A traffic supervision system applied to intelligent street lamps and an intelligent supervision method thereof

Abstract

The application discloses a traffic supervision system applied to intelligent street lamps and an intelligent supervision method thereof, relates to the technical field of intelligent traffic, and solves the problems of the lack of physical-digital mapping relationship, the low efficiency of edge computing resource allocation and the high delay of cloud computing of the existing intelligent street lamp system; the scheme is based on multi-sensor data fusion, adopts atomic clocks and GNSS to unify the space-time reference, and constructs a dynamic causal graph through a graph neural network to optimize abnormal event detection; an improved Jaccard space-time similarity algorithm is used to optimize the calculation task allocation, and an edge computing cluster is constructed based on 5G-V2X to identify high-risk areas and predict traffic flow; a multi-modal fusion reinforcement learning algorithm is used to adaptively select LiFi or 5G-UWB communication media to realize efficient early warning information synchronization; the application significantly improves the multi-source data fusion value and early warning accuracy, optimizes the utilization rate of computing power resources, and enhances the real-time performance of instructions and the self-adaptive ability of the system in complex environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation, and more particularly to a traffic supervision system applied to intelligent street lamps and an intelligent supervision method thereof. BACKGROUND

[0002] With the rapid development of smart city and Internet of Things technology, urban traffic supervision is gradually transforming towards digitization and intelligentization. As a core node of urban infrastructure, intelligent street lamps, with their wide distribution, stable power supply, and uniform spatial coverage, have become a multifunctional carrier integrating traffic perception, environmental monitoring, and communication interaction. Traditional traffic management relies on manual patrols and isolated information systems, making it difficult to respond to complex road conditions in real time. However, the traffic supervision method based on intelligent street lamps, by integrating sensor networks, edge computing, AI analysis, and other technologies, can achieve accurate perception and dynamic warning of traffic flow, accident risk, and environmental anomalies, providing full-time, multi-dimensional decision support for urban traffic governance, and becoming an important direction for the evolution of intelligent transportation systems.

[0003] Currently, intelligent street lamp traffic supervision technology mainly realizes road state monitoring through multi-sensor integration and edge computing. For example, patent CN116092287A deploys network cameras, air quality monitors, weather sensors, and 5G base stations in street lamps, connects to the municipal platform and police network through "two networks," and realizes data distribution. Patent CN118328359A further proposes an abnormal warning component that uses cameras to collect road images, identifies abnormal road conditions using AI algorithms, and triggers warnings through dynamic aperture lamps and lighting brightness adjustments. At the data processing level, patent CN118506582A discloses a dangerous judgment model based on cloud and edge collaboration, which collects local road information from street lamps and generates a global risk map by fusing multi-node data in the cloud. In traffic event detection, patent CN118334866A uses a deep learning model to detect and track targets in surveillance videos, and predicts traffic changes based on historical data. In addition, patent CN119068682A optimizes traffic signal control strategies through accident index analysis and road condition correction coefficient calculation. Existing technologies generally follow the technical path of "multi-sensor collection → edge / cloud processing → rule-based warning," with core features including heterogeneous device integration, AI model-driven decision-making, and multi-platform data interconnection.

[0004] However, although intelligent street lamps have made some progress in traffic regulation, there are still some obvious defects in the existing technology: first, although the existing system piles up various sensors (such as CN116092287A integrates camera, meteorological monitoring equipment), but it lacks the ability to construct the physical-digital mapping relationship, resulting in the loss of data value. For example, when the camera detects the vehicle emergency braking behavior, the meteorological sensor data is not associated with the road friction coefficient model, and it is not possible to determine whether it is caused by wet and slippery, and only rely on single mode data to trigger warning (such as CN118506582A), the misjudgment rate is high. Second, the existing scheme emphasizes edge data processing (such as CN118506582A's street lamp end danger judgment), but does not optimize the calculation task allocation according to the spatio-temporal correlation of events, resulting in waste of resources. For example, adjacent street lamps repeatedly calculate the trajectory of the same moving target (such as CN118334866A), instead of sharing intermediate results through distributed collaboration, resulting in redundant computing power. At the same time, cloud computing tasks still rely on centralized processing mode, there is delay in data transmission and decision execution, also reduces the overall system efficiency. Therefore, a new type of intelligent street lamp traffic regulation system and its intelligent regulation method are needed to solve the above problems. SUMMARY

[0005] In view of the defects of the prior art, the present application discloses a traffic regulation system applied to intelligent street lamps and an intelligent regulation method thereof, which aims to solve the problems of weak multi-modal data correlation, redundant edge computing resources and serious cloud decision delay in the prior art.

[0006] In order to achieve the above technical effects, the present application adopts the following technical solutions:

[0007] A traffic intelligent regulation method applied to intelligent street lamps, comprising:

[0008] Step 1, based on the original data of multiple sensors, the space-time reference is unified and the space mapping is associated through atomic clock and GNSS positioning module, and the multi-modal fusion data set is generated;

[0009] Step 2, calling the road physical model to constrain the multi-modal fusion data set, and using graph neural network to construct dynamic causal graph and execute back propagation reasoning, outputting warning event with causal chain evidence;

[0010] Step 3, according to the warning event, using improved Jaccard spatio-temporal similarity algorithm to detect the spatio-temporal correlation of adjacent street lamp event cluster, if the overlap degree is higher than or equal to the preset threshold, through 5G-V2X networking to construct edge computing cluster, calculate the boundary parameters of high risk area and the evolution trend of traffic flow; if the overlap degree is lower than the preset threshold, upload the compressed feature vector of the node to the cloud to execute incremental model training, output the optimized global traffic situation atlas and model iteration instruction;

[0011] Step 4, extracting structured elements from the early warning event, including event location, risk level and recommended action, encapsulating binary instruction stream according to ISO 20078 standard and attaching digital signature to generate early warning instruction;

[0012] Step 5, based on the environmental state, dynamically selecting LiFi or 5G-UWB communication medium through multi-modal fusion reinforcement learning algorithm, synchronously conveying the early warning instruction to the vehicle terminal, AR-HUD and traffic control platform.

[0013] As a further technical solution of the application, the working method for improving the Jaccard spatiotemporal similarity algorithm is as follows: first, the event cluster set C of adjacent street lamp nodes is extracted through a sliding time window i =(e1,e2,...e n ),wherein each event e k contains a timestamp t k , spatial coordinates (x k , y k ) and event type code T k ; then, a spatiotemporal similarity weight function w(t, d) is defined wherein t diff is the event time difference, d is the spatial Euclidean distance, wherein d = 0.6; β = 0.1 and γ = 0.05 are respectively the time decay coefficient and the spatial decay coefficient; a is the balance factor of the time and spatial weights; and the Jaccard coefficient is improved based on the spatiotemporal similarity weight function w(t, d) as follows:

[0014]

[0015] In formula (1), δ(T p ,T q ) is an event type matching function; ∑e p ∈C i, e q ∈C j w(t diff ,d)δ(T p ,T q ) represents the sum of the weighted similarities of all event pairs in the two event clusters, only considering events of the same type; ∑e p ∈C i w p -∑e p ∈C i ,e q ∈C j w(t diff ,d)δ(T p ,T q ) is used to avoid bias caused by the difference in the number of events; and if J st≥θ, triggering the 5G-V2X PC5 interface to establish an edge computing cluster, and distributing the event cluster to the master computing node through a distributed consistent hashing algorithm; wherein θ is a dynamic threshold, and the calculation formula is:

[0016] θ=(0.7+0.1log(1+Q) (2)

[0017] In formula (2), Q is the current regional event density, then the master node calls a trajectory prediction algorithm to model the vehicle motion trend in the abnormal event influence area, and combines a dynamic traffic state regression analysis to calculate the secondary accident probability, the risk area boundary parameter and the traffic flow evolution gradient field; if J st <θ, PCA dimensionality reduction is performed on each node event cluster to generate a compressed feature vector, which is uploaded to the cloud federated learning framework through the MQTT protocol, the global traffic situation atlas model parameters are updated based on the incremental Adam optimizer, and the model iteration instruction and risk probability distribution matrix are output.

[0018] As a further technical solution of the application, the calculation principle of the edge computing cluster for calculating the high-risk area boundary parameter and the traffic flow evolution trend is as follows: first, a low-latency communication link is established through the 5G-V2X PC5 interface, the master node collects the spatio-temporal coordinate data of adjacent street lamp event clusters, inputs the data into an LSTM-Kalman filter fusion model to predict the vehicle trajectory distribution, and identifies the potential collision point set based on a trajectory conflict detection algorithm; then, the collision point set is topologically reconstructed through an alpha-shape algorithm to generate a non-convex polygon risk area boundary parameter; a spatio-temporal convolution network is constructed to extract the spatio-temporal correlation mode of historical traffic flow tensor and real-time event features through a three-dimensional hollow convolution layer, and output the vector field distribution of the traffic flow evolution gradient field; a generalized additive model is combined to fit a nonlinear regression equation based on the vehicle density, average speed and event intensity features to calculate the secondary accident probability weight; finally, an exponential decay function of a risk diffusion model is used to map the accident probability to a gradient field correction factor to drive the dynamic adjustment of the warning area radius; and the consistent hashing protocol in the edge cluster is used to synchronize the update of the warning parameters of each node to realize the distributed collaborative calculation of the regional risk boundary and the traffic flow trend.

[0019] As a further technical solution of the application, the working principle of the multi-modal fusion reinforcement learning algorithm is as follows: according to the light intensity L lux and the UWB channel state information collected by the environment perception module in real time, a dynamic selection strategy is constructed, if L lux ≥2000 and the terminal relative position is within the LiFi line-of-sight coverage cone angle, the LiFi module is triggered to send instructions to the vehicle terminal and AR-HUD in an orthogonal frequency division multiplexing modulation mode; if the multipath fading index F MPIf the terminal moving speed v is greater than or equal to 80 km / h, the system switches to 5G-UWB, calls the 3GPP NR-UMAC layer protocol, allocates time slot resources based on a dynamic TDD frame structure, enhances transmission reliability through a Polar code, and realizes beamforming by combining an SRS channel sounding method to multicast to the AR-HUD and the traffic management platform at a rate of 10 Gbps; the multi-modal fusion reinforcement learning algorithm maintains the LiFi and UWB link hot standby state through the dual stack protocol stack IPv6 and TSN, guarantees the integrity of the instruction transmission based on the forward error correction and the hybrid automatic repeat request mechanism, and dynamically selects the channel with the lowest packet loss rate to perform the main data stream transmission.

[0020] As a further technical solution of the application, a traffic supervision system applied to intelligent street lamps comprises a space-time synchronization module, a cross-modal causal reasoning module, a distributed collaborative computing optimization module, a structured early warning coding module and a multi-modal interaction module.

[0021] The space-time synchronization module is used for space-time reference unification and fusion preprocessing of the multi-sensor data of the street lamp end.

[0022] The cross-modal causal reasoning module is used for causal reasoning of the multi-modal data in combination with a road physical model and a graph neural network, and outputs an interpretable early warning event.

[0023] The distributed collaborative computing optimization module is used for dynamic allocation of the computing tasks of the intelligent street lamp edge node and the cloud based on the event space-time correlation.

[0024] The structured early warning coding module is used for coding the early warning event into a structured instruction supporting machine analysis.

[0025] The multi-modal interaction module is used for transmitting the structured early warning instruction through an adaptive channel.

[0026] Based on the above technical solution, the application has the following positive and beneficial effects:

[0027] 1. By constructing a physical-digital mapping relationship, the sensor data is dynamically coupled with the road physical model (such as the friction coefficient model), solving the misjudgment problem caused by the isolation of multi-source data in the prior art. The graph neural network is used for backward propagation reasoning to generate causal chain evidence, so that the early warning event has physical interpretability, significantly reduces the false positive rate caused by the one-sidedness of single-modal data, and enhances the trust of the supervision department to the system decision.

[0028] 2. Based on the improved Jaccard spatiotemporal similarity algorithm, the spatiotemporal correlation of the event cluster is accurately identified, the edge computing cluster and the cloud incremental learning task are dynamically divided, and the redundant calculation of adjacent street lamps on the same target is avoided. Through distributed cooperation and task chain splitting, the on-demand allocation of computing resources is realized, the invalid energy consumption is reduced, the data transmission delay of centralized processing in the cloud is shortened, and the overall response real-time performance of the system is improved.

[0029] 3. Structured element extraction and standardized coding (ISO 20078) are adopted, combined with digital signature technology, to ensure the semantic consistency and transmission security of the early warning instructions. Through a multi-modal communication dynamic selection mechanism, LiFi / 5G-UWB communication media are adaptively switched according to the environmental state, solving the signal attenuation or interference problem of traditional single communication link in complex scenarios, and ensuring the synchronous and reliable communication of instructions in vehicle terminals, AR-HUD and management platforms.

[0030] 4. Through real-time calculation of the edge end, the boundary of the high-risk area and the evolution trend of the traffic flow are calculated, combined with the global traffic situation atlas generated by the cloud incremental training, forming a double-layer decision system of "local rapid response-global continuous optimization". The combination of dynamic causal reasoning and physical model constraints enables the system to quickly handle sudden risks and iteratively improve long-term prediction accuracy based on historical data, breaking through the limitations of the existing technology of local response and global optimization. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings, wherein:

[0032] Figure 1 The figure is a structural diagram of a traffic intelligent monitoring method applied to intelligent street lamps of the present application;

[0033] Figure 2 The figure is a working principle diagram of step 1 of the present application;

[0034] Figure 3 The figure is a working principle diagram of step 2 of the present application;

[0035] Figure 4 The figure is a working principle framework diagram of the edge computing cluster of the present application;

[0036] Figure 5 The figure is a structural framework diagram of the adaptive risk assessment model of the present application;

[0037] Figure 6This is a schematic diagram of the traffic monitoring system of the present invention applied to smart streetlights. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] In the embodiments, the traffic monitoring method applied to smart streetlights, such as... Figure 1 As shown, it includes:

[0040] Step 1: Based on the raw data from multiple sensors, a multimodal fusion dataset is generated by unifying the spatiotemporal reference and spatial mapping association between the atomic clock and the GNSS positioning module; for example... Figure 2 As shown, the method is as follows: Based on the real-time GNSS positioning coordinates and laser rangefinder measurements, a sensor spatial mapping matrix is ​​constructed using the Lie group spatial transformation algorithm to align the radar point cloud coordinate system with the camera imaging plane; for asynchronously arriving sensor data streams, a cubic spline interpolation algorithm is used to resample under a spatiotemporal reference to eliminate time offset, and a dynamic time warping algorithm is used to detect the time axis distortion of the multimodal data stream. If the maximum path deviation exceeds a preset threshold, a Kalman filter compensation mechanism is used to reconstruct the temporal consistency; finally, a multimodal fusion dataset is output, which includes timestamps, spatial coordinates, and sensor feature vectors normalized to physical dimensions.

[0041] Step 2: Invoke the road physical model to constrain the multimodal fusion dataset, and use a graph neural network to construct a dynamic causal graph and perform backpropagation inference to output early warning events with causal chain evidence; the specific working principle is as follows: Figure 3The road physical model adopts a Lagrange-Euler mixed coordinate transformation method to analyze vehicle dynamics parameters, and combines a Poisson-Paschen coupled fluid model to calculate road environment constraints, so that the motion state and the road topology form a dynamic constraint relationship; in the constraint solving process, the road physical model uses a sparse tensor decomposition method to extract key influence factors in the physical model, and selects abnormal driving behavior influence parameters through a normalized mutual information calculation method, to generate a road state causal constraint matrix. The graph neural network calculates the spatio-temporal correlation weight of the road state, vehicle behavior and environmental factors based on the road state causal constraint matrix, and combines a Bayesian structure learning algorithm to perform causal relationship reasoning, to generate a dynamic causal graph; in the causal reasoning process, the graph neural network calls a back propagation gradient constraint optimization to correct the causal relationship weight, and combines an optimal transport mapping method to evaluate the reliability of the causal chain evidence, to exclude low confidence causal paths.

[0042] Step 3, according to the early warning event, an improved Jaccard spatio-temporal similarity algorithm is used to detect the spatio-temporal correlation of adjacent street lamp event clusters, if the overlap degree is higher than or equal to the preset threshold, an edge computing cluster is constructed through 5G-V2X networking to calculate the high-risk area boundary parameter and the traffic flow evolution trend; if the overlap degree is lower than the preset threshold, the compressed feature vector of the node is uploaded to the cloud to perform incremental model training, and an optimized global traffic situation map and model iteration instruction are output; the working method of the improved Jaccard spatio-temporal similarity algorithm is as follows: first, the event cluster set C i =(e1,e2,...e n ) of adjacent street lamp nodes is extracted through a sliding time window, wherein each event e k contains a timestamp t k , spatial coordinates (x k , y k ) and event type code T k ; then, the spatio-temporal similarity weight function is defined, wherein t diff is the event time difference, d is the spatial Euclidean distance, wherein d=0.6; β=0.1, γ=0.05 are respectively the time attenuation coefficient and the space attenuation coefficient; a is the balance factor of time and space weight; based on the spatio-temporal similarity weight function w(t, d), the Jaccard coefficient is improved as:

[0043]

[0044] In formula (1), δ(T p , T q ) is an event type matching function; ∑e p ∈C i , e q ∈C jw(t diff ,d)δ(T p ,T q ) represents the sum of the weighted similarity of all event pairs in the two event clusters, only considering events of the same type;∑e p ∈C i w p -∑e p ∈C i ,e q ∈C j w(t diff ,d)δ(T p ,T q ) is used to avoid the deviation caused by the difference in the number of events; if J st ≥θ, the PC5 interface of 5G-V2X is triggered to establish an edge computing cluster, and the event cluster is distributed to the master computing node through a distributed consistent hashing algorithm; wherein θ is a dynamic threshold, and the calculation formula is:

[0045] θ=(0.7+0.1log(1+Q) (2)

[0046] In formula (2), Q is the current regional event density, then the master node calls the trajectory prediction algorithm to model the vehicle motion trend in the abnormal event influence area, and combines the dynamic traffic state regression analysis to calculate the secondary accident occurrence probability, calculates the risk area boundary parameter and traffic flow evolution gradient field; if J st <θ, PCA dimension reduction is performed on each node event cluster to generate a compressed feature vector, which is uploaded to the cloud federated learning framework through the MQTT protocol, the global traffic situation atlas model parameters are updated based on the incremental Adam optimizer, and the model iteration instruction and risk probability distribution matrix are output. For example Figure 4As shown, the edge computing cluster calculates the calculation principle of high-risk area boundary parameters and traffic flow evolution trend: first, the low-latency communication link is established through the PC5 interface of 5G-V2X, the master node collects the spatio-temporal coordinate data of adjacent streetlight event clusters, inputs into the LSTM-Kalman filter fusion model to predict the vehicle trajectory distribution, and identifies the potential collision point set based on the trajectory conflict detection algorithm; then, the topological reconstruction is performed on the collision point set through the alpha-shape algorithm to generate the non-convex polygon risk area boundary parameters; at the same time, the spatio-temporal convolution network is constructed, the spatio-temporal correlation mode of historical traffic flow tensor and real-time event features is extracted through the three-dimensional hollow convolution layer, and the vector field distribution of traffic flow evolution gradient field is output; combined with the generalized additive model, based on the vehicle density, average speed and event intensity features, the nonlinear regression equation is fitted to calculate the secondary accident probability weight; finally, through the exponential decay function of the risk diffusion model, the accident probability is mapped to the gradient field correction factor to drive the dynamic adjustment of the warning area radius; and through the consistent hashing protocol in the edge cluster, the warning parameters of each node are updated synchronously to realize the distributed collaborative calculation of regional risk boundary and traffic flow trend.

[0047] Step 4, structured elements are extracted from the warning event, including event location, risk level and recommended action, binary instruction stream is packaged according to ISO 20078 standard and digital signature is attached to generate warning instruction; specifically, adaptive risk assessment model is used to perform risk classification calculation, and historical event database is used to dynamically adjust risk classification threshold, fuzzy decision method is used to calculate optimal disposal strategy; for example Figure 5As shown, the adaptive risk assessment model includes a data layer, a dynamic risk modeling layer, an adaptive threshold optimization layer, a multi-level risk classification layer, and a decision feedback adjustment layer; the data layer is used to receive multi-modal fusion dataset data, extract time dependence using long short-term memory network time series encoding, and combine spatial histogram projection to construct spatial distribution mapping of road events, generating a spatio-temporal feature vector; the dynamic risk modeling layer is used to call the Gaussian process regression method to perform continuous risk estimation on the spatio-temporal feature vector, and use the conditional variational autoencoder to derive the event influence range to construct a dynamic risk function; the adaptive threshold optimization layer is used to calculate the historical event risk level distribution using the kernel density estimation method, and call the quantile regression to dynamically adjust the risk classification threshold, generating real-time updated risk boundary parameters; the multi-level risk classification layer is used to classify the early warning events using hierarchical Bayesian clustering, and combine Markov random field to optimize the classification label, generating a risk classification matrix; the decision feedback adjustment layer optimizes the classification model by incremental Bayesian update combined with the latest accident data, and uses the exponentially weighted moving average to smooth the risk level change, adjusting the final risk classification decision. The fuzzy decision method performs multi-objective optimization on the event location, risk level and road network topology based on the fuzzy decision tree, generating a set of disposal strategies, including speed limit value, detour path and emergency lane activation; according to the ISO 20078 standard, the longitude, latitude, risk level and strategy are packaged into a TLV format binary stream, and the ECDSA algorithm is used to add a digital signature, generating a tamper-resistant early warning instruction.

[0048] Step 5, based on the environmental state, dynamically select LiFi or 5G-UWB communication medium through multi-modal fusion reinforcement learning algorithm, synchronize the early warning instruction to the vehicle terminal, AR-HUD and traffic control platform. The working principle of the multi-modal fusion reinforcement learning algorithm is: according to the light intensity L lux and UWB channel state information collected by the environmental perception module in real time, construct a dynamic selection strategy, if L lux ≥2000 and the terminal relative position is within the LiFi line-of-sight coverage cone angle, trigger the LiFi module to send instructions to the vehicle terminal and AR-HUD in orthogonal frequency division multiplexing modulation mode; if the multipath fading index F MP≥3dB or terminal moving speed v≥80km / h, switch to 5G-UWB, call 3GPP NR-U MAC layer protocol, allocate time slot resources based on dynamic TDD frame structure, enhance transmission reliability through Polar code, and realize beamforming combined with SRS channel sounding method, multicast to AR-HUD and traffic management platform at a rate of 10Gbps; the multi-modal fusion reinforcement learning algorithm maintains the LiFi and UWB link hot standby state through the dual stack protocol stack IPv6 and TSN, and guarantees the integrity of the instruction transmission based on the forward error correction and hybrid automatic repeat request mechanism, and dynamically selects the channel with the lowest packet loss rate to perform the main data stream transmission.

[0049] In step 1 of the above embodiment, the Lie group space transformation algorithm uses the mathematical structure of the special Euclidean group SE(3) to map the polar coordinate point cloud of the radar and the pixel coordinate system of the camera to the unified GNSS world coordinate system through rigid transformation (rotation matrix and translation vector). The principle is to solve the geometric transformation parameters through feature point matching (such as corner points of the calibration board), minimize the re-projection error, and ensure millimeter-level spatial alignment accuracy. The cubic spline interpolation algorithm is used to solve the sampling rate difference of asynchronous data streams (such as weather sensors 1Hz and cameras 30Hz). In the time axis, a piecewise cubic polynomial function is constructed to generate a continuous and smooth synchronous data sequence with camera frame time as the interpolation node, and the time discreteness of low-frequency sensors is eliminated. Dynamic time warping (DTW) calculates the cumulative deviation of the minimum bending path between different data streams to detect timing misalignment phenomena (such as radar scan period jitter). When the deviation exceeds the preset threshold, Kalman filter compensation is triggered. The state space model (process model and observation model) is used to predict the target motion trajectory, and the covariance matrix is updated combined with sensor observation values to dynamically correct the time axis distortion of the data stream and reconstruct the timing consistency. The physical dimension normalization adopts Z-score standardization method to map different dimension data (such as speed m / s, pixel coordinates, humidity percentage) to zero mean unit variance space, eliminate the influence of dimension difference on the input bias of subsequent machine learning model, and improve the model generalization ability.

[0050] In implementation, the hardware level for realizing this step adopts a dual-frequency RTK-GNSS module (U-blox ZED-F9P, positioning accuracy ±2 cm) to obtain the streetlight reference coordinates, a laser rangefinder (VL53L5CX, ranging error ±1 mm) to measure the installation offset of the radar and the camera (Δx=0.5 m, Δy=0.2 m, Δz=0.3 m), and a multi-line radar (Velodyne VLP-16, horizontal resolution 0.1°) and a global shutter camera (Basler acA2440-75um, 30 fps) to respectively collect point cloud and image data. In software implementation, the radar point cloud and image feature points (checkerboard corner error <0.1 pixel) are extracted through a calibration board, the Lie group transformation parameters (rotation matrix R and translation vector t) are solved by using the Levenberg-Marquardt algorithm, and the re-projection error is optimized to a sub-pixel level; the 1 Hz data stream of the weather sensor is generated into a 30 Hz synchronous sequence by using a cubic spline interpolation, the dynamic time warping window is set to 10 frames (about 333 ms), the Kalman filter compensation is triggered when the time offset of the radar-camera data stream is ≥3 frames (100 ms), the process noise covariance Q is configured as diag(0.1, 0.1, 0.01) and the observation noise covariance R is configured as diag(0.5, 0.5, 0.1), and the output fusion data set is stored in a Parquet columnar format (containing WGS84 coordinates, speed, pixel coordinates and normalized environment parameters) for calling by downstream modules.

[0051] In the road physical model of step 2 of the above embodiment, the vehicle dynamics generally adopts a Lagrangian description to track the motion trajectory of a single vehicle, and the road environment modeling is based on an Eulerian description, i.e., describing the fluid state change on a fixed spatial grid. In order to unify the two, the present application adopts a Lagrangian-Eulerian coordinate transformation method to form a dynamic constraint relationship between the motion state and the road topology.

[0052] The motion equation of the vehicle is expressed in the Lagrangian coordinate system as:

[0053]

[0054] where (x, y) is the vehicle position, θ is the heading angle, v is the speed, a is the acceleration, δ is the steering angle, and L is the wheelbase. In the Eulerian coordinate system, the state change of the road is described by the fluid conservation equation:

[0055]

[0056] where p is the local traffic density and v is the fluid velocity. By constructing a state variable transformation matrix, the speed, acceleration, and direction change of the vehicle are mapped in the coordinate, and the state differential is performed in different coordinate systems, so that the system can track the individual behavior locally while also calculating the evolution trend of the overall traffic flow. In addition, based on the nonlinear differential equation, the motion state of the vehicle under the conditions of road slope, curve curvature, and different friction coefficients is analyzed, and a complete set of dynamic constraint equations is established, and the motion prediction model is corrected combined with the vehicle characteristics (such as mass, tire parameters, etc.), so that the calculation result is more consistent with the real traffic environment. This method enables the individual motion trajectory of the vehicle to be optimized and modeled in the global flow field, and the feasible vehicle motion range is calculated under the constraint of road topology.

[0057] In the modeling of road environment constraints, a Poisson-Poisson coupled fluid model is used. This model introduces the Poisson distribution characteristics of road traffic based on traditional fluid dynamics, and models the change of vehicle flow in the road grid through double Poisson distribution, so that the traffic fluid model can adapt to changing road conditions. Specifically, the first layer of Poisson equation is used to describe the change of traffic density of random traffic flow, and the second layer of Poisson equation is used to calculate the flow velocity gradient of local fluid to dynamically evaluate congestion, flow velocity change and driving safety distance constraints. In the solving process, the finite volume method is used to discretize the road grid, so that the flow conservation constraint can maintain calculation stability under any complex road topology, and the adaptive grid refinement method is used to locally improve the calculation accuracy in the congestion area to optimize the calculation efficiency and accuracy of the fluid solution.

[0058] In the constraint solving process, in order to extract the key influencing factors of road environment on vehicle motion, a sparse tensor decomposition method is used for dimensionality reduction analysis of the physical model. Road environment variables include multiple high-dimensional characteristics such as slope, friction coefficient, lane width, and traffic density, and direct modeling will lead to high computational complexity, so the original high-dimensional data is decomposed into a low-rank subspace by tensor decomposition technology to extract the main factors affecting vehicle motion. Then, the normalized mutual information calculation method is called to evaluate the information correlation between variables, so as to screen out key parameters that have a greater impact on abnormal driving behavior, and further reduce the dimension through principal component analysis to generate a road state causal constraint matrix. This matrix represents the influence of road environmental factors on traffic state, and is used to construct a causal relationship graph in the graph neural network to provide explainable warning event analysis.

[0059] In implementation, the road physics model runs on the NVIDI A Jetson AGX Xavier module (512 CUDA cores) of the edge computing cluster, with GPU-accelerated tensor decomposition (cuTENSOR library) and fluid model solving (OpenFOAM coupled solver). In the software implementation of the road physics model, the vehicle dynamics parameters (mass m = 1500 kg, rolling friction coefficient μ = 0.015) of the Lagrangian-Eulerian hybrid model are fused with sensor data (millimeter wave radar speed measurement error ± 0.1 m / s) through Kalman filtering, and the fluid viscosity coefficient of the Poisson-Poisson model is set to v = 0.2 m 2 / s to match the traffic flow characteristics of urban roads. Sparse tensor decomposition uses the alternating least squares (ALS) method for 10 iterations, with a decomposition rank of 8 and the top 5% significant factors (such as friction coefficient weight accounting for 32%) retained. The normalized mutual information calculation is based on a historical accident data set (100,000 samples), and the joint probability distribution of sudden braking events (deceleration ≥ 3.5 m / s 2 ) and wet road surfaces is calculated using a sliding window, and parameters with NMI ≥ 0.3 (such as rainfall intensity-braking distance) are selected to construct a 64x64 causal constraint matrix. The GNN of the dynamic causal graph uses the GraphSAGE architecture, with a node embedding dimension of 256, a message passing layer L = 3, and a loss function with physical constraints (mean square error + mutual information regularization term) used during training, with a learning rate of 1e-4. After 50 iterations, the causal chain evidence is output (such as "rainfall 0.5 mm / min → friction coefficient decrease 40% → rear-end collision risk increase 2.6 times").

[0060] In the graph neural network of step 2 of the above embodiment, the heterogeneous graph attention mechanism designs a multi-head relational attention weight calculation model for three types of heterogeneous nodes (road topology attributes, vehicle motion parameters, and weather data), extracts spatio-temporal correlation features through node type-sensitive query-key value projection matrices, and the mathematical expression is where e i ,e j are node embedding vectors, is the projection matrix of the h-th attention head, d is the embedding dimension, and the spatio-temporal correlation weight matrix is generated by aggregating the interaction weights of different relationship types (such as "vehicle-road friction constraint" and "environment-vehicle visibility influence") through multi-head attention; the Bayesian structure learning algorithm infers the latent causal graph structure from the correlation weight matrix based on the Markov Chain Monte Carlo sampling method, optimizes the graph topology through the Bayesian posterior probability distribution P(G|D) ∝ P(D|G)P(G), where P(D|G) is the data likelihood and P(G) is the prior probability (such as the sparsity constraint), dynamically adjusts the edge connection direction and strength, and captures non-explicit causal relationships (such as the lagging effect of humidity mutation on vehicle skidding); the backpropagation gradient constraint optimizes the update direction of the attention weight matrix during the causal graph reasoning process by calculating the gradient backpropagation through the differentiable loss function (such as the causal effect mean square error), filters the high-contribution causal paths, and is mathematically expressed as where W is the attention weight matrix, γ is the sparsification coefficient, and low-correlation edges are removed through L1 regularization; the optimal transport mapping method evaluates the reliability of causal chain evidence based on the Wasserstein distance, optimally transports the probability distribution of the causal path to the historical benchmark distribution, calculates the transport cost W(P,Q) = inf γ∈(P,Q) ∫x(c,y)dγ(x,y), filters out low-confidence causal chains with transport costs exceeding a threshold (such as KL divergence difference ≥ 1.5), and retains high-reliability evidence for early warning decision-making.

[0061] In implementation, the graph neural network is deployed on the NVIDIA Jetson AGX Xavier platform of the edge computing node, and CUDA is used to accelerate attention calculation (cuBLAS library) and Bayesian inference (Pyro framework). In software implementation, the heterogeneous graph node feature dimension is set to 64, the vehicle node input includes speed (m / s), acceleration (m / s 2), yaw angle (rad), road node input friction coefficient (m), curvature radius (m), environmental node input rainfall intensity (mm / h) and wind speed (m / s). The dimensions of the query matrix Q, key matrix K, and value matrix V in the attention mechanism are 64x64, the time sliding window covers 500ms (5 frames @ 10Hz), and the spatial adjacency matrix is generated based on the road network topology (maximum hop count k = 3). Bayesian structure learning uses the Metropolis-Hastings algorithm to sample 1000 times, with an acceptance rate a = 0.3, and filters causal edges with BF > 3 (e.g. “friction coefficient decrease -> braking distance increase”). During backpropagation, the gradient mask retains the top 20% of edges in the weight ranking, the L1 regularization coefficient is l = 0.01, and the Jacobian matrix is calculated by automatic differentiation (PyTorch Autograd) to compute the partial derivative of the vehicle motion equation with respect to the input parameters. The optimal transport mapping uses the Sinkhorn algorithm for 50 iterations, with an entropy regularization coefficient e = 0.1, and dynamically adjusts the Wasserstein distance threshold for causal chains (W = 0.3). The final output is a causal chain evidence (e.g. “rainfall 0.8mm / h -> friction coefficient m = 0.3 -> braking distance extended by 3m -> rear-end collision risk level B”), formatted as a JSON-LD semantic description, for traffic management platform visualization decision-making.

[0062] In the actual implementation of step 2, a multi-line laser radar (Velodyne VLP-16, horizontal angle resolution 0.1°) and a global shutter industrial camera (Basler acA2440-75um, 30fps) are installed on the top of the street lamp pole to collect vehicle trajectories and road images; a dual-frequency RTK-GNSS module (U-blox ZED-F9P) and a laser range finder (VL53L5CX) are fixed to the base of the lamp pole to achieve centimeter-level positioning and sensor calibration; an edge computing node (NVIDIA Jetson AGX Xavier) is deployed in the middle case of the lamp pole to run physical models and graph neural network algorithms; a 5G communication module (Huawei MH5000-871) and a LiFi transceiver (PureLiFi GigaDock) are integrated on the side of the lamp pole to support multi-channel data transmission;

[0063] To verify the effectiveness of step 2 (dynamic causal graph reasoning) in complex traffic scenarios, a virtual environment of multi-level risk events is constructed and the accuracy of causal reasoning is quantified. The experiment is based on the SUMO-CARLA joint simulation platform to generate traffic scenarios containing road topology, vehicle dynamics and environmental parameters, and sets up three types of progressive test scenarios:

[0064] The basic scenario simulates a single risk (such as a sudden stop by the preceding vehicle or illegal lane changing), the compound scenario combines environmental and behavioral risks (such as wet road surface + speeding through a curve), and the chain scenario designs a causal transmission chain (such as "rain → reduced visibility → misjudgment of the distance of the preceding vehicle → sudden stop → rear-end collision").

[0065] The CARLA is driven by a Python script to generate 100 intelligent agent vehicles, randomly trigger preset risk events (a total of 50 times, including 20 basic, 20 compound, and 10 chain), and synchronize the simulation of cameras (OpenCV generates 1280×720 resolution images, YOLOv5 real-time detects vehicles / lanes), millimeter wave radars (outputs 0.1 m precision point cloud, superimposes Gaussian noise with σ = 0.15), and meteorological sensors (dynamically adjusts the intensity of rainfall / fog). The causal reasoning module in Step 2 accesses the sensor data stream in real time, analyzes the risk transmission path by dynamically constructing a Bayesian causal network (nodes include 20 variables such as vehicle speed, friction coefficient, and visibility, and edge weights are trained by historical accident data), and triggers an early warning when a key causal chain (such as "rain intensity > 1.5 mm / h → road friction coefficient < 0.4 → braking distance > 8 m") is detected. The experiment is repeated 3 times, the preset true value label (risk event type, trigger time, causal chain structure) is used to calculate the early warning accuracy (True Risk Detection Rate, TRDR), the completeness of the causal chain (the proportion of nodes / edges matching the preset causal path), and the average response delay (the time from the first risk signal input to the complete causal chain generation), and finally the t-SNE is used to visualize the clustering characteristics of the causal graph in the high-dimensional parameter space to verify its physical interpretability. The experimental data recording table is shown in Table 1:

[0066] Table 1 Experimental data table

[0067]

[0068]

[0069] In Table 1, in the TRDR calculation, the accuracy rate = (the number of correct early warning times / the total number of actual risk events) x 100%; 3 false negatives in the composite scenario are caused by environmental noise (radar point cloud distortion) to cause misjudgment of the causal chain; 2 false negatives in the chain scenario are caused by the causal transmission level exceeding the preset network depth (maximum 4 layers). In the completeness of the causal chain, the average number of nodes in the causal chain of the basic scenario: 3.2 (preset 3 nodes); the average number of nodes in the causal chain of the chain scenario: 5.1 (preset 6 nodes), the missing nodes are mostly secondary environmental parameters (such as the second-order influence of humidity on visibility). In the response delay distribution, 90% of the early warnings are completed within 700ms, and the maximum delay in the chain scenario is 1.2 seconds (the causal chain needs to be iteratively updated 4 times); as can be seen from Table 1, the dynamic causal diagram reasoning module (step 2) verifies its effectiveness in multi-level risk scenarios in the simulation experiment, and can balance the reasoning accuracy and real-time performance, providing an interpretable causal support for the risk decision of the autonomous driving system.

[0070] In step 3 of the above embodiment, the improved Jaccard spatiotemporal similarity algorithm reconstructs the quantization method of the traditional Jaccard coefficient on event correlation by introducing a spatiotemporal decay weight function and a dynamic adaptive threshold mechanism, solving the false aggregation problem caused by fixed threshold and single spatial dimension. The spatiotemporal similarity weight function couples the influence of time difference and spatial distance in the form of exponential decay, where the time decay coefficient (β = 0.1) controls the decay rate of the influence of event time interval on the weight, and the spatial decay coefficient (γ = 0.05) adjusts the decay speed of the contribution of spatial distance, the time decay term ensures that the events occurring at a closer time have a higher correlation than the events far away from the time window, and the spatial decay term reduces the influence of distant events on the similarity calculation, so that the calculation result is more consistent with the actual traffic flow rule, in order to balance between time and space, the system introduces a weight factor (a = 0.6) to adjust the influence degree of the two; ensure the dominant role of recent and adjacent events in similarity calculation. The improved Jaccard coefficient numerator part only accumulates the weighted similarity of events of the same type, and the event type matching function (δ(T p ,T q))Invalid superposition of heterogeneous events is constrained, and the denominator eliminates the deviation caused by the difference in the number of events through normalization processing. The dynamic threshold θ is adaptively adjusted based on the logarithmic function of event density (Q). In high-density areas (such as congested road sections), the threshold is automatically increased to reduce low-value aggregation, and in low-density areas, the threshold is reduced to improve sensitivity. This calculation method ensures that in high-density traffic areas, adjacent events are more likely to be determined as related, improving computational efficiency, while in low-density areas, the threshold is increased to reduce unnecessary computational burden. If the spatio-temporal overlap exceeds the threshold, the edge computing cluster is triggered to build. The master node distributes event clusters to computing nodes through a distributed consistent hashing algorithm, combines trajectory prediction algorithms (such as LSTM-Kalman fusion models) and dynamic traffic state regression analysis, models vehicle motion trends and secondary accident probabilities, and outputs risk area boundary parameters (α-shape algorithm) and traffic flow evolution gradient field (spatio-temporal convolution network); if the overlap is insufficient, the event cluster feature vector is compressed through PCA dimensionality reduction, and incremental model training is performed based on the cloud federated learning framework to optimize the spatio-temporal coverage ability and risk prediction accuracy of the global traffic situation atlas.

[0071] In actual deployment, the time window sliding step is 2 seconds to ensure real-time event capture, and the spatio-temporal similarity calculation delay is controlled within 50 ms; the dynamic threshold θ changes dynamically with the event density Q, when Q≥5 events / square kilometer, θ≥0.75, and when Q≤1, θ=0.7; the edge cluster master node distributes event clusters to the computing node with the lowest load through consistent hashing to ensure balanced computing resources; the trajectory prediction model has an average displacement error of ≤0.3 meters on the test set, and the non-convex boundary generated by the α-shape algorithm has a coincidence degree with the real risk area of ≥90%; the cloud federated learning performs global model aggregation every 30 minutes, and the coverage error of the traffic situation atlas after incremental update is reduced by 15%.

[0072] Compared with existing technologies, the algorithm significantly improves the accuracy and scene adaptability of event correlation detection through the cooperative design of spatio-temporal decay weight and dynamic threshold; the edge-cloud collaborative architecture optimizes the allocation of computing resources, reduces redundant calculations while enhancing global situation awareness capabilities; the combination of PCA dimensionality reduction and federated learning realizes continuous model optimization while ensuring data privacy, providing efficient and reliable technical support for real-time warning and long-term decision-making in complex traffic scenarios.

[0073] To verify the beneficial effects of the improved Jaccard algorithm compared to the Jaccard algorithm, a simulation experiment was designed, based on the CARLA automatic driving simulation platform, with the following hardware virtualization configuration:

[0074] Virtual sensors: CARLA's built-in lidar (32 lines, 10 Hz) and RGB camera (30 fps, 1920x1080 resolution) collect event data from simulated streetlight nodes.

[0075] Edge computing simulation: NVIDIA CUDA environment (simulating Jetson AGXXavier computing power) is deployed in a local Docker container to run the improved Jaccard algorithm.

[0076] Communication module: Simulate 5G-V2X (bandwidth 100 Mbps / delay 50 ms) and LiFi (bandwidth 1 Gbps / delay 10 ms) transmission through Socket.

[0077] High-precision map: Call CARLA Town07 map, road topology error ≤0.1m.

[0078] Experimental method: Build a two-way 6-lane urban expressway scene (Town07) in CARLA, set traffic flow to 2000 vehicles / hour, randomly inject 5 types of events (emergency braking, skidding, reverse driving, congestion, and anchor throwing), repeat each experiment 5 times, and run for 10 minutes (simulation time) each time. Adjust the road friction coefficient (0.3-0.8) and weather (alternating between sunny and rainy) dynamically through CARLA Python API to simulate complex traffic environments. Group A runs the improved Jaccard algorithm, and Group B runs the traditional Jaccard algorithm; use ZeroMQ to transmit event cluster data, limit bandwidth and delay parameters (5G: 100 Mbps / 50 ms, LiFi: 1 Gbps / 10 ms). The experimental data is shown in Table 2:

[0079] Table 2 Algorithm Comparison Experimental Data Record

[0080]

[0081] As shown in Table 2, Group A (improved Jaccard) shows significant advantages in simulation: average correlation accuracy 94.2% (Group B 72.2%), false positive rate only 4.8% (Group B 25.9%), verifying the effectiveness of the spatiotemporal weight and dynamic threshold mechanism for cross-region event correlation; calculation delay reduced by 38% (8.5ms vs. 13.4ms), network load reduced by 35% (42.1MB vs. 65.8MB), indicating the algorithm's advantage in reducing redundant computation and communication overhead. Simulation results prove that the improved algorithm has higher reliability and resource efficiency in complex traffic scenarios.

[0082] In addition, in the edge computing cluster and cloud execution task of step 3, the edge computing cluster calculates the high-risk area boundary parameters and the traffic flow evolution trend based on the coupling mechanism of multi-modal data fusion and dynamic spatio-temporal modeling. The core technical features include: first, the PC5 interface of 5G-V2X adopts a straight-through link (Sidelink) mode, bypassing the base station to realize end-to-end communication between streetlight nodes. The physical layer supports a 30 kHz subcarrier spacing and dynamic TDD time slot allocation, ensuring a single-hop delay of less than 5 ms, providing a synchronous spatio-temporal coordinate flow for the LSTM-Kalman filter fusion model. The LSTM network captures the time sequence dependence of vehicle trajectories through the gating mechanism, and the Kalman filter corrects the sensor noise based on the kinematics equation (such as the acceleration covariance matrix Q = diag(0.1, 0.1)). The combination of the two outputs the confidence interval (such as the 95% probability ellipse region) of the vehicle's position within the next 5 seconds. Second, the trajectory conflict detection algorithm introduces the Minkowski Sum theory, models the vehicle's predicted trajectory as a time-varying polygon, and calculates the trajectory overlap area in real time through the Separating Axis Theorem (SAT), extracting the core density distribution of the collision point set. Third, the alpha-shape algorithm performs Delaunay triangulation on the discrete collision points by adjusting the radius parameter alpha (such as alpha = 8 meters), retains the triangle edges with an inscribed circle radius less than alpha, and generates a non-convex polygon risk boundary. Compared with the convex hull algorithm, it can accurately fit complex terrain such as road bends and intersections. Fourth, the spatio-temporal convolutional network (STCN) uses a three-dimensional hollow convolution kernel (size 3x3x3, expansion rate 2) to span 10 frames of historical traffic flow tensors (including traffic volume, speed field, and event heat map) in the time dimension and cover a 50-meter radius of the streetlight perception domain in the spatial dimension. Through multi-scale feature fusion, it extracts traffic flow mutation patterns (such as congestion propagation wave speed). Fifth, the generalized additive model (GAM) takes vehicle density (vehicles / km), average speed (m / s), and event intensity (times / min) as nonlinear base function inputs, uses spline interpolation to fit interaction terms (such as the U-shaped curve of density-speed), outputs the Logit value of the secondary accident probability, and maps the probability field to geographical space through the exponential decay function of the risk diffusion model (decay coefficient λ = 0.2 / s), driving the dynamic adjustment of the warning radius (such as expanding to 1.5 times the standard deviation when the probability is greater than 30%). Sixth, the consistent hashing protocol uses virtual node multiplication technology (each physical node maps 256 virtual nodes), achieving O(1) time complexity data positioning when allocating computing tasks within the edge cluster, ensuring load balancing and millisecond-level synchronization of risk parameter updates.

[0083] The core principle of cloud incremental model training lies in the global alignment of feature space and privacy protection optimization under the federated learning framework. When the event cluster overlap is lower than the threshold, the low-dimensional feature vector generated by PCA dimension reduction (retaining the first 3 principal components, variance contribution rate ≥95%) is pushed to the cloud through the MQTT protocol. The cloud uses the differential privacy (DP) mechanism to add Laplace noise (ε=0.5, δ=1e-5) in the gradient aggregation stage to prevent reverse deduction of the original event data. The incremental Adam optimizer introduces a momentum compensation factor (β1=0.9, β2=0.999), dynamically adjusts the learning rate (initial value 0.001) for sparse updated feature vectors, and uses the elastic weight consolidation (EWC) algorithm to constrain the update amplitude of important parameters (Fisher information matrix diagonal value >0.1) to avoid catastrophic forgetting. The global traffic situation atlas model uses graph attention network (GAT) as the backbone, node embedding contains regional risk probability and traffic flow gradient field direction, edge weight represents cross-regional risk transmission intensity, and the model parameters are updated every 30 minutes and distributed to the edge node through the parameter server to update the local reasoning engine.

[0084] In implementation, the hardware support of this step adopts Huawei MH5000-871 5G-V2X module (PC5 interface, delay ≤20ms) to establish edge cluster communication link, edge node (NVIDIA Jetson AGX Xavier, 32GB memory) runs LSTM-Kalman filter model (TensorFlow Lite framework, time step 10, prediction step 5), α-shape algorithm α value is set to 0.5m, ST-CNN model input is historical 30-minute traffic flow tensor (5-minute slice, resolution 2m×2m grid), training parameters: learning rate 0.001, batch size 16; GAM model uses PyGAM library, basis function is cubic spline (freedom 5), fitting quadratic accident probability weight; risk diffusion model decay coefficient k=0.1, initial radius λ0=50m. In the experimental scene, when the adjacent street lamps detect vehicle reverse (similarity J=0.82≥threshold 0.7), the main node collects trajectory data through 5G-V2X, LSTM-Kalman prediction error ±0.3m, identifies the collision point set, and α-shape generates non-convex boundary (vertex number 15-20), ST-CNN outputs gradient field divergence The GAM calculates the secondary accident probability P=0.72, triggers the risk diffusion model to expand the warning radius to 65m; if the similarity J=0.62< threshold, PCA reduces dimensionality (retaining 95% variance) to generate a 32-dimensional feature vector, which is uploaded to the cloud AWS IoT Core through the MQTT protocol (QoS=1), and the federated learning framework (TensorFlow Federated) uses the incremental Adam optimizer (learning rate 0.002, beta1=0.9, beta2=0.999) to update the global model, the end-to-end delay is ≤200ms, and the single-node power consumption is ≤30W.

[0085] Compared with traditional traffic monitoring methods, the application improves the accuracy of trajectory prediction through LSTM-Kalman filter fusion, and realizes high-risk area modeling more consistent with the real road environment through alpha-shape topology reconstruction. The use of spatiotemporal convolution network combined with three-dimensional hollow convolution layer improves the calculation accuracy of traffic flow evolution trend, enabling the system to predict traffic congestion and potential risks earlier. Combined with the generalized additive model and the risk diffusion model, the application can dynamically adjust the warning area radius, improving the pertinence and timeliness of the warning. In addition, the introduction of consistent hashing protocol enables edge computing clusters to efficiently collaborate in computing, avoiding waste of computing power. The use of federated learning and incremental Adam optimization enables continuous optimization of the cloud, enabling the global traffic situation atlas to adaptively adjust and provide more accurate traffic prediction capabilities, providing a more intelligent and efficient solution for smart city traffic management.

[0086] In step 4 of the above embodiment, the adaptive risk assessment model realizes dynamic risk quantification and classification decision through a multi-level cascade architecture, whose technical principles cover: the data layer adopts long short-term memory network (LSTM) to perform time series encoding (hidden state dimension 128) on the multi-modal fusion data set, capture the time dependence of the event sequence, and simultaneously combine spatial histogram projection (grid resolution 0.5 m x 0.5 m) to construct a spatial distribution heat map of road events, generating a spatio-temporal feature vector (dimension 256); the dynamic risk modeling layer calls Gaussian process regression (GPR) to perform nonlinear risk estimation on the spatio-temporal feature vector, uses a radial basis kernel function (RBF) to fit the continuous distribution of risk values, and derives the probability density function (KL divergence loss weight 0.1) of the event influence range in the latent space (dimension 32) through the conditional variational autoencoder (CVAE), generating a dynamic risk function (output risk value interval [0, 1]); the adaptive threshold optimization layer calculates the risk level distribution of historical events based on kernel density estimation (KDE, bandwidth selection Silverman criterion), dynamically adjusts the risk classification threshold boundary using quantile regression (quantile point τ = 0.75), and realizes the nonlinear contraction or expansion of the threshold with the event density; the multi-level risk classification layer adopts hierarchical Bayesian clustering (HBC, Dirichlet process hyperparameter α = 1.0) to perform multi-granularity classification on the warning events, combines Markov random field (MRF, neighborhood system radius 10 m) to optimize the spatial consistency of the classification labels, and generates a risk classification matrix (dimension 4 x 4, corresponding to low / medium / high / extreme high risk levels); the decision feedback adjustment layer optimizes the classification model parameters by incremental Bayesian update (conjugate prior distribution Beta(2, 2)) combined with real-time accident data, and uses exponential weighted moving average (EWMA, smoothing factor λ = 0.3) to suppress short-term fluctuations in risk levels, ensuring decision stability. The fuzzy decision method performs multi-objective optimization (weight vector [0.6, 0.3, 0.1]) on the event location (WGS84 coordinates), risk level (EN 302 637-2), and road network topology (lane connection graph) based on fuzzy decision tree (FDT, membership function triangular, number of fuzzy rules 50), generates a set of disposal strategies (speed limit value ±20%, detour path topology optimization, emergency lane activation sign), and finally encapsulates the TLV structure instruction stream (Type-Length-Value encoding) according to the ISO 20078-3 standard, adds digital signature (SHA-256 hash) and X.509 certificate chain using the ECDSA algorithm (secp256k1 curve), and constructs a replay attack-resistant warning instruction package.

[0087] In implementation, the data acquisition part includes millimeter wave radar, high-definition camera, V2X communication module, IMU sensor and environmental monitoring sensor. All sensor data are transmitted to the edge computing unit through the 5G-V2X network and processed in real time on the NVIDIA Jetson AGX Orin or other embedded computing devices.

[0088] In the data processing stage, the edge computing unit extracts the time dependence of the event by using LSTM time series encoding, and calculates the spatial distribution characteristics of the road event by spatial histogram projection. Subsequently, the system calculates the risk level of the event by Gaussian process regression, and derives the influence range of the event by combining the CVAE model to construct a dynamic risk assessment function. The adjustment of the risk threshold is performed in the cloud server. The cloud uses the TensorFlow Extended (TFX) framework to manage the dynamic threshold update of the Kernel Density Estimation (KDE) and Quantile Regression (QR), and optimizes the risk level classification model in real time through the Federated Learning Framework. The edge computing unit performs hierarchical Bayesian clustering and optimizes the classification label through MRF to ensure that the classification result conforms to the road topology. The decision optimization part is executed by the Fuzzy Decision Tree (FDT), which calculates the speed limit value, detour path and emergency lane activation strategy based on the event characteristics and historical treatment schemes, and encapsulates the warning instruction in the ISO20078 TLV format, calls the ECDSA encryption signature to ensure data security. Finally, the generated warning instruction is issued to the vehicle terminal (AR-HUD) and the traffic management platform through the 5G-V2X network to provide intelligent traffic supervision services.

[0089] In implementation, the fuzzy decision method is used to handle the uncertainty of event location, risk level and road network topology in traffic scenarios. The core is to construct membership functions and rule base through fuzzy decision tree (FDT), which maps discrete traffic parameters to continuous possibility space to quantify the fitness of different disposal strategies. The fuzzy processing of event location uses Gaussian membership function to convert latitude and longitude coordinates into membership probability of key areas (such as intersections and curves), for example, the membership degree of intersection center point is 1.0, which decays to 0 with the increase of Euclidean distance at a standard deviation σ = 15 meters. The risk level is divided into five levels (Level 1 to Level 5) by triangular membership function, each level covers overlapping intervals (such as Level 3 corresponds to risk value 30-70, peak at 50), ensuring smooth transition between adjacent levels. The road network topology features are extracted by graph embedding technology (such as Node2Vec) to extract topological attributes such as lane connectivity, lane number and traffic flow direction, and mapped to low-dimensional vectors, and then quantified their influence weight on strategy through trapezoidal membership function (such as lane number > 4, emergency lane activation weight is raised to 0.8). The multi-objective optimization module uses the Pareto front search algorithm to take the weighted sum of membership degrees as the objective function, combined with the constraint conditions (such as the length of the detour path is not more than 1.5 times the original route) to generate a set of non-dominated solutions, and finally selects the optimal strategy through the maximum satisfaction principle. The logic generation of disposal strategy depends on the pre-defined rule base, for example, “intersection membership > 0.7 and risk level ≥ Level 4 → speed limit value = 30 km / h + open emergency lane”, the confidence of each rule is calibrated by historical accident data. The TLV (Type-Length-Value) packaging of ISO 20078 standard uses little-endian byte order, arranges longitude (32-bit floating point), latitude (32-bit floating point), risk level (8-bit unsigned integer) and strategy code (16-bit bitmask, such as 0x0001 represents speed limit) according to priority, and ensures the integrity and anti-tamper of the instruction through the elliptic curve digital signature (using secp256k1 curve and SHA-256 hash) of ECDSA algorithm, and the signature process uses ephemeral key to avoid private key leakage.

[0090] In implementation, the hardware of the fuzzy decision method is equipped with an NVIDIA Jetson Orin edge computing module (64 GB memory, 2048 CUDA cores), an integrated dual-frequency GNSS module (positioning accuracy ± 0.1 meters), a laser radar (128 lines, horizontal field of view 360°), and a 5G-V2X communication unit (supporting NR FR1 frequency band). On the software level, the fuzzy decision tree is implemented based on the scikit-fuzzy library, the event location membership function parameters (σ = 15 meters) are calibrated by the intersection accident heat map, and the risk level triangular function vertex is dynamically adjusted by the historical data quantile (25%, 50%, 75%). The Node2Vec algorithm (walk length 80, dimension 16) is used for road network topology embedding, and the output vector is normalized by Min-Max and input into the trapezoidal membership function (threshold interval [0.3, 0.7]). The multi-objective optimization module calls the NSGA-II algorithm of the PlatypUS framework, with a population size of 100, 50 iterations, a crossover probability of 0.8, and a mutation probability of 0.1. The bit mask design of the strategy code is as follows: the low 4 bits represent the speed limit value (such as 0b0011 corresponding to 30 km / h), the middle 4 bits identify the detour path number, and the high 8 bits control the emergency lane and signal light state. When TLV is packaged, the longitude and latitude are kept to 6 decimal places (accuracy about 0.11 meters), the risk level is mapped to 0x01 to 0x05 from Level 1 to Level 5, and the strategy code is generated by bit operation combination. The ECDSA signature is implemented using the OpenSSL library, the temporary key pair is destroyed after each signature, and the signature result is attached to the end of the TLV (64-byte R+S value). The instructions are broadcast through the PC5 interface of 5G-V2X, the receiving end verifies the signature through the pre-set CA certificate, and if the verification fails, the retransmission mechanism is triggered (maximum retry 3 times). The edge node performs decision update every 5 seconds, and the strategy takes effect within 200 ms.

[0091] Compared with the traditional threshold judgment method, the flexibility and scene adaptability of the treatment strategy are significantly improved; the combination of TLV packaging and ECDSA signature ensures the standardization and security of the warning instructions, overcoming the defect that plaintext transmission is easy to tamper; the introduction of dynamic membership parameters and topology embedding technology enables the road network characteristics to participate deeply in the decision-making process, providing the intelligent street lamp system with high robustness and explainable real-time decision-making ability.

[0092] To verify the effectiveness of step 4 in intelligent streetlight traffic intelligent supervision, the experiment uses SUMO (Simulation of Urban Mobility) + OMNeT++ for software simulation experiment, simulates urban traffic environment, and runs the fuzzy decision model in Python + PyTorch environment. The performance of group A (using fuzzy decision tree + ISO 20078 standard + ECDSA encryption) and group B (traditional decision method based on fixed rules) is compared.

[0093] The experimental method is: load the road topology of Beijing Sanhuang in SUMO, set high (3000 vehicles / h), medium (1800 vehicles / h), and low (1000 vehicles / h) three traffic modes. Vehicles are generated according to Poisson distribution, and abnormal event (sudden braking, overspeed, lane conflict, etc.) data is collected. Each event records timestamp, latitude and longitude coordinates, event type, and impact range, and is stored in the database. Group A uses fuzzy decision tree + ISO 20078 TLV format packaging + ECDSA encryption, and group B uses fixed rule-based warning system (IF-THEN rule decision + plaintext JSON packaging). Warning information is transmitted through SUMO-OMNeT++ simulation V2X communication network, and delay and packet loss rate are tested in 5G-V2X environment. The experimental data table is shown in Table 3:

[0094] Table 3 Experimental data recording table of step 4

[0095]

[0096]

[0097] The experimental results show that the calculation delay of group A (using fuzzy decision tree + ISO 20078 standard + ECDSA encryption) is reduced by about 50%, the misjudgment rate is reduced by about 70%, and the execution rate of warning instruction is increased by about 10%-15%, indicating that the fuzzy decision tree can provide more accurate event evaluation and reduce unnecessary warning triggering. The data integrity of group A is close to 100%, while group B uses plaintext JSON packaging, which has a certain risk of data loss. In terms of network bandwidth occupation, group A is reduced by about 50% compared with group B, indicating that TLV format packaging data transmission is more efficient, reducing network load. Therefore, this method can improve the reliability of warning instruction, reduce calculation delay, and optimize the response speed of intelligent traffic system, so that intelligent streetlight can more accurately and stably manage urban traffic.

[0098] In step 5 of the above embodiment, the environment perception module collects the light intensity (unit: Lux) and UWB channel impulse response (CIR) in real time through the light intensity sensor (such as a photodiode array) and the UWB channel probe (based on the IEEE 802.15.4z standard), wherein the LiFi line-of-sight coverage cone angle is determined by the LED half-power angle (such as 60°) and the spatial geometric relationship of the relative position of the terminal, and if the terminal is in the normal direction ± 30° of the transmitting end and the distance is less than 15 meters, it is determined as the effective coverage area of LiFi. Orthogonal frequency division multiplexing (OFDM) modulation divides the LiFi data stream into 512 subcarriers, matches the channel state through adaptive bit loading (such as QAM-64 / QPSK dynamic switching), and maximizes the spectral efficiency. When the multipath fading index (MPF) is detected to exceed the threshold (such as MPF>2.0) or the terminal moving speed is >60km / h, the reinforcement learning strategy network (based on the DQN algorithm, the state space includes channel quality, terminal speed, and interference power) triggers the switch to the 5G-UWB link, calls the dynamic TDD frame structure (frame length 1ms, uplink and downlink time slot ratio 1:3) of the 3GPP NR-U protocol, and avoids adjacent interference through flexible time slot allocation (such as reserving 20% protection interval per frame). Polar code (code length 1024, code rate 0.5) is used as the channel coding scheme to improve the decoding success rate in high noise environment by utilizing the channel polarization effect, and combined with the wideband scanning (scanning interval 5ms) of SRS (sounding reference signal) to realize millimeter wave beamforming (beam width 5°), and the terminal receiving direction is aligned through the singular value decomposition (SVD) algorithm. The dual stack protocol stack maintains the seamless hot standby of LiFi and UWB links through IPv6 over TSN (time sensitive network), and the 802.1Qbv scheduler of TSN synchronizes the dual-link clock with microsecond-level precision to ensure service continuity during switching; the forward error correction (FEC) works with the Reed-Solomon code (RS(255,223)) and the Type-II incremental redundancy mechanism of the hybrid automatic repeat request (HARQ) to dynamically adjust the number of redundant packets according to the real-time packet error rate (PER), and to ensure that the end-to-end transmission bit error rate (BER) is lower than 1e-6.

[0099] In the implementation of this step, on the hardware, the intelligent street lamp integrates the LiFi transmission module (based on Osram SFH 4715SLED, wavelength 850nm, modulation bandwidth 120MHz), UWB transceiver (Decawave DW3000, frequency band 6.5GHz, bandwidth 500MHz), environment perception unit (AMS TSL2591 light sensor + Infineon DEMO BGT60LTR11AIP radar) and 5G-UWB baseband processor (Qualcomm QCA6391). On the software implementation, the reinforcement learning strategy network is deployed on the edge computing unit using TensorFlowLite, the input state vector dimension is 8 (including UWB signal-to-noise ratio, LiFi RSSI, terminal speed, interference power, etc.), the output action space is 2-dimensional (LiFi / UWB selection), and the deep Q network is used in the training stage (γ=0.95, ε-greedy exploration rate 0.1). The OFDM modulation in the LiFi link is realized by FPGA (Xilinx Zynq UltraScale+), the subcarrier spacing is 2MHz, and the cyclic prefix length is 1 / 4 symbol period; the 5G-UWB dynamic TDD frame configuration is 0.7ms downlink time slot, 0.3ms uplink, SRS detection period 5ms, and beam codebook contains 64 predefined directions. The dual protocol stack realizes link redundancy through the IPv6 dual stack module of the Linux kernel and the TSN shaper (gPTP clock synchronization accuracy ±100ns), the FEC encoder uses the libfec library, and the HARQ buffer size is set to 10 transport blocks. In actual deployment, the LiFi transmission rate in the static scene (vehicle speed <30km / h) reaches 3Gbps (BER=5e-7), the UWB maintains 1.5Gbps (PER<0.1%) in the mobile scene (vehicle speed >60km / h), and the switching delay is less than 5ms.

[0100] In the implementation, the scheme significantly improves the transmission robustness and real-time performance of the early warning instruction in complex traffic scenes through the multi-modal communication dynamic selection mechanism driven by reinforcement learning; the cooperation and complementation of LiFi and UWB overcome the coverage blind area and mobility limitation of a single link; the dual stack protocol and hybrid error correction mechanism ensure the continuity and integrity of data transmission, and build a high-reliability, self-adaptive communication infrastructure for the intelligent street lamp system, effectively supporting the vehicle-road cooperation and real-time traffic control needs.

[0101] In the implementation of the traffic monitoring system applied to the intelligent street lamp, the space-time synchronization module, the cross-modal causal reasoning module, the distributed collaborative computing optimization module, the structured early warning coding module and the multi-modal interaction module are included.

[0102] In implementation, the system hardware takes smart street lamps as the physical carrier, and each street lamp is integrated with a multi-sensor module (camera, millimeter wave radar, laser radar, temperature and humidity sensor), a communication unit (LiFi / 5G-UWB dual-mode module, 5G-V2X PC5 interface), and an edge computing node (NVIDIA Jetson Orin) at the top. The sensor module is connected with the edge node through an RS-485 bus, and the power supply adopts a POE (Power over Ethernet) and solar complementary scheme. A POE switch (Cisco CBS110-8PP) provides 48V direct current for high-power devices such as cameras and radars, and low-power devices such as weather sensors are jointly powered by a solar panel (EcoFlow 400W) and a lithium battery (Tesla Powerwall 2). The edge node is connected with a GNSS positioning module (u-blox ZED-F9P) and an atomic clock (Symmetricom SA.45s) through a PCIe interface. The GNSS antenna (Taoglas MA240) and the 10MHz reference clock signal of the atomic clock are connected with the edge node synchronization board card through a coaxial cable to ensure the unity of the space-time reference. The LiFi transmitter (VLNComm LumiNex) and the UWB radio frequency unit (Decawave DW3000) communicate with the edge node through a MIPI interface. The LED array (Cree XLamp XP-E2) of the LiFi is installed at the top of the lamp pole and tilted downward by 30°, covering a radius of 50 meters. The UWB phased array antenna (8x8 unit) has a horizontal radiation angle of ±60° and a vertical angle of ±15°. The cloud data center deploys a federated learning cluster based on Kubernetes (Intel Xeon Platinum 8380+NVIDIA A100), and is interconnected with the edge node through a 10G optical fiber.

[0103] In implementation, the working logic of the system is as follows Figure 6As shown, first, the camera (Sony IMX585) collects a video stream at 30 fps, which is encoded by H.265 and transmitted to the edge node through the GigE interface; the millimeter wave radar (TI AWR1843) outputs point cloud data (10 Hz, JSON format) through the CAN bus; and the 16-line scanning data (20 Hz) of the laser radar (Velodyne VLP-16) is transmitted through Ethernet. The space-time synchronization module first preprocesses the original data: the 10 MHz clock signal generated by the atomic clock drives the hardware of each sensor to synchronize sampling, and the PPP (Precise Point Positioning) coordinates output by the GNSS module are fused with the laser range finder data through the SE(3) Lie group transformation algorithm, which maps the radar point cloud coordinate system (polar coordinates) to the camera imaging plane (pixel coordinates), and the spatial alignment residual is ≤0.1 pixels; the asynchronous data stream (such as 1 Hz sampling of the weather sensor) is resampled to 100 Hz through cubic spline interpolation, and the dynamic time warping (DTW) algorithm is used to detect the time deviation, and when the deviation exceeds 5 ms, the Kalman filter compensation is triggered.

[0104] The preprocessed multi-modal data (time stamp, spatial coordinate, physical dimension normalized feature) is input into the cross-modal causal inference module: the graph neural network (GNN) constructs a dynamic causal graph containing nodes (vehicles, roads, environmental factors) and edges (causal relationship), the node feature dimension is 128, and the edge weight is calculated through the multi-head attention mechanism (number of heads 8, temperature coefficient τ=0.5). The road physical model (such as Pacejka tire equation, Navier-Stokes traffic flow model) is encoded as a physical regularization term of GNN, which constrains the vehicle acceleration prediction error ≤0.2 m / s 2 . The causal inference result (such as “wet road surface↑→friction coefficient↓→braking distance↑”) is filtered through Bayesian structure learning to obtain high-confidence paths (confidence ≥0.7), and an early warning event with evidence chain (JSON format, containing event type, location, risk level) is output.

[0105] The early warning event enters the distributed collaborative computing optimization module: the Jaccard algorithm is improved to calculate the space-time similarity of adjacent streetlight event clusters (threshold θ=0.7+0.1·log(1+Q)), and if the similarity meets the standard, the 5G-V2X networking (PC5 interface, delay ≤10 ms) is triggered to build an edge computing cluster. The master node (NVIDIA Jetson AGX Xavier) allocates computing tasks through the consistent hashing algorithm (virtual node number 1024), and the slave node only provides original data or intermediate features. The master node runs the LSTM-Kalman filter model (hidden layer 64 dimensions, prediction step 5 seconds) to predict the vehicle trajectory, combines the α-shape algorithm (α=0.5) to generate the boundary of the non-convex risk area, and the space-time convolution network (ST-ConvNet, three-dimensional hollow convolution kernel 3×3×3) outputs the traffic flow evolution gradient field F. If the similarity is insufficient, the event cluster is compressed to a 32-dimensional feature vector by PCA dimensionality reduction (retaining 95% variance), uploaded to the cloud federated learning platform through the MQTT protocol (QoS = 1, topic "edge / alert"), and the global traffic situation map model parameters ΔΘ are updated by the incremental Adam optimizer (learning rate 0.001, β1 = 0.9) and issued to the edge node for synchronous update.

[0106] The structured warning encoding module receives the warning event, extracts the structured elements (longitude, latitude, risk level, strategy code), and encapsulates it into a TLV format binary stream according to the ISO 20078-3 standard: the Type field is 4 bytes to identify the data type (e.g. 0x01 for location, 0x02 for risk level), the Length field is 2 bytes to define the data length (longitude / latitude each occupies 8 bytes, risk level 1 byte), and the Value field is dynamically filled. The encoded instruction stream is appended with a digital signature (64 bytes) by ECDSA algorithm (elliptic curve secp256k1, SHA-256 hash) and embedded with X.509 certificate chain (root certificate, intermediate CA) to form a tamper-resistant instruction package.

[0107] The multi-modal interaction module dynamically selects the transmission channel according to the environmental state: the Q-learning reinforcement learning model (state space S = {Lux, CSI, speed v}, reward function R = 0.6 · throughput + 0.3 · reliability - 0.1 · energy consumption) makes real-time decisions on LiFi or UWB channels. If the light intensity ≥2000Lux and the terminal is in the LiFi coverage cone angle (±60°), start LiFi OFDM modulation (subcarrier 256, modulation depth 0.8), and the optical signal is transmitted at a rate of 10Gbps to the vehicle-mounted terminal (Hamamatsu S1223 photodiode) and AR-HUD (HoloLens 2); if the multipath fading ≥3dB or the vehicle speed ≥80km / h is detected, switch to 5G-UWB (3GPP NR-U standard), use Polar code (code length 512) and dynamic TDD frame structure (DL:UL = 4:1) to multicast instructions at 3.5GHz frequency band. The dual stack protocol (IPv6 / TSN) maintains link hot standby, forward error correction (FEC, redundancy 20%) and hybrid automatic repeat request (HARQ, maximum retransmission 3 times) ensure transmission integrity, and dynamically select the channel with the lowest packet loss rate as the main link (switching delay ≤10ms).

[0108] The logical processing of the system presents the characteristics of hierarchical progression and closed-loop feedback: the space-time synchronization module is the bottom layer data entrance, ensuring the space-time consistency of multi-source data; the cross-modal causal reasoning module realizes event root cause tracing and explainable early warning through the fusion of physical constraints and graph neural networks; the distributed collaborative module dynamically allocates computing power based on space-time correlation, optimizing the balance between local response and global learning; the structured coding module converts the decision results into machine interpretable instructions, ensuring cross-platform compatibility; the multi-modal interaction module ensures the high reliable transmission of instructions through adaptive channel selection and error correction mechanism. The modules interact with each other through standardized interfaces (such as ROS topics and MQTT messages) to realize loose coupling, and rely on the data bidirectional synchronization (model parameter ΔΘ, risk map P_risk) between edge and cloud to form a closed-loop optimization system.

[0109] Although the specific embodiments of the present application are described above, those skilled in the art should understand that these specific embodiments are only illustrative, and those skilled in the art can make various omissions, substitutions and changes to the details of the above-mentioned method and system without departing from the principles and essence of the present application. For example, combining the above method steps, performing substantially the same function in substantially the same manner to achieve substantially the same result according to the same method is within the scope of the present application. Therefore, the scope of the present application is only limited by the appended claims.

Claims

1. A traffic intelligent monitoring method applied to smart streetlights; characterized in that: include: Step 1: Based on the raw data from multiple sensors, a multimodal fusion dataset is generated by unifying the spatiotemporal reference and spatial mapping association between the atomic clock and the GNSS positioning module. Step 2: Use the road physical model to constrain the multimodal fusion dataset, and use a graph neural network to construct a dynamic causal graph and perform backpropagation inference to output early warning events with causal chain evidence; Step 3: Based on the warning event, use the improved Jaccard spatiotemporal similarity algorithm to detect the spatiotemporal correlation of adjacent street light event clusters. If the overlap is higher than or equal to the preset threshold, construct an edge computing cluster through 5G-V2X networking to calculate the boundary parameters of high-risk areas and traffic flow evolution trends. If the overlap is lower than the preset threshold, the compressed feature vectors of each node are uploaded to the cloud to perform incremental model training, and the optimized global traffic situation map and model iteration instructions are output. Step 4: Extract structured elements from the warning event, including event location, risk level, and recommended action; encapsulate the binary instruction stream according to ISO 20078 standard and attach a digital signature to generate the warning instruction; Step 5: Based on the environmental conditions, dynamically select LiFi or 5G-UWB communication medium through a multimodal fusion reinforcement learning algorithm, and synchronously transmit the warning instructions to the vehicle terminal, AR-HUD and traffic management platform.

2. The traffic intelligent monitoring method applied to smart streetlights according to claim 1, characterized in that: The method for unifying the spatiotemporal reference and associating spatial mapping in step 1 is as follows: based on the real-time GNSS positioning coordinates and the laser rangefinder measurement values, a sensor spatial mapping matrix is ​​constructed through the Lie group spatial transformation algorithm, and the radar point cloud coordinate system is aligned with the camera imaging plane. For asynchronously arriving sensor data streams, a cubic spline interpolation algorithm is used to resample the data under a spatiotemporal reference to eliminate time offset. A dynamic time warping algorithm is used to detect time axis distortion of the multimodal data stream. If the maximum path deviation exceeds a preset threshold, a Kalman filter compensation mechanism is used to reconstruct the temporal consistency. Finally, a multimodal fusion dataset is output, which includes timestamps, spatial coordinates, and sensor feature vectors normalized to physical dimensions.

3. The traffic intelligent monitoring method applied to smart streetlights according to claim 1, characterized in that: The road physical model uses the Lagrange-Euler hybrid coordinate transformation method to analyze vehicle dynamic parameters and combines it with the Passon-Poisson coupled fluid model to calculate road environmental constraints, so that the motion state and the road topology form a dynamic constraint relationship. In the constraint solution process, the road physical model uses the sparse tensor decomposition method to extract key influencing factors in the physical model and uses the normalized mutual information calculation method to screen the parameters affecting abnormal driving behavior, generating a road state causal constraint matrix.

4. The traffic intelligent monitoring method applied to smart streetlights according to claim 1, characterized in that: The graph neural network is based on the road state causal constraint matrix. It uses a heterogeneous graph attention mechanism to calculate the spatiotemporal correlation weights of road state, vehicle behavior and environmental factors, and combines a Bayesian structure learning algorithm to perform causal relationship inference and generate a dynamic causal graph. During the causal inference process, the graph neural network calls backpropagation gradient constraint optimization to correct the causal relationship weights, and combines the optimal transport mapping method to evaluate the credibility of causal chain evidence and filter out low-confidence causal paths.

5. A traffic intelligent monitoring method applied to smart streetlights according to claim 1, characterized in that: The improved Jaccard spatiotemporal similarity algorithm works as follows: First, it extracts the event cluster set C of adjacent street light nodes through a sliding time window. i =(e1,e2,...e n ), where each event e k Includes timestamp t k Spatial coordinates (x) k ,y k ) and event type code T k Next, the spatiotemporal similarity weight function is defined. Where t diff Let be the event time difference, d be the spatial Euclidean distance, where d = 0.6; β = 0.1 and γ = 0.05 are the time decay coefficient and spatial decay coefficient, respectively; a is the balance factor for time and spatial weights; the Jaccard coefficient is improved based on the aforementioned spatiotemporal similarity weight function w(t,d) as follows: In formula (1), δ(T) p ,T q ) is the event type matching function; ∑e p ∈C i, e q ∈C j w(t diff ,d)δ(T p ,T q ) represents the weighted similarity sum of all event pairs in two event clusters, considering only events of the same type; ∑e p ∈C i w p -∑e p ∈C i, e q ∈C j W(t diff ,d)δ(T p ,T q This is used to avoid deviations caused by differences in the number of events; if J st When ≥θ, the 5G-V2X PC5 interface is triggered to establish an edge computing cluster, and the event cluster is distributed to the master computing node through a distributed consistent hashing algorithm; where θ is a dynamic threshold, calculated using the following formula: θ=(0.7+0.1log(1+Q) (2) In formula (2), Q represents the current regional event density. Next, the master node calls the trajectory prediction algorithm to model the vehicle movement trend within the affected area of ​​the abnormal event, and combines this with dynamic traffic state regression analysis to calculate the probability of a secondary accident, as well as the boundary parameters of the risk area and the traffic flow evolution gradient field. If J st If θ < , then PCA dimensionality reduction is performed on each node event cluster to generate compressed feature vectors, which are then uploaded to the cloud federated learning framework via the MQTT protocol. The global traffic situation map model parameters are updated based on the incremental Adam optimizer, and the model iteration instructions and risk probability distribution matrix are output.

6. A traffic intelligent monitoring method applied to smart streetlights according to claim 1, characterized in that: The calculation principle of the edge computing cluster for calculating the boundary parameters of high-risk areas and traffic flow evolution trends is as follows: First, a low-latency communication link is established through the PC5 interface of 5G-V2X. The master node collects the spatiotemporal coordinate data of adjacent street light event clusters and inputs it into the LSTM-Kalman filter fusion model to predict the vehicle trajectory distribution. The potential collision point set is identified based on the trajectory conflict detection algorithm. Next, the collision point set is reconstructed topologically using the α-shape algorithm to generate non-convex polygonal risk area boundary parameters. At the same time, a spatiotemporal convolutional network is constructed. The spatiotemporal correlation pattern between historical traffic flow tensors and real-time event features is extracted through a three-dimensional dilated convolutional layer, and the vector field distribution of the traffic flow evolution gradient field is output. Combined with a generalized additive model, based on vehicle density, average speed, and event intensity characteristics, a nonlinear regression equation is fitted to calculate the secondary accident probability weight. Finally, the accident probability is mapped to a gradient field correction factor through the exponential decay function of the risk diffusion model, driving the dynamic adjustment of the warning area radius. The warning parameters of each node are updated synchronously through the consistent hashing protocol within the edge cluster, realizing the distributed collaborative calculation of regional risk boundaries and traffic flow trends.

7. A traffic intelligent monitoring method applied to smart streetlights according to claim 1, characterized in that: Step 4 employs an adaptive risk assessment model to perform risk grading calculations and combines this with a historical event database to dynamically adjust the risk classification threshold. The optimal response strategy is then calculated using a fuzzy decision-making method. The adaptive risk assessment model comprises a data layer, a dynamic risk modeling layer, an adaptive threshold optimization layer, a multi-level risk classification layer, and a decision feedback adjustment layer. The data layer receives multimodal fusion dataset data, uses temporal encoding with a long short-term memory network to extract time dependencies, and combines spatial histogram projection to construct a spatial distribution mapping of road events, generating a spatiotemporal feature vector. The dynamic risk modeling layer uses a Gaussian process regression method to perform continuous risk estimation on the spatiotemporal feature vector and employs a conditional variational autoencoder to derive the event impact range, thereby constructing a dynamic risk function. The adaptive threshold optimization layer is used to calculate the risk level distribution of historical events using kernel density estimation and dynamically adjusts the risk classification threshold by calling quantile regression to generate real-time updated risk boundary parameters. The multi-level risk classification layer is used to classify early warning events using hierarchical Bayesian clustering and optimizes the classification labels by combining Markov random fields to generate a risk classification matrix. The decision feedback adjustment layer optimizes the classification model by combining incremental Bayesian updates with the latest accident data and uses exponentially weighted moving averages to smooth changes in risk levels and adjust the final risk classification decision.

8. A traffic intelligent monitoring method applied to smart streetlights according to claim 7, characterized in that: The fuzzy decision-making method is based on fuzzy decision trees to perform multi-objective optimization of event location, risk level and road network topology, and generates a set of handling strategies, including speed limit, detour route and emergency lane activation; longitude, latitude, risk level and strategy are encoded and encapsulated into TLV format binary stream according to ISO 20078 standard, and digital signature is added using ECDSA algorithm to generate tamper-proof warning instructions.

9. A traffic intelligent monitoring method applied to smart streetlights according to claim 1, characterized in that: The working principle of the multimodal fusion reinforcement learning algorithm is as follows: based on the real-time light intensity L collected by the environment perception module... lux Based on UWB channel state information, a dynamic selection strategy is constructed. If L lux If the value is ≥2000 and the relative position of the terminal is within the LiFi line-of-sight coverage cone angle, the LiFi module is triggered to send commands to the vehicle terminal and AR-HUD in orthogonal frequency division multiplexing modulation mode; if multipath fading index F is detected... MP If the signal strength is ≥3dB or the terminal's mobile speed v≥80km / h, the system switches to 5G-UWB, invokes the 3GPP NR-U MAC layer protocol, allocates time slot resources based on the dynamic TDD frame structure, enhances transmission reliability through Polar codes, and achieves beamforming by combining the SRS channel detection method, multicasting to the AR-HUD and traffic management platform at a rate of 10Gbps. The multimodal fusion reinforcement learning algorithm maintains the hot standby state of the LiFi and UWB links through the dual-stack protocol stacks IPv6 and TSN, and ensures the integrity of instruction transmission based on forward error correction and hybrid automatic repeat request mechanisms, dynamically selecting the channel with the lowest packet loss rate to execute the main data stream transmission.

10. A traffic monitoring system applied to smart streetlights, characterized in that: A traffic intelligent monitoring method for smart streetlights as described in any one of claims 1-9, the system comprising: a spatiotemporal synchronization module, a cross-modal causal reasoning module, a distributed collaborative computing optimization module, a structured early warning coding module, and a multimodal interaction module; The spatiotemporal synchronization module is used to perform spatiotemporal benchmark unification and fusion preprocessing on multi-sensor data from streetlights. The cross-modal causal reasoning module is used to combine road physical models and graph neural networks to perform causal reasoning on multimodal data and output interpretable early warning events. The distributed collaborative computing optimization module is used to dynamically allocate computing tasks between smart street light edge nodes and the cloud based on the spatiotemporal correlation of events. The structured early warning coding module is used to encode early warning events into structured instructions that support machine parsing; The multimodal interaction module is used to transmit structured early warning instructions through an adaptive channel.

Images