors and the traffic flow mutation features, road topology constraint features, and environmental interference factors, determine the tightness of the correlation between each type of feature vector and these key features and factors respectively, clarify which types of feature vectors are greatly affected by traffic flow mutations, which are significantly restricted by the road topology structure, and which are closely related to environmental interference, so as to provide a basis for dynamically adjusting the weights.

[0050] To dynamically calculate the weight distribution coefficients of traffic flow mutation characteristics, road topology constraint characteristics, and environmental interference factors based on the degree of association, the gradient descent method, an adaptive weight adjustment algorithm, is adopted. First, a target function is defined. For example, the sum of the squares of the errors between the predicted traffic risk and the actual traffic risk can be used as the target function, and the goal is to minimize this sum of squared errors. In the initial stage, an initial weight value is assigned to the traffic flow mutation characteristics, road topology constraint characteristics, and environmental interference factors respectively. Then, based on the degree of association between the different category feature vectors obtained in the previous steps and these three key characteristics and factors, the weights are iteratively adjusted. In each iteration, the gradient of the target function with respect to each weight is calculated. The gradient represents the rate of change of the target function at the current weights. According to the direction and magnitude of the gradient, the weights are updated according to a certain learning rate. If the degree of association shows that a certain characteristic or factor has a greater impact on traffic risk, then in the iterative process, the weight update amplitude corresponding to this characteristic or factor will be relatively large; conversely, if the degree of association is weak, the weight update amplitude will be small. By continuously iteratively calculating and updating the weights, the target function gradually converges to a smaller value. At this time, the obtained weights are the optimal weight distribution coefficients of the traffic flow mutation characteristics, road topology constraint characteristics, and environmental interference factors. The entire process dynamically adjusts the weights according to the actual situation of the data to ensure that the final weight distribution can accurately reflect the importance of each characteristic and factor in traffic risk prediction.

[0051] According to the calculated weight distribution coefficients, weighted summation is performed on the traffic flow mutation characteristics, road topology constraint characteristics, and environmental interference factors. Multiply each characteristic and factor by its corresponding weight and then sum them up. Integrate these characteristic information from different sources and of different natures organically to generate a fused feature vector. This fused feature vector comprehensively considers the impacts of various key factors under different spatio-temporal conditions, and more comprehensively and accurately reflects the actual situation of the traffic scene, providing more valuable characteristic data for subsequent risk assessment and prediction.

[0052] In a possible implementation manner, step S230 further includes:

[0053] Step S231: Perform multiple random samplings on the fused feature vector to obtain multiple sampled subsets.

[0054] Step S232: For the multiple sampled subsets, use the principal component analysis algorithm for dimensionality reduction processing to extract multiple main feature components.

[0055] Step S233: Calculate the similarity between the multiple main feature components and construct a similarity matrix.

[0056] Step S234: According to the similarity matrix, use a threshold-based anomaly detection algorithm to evaluate the confidence of each feature in the fused feature vector.

[0057] Step S235: Remove the abnormal features with confidence lower than the predetermined threshold from the fused feature vector to generate the multi-dimensional risk feature vector.

[0058] Specifically, perform multiple random samplings on the fused feature vector. Random sampling can increase the diversity of samples and avoid misjudgment of data features caused by sampling bias. By repeating random sampling multiple times, multiple sampled subsets are obtained, and each sampled subset contains some features in the fused feature vector.

[0059] For these sampled subsets, use the principal component analysis algorithm for dimensionality reduction. The principal component analysis algorithm can convert high-dimensional data into low-dimensional data and extract multiple main feature components on the premise of retaining the main information of the data as much as possible. This is because the fused feature vector may contain many features, and there is a correlation between some of these features, resulting in information redundancy. Through principal component analysis, these related features are combined into a few comprehensive main feature components, which not only reduces the data dimension and computational complexity, but also highlights the core features of the data more prominently.

[0060] Calculate the similarity between these main feature components and construct a similarity matrix. By calculating the similarity, the degree of association between each main feature component can be understood. Using calculation methods such as cosine similarity, the similarity values between each main feature component and other components are obtained, and these values are arranged in a certain order to form a similarity matrix, which intuitively shows the similarity relationship between the main feature components.

[0061] Based on the similarity matrix constructed above, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise, a density-based spatial clustering algorithm) threshold-based anomaly detection algorithm is used to evaluate the confidence of each feature in the fused feature vector. First, two key thresholds are determined: the neighborhood radius (eps) and the minimum number of points (MinPts). For each feature in the fused feature vector, based on the similarity data of its corresponding main feature component in the similarity matrix, in the data space, the region centered on this feature with a radius of eps is defined as the neighborhood. Count the number of points in the neighborhood. If the number of points in the neighborhood is greater than or equal to MinPts, then this feature is marked as a core point; if the number of points is less than MinPts but this point falls within the neighborhood of a core point, it is marked as a border point; the remaining points are marked as noise points. The features where the core points and border points are located are considered normal features with relatively high confidence because they are in areas with relatively dense data distributions and the similarity with surrounding features meets expectations; while the features corresponding to the noise points are regarded as abnormal features with low confidence because they are far from the data-dense areas and are significantly different from other features in terms of similarity. Through the clustering and noise point identification process of the DBSCAN algorithm, the confidence of each feature in the fused feature vector can be effectively evaluated, providing a basis for removing abnormal features in the subsequent steps.

[0062] Remove the abnormal features with confidence lower than the predetermined threshold from the fused feature vector. After such screening and processing, the possible incorrect or abnormal data are removed, generating a more accurate and reliable multi-dimensional risk feature vector. These multi-dimensional risk feature vectors exclude the interference of abnormal features and can more truly reflect the risk features in the traffic scene, providing high-quality data support for the subsequent dynamic risk prediction in complex traffic scenes and helping to improve the accuracy and reliability of risk prediction.

[0063] In a possible implementation manner, step S300 further includes:

[0064] Step S310: Collect historical traffic data and construct a training data set, where the training data set includes sample multi-source traffic dynamic data and sample risk heat map labels.

[0065] Step S320: Perform spatio-temporal alignment processing, feature extraction, and feature matrix construction operations on the sample multi-source traffic dynamic data to obtain a sample spatio-temporal coupling feature matrix.

[0066] Step S330: Input the sample spatio-temporal coupling feature matrix into the initial deep spatio-temporal prediction model, output the initial risk heat map, and iteratively optimize the model parameters by calculating the mean square error loss function between the initial risk heat map and the sample risk heat map label to complete the pre-training of the deep spatio-temporal prediction model.

[0067] Specifically, first, historical traffic data is widely collected from various channels such as traffic monitoring devices and databases. These data cover multi-source traffic dynamic information in different time periods and regions, such as video image data, traffic flow monitoring data, and environmental data, etc., constituting the sample multi-source traffic dynamic data. At the same time, combined with the actual traffic conditions, the corresponding sample risk heat map labels are marked to clarify the real risk distribution under different spatio-temporal conditions, and the two are integrated to construct a training data set to provide rich sample data for model training.

[0068] For the sample multi-source traffic dynamic data, first, a multi-modal feature alignment network is used for spatio-temporal alignment processing. For video image data, convolutional neural networks are used to extract image features, such as vehicle position, speed, quantity, etc.; for traffic flow monitoring data, through wavelet transform decomposition and reconstruction algorithms, the flow change features in different frequency bands are obtained; environmental data is normalized. Then, these processed feature sub-vectors are input into the multi-modal feature alignment network, and with the help of the attention mechanism in the network, they are weighted and fused according to their importance in different spatio-temporal dimensions to generate a standardized feature vector set. Then feature extraction is carried out. Based on the standardized feature vector set, the time series correlation and spatial distance correlation between vectors are calculated to construct a spatio-temporal correlation matrix, and then the hierarchical clustering algorithm is used for analysis to determine the correlation degree between different category feature vectors and traffic flow mutation features, road topology constraint features, and environmental interference factors. The adaptive weight adjustment algorithm is used to calculate the weight distribution coefficient, and these features and factors are weighted and summed to obtain a fused feature vector. Finally, using the redundancy check mechanism, the fused feature vector is randomly sampled multiple times, the sampled subset is dimension-reduced by the principal component analysis algorithm, the similarity of the main feature components is calculated to construct a similarity matrix, and the confidence level is evaluated using a threshold-based anomaly detection algorithm, and the abnormal features below the predetermined threshold are removed to generate a multi-dimensional risk feature vector. According to its spatio-temporal correlation, the time stamp index and spatial coordinate mapping relationship are incorporated to construct the sample spatio-temporal coupling feature matrix.

[0069] Input the constructed sample spatio-temporal coupling feature matrix into the initial deep spatio-temporal prediction model, which is composed of a spatio-temporal convolution module, a long short-term memory network module, and an attention mechanism module working together. The spatio-temporal convolution module performs convolution operations on the input sample spatio-temporal coupling feature matrix to capture the local feature patterns of the data in the spatio-temporal dimension, such as the variation law of traffic flow in a specific time and space area; the long short-term memory network module processes time series data, and with its special memory unit structure, it can effectively capture the long-term dependencies of traffic data over time, such as the periodic changes during morning and evening rush hours; the attention mechanism module dynamically assigns weights according to the importance of features, focusing on the key features that have a greater impact on risk prediction, such as the feature information of accident-prone sections. After being processed by these three modules in sequence, the model outputs the initial risk heat map. Subsequently, by calculating the mean square error loss function between the initial risk heat map and the sample risk heat map label, the deviation degree between the model prediction result and the real situation is measured. The stochastic gradient descent algorithm is used. In each iteration process, calculate the gradient of the loss function with respect to the model parameters, and according to the direction and magnitude of the gradient, adjust the model parameters according to the preset learning rate to gradually reduce the loss value. Continuously perform iterative optimization until the loss function converges, that is, the loss value no longer has an obvious downward trend. At this time, the pre-training of the deep spatio-temporal prediction model is completed, enabling the model to have the ability to accurately predict the dynamic risks in complex traffic scenarios.

[0070] In a possible implementation manner, step S400 further includes:

[0071] Step S410: Divide the risk heat map into multiple sub-region heat maps according to a preset traffic sub-region division rule.

[0072] Step S420: Each edge computing node is responsible for the parsing task of one or more sub-region heat maps, and uses the local computing resources of the node to statistically analyze the risk data in the sub-region heat maps it is responsible for.

[0073] Step S430: Calculate the dynamic risk index of each sub-region according to a preset risk assessment model.

[0074] Step S440: Sort the dynamic risk indices of each traffic sub-region calculated by each edge computing node in chronological order to generate the risk level distribution sequence.

[0075] Specifically, in order to analyze the risk heat map more meticulously and accurately, it is necessary to segment it according to the preset traffic sub-region division rules. These division rules are formulated based on a comprehensive study of the target traffic area, fully considering various factors such as geographical information, road layout, and traffic flow distribution. For example, taking the main roads as boundaries, the traffic area is divided into different block-shaped areas; or according to the density of traffic flow, areas with similar traffic flows are merged into one sub-region. When performing the segmentation operation, according to these established rules, precise division is carried out on the risk heat map, cutting the original overall map into multiple relatively independent sub-region heat maps. Each sub-region heat map focuses on a specific traffic sub-region, clearly presenting the risk distribution within that sub-region, which facilitates subsequent more in-depth risk assessment and analysis of each sub-region, enabling traffic managers to pay more targeted attention to the risk situations in different regions, thereby improving the efficiency and accuracy of traffic risk management.

[0076] Each edge computing node is responsible for statistically analyzing the risk data in one or more sub-region heat maps. For the sub-region heat map it is responsible for, the node first uses the histogram statistical algorithm to divide the risk data into several intervals according to the risk level, and counts the frequency of risk data in each interval, so as to visually present the distribution of risk data at different levels. For example, it can be clearly seen the respective proportions of low, medium, and high risks. Then, the sliding window algorithm is used to analyze the time series characteristics of the risk data. Taking a certain time interval as the window, the window slides on the time axis, and statistical quantities such as the mean and variance of the risk data in each window are calculated to capture the dynamic changes of the risk data over time, such as judging whether the risk is gradually increasing, decreasing, or remaining stable. At the same time, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is used to perform spatial clustering analysis on the risk data. This algorithm will divide the data points that are closely connected in space and have similar risk characteristics into one cluster according to the distance and density between the risk data points, so as to identify the high-risk aggregation areas within the sub-region. The node will also use the association rule mining algorithm, such as the Apriori algorithm, to explore the potential associations between different risk factors, such as analyzing the association between road congestion degree and accident occurrence frequency. By comprehensively using these algorithms, the edge computing node can deeply and comprehensively statistically analyze the risk data in the sub-region heat map it is responsible for, providing a solid foundation for subsequent calculation of the sub-region dynamic risk index.

[0077] When constructing this neural network risk assessment model for calculating the dynamic risk index of sub-regions, the traffic flow, road congestion conditions, accident historical data, etc. within the sub-regions are first used as key input information. The traffic flow is quantified by the number of vehicle flows at different time intervals (such as 5 minutes, 15 minutes, etc.), the road congestion conditions are represented by indicators such as the proportion of the length of congested sections and the average congestion duration, and the accident historical data is refined into the accident occurrence frequency, the degree of casualties and property losses caused by the accidents, etc. These data are normalized to be within the range of 0-1 to eliminate the influence of dimensional differences. The model adopts a multi-layer perceptron (MLP) neural network structure. The number of neurons in the input layer is determined according to the number of features of the input data, and the processed multi-source data are respectively connected to the corresponding neurons. The hidden layer is set with 2-3 layers, and the number of neurons in each layer is determined based on experience and multiple experiments, generally ranging from dozens to hundreds. The neurons in the hidden layer are connected by weights, and activation functions such as Sigmoid or ReLU are used to introduce non-linearity, enabling the model to learn complex non-linear relationships. For example, the ReLU function can effectively solve the problem of gradient disappearance and make the model training more efficient. The output layer is a single neuron, and the output result is the dynamic risk index of the sub-region. When training the model, a large number of multi-source data samples and their corresponding true risk index labels in different sub-regions and different time periods are collected to form a training data set. The random gradient descent algorithm is used to optimize the model parameters, and the mean squared error (MSE) is selected as the loss function. During the training process, the model performs forward propagation calculation on each training sample to predict the risk index, then calculates the MSE loss between the predicted value and the true label, calculates the gradient of the loss with respect to the weights of each layer through the backpropagation algorithm, and adjusts the weights according to the gradient direction to continuously reduce the loss value. After multiple rounds of training, when the loss value converges and performs well on the validation set, the model training is completed. When actually calculating the dynamic risk index of the sub-region, the relevant data of the sub-region collected in real time are input into the trained model, and the model can output the current accurate dynamic risk index of the sub-region, providing a key quantitative basis for traffic risk control.

[0078] In chronological order, collect the dynamic risk indices of each traffic sub-region calculated by each edge computing node and sort them. Arrange the dynamic risk indices of each sub-region at different time points in sequence to generate a risk level distribution sequence. This sequence intuitively shows the changes in the risks of each traffic sub-region within the target area over time, provides a key basis for the traffic management department to timely discover high-risk areas, predict the development trend of risks, and take targeted traffic management measures, and helps to improve the risk management efficiency and safety in complex traffic scenarios.

[0079] In a possible implementation manner, step S500 further includes:

[0080] Step S510: Set warning thresholds and warning methods corresponding to different risk levels. When the risk level in the risk level distribution sequence reaches or exceeds the warning threshold, trigger the corresponding warning information for pushing.

[0081] Step S520: According to the changing trend of the risk level distribution sequence, adopt a reinforcement learning algorithm to optimize the display content of the variable message sign.

[0082] Step S530: Apply a traffic signal optimization algorithm to dynamically adjust the signal timing plan of the traffic signal controller according to the risk level and traffic flow conditions.

[0083] Specifically, it is necessary to preset warning thresholds and warning methods corresponding to different risk levels in advance. The warning thresholds are determined based on a large amount of historical traffic data and actual management experience. For example, the risk level is divided into three levels: low, medium, and high, and corresponding specific values are set as warning thresholds respectively. The warning methods cover various forms, such as SMS reminders, APP push notifications, radio announcements, etc. When the risk level in the risk level distribution sequence reaches or exceeds the warning threshold, quickly trigger the corresponding warning information for pushing. This process can timely inform the traffic management department, drivers, and relevant personnel of the current traffic risk situation, enabling them to make preparations in advance, thereby effectively reducing the possibility and impact range of accidents.

[0084] By analyzing the changing trend of the risk level distribution sequence, including real-time display of information such as congested sections and accident locations, adopt a reinforcement learning algorithm to optimize the display content of the variable message sign. The core of the reinforcement learning algorithm is to continuously interact with the environment to learn the optimal behavior strategy. In this scenario, the algorithm will dynamically adjust the content displayed on the variable message sign according to the current risk level distribution and traffic conditions. For example, when a high-risk congestion occurs on a certain section, the variable message sign will promptly display detour suggestions, estimated congestion duration, etc., guiding drivers to reasonably plan their travel routes, improve traffic efficiency, and relieve the congestion situation.

[0085] Apply a traffic signal optimization algorithm to dynamically adjust the signal timing plan of the traffic signal controller according to the risk level and traffic flow conditions. The traffic signal optimization algorithm will comprehensively consider the risk levels of different sections and real-time traffic flow data to flexibly adjust the signal timing. On sections with high risk and large traffic flow, appropriately extend the green light time to increase the vehicle throughput; while on sections with low risk and small traffic flow, shorten the green light time to improve the overall traffic signal allocation efficiency. By dynamically adjusting the signal timing plan in this way, it can better adapt to the changes in traffic conditions, reduce vehicle waiting time, and lower the risks of traffic congestion and accidents.

[0086] Through the synergistic effect of these three steps, an intelligent and dynamic traffic risk management system can be constructed, effectively improving traffic safety and traffic efficiency in complex traffic scenarios.

[0087] It should be noted that the above order of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above specific embodiments of this specification have been described. In addition, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0088] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0089] This specification and the drawings are only exemplary descriptions of the present application and are considered to have covered any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these changes and modifications.

Claims

1. A dynamic risk prediction method for complex traffic scenarios driven by perception intelligence, characterized in that: The method comprises: Obtain multi-source traffic dynamic data of the target area in real time; Performing spatiotemporal alignment processing on the multi-source traffic dynamic data, extracting a multi-dimensional risk feature vector including traffic flow mutation characteristics, road topology constraint characteristics and environmental interference factors to construct a spatiotemporal coupling feature matrix; Inputting the spatiotemporal coupling feature matrix into a pre-trained deep spatiotemporal prediction model to generate a risk heat map; Based on the edge computing node cluster, the risk heat map is analyzed in real time, the dynamic risk index of each traffic sub-area is calculated, and a risk level distribution sequence is generated; An adaptive early warning response mechanism is triggered according to the risk level distribution sequence to dynamically adjust the operating parameters of the variable information signs and traffic signal controllers in the target area.

2. The method for dynamic risk prediction of complex traffic scenarios driven by perception intelligence as claimed in claim 1, characterized in that: Construct the spatiotemporal coupling feature matrix, including: Performing spatiotemporal alignment processing on the multi-source traffic dynamic data through a multimodal feature alignment network to generate a standardized feature vector set; Based on the spatiotemporal correlation distribution of the standardized feature vector set, dynamically adjusting the weight distribution coefficients of the traffic flow mutation characteristics, road topology constraint characteristics and environmental interference factors to generate a fused feature vector; Using a redundant check mechanism to perform confidence assessment on the fused feature vector, eliminating abnormal features with confidence levels lower than a predetermined threshold, and generating the multi-dimensional risk feature vector; According to the spatiotemporal correlation of the multidimensional risk feature vector, a spatiotemporal coupling feature matrix including a mapping relationship between timestamp index and spatial coordinates is reconstructed.

3. The method for dynamic risk prediction of complex traffic scenarios driven by perception intelligence as claimed in claim 2, characterized in that: The multi-source traffic dynamic data is subjected to spatiotemporal alignment processing by a multimodal feature alignment network to generate a standardized feature vector set, including: Using a convolutional neural network to extract features from the video image data in the multi-source traffic dynamic data to obtain an image feature sub-vector; Using wavelet transform decomposition and reconstruction algorithm for traffic flow monitoring data in the multi-source traffic dynamic data, the traffic flow variation characteristics in different frequency bands are extracted to generate traffic flow characteristic sub-vectors; Normalizing the environmental data in the multi-source traffic dynamic data to form an environmental feature sub-vector; The image feature sub-vector, traffic feature sub-vector and environment feature sub-vector are input into a multimodal feature alignment network, and weighted fusion is performed according to the importance of each sub-vector in different spatiotemporal dimensions through the attention mechanism in the network to generate a standardized feature vector set.

4. The method for dynamic risk prediction of complex traffic scenarios driven by perception intelligence as claimed in claim 2, characterized in that: Based on the spatiotemporal correlation distribution of the standardized feature vector set, the weight distribution coefficients of the traffic flow mutation characteristics, road topology constraint characteristics and environmental interference factors are dynamically adjusted to generate a fusion feature vector, including: Calculating the time series correlation and spatial distance correlation between the vectors in the standardized feature vector set to construct a spatiotemporal correlation matrix; Based on the spatiotemporal correlation matrix, a hierarchical clustering algorithm is used to perform cluster analysis on the standardized feature vectors to determine the degree of correlation between the feature vectors of different categories and the traffic flow mutation characteristics, road topology constraint characteristics and environmental interference factors; According to the degree of association, an adaptive weight adjustment algorithm is used to dynamically calculate the weight distribution coefficients of the traffic flow mutation characteristics, road topology constraint characteristics and environmental interference factors; According to the weight distribution coefficient, the traffic flow mutation characteristics, road topology constraint characteristics and environmental interference factors are weighted and summed to generate a fusion feature vector.

5. The method for dynamic risk prediction of complex traffic scenarios driven by perception intelligence as claimed in claim 2, characterized in that: The confidence of the fused feature vector is evaluated by using a redundant check mechanism, abnormal features with a confidence lower than a predetermined threshold are eliminated, and the multi-dimensional risk feature vector is generated, including: Performing multiple random sampling on the fused feature vector to obtain multiple sampling subsets; For the multiple sampling subsets, a principal component analysis algorithm is used to perform dimensionality reduction processing to extract multiple main characteristic components; Calculating the similarities between the plurality of main feature components and constructing a similarity matrix; According to the similarity matrix, using a threshold-based anomaly detection algorithm to evaluate the confidence of each feature in the fused feature vector; Abnormal features with confidence levels lower than a predetermined threshold are removed from the fused feature vector to generate the multi-dimensional risk feature vector.

6. The method for dynamic risk prediction of complex traffic scenarios driven by perception intelligence as claimed in claim 1, characterized in that: The training process of the pre-trained deep spatiotemporal prediction model includes: Collect historical traffic data and construct a training data set, wherein the training data set includes sample multi-source traffic dynamic data and sample risk heat map labels; Performing spatiotemporal alignment processing, feature extraction and feature matrix construction operations on the sample multi-source traffic dynamic data to obtain a sample spatiotemporal coupling feature matrix; The sample spatiotemporal coupling feature matrix is ​​input into the initial deep spatiotemporal prediction model, and an initial risk heat map is output. The model parameters are iteratively optimized by calculating the mean square error loss function between the initial risk heat map and the sample risk heat map labels to complete the pre-training of the deep spatiotemporal prediction model.

7. The method for dynamic risk prediction of complex traffic scenarios driven by perception intelligence as claimed in claim 1, characterized in that: Based on the edge computing node cluster, the risk heat map is analyzed in real time, the dynamic risk index of each traffic sub-area is calculated, and a risk level distribution sequence is generated, including: Dividing the risk heat map into a plurality of sub-area heat maps according to a preset traffic sub-area division rule; Each edge computing node is responsible for parsing the heat map of one or more sub-regions, and uses the local computing resources of the node to perform statistical analysis on the risk data in the heat map of the sub-region it is responsible for; Calculate the dynamic risk index of each sub-area based on the preset risk assessment model; The dynamic risk indexes of each traffic sub-area calculated by each edge computing node are sorted in chronological order to generate the risk level distribution sequence.

8. The method for dynamic risk prediction of complex traffic scenarios driven by perception intelligence as claimed in claim 1, characterized in that: According to the risk level distribution sequence, an adaptive early warning response mechanism is triggered to dynamically adjust the operating parameters of variable information signs and traffic signal controllers in the target area, including: Set warning thresholds and warning methods corresponding to different risk levels. When the risk level in the risk level distribution sequence reaches or exceeds the warning threshold, the corresponding warning information is triggered for push; According to the changing trend of the risk level distribution sequence, a reinforcement learning algorithm is used to optimize the display content of the variable information sign; The traffic signal optimization algorithm is used to dynamically adjust the signal light timing plan of the traffic signal controller according to the risk level and traffic flow conditions.

Images