Traffic supervision system applied to intelligent street lamp and intelligent supervision method thereof

Abstract

The invention discloses a traffic supervision system applied to an intelligent street lamp and an intelligent supervision method thereof, relates to the technical field of intelligent traffic, and solves the problems that an existing intelligent street lamp system lacks a physical-digital mapping relation, edge computing resource allocation is low in efficiency and cloud computing delay is high. According to the scheme, on the basis of multi-sensor data fusion, space-time reference unification is carried out by adopting an atomic clock and a GNSS, and a dynamic causal graph is constructed through a graph neural network, so that abnormal event detection is optimized; an improved Jaccard space-time similarity algorithm is adopted to optimize calculation task allocation, an edge calculation cluster is constructed based on 5G-V2X, and high-risk region identification and traffic flow prediction are carried out; a LiFi or 5G-UWB communication medium is adaptively selected through a multi-modal fusion reinforcement learning algorithm, and efficient early warning information synchronization is realized; according to the method, the multi-source data fusion value and the early warning precision are remarkably improved, the computing power resource utilization rate is optimized, and the instruction real-time performance and the system self-adaptive capability in a complex environment are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation, and more particularly to a traffic supervision system applied to intelligent street lights and an intelligent supervision method thereof. Background Art

[0002] With the rapid development of smart city and Internet of Things technologies, urban traffic supervision is gradually transforming towards digitalization and intelligence. As a core node of urban infrastructure, intelligent street lights have become a multi-functional carrier integrating traffic perception, environmental monitoring, and communication interaction due to their wide distribution, stable power supply, and uniform spatial coverage. The traditional traffic management mode relies on manual inspections and isolated information systems, making it difficult to meet the real-time response requirements for complex road conditions. In contrast, the traffic supervision method based on intelligent street lights can achieve precise perception and dynamic early warning of traffic flow, accident risks, and environmental anomalies by integrating technologies such as sensor networks, edge computing, and AI analysis, providing decision-making support in the full time domain and multiple dimensions for urban traffic governance, and becoming an important direction for the evolution of intelligent transportation systems.

[0003] Currently, intelligent street light traffic supervision technology mainly realizes road condition monitoring through multi-sensor integration and edge computing. For example, in patent CN116092287A, network cameras, air quality monitors, meteorological sensors, and 5G base stations are deployed in street lights, and data is separately sent through "two networks" to the municipal platform and the traffic police private network. Patent CN118328359A further proposes an abnormal warning component, which uses a camera to collect road images, combines AI algorithms to identify abnormal road conditions, and triggers an early warning through dynamic aperture lights and lighting brightness adjustment. At the data processing level, patent CN118506582A discloses a risk judgment model based on the collaboration between the cloud and the edge. The street light end collects local road information, and the cloud fuses multi-node data to generate a global risk map. In traffic event detection, patent CN118334866A uses a deep learning model to perform object detection and tracking on surveillance videos and predicts traffic flow changes in combination with historical data. In addition, patent CN119068682A optimizes traffic signal control strategies through accident index analysis and road condition correction coefficient calculation. Existing technologies generally adopt the technical path of "multi-sensor collection → edge / cloud processing → rule-based early warning", and the core features include heterogeneous device integration, AI model-driven decision-making, and multi-platform data interconnection.

[0004] However, despite the certain progress made by intelligent street lights in the field of traffic supervision, there are still some obvious defects in the existing technologies: First, although the existing systems stack a variety of sensors (such as the integrated camera and meteorological monitoring equipment in CN116092287A), they lack the ability to construct the physical-digital mapping relationship, resulting in the loss of data value. For example, when the camera detects a vehicle's emergency braking behavior, the meteorological sensor data is not associated with the road surface friction coefficient model, and it is impossible to determine whether it is caused by slipperiness. Only relying on single-modal data to trigger warnings (such as in CN118506582A), the misjudgment rate remains high. Second, the existing solutions emphasize edge-side data processing (such as the risk judgment at the street light end in CN118506582A), but do not optimize the calculation task allocation according to the spatio-temporal correlation of events, resulting in waste of resources. For example, adjacent street lights repeatedly calculate the trajectory of the same moving target (such as in CN118334866A), rather than sharing intermediate results through distributed collaboration, causing redundant computing power. At the same time, the cloud computing tasks still rely on the centralized processing mode, resulting in delays in data transmission and decision execution, and also reducing the overall system efficiency. Therefore, a traffic supervision system for a new type of intelligent street light and its intelligent supervision method are needed to solve the above problems. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technologies, the present invention discloses a traffic supervision system for intelligent street lights and its intelligent supervision method, aiming to solve the problems of weak multi-modal data correlation, redundant edge computing resources, and serious cloud decision-making delays in the existing technologies.

[0006] To achieve the above technical effects, the present invention adopts the following technical solutions:

[0007] A traffic intelligent supervision method for intelligent street lights, including:

[0008] Step 1: Based on the original multi-sensor data, unify the spatio-temporal reference and perform spatial mapping association through an atomic clock and a GNSS positioning module to generate a multi-modal fusion data set;

[0009] Step 2: Invoke the road physical model to constrain the multi-modal fusion data set, and use a graph neural network to construct a dynamic causal graph and perform backpropagation inference to output a warning event with causal chain evidence;

[0010] Step 3: According to the warning event, use an improved Jaccard spatio-temporal similarity algorithm to detect the spatio-temporal correlation of adjacent street light event clusters. If the overlap degree 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 the evolution trend of traffic flow; if the overlap degree is lower than the preset threshold, upload the compressed feature vectors of each node to the cloud to perform incremental model training, and output an optimized global traffic situation map and a model iteration instruction;

[0011] Step 4: Extract structured elements from the warning events, including event location, risk level, and recommended actions, encapsulate the binary instruction stream according to the ISO 20078 standard and append a digital signature to generate a warning instruction;

[0012] Step 5: Based on the environmental status, dynamically select the LiFi or 5G-UWB communication medium through a multi-modal fusion reinforcement learning algorithm, and synchronously transmit the warning instruction to the vehicle-mounted terminal, AR-HUD, and traffic management platform.

[0013] As a further technical solution of the present invention, the working method of the improved Jaccard spatio-temporal similarity algorithm is as follows: First, extract the event cluster set C i =(e1, e2,... e n ) of adjacent street lamp nodes through a sliding time window, where each event e k includes a timestamp t k , spatial coordinates (x k , y k ) and an event type code T k ; Then, define the spatio-temporal similarity weight function where t diff is the event time difference, d is the spatial Euclidean distance, where d = 0.6; β = 0.1 and γ = 0.05 are the time decay coefficient and the spatial decay coefficient respectively; a is the balance factor of the time and space weights; Improve the Jaccard coefficient based on the spatio-temporal similarity weight function w(t, d) as:

[0014]

[0015] 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 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 differences in the number of events; If J st≥θ, trigger the establishment of an edge computing cluster for the PC5 interface of 5G-V2X, and allocate the event cluster to the main computing node through the distributed consistent hashing algorithm; where 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 event density in the current area. Then, the main node calls the trajectory prediction algorithm to model the vehicle movement trend in the abnormal event impact area, and combines the dynamic traffic state regression analysis to calculate the probability of a secondary accident, calculate the boundary parameters of the risk area and the traffic flow evolution gradient field; if J st <θ, then perform PCA dimensionality reduction on each node event cluster to generate compressed feature vectors, upload them to the cloud federated learning framework through the MQTT protocol, update the global traffic situation map model parameters based on the incremental Adam optimizer, and output the model iteration instruction and the risk probability distribution matrix.

[0018] As a further technical solution of the present invention, the calculation principle of the edge computing cluster for calculating the boundary parameters of the high-risk area and the traffic flow evolution trend is: First, establish a low-latency communication link through the PC5 interface of 5G-V2X. The main node collects the spatio-temporal coordinate data of the adjacent street lamp event clusters, inputs them 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, perform topological reconstruction on the collision point set through the α-shape algorithm to generate the boundary parameters of the non-convex polygon risk area; at the same time, construct a spatio-temporal convolutional network, extract the spatio-temporal correlation pattern of the historical traffic flow tensor and the real-time event features through the three-dimensional dilated convolutional layer, and output the vector field distribution of the traffic flow evolution gradient field; combine the generalized additive model, and based on the vehicle density, average speed and event intensity features, fit the non-linear regression equation to calculate the probability weight of the secondary accident; finally, through the exponential decay function of the risk diffusion model, map the accident probability to the gradient field correction factor, drive the dynamic adjustment of the early warning area radius; and synchronously update the early warning parameters of each node through the consistent hashing protocol within the edge cluster to achieve the distributed collaborative calculation of the regional risk boundary and the traffic flow trend.

[0019] As a further technical solution of the present invention, 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 environment perception module, construct a dynamic selection strategy. If L lux ≥2000 and the relative position of the terminal is within the LiFi line-of-sight coverage cone angle, then trigger the LiFi module to send instructions to the in-vehicle terminal and AR-HUD in the orthogonal frequency division multiplexing modulation mode; if the multipath fading exponent F MPIf the signal-to-noise ratio ≥ 3dB or the terminal moving speed v ≥ 80km / h, then switch to 5G-UWB, call the 3GPP NR-UMAC layer protocol, allocate time slot resources based on the dynamic TDD frame structure, enhance the transmission reliability through Polar codes, and combine the SRS channel sounding method to achieve beamforming, and multicast to the AR-HUD and traffic management platform at a rate of 10Gbps; the multi-modal fusion reinforcement learning algorithm maintains the hot standby state of the LiFi and UWB links through the dual-stack protocol stack IPv6 and TSN, and ensures 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 execute the main data stream transmission.

[0020] As a further technical solution of the present invention, a traffic supervision system applied to smart street lights includes a spatio-temporal synchronization module, a cross-modal causal reasoning module, a distributed collaborative computing optimization module, a structured warning coding module, and a multi-modal interaction module;

[0021] The spatio-temporal synchronization module is used to unify the spatio-temporal reference and perform fusion preprocessing on the multi-sensor data at the street light end;

[0022] The cross-modal causal reasoning module is used to perform causal reasoning on multi-modal data by combining the road physical model and the graph neural network, and output an interpretable warning event;

[0023] The distributed collaborative computing optimization module is used to dynamically allocate the computing tasks of the smart street light edge nodes and the cloud based on the spatio-temporal correlation of events;

[0024] The structured warning coding module is used to encode the warning event into a structured instruction that supports machine parsing;

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

[0026] Based on the above technical solutions, the positive and beneficial effects of the present invention are:

[0027] 1. By constructing a physical-digital mapping relationship, dynamically coupling the sensor data with the road physical model (such as the friction coefficient model), the problem of misjudgment caused by the isolation of multi-source data in the prior art is solved. Using the backpropagation inference of the graph neural network to generate causal chain evidence, making the warning event physically interpretable, significantly reducing the false alarm rate caused by the one-sidedness of single-modal data, and at the same time enhancing the trust of the supervision department in the system decision-making.

[0028] 2. Based on the improved Jaccard spatio-temporal similarity algorithm, accurately identify the spatio-temporal correlation of event clusters, dynamically partition the edge computing cluster and cloud incremental learning tasks, and avoid redundant calculations of adjacent street lights on the same target. Through distributed collaboration and task chain splitting, achieve on-demand allocation of computing resources, reduce ineffective energy consumption, and at the same time shorten the data transmission delay of cloud centralized processing, and improve the overall real-time response of the system.

[0029] 3. Adopt structured element extraction and standard coding (ISO 20078), combined with digital signature technology, to ensure the semantic consistency and transmission security of early warning instructions. Through the multi-modal communication dynamic selection mechanism, adaptively switch the LiFi / 5G-UWB communication medium according to the environmental state, solve the signal attenuation or interference problem of traditional single communication links in complex scenarios, and ensure the synchronous and reliable transmission of instructions among in-vehicle terminals, AR-HUDs, and management platforms.

[0030] 4. By calculating the boundaries of high-risk areas and the evolution trends of traffic flow in real time at the edge, combined with the global traffic situation map generated by cloud incremental training, form a two-layer decision-making system of "local rapid response - global continuous optimization". The combination of dynamic causal reasoning and physical model constraints enables the system to not only quickly handle sudden risks but also iteratively improve the long-term prediction accuracy based on historical data, breaking through the limitations of the separation between local response and global optimization in existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings, where:

[0032] Figure 1 is the architecture diagram of a traffic intelligent supervision method applied to intelligent street lights of the present invention;

[0033] Figure 2 is the working principle diagram of step 1 of the present invention;

[0034] Figure 3 is the working principle diagram of step 2 of the present invention;

[0035] Figure 4 is the working principle framework diagram of the edge computing cluster of the present invention;

[0036] Figure 5 is the structural framework diagram of the adaptive risk assessment model of the present invention;

[0037] Figure 6This is the principle framework diagram of the traffic supervision system applied to intelligent street lights in the present invention. Detailed implementation manners

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0039] In the embodiment, a traffic supervision method applied to intelligent street lights, as Figure 1 shown, includes:

[0040] Step 1: Based on the original data of multiple sensors, unify the spatio-temporal reference and perform spatial mapping association through an atomic clock and a GNSS positioning module to generate a multi-modal fusion data set; as Figure 2 shown, the method is: based on the GNSS real-time positioning coordinates and the measurement values of the laser rangefinder, construct a sensor spatial mapping matrix through the Lie group space transformation algorithm, and align the data of the radar point cloud coordinate system and the camera imaging plane; for the sensor data streams arriving asynchronously, use the cubic spline interpolation algorithm to resample under the spatio-temporal reference, eliminate the time offset, and detect the time-axis distortion of the multi-modal data streams through the dynamic time warping algorithm. If the maximum path deviation exceeds the preset threshold, use the Kalman filter compensation mechanism to reconstruct the temporal consistency; finally, output a multi-modal fusion data set, including a timestamp, a spatial coordinate, and a sensor feature vector after physical dimension normalization.

[0041] Step 2: Invoke the road physical model to constrain the multi-modal fusion data set, and use a graph neural network to construct a dynamic causal graph and perform backpropagation inference to output a warning event with causal chain evidence; the specific working principle is as Figure 3As shown: The road physical model uses the Lagrangian-Eulerian mixed coordinate transformation method to analyze vehicle dynamics parameters, and combines the Poisson-Poisson coupled fluid model to calculate road environment constraints, so that the motion state and the road topology form a dynamic constraint relationship; in the process of constraint solving, the road physical model uses the sparse tensor decomposition method to extract the key influencing factors in the physical model, and screens the abnormal driving behavior influencing parameters through the normalized mutual information calculation method to generate a road state causal constraint matrix. The graph neural network is based on the road state causal constraint matrix, uses the heterogeneous graph attention mechanism to calculate the spatio-temporal correlation weights of road states, vehicle behaviors, and environmental factors, and combines the Bayesian structure learning algorithm to perform causal relationship reasoning to generate a dynamic causal graph; in the process of causal reasoning, the graph neural network calls the backpropagation gradient constraint to optimize and correct the causal relationship weights, and combines the optimal transport mapping method to evaluate the credibility of the causal chain evidence and screen out low-confidence causal paths.

[0042] Step 3: According to the warning event, use the improved Jaccard spatio-temporal similarity algorithm 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, construct an edge computing cluster through 5G-V2X networking to calculate the boundary parameters of the high-risk area and the traffic flow evolution trend; if the overlap degree is lower than the preset threshold, upload the compressed feature vectors of each node to the cloud to perform incremental model training, and output the optimized global traffic situation map and model iteration instructions; The working method of the improved Jaccard spatio-temporal similarity algorithm is: First, extract the event cluster set C i =(e1,e2,...e n ) of adjacent street lamp nodes through a sliding time window, where each event e k contains the time stamp t k , the spatial coordinates (x k ,y k ) and the event type code T k ; Then, define the spatio-temporal similarity weight function where t diff is the event time difference, d is the spatial Euclidean distance, where d = 0.6; β = 0.1, γ = 0.05 are the time decay coefficient and the spatial decay coefficient respectively; a is the balance factor of the time and space weights; Based on the spatio-temporal similarity weight function w(t,d), the improved Jaccard coefficient is:

[0043]

[0044] In formula (1), δ(T p ,T q ) is the 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 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 the deviation caused by the difference in the number of events; if J st ≥θ, trigger the establishment of an edge computing cluster for the PC5 interface of 5G-V2X, and allocate the event cluster to the main computing node through the distributed consistent hashing algorithm; where 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 event density in the current area. Then, the main node calls the trajectory prediction algorithm to model the vehicle movement trend in the abnormal event impact area, and combines the dynamic traffic state regression analysis to calculate the probability of secondary accidents, calculate the boundary parameters of the risk area and the traffic flow evolution gradient field; if J st <θ, perform PCA dimensionality reduction on each node event cluster to generate compressed feature vectors, upload them to the cloud federated learning framework through the MQTT protocol, update the global traffic situation map model parameters based on the incremental Adam optimizer, and output the model iteration instruction and the risk probability distribution matrix. As Figure 4As shown in the figure, the calculation principle of the edge computing cluster for calculating the boundary parameters of high-risk areas and the evolution trend of traffic flow is as follows: First, a low-latency communication link is established through the PC5 interface of 5G-V2X. The master node collects the spatio-temporal coordinate data of adjacent street lamp event clusters, inputs it into the LSTM-Kalman filter fusion model to predict the vehicle trajectory distribution, and identifies potential collision point sets based on the trajectory conflict detection algorithm; then, the α-shape algorithm is used to perform topological reconstruction on the collision point sets to generate non-convex polygon risk area boundary parameters; at the same time, a spatio-temporal convolutional network is constructed, and the spatio-temporal correlation pattern of historical traffic flow tensors and real-time event features is extracted through three-dimensional dilated convolutional layers, and the vector field distribution of the traffic flow evolution gradient field is output; combined with the generalized additive model, based on vehicle density, average speed and event intensity features, a non-linear 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 into a gradient field correction factor to drive the dynamic adjustment of the warning area radius; and the warning parameters of each node are synchronized and updated through the consistent hashing protocol within the edge cluster to achieve distributed collaborative calculation of the regional risk boundary and traffic flow trend.

[0047] Step 4: Extract structured elements from the warning events, including event location, risk level and recommended actions, encapsulate the binary instruction stream according to the ISO 20078 standard and attach a digital signature to generate a warning instruction; specifically, an adaptive risk assessment model is used to perform risk grading calculations, and the dynamic risk classification threshold is adjusted in combination with the historical event database, and the optimal disposal strategy is calculated through a fuzzy decision-making method; such as 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 data of a multi-modal fusion data set, extract time dependence by using the time series coding of a long short-term memory network, and combine with a spatial histogram projection to construct a spatial distribution mapping of road events, and generate a spatio-temporal feature vector; the dynamic risk modeling layer is used to call a Gaussian process regression method to perform continuous risk estimation on the spatio-temporal feature vector, and use a conditional variational autoencoder to deduce 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 by using a kernel density estimation method, and call a quantile regression to dynamically adjust the risk classification threshold to generate a risk boundary parameter updated in real time; the multi-level risk classification layer is used to classify early warning events by using hierarchical Bayesian clustering, and optimize classification labels in combination with a Markov random field to generate a risk classification matrix; the decision feedback adjustment layer optimizes the classification model by combining incremental Bayesian update with the latest accident data, and uses an exponentially weighted moving average to smooth the change of the risk level to adjust the final risk classification decision. The fuzzy decision-making method performs multi-objective optimization on the event location, risk level, and road network topology based on a fuzzy decision tree to generate a set of disposal strategies, including speed limit values, detour routes, and emergency lane activation; according to the ISO 20078 standard, longitude, latitude, risk level, and strategy codes are encapsulated into a binary stream in TLV format, and a digital signature is added by using the ECDSA algorithm to generate a tamper-proof early warning instruction.

[0048] Step 5: Based on the environmental state, dynamically select the LiFi or 5G-UWB communication medium through a multi-modal fusion reinforcement learning algorithm, and synchronously convey the early warning instruction to in-vehicle terminals, AR-HUDs, and traffic management platforms. The working principle of the multi-modal fusion reinforcement learning algorithm is as follows: 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 relative position of the terminal is within the LiFi line-of-sight coverage cone angle, trigger the LiFi module to send instructions to in-vehicle terminals and AR-HUDs in an orthogonal frequency division multiplexing modulation manner; if the multipath fading index F MPIf ≥3dB or the terminal moving speed v≥80km / h, then switch to 5G-UWB, call the 3GPP NR-U MAC layer protocol, allocate time slot resources based on the dynamic TDD frame structure, enhance the transmission reliability through Polar codes, and implement beamforming by combining the SRS channel sounding method, and multicast to the AR-HUD and the traffic management platform at a rate of 10Gbps; the multi-modal fusion reinforcement learning algorithm maintains the hot standby state of the LiFi and UWB links through the dual-stack protocol stack IPv6 and TSN, and ensures 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 execute 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 body transformation (rotation matrix and translation vector). The principle is to solve the geometric transformation parameters through feature point matching (such as calibration board corner points), minimize the reprojection error, and ensure the millimeter-level spatial alignment accuracy. The cubic spline interpolation algorithm constructs a piecewise cubic polynomial function on the time axis for the sampling rate difference of asynchronous data streams (such as 1Hz for the meteorological sensor and 30Hz for the camera), and generates a continuous and smooth synchronous data sequence with the camera frame time as the interpolation node to eliminate the time discreteness of the low-frequency sensor. The dynamic time warping (DTW) detects the time series misalignment phenomenon (such as radar scan cycle jitter) by calculating the cumulative deviation of the minimum bending path between different data streams. When the deviation exceeds the preset threshold, the Kalman filter compensation is triggered. The target motion trajectory is predicted using the state space model (process model and observation model), and the covariance matrix is updated by combining the sensor observations to dynamically correct the time axis distortion of the data stream and reconstruct the time series consistency. The physical dimension normalization uses the Z-score standardization method to map different dimension data (such as speed m / s, pixel coordinates, humidity percentage) to the zero-mean unit-variance space, eliminate the input deviation influence of the dimension difference on the subsequent machine learning model, and improve the model generalization ability.

[0050] During implementation, at the hardware level to achieve this step, a dual-frequency RTK-GNSS module (U-blox ZED-F9P, positioning accuracy ±2 cm) is used to obtain the reference coordinates of street lights, and a laser rangefinder (VL53L5CX, ranging error ±1 mm) measures the installation offset of the radar and camera (Δx = 0.5 m, Δy = 0.2 m, Δz = 0.3 m). A multi-line radar (Velodyne VLP-16, horizontal resolution 0.1°) and a global shutter camera (Basler acA2440-75um, 30 fps) are used to collect point cloud and image data respectively. In software implementation, calibration plates are used to extract the feature points of the radar point cloud and the image (the checkerboard corner error < 0.1 pixel), and the Levenberg-Marquardt algorithm is used to solve the Lie group transformation parameters (rotation matrix R and translation vector t), optimizing the reprojection error to the sub-pixel level; for the 1 Hz data stream of the meteorological sensor, a 30 Hz synchronous sequence is generated by cubic spline interpolation, and the dynamic time warping window is set to 10 frames (about 333 ms). When the time offset of the radar-camera data stream ≥ 3 frames (100 ms), Kalman filter compensation is triggered, and the process noise covariance Q = diag(0.1, 0.1, 0.01) and the observation noise covariance R = diag(0.5, 0.5, 0.1) are configured. The output fusion dataset is stored in the Parquet columnar format (including WGS84 coordinates, speed, pixel coordinates, and normalized environmental parameters) for downstream module calls.

[0051] In the road physical model of step 2 of the above embodiment, vehicle dynamics usually use Lagrangian description to track the motion trajectory of a single vehicle, while road environment modeling is based on Eulerian description, that is, the fluid state change is described on a fixed spatial grid. To unify the two, the present invention adopts the 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] Among them, ρ is the local traffic flow density and v is the fluid velocity. By constructing a state variable transformation matrix, the coordinate mapping of the vehicle's speed, acceleration, and direction changes is carried out, and the state differentiation is performed in different coordinate systems, enabling the system to track individual behaviors locally while also calculating the evolution trend of the overall traffic flow. In addition, based on the nonlinear differential equations, the motion states of vehicles under road gradients, curve curvatures, and different friction coefficients are analyzed, a complete set of dynamic constraint equations is established, and the motion prediction model is corrected in combination with vehicle characteristics (such as mass, tire parameters, etc.) to make the calculation results more consistent with the real traffic environment. This method enables the individual motion trajectories of vehicles to be optimized and modeled in the global flow field and calculates the feasible vehicle motion range under road topology constraints.

[0057] In the modeling of road environment constraints, the Poisson-Poisson coupled fluid model is adopted. Based on traditional hydrodynamics, this model introduces the Poisson distribution characteristics of road traffic. By using a double Poisson distribution to model the vehicle flow changes within the road grid, the traffic fluid model can adapt to changing road conditions. Specifically, the first Poisson equation is used to describe the change in the traffic flow density of random traffic, and the second Poisson equation is used to calculate the velocity gradient of the local fluid to dynamically evaluate congestion, velocity changes, and driving safety distance constraints. During the solution process, the finite volume method is used to discretize the road grid, enabling the flow conservation constraint to maintain computational stability under any complex road topology, and the computational efficiency and accuracy of the fluid solution are optimized by using an adaptive grid refinement method to locally improve the computational accuracy in congested areas.

[0058] During the constraint solution process, to extract the core influencing factors of the road environment on vehicle motion, the sparse tensor decomposition method is used to perform dimensionality reduction analysis on the physical model. Road environment variables contain multiple high-dimensional features, such as gradient, friction coefficient, lane width, traffic flow density, etc. Direct modeling will lead to too high computational complexity. Therefore, the original high-dimensional data is decomposed into low-rank subspaces through 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, thereby screening out the key parameters that have a greater impact on abnormal driving behaviors, and further reducing the dimension through principal component analysis to generate a causal constraint matrix for road state. This matrix characterizes the influence degree of road environment factors on traffic states and is used to construct a causal relationship graph in the graph neural network to provide interpretable early warning event analysis.

[0059] In implementation, the road physical model runs on the NVIDIA Jetson AGX Xavier module (with 512 CUDA cores) of the edge computing cluster, using GPU to accelerate tensor decomposition (cuTENSOR library) and fluid model solution (OpenFOAM coupled solver). In the software implementation of the road physical model, the vehicle dynamics parameters (mass m = 1500 kg, rolling friction coefficient μ = 0.015) of the Lagrangian-Euler hybrid model are fused through Kalman filtering and sensor data (millimeter-wave radar speed measurement error ±0.1 m / s). The fluid viscosity coefficient of the Poisson-Poisson model is set to ν = 0.2 m 2 / s to match the characteristics of urban road traffic flow. Sparse tensor decomposition uses the alternating least squares method (ALS) for 10 iterations, the decomposition rank is set to 8, and the top 5% of significant factors are retained (such as the friction coefficient weight ratio is 32%). The normalized mutual information calculation is based on the historical accident dataset (100,000 samples), and the sliding window statistics the joint probability distribution of hard braking events (deceleration ≥ 3.5 m / s 2 ) and wet road surfaces, and filters out the parameters with NMI ≥ 0.3 (such as rainfall intensity - braking distance) to construct a 64×64 causal constraint matrix. The GNN of the dynamic causal graph adopts the GraphSAGE architecture, the node embedding dimension is 256, the number of message passing layers L = 3, and a loss function with physical constraints (mean square error + mutual information regularization term) is used during training. The learning rate is set to 1e-4, and after 50 rounds of iteration, the causal chain evidence is output (such as "rainfall 0.5 mm / min → friction coefficient drops by 40% → rear-end collision risk increases by 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 of road state, vehicle behavior, and environmental factors (road topology attributes, vehicle motion parameters, meteorological data), and extracts spatio-temporal correlation features through a node type-sensitive query-key value projection matrix, and the mathematical expression is where e i , e j are node embedding vectors, $W_h$ is the projection matrix for the $h$-th attention head, and $d$ is the embedding dimension. By aggregating the interaction weights of different relationship types (such as "vehicle-road friction constraint", "environment-vehicle visibility impact") through multi-head attention, a spatio-temporal correlation weight matrix is generated. The Bayesian structure learning algorithm is based on the Markov chain Monte Carlo sampling method to infer the latent causal graph structure from the correlation weight matrix, and optimizes the graph topology through the Bayesian posterior probability distribution $P(G|D)\propto P(D|G)P(G)$, where $P(D|G)$ is the data likelihood and $P(G)$ is the prior probability (such as sparsity constraint), dynamically adjusting the edge connection direction and strength to capture non-explicit causal relationships (such as the lag effect of humidity mutation on vehicle side slip). The backpropagation gradient constraint optimization calculates the gradient backpropagation through a differentiable loss function (such as causal effect mean square error) during the causal graph inference process, constraining the update direction of the attention weight matrix and screening high-contribution causal paths. The mathematical expression is where $W$ is the attention weight matrix and $\gamma$ is the sparsification coefficient, and low-correlation edges are removed through L1 regularization. The optimal transport mapping method evaluates the credibility of causal chain evidence based on the Wasserstein distance, performs optimal transport matching between the probability distribution of the causal path and the historical benchmark distribution, and calculates the transport cost $W(P,Q)=\inf$ γ∈(P,Q) $\int x(c,y)d\gamma(x,y)$, screening out low-confidence causal chains with transport costs exceeding the threshold (such as KL divergence difference $\geq1.5$), and retaining high-reliability evidence for early warning decisions.

[0061] During implementation, the graph neural network is deployed on the NVIDIA Jetson AGX Xavier platform of the edge computing node, using CUDA to accelerate attention calculation (cuBLAS library) and Bayesian inference (Pyro framework). In the software implementation, the heterogeneous graph node feature dimension is set to 64, and the vehicle node input includes speed (m / s), acceleration (m / s 2) Yaw angle (rad), road node input friction coefficient (μ), radius of curvature (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 64×64. The time sliding window covers 500 ms (5 frames @ 10 Hz). 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 α = 0.3, and causal edges with BF ≥ 3 are selected (such as "friction coefficient decreases → braking distance increases"). During backpropagation, the gradient mask retains the top 20% of the edges by weight. The L1 regularization coefficient λ = 0.01, and the Jacobian matrix calculates the partial derivatives of the vehicle motion equation with respect to the input parameters through automatic differentiation (PyTorch Autograd). The optimal transport map uses the Sinkhorn algorithm to iterate 50 times, with an entropy regularization coefficient ε = 0.1, and dynamically adjusts the Wasserstein distance threshold of the causal chain (W = 0.3). Finally, the causal chain evidence is output (such as "rainfall 0.8 mm / h → friction coefficient μ = 0.3 → braking distance extended by 3 m → rear-end collision risk level B"), formatted as a JSON-LD semantic description for visual decision-making by the traffic management platform.

[0062] When actually implementing step 2, a multi-line lidar (Velodyne VLP-16, horizontal angular resolution 0.1°) and a global shutter industrial camera (Basler acA2440-75um, 30 fps) 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 rangefinder (VL53L5CX) are fixed to the lamp pole base to achieve centimeter-level positioning and sensor calibration; an edge computing node (NVIDIA Jetson AGX Xavier) is deployed in the middle chassis of the lamp pole to run the physical model and the graph neural network algorithm; 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 operational effectiveness of step 2 (dynamic causal graph reasoning) in complex traffic scenarios, a virtual environment with multi-level risk events is constructed and the causal reasoning accuracy is quantified. The experiment generates traffic scenarios including road topology, vehicle dynamics, and environmental parameters based on the SUMO-CARLA joint simulation platform, and sets three types of progressive test scenarios:

[0064] The basic scenario simulates a single risk (such as sudden braking of the vehicle in front or illegal lane change), the composite scenario combines environmental and behavioral risks (such as slippery road surface + speeding through a curve), and the chain scenario designs a causal transmission chain (such as "rainfall → decreased visibility → misjudgment of the distance to the vehicle in front → sudden braking → rear-end collision of the following vehicle").

[0065] Drive 100 intelligent vehicles through Python scripts to trigger preset risk events randomly (50 times in total, including 20 basic events, 20 composite events, and 10 chain events). Synchronously simulate the camera (generate images with a resolution of 1280×720 using OpenCV, and detect vehicles / lane lines in real time with YOLOv5), millimeter-wave radar (output point clouds with an accuracy of 0.1m, and superimpose Gaussian noise with σ = 0.15), and meteorological sensors (dynamically adjust the intensity of rainfall / haze). The causal inference module in Step 2 accesses the sensor data stream in real time, and analyzes the risk conduction path by dynamically constructing a Bayesian causal network (the nodes include 20 variables such as vehicle speed, friction coefficient, visibility, etc., and the edge weights are trained by historical accident data). When a critical causal chain is detected (such as "rainfall intensity > 1.5mm / h → road surface friction coefficient < 0.4 → braking distance > 8m"), a warning is triggered. The experiment is repeated 3 times, and the warning accuracy rate (True Risk Detection Rate, TRDR), the integrity of the causal chain (the ratio of nodes / edges that match the preset causal path), and the average response delay (the time from the input of the first risk signal to the generation of the complete causal chain) are calculated through preset true value labels (risk event type, trigger time, causal chain structure). Finally, use t-SNE 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, accuracy rate = (number of correct warnings / total number of actual risk events) × 100%; in the composite scenario, 3 missed warnings are caused by misjudgment of the causal chain due to environmental noise (radar point cloud distortion); the 2 missed warnings in the chain scenario are due to the causal transmission level exceeding the preset network depth (maximum 4 layers). In terms of causal chain integrity, 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), and 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 warnings are completed within 700 ms, 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 graph reasoning module (step 2) verifies its effectiveness in multi-level risk scenarios in the simulation experiment, can balance reasoning accuracy and real-time performance, and provides interpretable causal support for the risk decision-making of the autonomous driving system.

[0070] In step 3 of the above embodiment, the improved Jaccard spatio-temporal similarity algorithm reconstructs the quantization method of event correlation by the traditional Jaccard coefficient by introducing a spatio-temporal decay weight function and a dynamic adaptive threshold mechanism, and solves the problem of misaggregation caused by a fixed threshold and a single spatial dimension. The spatio-temporal similarity weight function couples the influence of time difference and spatial distance in an exponential decay form, where the time decay coefficient (β = 0.1) controls the decay rate of the influence of the event time interval on the weight, and the spatial decay coefficient (γ = 0.05) adjusts the decay speed of the contribution of the spatial distance. The time decay term ensures that events occurring in a closer time have a higher correlation than events far from the time window, while the spatial decay term reduces the influence of distant events on the similarity calculation, making the calculation result more in line with the actual traffic flow law. 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; ensuring the dominant role of recent and adjacent events in the similarity calculation. The numerator part of the improved Jaccard coefficient only accumulates the weighted similarities of the same type of events, through the event type matching function (δ(T p ,T q)) Constraining the invalid superposition of heterogeneous events, the denominator eliminates the bias caused by the difference in the number of events through normalization. The dynamic threshold θ is adaptively adjusted based on the logarithmic function of the event density (Q). In high-density areas (such as congested sections), the threshold is automatically increased to reduce low-value aggregations, and in low-density areas, the threshold is decreased to improve sensitivity. This calculation method ensures that in high-density traffic areas, adjacent events are more likely to be determined as relevant, improving the calculation efficiency, while in low-density areas, the threshold is increased to reduce unnecessary computational burdens. If the spatio-temporal overlap degree exceeds the threshold, it triggers the construction of an edge computing cluster. The master node distributes event clusters to computing nodes through the distributed consistent hashing algorithm, combines trajectory prediction algorithms (such as the LSTM-Kalman fusion model) with dynamic traffic state regression analysis, models the vehicle movement trend and the probability of secondary accidents, and outputs the boundary parameters of the risk area (α-shape algorithm) and the traffic flow evolution gradient field (spatio-temporal convolutional network); if the overlap degree is insufficient, the feature vectors of the event clusters are compressed by 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 map.

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

[0072] Compared with the existing technology, this algorithm significantly improves the accuracy and scenario adaptability of event correlation detection through the collaborative design of spatio-temporal decay weights and dynamic thresholds; the edge-cloud collaborative architecture optimizes the allocation of computing resources, reduces redundant calculations while enhancing the global situation awareness ability; the combination of PCA dimensionality reduction and federated learning realizes continuous model optimization on the premise of ensuring data privacy, providing an 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 with the Jaccard algorithm, a simulation experiment was designed. The experiment was based on the CARLA autonomous driving simulation platform, and the hardware virtualization configuration was as follows:

[0074] Virtual sensor: CARLA’s built-in lidar (32 lines, 10Hz) and RGB camera (30fps, 1920×1080 resolution) simulate street light nodes to collect event data;

[0075] Edge computing simulation: Deploy the NVIDIA CUDA environment in a local Docker container (simulating the computing power of the Jetson AGXXavier) and run the improved Jaccard algorithm;

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

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

[0078] The experimental method is as follows: a two-way 6-lane urban expressway scenario (Town07) was constructed in CARLA, the traffic volume was set to 2000 vehicles / hour, and 5 types of events (sudden braking, skidding, reverse driving, congestion, and breakdown) were randomly injected. Each group of experiments was repeated 5 times, each for 10 minutes (simulation time). The road friction coefficient (0.3-0.8) and weather (alternating sunny / rainy days) were dynamically adjusted through the CARLA Python API to simulate a complex traffic environment. Group A ran the improved Jaccard algorithm, and Group B ran the traditional Jaccard algorithm; ZeroMQ was used to transmit event cluster data, and bandwidth and delay parameters were limited (5G: 100Mbps / 50ms, LiFi: 1Gbps / 10ms). The experimental data are shown in Table 2:

[0079] Table 2 Algorithm comparison experimental data record table

[0080]

[0081] As shown in Table 2, Group A (Improved Jaccard) showed significant advantages in the simulation: the average association accuracy was 94.2% (Group B 72.2%), and the false alarm rate was only 4.8% (Group B 25.9%), which verified the effectiveness of the spatiotemporal weight and dynamic threshold mechanism for cross-regional event association; the calculation delay was reduced by 38% (8.5ms vs. 13.4ms), and the network load was reduced by 35% (42.1MB vs. 65.8MB), indicating the advantage of the algorithm in reducing redundant calculations and communication overhead. The simulation results show that the improved algorithm has higher reliability and resource efficiency in complex traffic scenarios.

[0082] In addition, in the task execution between the edge computing cluster and the cloud in Step 3, the computing principle of the edge computing cluster for the boundary parameters of high-risk areas and the evolution trend of traffic flow is based on the coupling mechanism of multi-modal data fusion and dynamic spatio-temporal modeling. Its core technical features include: First, the PC5 interface of 5G-V2X adopts the Sidelink mode to bypass the base station to achieve end-to-end communication between street lamp nodes. The physical layer supports a 30kHz subcarrier spacing and dynamic TDD time slot allocation to ensure that the single-hop delay is less than 5ms, providing a synchronous spatio-temporal coordinate stream for the LSTM-Kalman filter fusion model. The LSTM network captures the temporal dependence of vehicle trajectories through a gating mechanism, and the Kalman filter corrects sensor noise based on kinematic equations (such as the acceleration covariance matrix Q = diag(0.1, 0.1)). The two jointly output the confidence interval of the vehicle position within the next 5 seconds (such as a 95% probability ellipse area). Second, the trajectory conflict detection algorithm introduces the Minkowski Sum theory, models the vehicle prediction trajectory as a time-varying polygon, and calculates the overlapping area of the trajectories in real time through the Separating Axis Theorem (SAT) to extract the core density distribution of the collision point set. Third, the α-shape algorithm performs Delaunay triangulation on discrete collision points by adjusting the radius parameter α (such as α = 8 meters), retains the triangle sides with an inscribed circle radius less than α, and generates a non-convex polygon risk boundary, which can accurately fit complex terrains such as road curves and intersections compared to the convex hull algorithm. Fourth, the Spatio-Temporal Convolutional Network (STCN) uses a three-dimensional dilated convolutional kernel (size 3×3×3, dilation rate 2), spans 10 frames of historical traffic flow tensors (including traffic volume, velocity field, event heat map) in the time dimension, and covers a street lamp sensing area with a radius of 50 meters in the spatial dimension, and extracts traffic flow mutation patterns (such as congestion propagation wave speed) through multi-scale feature fusion. Fifth, the Generalized Additive Model (GAM) takes vehicle density (vehicles / km), average speed (m / s), and event intensity (times / minute) as inputs of non-linear basis functions, fits interaction terms (such as the U-shaped curve of density-speed) using spline interpolation, outputs the Logit value of the secondary accident probability, and maps the probability field to the geographical space through the exponential decay function of the risk diffusion model (decay coefficient λ = 0.2 / s) to drive the dynamic adjustment of the warning radius (such as when the probability > 30%, the radius expands to 1.5 times the standard deviation). Sixth, the consistent hashing protocol adopts the virtual node doubling technology (each physical node maps 256 virtual nodes) to achieve O(1) time complexity data localization 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-based incremental model training lies in the global alignment of the feature space and privacy protection optimization under the federated learning framework. When the overlap degree of event clusters is lower than the threshold, the low-dimensional feature vectors generated by PCA dimensionality reduction (retaining the first 3 principal components, variance contribution rate ≥ 95%) are 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) during the gradient aggregation stage to prevent reverse derivation 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 the sparsely updated feature vectors, and uses the elastic weight consolidation (EWC) algorithm to constrain the update amplitude of important parameters (diagonal value of the Fisher information matrix > 0.1) to avoid catastrophic forgetting. The global traffic situation map model uses the graph attention network (GAT) as the backbone. The node embedding includes the regional risk probability and the direction of the traffic flow gradient field. The edge weight represents the cross-regional risk conduction intensity. The model parameters are iterated every 30 minutes and sent to the edge nodes through the parameter server to update the local inference engine.

[0084] In implementation, the hardware support deployment of this step uses the Huawei MH5000-871 5G-V2X module (PC5 interface, latency ≤ 20ms) to establish an edge cluster communication link. The edge nodes (NVIDIA Jetson AGX Xavier, 32GB memory) run the LSTM-Kalman filtering model (TensorFlow Lite framework, time step 10, prediction step 5). The α value of the α-shape algorithm is set to 0.5m. The input of the ST-CNN model is the traffic flow tensor of the past 30 minutes (5-minute slices, resolution 2m × 2m grid). Training parameters: learning rate 0.001, batch size 16; The GAM model uses the PyGAM library, and the basis function is a cubic spline (degree of freedom 5) to fit the quadratic accident probability weight; The decay coefficient k of the risk diffusion model is 0.1, and the initial radius λ0 is 50m. In the experimental scenario, when adjacent street lights detect a vehicle going in reverse (similarity J = 0.82 ≥ threshold 0.7), the master node collects trajectory data through 5G-V2X. The LSTM-Kalman prediction error is ±0.3m. After identifying the collision point set, the α-shape generates a non-convex boundary (number of vertices 15-20), and the ST-CNN outputs the divergence of the gradient field 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 < the threshold, PCA dimensionality reduction (retaining 95% variance) generates a 32-dimensional feature vector, which is uploaded to the cloud AWS IoT Core through the MQTT protocol (QoS = 1). The federated learning framework (TensorFlow Federated) uses an incremental Adam optimizer (learning rate 0.002, β1 = 0.9, β2 = 0.999) to update the global model, with end-to-end latency ≤ 200ms and single-node power consumption ≤ 30W.

[0085] Compared with traditional traffic monitoring methods, the present invention improves the accuracy of trajectory prediction through the fusion of LSTM-Kalman filtering, and realizes the modeling of high-risk areas that is more in line with the real road environment through α-shape topological reconstruction. By using a spatio-temporal convolutional network combined with a three-dimensional dilated convolutional layer, the calculation accuracy of the traffic flow evolution trend is improved, enabling the system to predict traffic congestion and potential risks earlier. Combining the generalized additive model and the risk diffusion model, the present invention can dynamically adjust the warning area radius, improving the pertinence and timeliness of the warning. In addition, the introduction of the consistent hashing protocol enables the edge computing cluster to perform efficient collaborative computing, avoiding waste of computing power. The use of federated learning and incremental Adam optimization achieves continuous optimization of the cloud, enabling the global traffic situation map to be adaptively adjusted, providing more accurate traffic prediction capabilities, and providing a more intelligent and efficient solution for smart city traffic management.

[0086] In step 4 of the above embodiments, the adaptive risk assessment model realizes dynamic risk quantification and classification decision-making through a multi-level cascade architecture. Its technical principles include: in the data layer, a long short-term memory network (LSTM) is used to perform temporal encoding (hidden state dimension 128) on the multi-modal fusion data set, capturing the time dependence of the event sequence. At the same time, a spatial histogram projection (grid resolution 0.5m×0.5m) is combined to construct a spatial distribution heat map of road events, generating a spatio-temporal feature vector (dimension 256); in the dynamic risk modeling layer, Gaussian process regression (GPR) is called to perform non-linear risk estimation on the spatio-temporal feature vector, using a radial basis kernel function (RBF) to fit the continuous distribution of risk values, and a conditional variational autoencoder (CVAE) is used to derive the probability density function of the event impact range in the latent space (dimension 32) (KL divergence loss weight 0.1), generating a dynamic risk function (output risk value interval [0,1]); in the adaptive threshold optimization layer, the risk level distribution of historical events is calculated based on kernel density estimation (KDE, bandwidth selection Silverman criterion), and the quantile regression (quantile τ = 0.75) is used to dynamically adjust the risk classification threshold boundary, realizing the non-linear contraction or expansion of the threshold adaptively with the event density; in the multi-level risk classification layer, hierarchical Bayesian clustering (HBC, Dirichlet process hyperparameter α = 1.0) is used to perform multi-granularity classification on early warning events, and the Markov random field (MRF, neighborhood system radius 10m) is combined to optimize the spatial consistency of classification labels, generating a risk classification matrix (dimension 4×4, corresponding to low / medium / high / extremely high risk levels); in the decision feedback adjustment layer, the classification model parameters are optimized by incremental Bayesian update (conjugate prior distribution is Beta(2,2)) combined with real-time accident data, and the exponentially weighted moving average (EWMA, smoothing coefficient λ = 0.3) is used to suppress short-term fluctuations in risk levels, ensuring decision-making stability. The fuzzy decision-making 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 map) based on a fuzzy decision tree (FDT, membership function is triangular, number of fuzzy rules 50), generating a set of disposal strategies (speed limit value ±20%, detour path topology optimization, emergency lane activation sign). Finally, according to the ISO 20078-3 standard, a TLV structure instruction stream (Type-Length-Value encoding) is encapsulated, and a digital signature (SHA-256 hash) and an X.509 certificate chain are added using the ECDSA algorithm (secp256k1 curve) to construct an early warning instruction packet resistant to replay attacks.

[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 is transmitted to the edge computing unit through the 5G-V2X network and is processed in real time on NVIDIA Jetson AGX Orin or other embedded computing devices.

[0088] In the data processing stage, the edge computing unit uses LSTM time series coding to extract the time dependencies of events and calculates the spatial distribution characteristics of road events through spatial histogram projection. Subsequently, the system calls Gaussian process regression to calculate the risk level of events and combines the CVAE model to deduce the impact range of events to construct a dynamic risk assessment function. The adjustment of the risk threshold is executed on the cloud server. The cloud uses the TensorFlow Extended (TFX) framework to manage the dynamic threshold update of 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 labels through MRF to ensure that the classification results conform to the road topology. The decision optimization part is executed by the fuzzy decision tree (FDT). Based on event characteristics and historical disposal plans, it calculates speed limit values, detour paths and emergency lane enabling strategies, encapsulates warning instructions in the ISO20078 TLV format, and calls the ECDSA encryption signature to ensure data security. Finally, the generated warning instructions are sent to in-vehicle terminals (AR-HUD) and traffic management platforms through the 5G-V2X network to provide intelligent traffic supervision services.

[0089] In implementation, the fuzzy decision-making method is used to handle the uncertainties of event location, risk level, and road network topology in traffic scenarios. Its core lies in constructing a membership function and a rule base through a Fuzzy Decision Tree (FDT), mapping discrete traffic parameters into a continuous possibility space to quantify the fitness of different disposal strategies. The fuzzy processing of event location uses a Gaussian membership function to convert longitude and latitude coordinates into membership probabilities for key areas (such as intersections and curves). For example, the membership degree of the intersection center point is 1.0, and it decays to 0 with an increase in the Euclidean distance according to the standard deviation σ = 15 meters. The risk level is divided into five levels (Level 1 to Level 5) through a triangular membership function, and each level covers an overlapping interval (such as Level 3 corresponding to risk values of 30 - 70, with a peak at 50) to ensure a smooth transition between adjacent levels. The topological features of the road network are extracted through graph embedding techniques (such as Node2Vec) to obtain topological attributes such as lane connectivity, number of lanes, and traffic flow direction, and are mapped into low-dimensional vectors. Then, the influence weight of these attributes on the strategy is quantified through a trapezoidal membership function (such as when the number of lanes > 4, the weight for enabling the emergency lane is increased to 0.8). The multi-objective optimization module uses the Pareto front search algorithm, with the weighted sum of membership degrees as the objective function, and generates a non-dominated solution set in combination with constraints (such as the length of the detour path not exceeding 1.5 times the original route). Finally, the optimal strategy is selected according to the principle of maximum satisfaction. The logical generation of the disposal strategy depends on a predefined rule base. For example, "intersection membership degree > 0.7 and risk level ≥ Level 4 → speed limit value = 30 km / h + turn on the emergency lane". The confidence of each rule is calibrated through training with historical accident data. The TLV (Type-Length-Value) encapsulation of the 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 encoding (16-bit bitmask, such as 0x0001 representing speed limit) according to priority, and ensures the integrity and anti-tampering of the instruction through the elliptic curve digital signature of the ECDSA algorithm (using the secp256k1 curve and SHA-256 hash). The signature process uses an ephemeral key pair to avoid private key leakage.

[0090] During implementation, in terms of the hardware of the fuzzy decision-making method, the intelligent street lamp is equipped with an NVIDIA Jetson Orin edge computing module (64GB memory, 2048 CUDA cores), integrated with a dual-frequency GNSS module (positioning accuracy ±0.1 m), a lidar (128 lines, horizontal field of view 360°), and a 5G-V2X communication unit (supporting the NR FR1 band). At the software level, the fuzzy decision tree is implemented based on the scikit-fuzzy library. The parameters of the event location membership function (σ = 15 m) are calibrated through the accident heat map at intersections, and the vertices of the risk level trigonometric function are dynamically adjusted by the historical data quantiles (25%, 50%, 75%). The road network topology embedding uses the Node2Vec algorithm (walk length 80, dimension 16), and the output vector is input into the trapezoidal membership function (threshold interval [0.3, 0.7]) after Min-Max normalization. The multi-objective optimization module calls the NSGA-II algorithm of the PlatypUS framework, with the population size set to 100, iterating 50 times, the crossover probability 0.8, and the mutation probability 0.1. The bit mask of the policy encoding is designed as follows: the lower 4 bits represent the speed limit value (e.g., 0b0011 corresponds to 30 km / h), the middle 4 bits identify the detour path number, and the upper 8 bits control the emergency lane and signal light status. When performing TLV encapsulation, the longitude and latitude are reserved to 6 decimal places (accuracy approximately 0.11 m), the risk level is mapped from Level 1 to Level 5 as 0x01 to 0x05, and the policy encoding is generated through bitwise operations. The ECDSA signature is implemented using the OpenSSL library, and the temporary key pair is destroyed after each signature. The signature result is appended to the end of the TLV (64-byte R+S value). The instruction is broadcast through the PC5 interface of 5G-V2X, and the receiving end verifies the signature through the pre-set CA certificate. If the verification fails, the retransmission mechanism is triggered (maximum retry 3 times). The edge node performs decision update every 5 seconds, and the policy effective delay is controlled within 200 ms.

[0091] Compared with the traditional threshold judgment method, it significantly improves the flexibility and scenario adaptability of the disposal strategy; the combination of TLV encapsulation and ECDSA signature ensures the standardization and security of the warning instruction, overcoming the defect of easy tampering in plaintext transmission; the introduction of dynamic membership parameters and topology embedding technology enables the road network characteristics to deeply participate in the decision-making process, providing the intelligent street lamp system with high-robustness and interpretable real-time decision-making capabilities.

[0092] To verify the effect of Step 4 in the intelligent supervision of smart street lights for traffic, the experiment uses SUMO (Simulation of Urban Mobility) + OMNeT++ for software simulation experiments to simulate the urban traffic environment, and runs the fuzzy decision-making model in the Python + PyTorch environment. The experiment compares the performance of Group A (using fuzzy decision tree + ISO 20078 standard + ECDSA encryption) and Group B (traditional fixed-rule-based decision-making method).

[0093] The experimental method is as follows: Load the road topology of the Third Ring Road in Beijing in SUMO, and set three traffic flow modes: high (3000 vehicles / h), medium (1800 vehicles / h), and low (1000 vehicles / h). Vehicles are generated according to Poisson distribution, and data on abnormal events (such as hard braking, speeding, and lane-changing conflicts) are collected. Each event records the timestamp, longitude and latitude coordinates, event type, and influence range, and is stored in the database. Group A uses a fuzzy decision tree + ISO 20078 TLV format encapsulation + ECDSA encryption, and Group B uses a fixed-rule-based early warning system (IF-THEN rule decision + plaintext JSON encapsulation). The early warning information is tested for latency and packet loss rate in a 5G-V2X environment through the SUMO-OMNeT++ simulated V2X communication network. The experimental data recording table is shown in Table 3:

[0094] Table 3 Experimental data recording table for Step 4

[0095]

[0096]

[0097] The experimental results show that for Group A (using fuzzy decision tree + ISO 20078 standard + ECDSA encryption), the calculation latency is reduced by about 50%, the misjudgment rate is reduced by about 70%, and the execution rate of early warning instructions is increased by about 10% - 15%, indicating that the fuzzy decision tree can provide more accurate event evaluation and reduce unnecessary early warning triggers. The data integrity of Group A is close to 100%, while Group B uses plaintext JSON encapsulation, which has a certain risk of data loss. In terms of network bandwidth occupancy, Group A reduces by about 50% compared to Group B, indicating that the data transmission in TLV format encapsulation is more efficient and reduces network load. Therefore, this method can improve the reliability of early warning instructions, reduce calculation latency, and optimize the response speed of the intelligent transportation system, enabling smart street lights to manage urban traffic more accurately and stably.

[0098] In step 5 of the above embodiment, the environmental perception module uses a light intensity sensor (such as a photodiode array) and a UWB channel probe (based on the IEEE 802.15.4z standard) to collect the light intensity (unit: Lux) and the UWB channel impulse response (CIR) in real time. The line-of-sight coverage cone angle of LiFi is determined by the spatial geometric relationship between the LED half-power angle (such as 60°) and the relative position of the terminal. If the terminal is within ±30° of the normal direction 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, and matches the channel state through adaptive bit loading (such as dynamic switching between QAM-64 / QPSK) to maximize the spectral efficiency. When it is detected that the multipath fading exponent (MPF) exceeds the threshold (such as MPF>2.0) or the terminal moving speed > 60 km / h, the reinforcement learning policy network (based on the DQN algorithm, the state space includes channel quality, terminal speed, and interference power) triggers a switch to the 5G-UWB link, and calls the dynamic TDD frame structure of the 3GPP NR-U protocol (frame length 1 ms, uplink and downlink time slot ratio 1:3), and avoids adjacent cell interference through flexible time slot allocation (such as reserving 20% guard interval per frame). The Polar code (code length 1024, code rate 0.5) is used as the channel coding scheme, which utilizes the channel polarization effect to improve the decoding success rate in a high-noise environment, and combines the wideband scanning of the SRS (sounding reference signal) (scanning interval 5 ms) to achieve millimeter wave beamforming (beam width 5°), and aligns the terminal receiving direction through the singular value decomposition (SVD) algorithm. The dual-stack protocol stack maintains seamless hot standby of the LiFi and UWB links through IPv6 over TSN (time-sensitive network). The 802.1Qbv scheduler of TSN synchronizes the dual-link clocks with microsecond-level accuracy to ensure service continuity during handover; forward error correction (FEC) uses the Reed-Solomon code (RS(255,223)) and the Type-II incremental redundancy mechanism of hybrid automatic repeat request (HARQ) to work together, and dynamically adjusts the number of redundant packets according to the real-time packet loss rate (PER) to ensure that the end-to-end transmission bit error rate (BER) is lower than 1e-6.

[0099] When implementing this step, in terms of hardware, the intelligent street lamp integrates a LiFi transmitting module (based on Osram SFH 4715S LED, wavelength 850 nm, modulation bandwidth 120 MHz), a UWB transceiver (Decawave DW3000, frequency band 6.5 GHz, bandwidth 500 MHz), an environmental perception unit (AMS TSL2591 optical sensor + Infineon DEMO BGT60LTR11AI radar), and a 5G-UWB baseband processor (Qualcomm QCA6391). In terms of software implementation, the reinforcement learning policy 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.), and the output action space is 2-dimensional (LiFi / UWB selection). In the training phase, a deep Q network (γ = 0.95, ε-greedy exploration rate 0.1) is used. OFDM modulation in the LiFi link is implemented through an FPGA (Xilinx Zynq UltraScale+), with a subcarrier spacing of 2 MHz and a cyclic prefix length of 1 / 4 symbol period; the 5G-UWB dynamic TDD frame configuration is 0.7 ms for the downlink slot and 0.3 ms for the uplink, the SRS detection period is 5 ms, and the beam codebook contains 64 predefined directions. The dual protocol stack achieves link redundancy through the IPv6 dual-stack module of the Linux kernel and the TSN shaper (gPTP clock synchronization accuracy ±100 ns). 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 reaches 3 Gbps (BER = 5e-7) in static scenarios (vehicle speed < 30 km / h), and the UWB maintains 1.5 Gbps (PER < 0.1%) in mobile scenarios (vehicle speed > 60 km / h), with a handover delay of less than 5 ms.

[0100] When implemented, the proposed solution significantly improves the transmission robustness and real-time performance of warning instructions in complex traffic scenarios through a multi-modal communication dynamic selection mechanism driven by reinforcement learning; the synergy and complementarity of LiFi and UWB overcome the coverage blind spots and mobility limitations of a single link; the dual-stack protocol and hybrid error correction mechanism ensure the continuity and integrity of data transmission, constructing a highly reliable and adaptive communication infrastructure for the intelligent street lamp system and effectively supporting the requirements of vehicle-road collaboration and real-time traffic control.

[0101] In the implementation of the traffic supervision system applied to intelligent street lamps, it includes a spatio-temporal synchronization module, a cross-modal causal inference module, a distributed collaborative computing optimization module, a structured warning coding module, and a multi-modal interaction module;

[0102] In implementation, the system hardware uses smart streetlights as the physical carrier. At the top of each streetlight, a multi-sensor module (camera, millimeter-wave radar, lidar, 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) are integrated. The sensor module is connected to the edge node through the RS-485 bus, and the power supply adopts a complementary scheme of POE (Power over Ethernet) and solar energy. The POE switch (Cisco CBS110-8PP) provides 48V direct current for high-power-consuming devices such as cameras and radars, and low-power-consuming devices such as meteorological sensors are jointly powered by a solar panel (EcoFlow 400W) and a lithium battery (Tesla Powerwall 2). The edge node is connected to a GNSS positioning module (u-blox ZED-F9P) and an atomic clock (Symmetricom SA.45s) through a PCIe interface. The 10MHz reference clock signals of the GNSS antenna (Taoglas MA240) and the atomic clock are connected to the edge node synchronization board through coaxial cables to ensure the unification of the spatio-temporal reference. The LiFi transmitter (VLNComm LumiNex) and the UWB radio frequency unit (Decawave DW3000) communicate with the edge node through the MIPI interface. The LED array (Cree XLamp XP-E2) of LiFi is installed at the top of the lamp post and tilted downward by 30°, with a coverage radius of 50 meters. The UWB phased array antenna (8×8 units) 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 10G optical fibers.

[0103] During implementation, the working logic of the system is as Figure 6As shown in the figure, first, the camera (Sony IMX585) captures the video stream at 30fps. After being encoded by H.265, it is transmitted to the edge node through the GigE interface; the millimeter-wave radar (TI AWR1843) outputs point cloud data (10Hz, JSON format) and sends it through the CAN bus; the 16-line scan data (20Hz) of the lidar (Velodyne VLP-16) is transmitted through the Ethernet. The spatio-temporal synchronization module first preprocesses the original data: the 10MHz clock signal generated by the atomic clock drives the hardware of each sensor to synchronously sample. The PPP (Precise Point Positioning) coordinates output by the GNSS module are fused with the lidar data through the SE(3) Lie group transformation algorithm, mapping the radar point cloud coordinate system (polar coordinates) to the camera imaging plane (pixel coordinates), and the spatial alignment residual ≤ 0.1 pixel; the asynchronous data stream (such as the 1Hz sampling of the meteorological sensor) is resampled to 100Hz through cubic spline interpolation, and the dynamic time warping (DTW) algorithm is used to detect the timing deviation. When it exceeds 5ms, the Kalman filter compensation is triggered.

[0104] The preprocessed multi-modal data (timestamp, spatial coordinates, physically dimension-normalized features) 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 relationships). The node feature dimension is 128, and the edge weights are calculated through the multi-head attention mechanism (the number of heads is 8, and the temperature coefficient τ = 0.5). The road physical models (such as the Pacejka tire equation, Navier-Stokes traffic flow model) are encoded as the physical regularization terms of the GNN, constraining the vehicle acceleration prediction error ≤ 0.2m / s 2 . The causal inference results (such as "road surface humidity↑ → friction coefficient↓ → braking distance↑") screen high-confidence paths (confidence ≥ 0.7) through Bayesian structure learning, and output warning events with evidence chains (JSON format, including event type, location, risk level).

[0105] The warning events enter the distributed collaborative computing optimization module: the improved Jaccard algorithm is used to calculate the spatio-temporal similarity of adjacent street lamp event clusters (threshold θ = 0.7 + 0.1·log(1 + Q)). If the similarity meets the standard, it triggers the 5G-V2X networking (PC5 interface, latency ≤ 10ms) to build an edge computing cluster. The master node (NVIDIA Jetson AGX Xavier) distributes computing tasks through the consistent hashing algorithm (the number of virtual nodes is 1024), and the slave nodes only provide the original data or intermediate features. The master node runs the LSTM-Kalman filter model (the hidden layer is 64-dimensional, and the prediction step is 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 spatio-temporal convolutional network (ST-ConvNet, three-dimensional dilated convolutional 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 through PCA dimensionality reduction (retaining 95% of the variance), and uploaded to the cloud federated learning platform via the MQTT protocol (QoS = 1, topic "edge / alert"). The incremental Adam optimizer (learning rate 0.001, β1 = 0.9) updates the global traffic situation map model parameters ΔΘ, which are then sent to the edge nodes for synchronous update.

[0106] The structured warning encoding module receives warning events, extracts structured elements (longitude, latitude, risk level, policy code), and encapsulates them into a binary stream in TLV format 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 occupy 8 bytes, and the risk level is 1 byte), and the Value field is filled dynamically. The encoded instruction stream is appended with a digital signature (64 bytes) through the ECDSA algorithm (elliptic curve secp256k1, SHA-256 hash) and embedded with an X.509 certificate chain (root certificate, intermediate CA) to form a tamper-resistant instruction packet.

[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 the LiFi or UWB channel. If the light intensity ≥ 2000 Lux and the terminal is within the LiFi coverage cone angle (±60°), LiFi OFDM modulation is initiated (256 subcarriers, modulation depth 0.8), and the optical signal is transmitted directionally to the in-vehicle terminal (Hamamatsu S1223 photodiode) and AR-HUD (HoloLens 2) at a rate of 10 Gbps; if multipath fading ≥ 3 dB or vehicle speed ≥ 80 km / h is detected, the system switches to 5G-UWB (3GPP NR-U standard), and uses Polar codes (code length 512) and a dynamic TDD frame structure (DL:UL = 4:1) to multicast instructions in the 3.5 GHz band. The dual-stack protocol (IPv6 / TSN) maintains link hot standby, and forward error correction (FEC, redundancy 20%) and hybrid automatic repeat request (HARQ, maximum retransmission 3 times) ensure transmission integrity, and the channel with the lowest packet loss rate is dynamically selected as the main link (switching delay ≤ 10 ms).

[0108] The logical processing of the system presents hierarchical progression and closed-loop feedback characteristics: the spatio-temporal synchronization module is the underlying data entry point to ensure the spatio-temporal consistency of multi-source data; the cross-modal causal reasoning module realizes the root cause tracing and interpretable early warning of events through the integration of physical constraints and graph neural networks; the distributed collaboration module dynamically allocates computing power based on spatio-temporal correlation to optimize the balance between local response and global learning; the structured encoding module converts the decision results into machine-parsable instructions to ensure cross-platform compatibility; the multi-modal interaction module ensures the highly reliable transmission of instructions through adaptive channel selection and error correction mechanisms. Loose-coupling interactions are achieved between modules through standardized interfaces (such as ROS topics, MQTT messages), and at the same time, a closed-loop optimization system is formed relying on the two-way data synchronization between the edge and the cloud (model parameters ΔΘ, risk map P_risk).

[0109] Although the specific implementation manners of the present invention have been described above, those skilled in the art should understand that these specific implementation manners are only illustrative. Without departing from the principles and essence of the present invention, those skilled in the art can make various omissions, substitutions, and changes to the details of the above methods and systems. For example, combining the above method steps so as to perform substantially the same function in a substantially the same manner to achieve substantially the same result belongs to the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. A traffic intelligent supervision method applied to smart street lights; characterized in that: Including: Step 1: Based on the raw data of multiple sensors, unify the spatio-temporal reference and perform spatial mapping association through an atomic clock and a GNSS positioning module to generate a multi-modal fusion data set; Step 2: Invoke the road physical model to constrain the multi-modal fusion data set, and use a graph neural network to construct a dynamic causal graph and perform backpropagation inference to output warning events with causal chain evidence; Step 3: According to the warning events, use an improved Jaccard spatio-temporal similarity algorithm 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, construct an edge computing cluster through 5G-V2X networking to calculate the boundary parameters of high-risk areas and the evolution trend of traffic flow; If the overlap degree is lower than the preset threshold, upload the compressed feature vectors of each node to the cloud to perform incremental model training, and output an optimized global traffic situation map and model iteration instructions; Step 4: Extract structured elements from the warning events, including event location, risk level, and recommended actions, encapsulate the binary instruction stream according to the ISO 20078 standard and append a digital signature to generate a warning instruction; Step 5: Based on the environmental state, dynamically select the LiFi or 5G-UWB communication medium through a multi-modal fusion reinforcement learning algorithm, and synchronously convey the warning instruction to in-vehicle terminals, AR-HUDs, and traffic management platforms.

2. The traffic intelligent supervision method applied to intelligent street lamps according to claim 1, wherein: The spatio-temporal reference unification and spatial mapping association method in Step 1 is as follows: Based on the GNSS real-time positioning coordinates and the measurement values of a laser rangefinder, construct a sensor spatial mapping matrix through the Lie group space transformation algorithm to align the data of the radar point cloud coordinate system and the camera imaging plane; For the sensor data streams that arrive asynchronously, use the cubic spline interpolation algorithm to resample them under the spatio-temporal reference to eliminate time offsets, and detect the time axis distortion of the multi-modal data streams through the dynamic time warping algorithm. If the maximum path deviation exceeds the preset threshold, use the Kalman filter compensation mechanism to reconstruct the temporal consistency; finally, output a multi-modal fusion data set containing the sensor feature vectors after timestamp, spatial coordinates, and physical dimension normalization.

3. The traffic intelligent supervision method applied to smart street lights according to claim 1, characterized in that: The road physical model uses the Lagrangian-Eulerian mixed coordinate transformation method to analyze vehicle dynamics parameters, and combines the Poisson-Poisson coupled fluid model to calculate road environment constraints, so that the motion state and the road topology form a dynamic constraint relationship; during the constraint solving process, the road physical model uses the sparse tensor decomposition method to extract the key influencing factors in the physical model, and screens the abnormal driving behavior influencing parameters through the normalized mutual information calculation method to generate a road state causal constraint matrix.

4. The traffic intelligent supervision method applied to smart street lights according to claim 1, wherein: Based on the road state causal constraint matrix, the graph neural network uses the heterogeneous graph attention mechanism to calculate the spatio-temporal correlation weights of road states, vehicle behaviors, and environmental factors, and combines the Bayesian structure learning algorithm to perform causal relationship reasoning to generate a dynamic causal graph; during the causal reasoning process, the graph neural network invokes the backpropagation gradient constraint to optimize and correct the causal relationship weights, and combines the optimal transport mapping method to evaluate the credibility of the causal chain evidence and screen out low-confidence causal paths.

5. A traffic intelligent supervision method applied to smart street lights according to claim 1, characterized in that: The working method of the improved Jaccard spatio-temporal similarity algorithm is as follows: First, extract the event cluster set C of adjacent street lamp nodes by sliding the time window i =(e1, e2,... e n ), where each event e k contains a timestamp t k , spatial coordinates (x k , y k ) and an event type code T k ; Next, define the spatio-temporal similarity weight function where t diff is the event time difference, d is the spatial Euclidean distance, where d = 0.6; β = 0.1 and γ = 0.05 are the time decay coefficient and the spatial decay coefficient respectively; a is the balance factor of the time and space weights; Based on the spatio-temporal similarity weight function w(t, d), the Jaccard coefficient is improved 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 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 the deviation caused by the difference in the number of events; if J st ≥θ, trigger the establishment of an edge computing cluster for the PC5 interface of 5G-V2X, and allocate the event cluster to the main computing node through the distributed consistent hashing algorithm; where θ is the dynamic threshold, and the calculation formula is: θ=(0.7+0.1log(1+Q) (2) In formula (2), Q is the current regional event density. Next, the master node calls a trajectory prediction algorithm to model the movement trend of vehicles in the abnormal event impact area, and combines dynamic traffic state regression analysis to calculate the probability of secondary accidents, calculate the boundary parameters of the risk area and the traffic flow evolution gradient field; if J st < θ, then perform PCA dimensionality reduction on each node event cluster to generate compressed feature vectors, upload them to the cloud federated learning framework through the MQTT protocol, update the global traffic situation map model parameters based on the incremental Adam optimizer, and output model iteration instructions and risk probability distribution matrices.

6. The traffic intelligent supervision method applied to smart street lamps according to claim 1, characterized in that: The computing principle for the edge computing cluster to calculate the boundary parameters of high-risk areas and the evolution trend of traffic flow is as follows: First, a low-latency communication link is established through the PC5 interface of 5G-V2X. The master node collects the spatio-temporal coordinate data of adjacent street lamp event clusters, inputs it 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 α-shape algorithm is used to perform topological reconstruction on the collision point set to generate the boundary parameters of the non-convex polygon risk area. At the same time, a spatio-temporal convolutional network is constructed, and the spatio-temporal correlation pattern between the historical traffic flow tensor and real-time event features is extracted through the three-dimensional dilated convolutional layer, and the vector field distribution of the traffic flow evolution gradient field is output. Combining with the generalized additive model, based on the vehicle density, average speed, and event intensity characteristics, a non-linear 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 the warning parameters of each node are synchronously updated through the consistent hashing protocol within the edge cluster to achieve the distributed collaborative calculation of the regional risk boundary and traffic flow trend.

7. A traffic intelligent supervision method applied to intelligent street lights according to claim 1, characterized in that: Step 4 uses an adaptive risk assessment model to perform risk grading calculations, combines with the historical event database to dynamically adjust the risk classification threshold, and calculates the optimal disposal strategy through the fuzzy decision-making method. 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 the data of the multi-modal fusion data set, extract the time dependence using the time series encoding of the long short-term memory network, and combine the spatial histogram projection to construct the spatial distribution mapping of road events to generate spatio-temporal feature vectors. The dynamic risk modeling layer is used to call the Gaussian process regression method to perform continuous risk estimation on the spatio-temporal feature vectors, and use the conditional variational autoencoder to deduce 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 grading threshold to generate real-time updated risk boundary parameters. The multi-level risk classification layer is used to classify warning events using hierarchical Bayesian clustering, and optimize the classification labels in combination with the Markov random field to generate a risk classification matrix. The decision feedback adjustment layer optimizes the classification model by combining the latest accident data through incremental Bayesian update, and uses the exponentially weighted moving average to smooth the change of the risk level to adjust the final risk classification decision.

8. The traffic intelligent supervision method applied to smart street lights according to claim 7, characterized in that: The fuzzy decision-making method performs multi-objective optimization on the event location, risk level, and road network topology based on the fuzzy decision tree to generate a set of disposal strategies, including speed limit values, detour routes, and emergency lane activation. According to the ISO 20078 standard, the longitude, latitude, risk level, and strategy encoding are encapsulated into a binary stream in TLV format, and a digital signature is attached using the ECDSA algorithm to generate a tamper-proof warning instruction.

9. The traffic intelligent supervision method applied to smart street lamps according to claim 1, characterized in that: The working principle of the multi-modal fusion reinforcement learning algorithm is as follows: Based on the light intensity L collected in real time by the environmental perception module lux and the UWB channel state information, a dynamic selection strategy is constructed. If L lux ≥ 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 instructions to the vehicle-mounted terminal and the AR-HUD in the orthogonal frequency division multiplexing modulation mode; If the multipath fading exponent F MP ≥ 3 dB or the terminal moving speed v ≥ 80 km / h, switch to 5G-UWB, call the 3GPP NR-U MAC layer protocol, allocate time slot resources based on the dynamic TDD frame structure, enhance the transmission reliability through Polar codes, and implement beamforming in combination with the SRS channel sounding method, and 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 hot standby state of the LiFi and UWB links through the dual-stack protocol stack IPv6 and TSN, and ensures 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 execute the main data stream transmission.

10. A traffic supervision system applied to smart street lights, characterized in that: Applied to a traffic intelligent supervision method for smart street lights described in any one of claims 1-9, the system includes: a spatio-temporal synchronization module, a cross-modal causal reasoning module, a distributed collaborative computing optimization module, a structured warning coding module, and a multi-modal interaction module; The spatio-temporal synchronization module is used to unify the spatio-temporal benchmark and perform fusion preprocessing on multi-sensor data at the street light end; The cross-modal causal reasoning module is used to perform causal reasoning on multi-modal data by combining the road physical model and the graph neural network, and output interpretable warning events; The distributed collaborative computing optimization module is used to dynamically allocate the computing tasks of the smart street light edge nodes and the cloud based on the spatio-temporal correlation of events; The structured warning coding module is used to encode warning events into structured instructions that support machine parsing; The multi-modal interaction module is used to transmit structured warning instructions through an adaptive channel.

Images