A lane guidance method and system based on cloud edge-end collaborative team fog monitoring

Abstract

This invention relates to the field of vehicle guidance technology, and more particularly to a lane guidance method and system based on cloud-edge-device collaborative fog monitoring. It acquires laser visibility, millimeter-wave radar, video, meteorological, and traffic flow data through edge sensors, and combines this with cloud data. Temporal alignment and spatial registration are performed at the edge to obtain a spatiotemporal dataset. Multi-source features are extracted in the cloud and subjected to temporal segmentation and fusion encoding to obtain fog spatial evolution data, enabling prediction of fog diffusion and migration processes. Furthermore, lane risk maps are constructed by combining lane topology, traffic flow, and historical accident data. Based on risk ranking and regional constraints, differentiated lane guidance control commands are generated through reinforcement learning. By collecting vehicle operation feedback data after guidance and calculating guidance deviation, the fog spatial evolution data and guidance strategy are corrected to form a closed-loop control, improving the accuracy, real-time performance, and adaptability of lane guidance in fog scenarios.

Description

Technical Field

[0001] This invention relates to the field of vehicle guidance technology, and more particularly to a lane guidance method and system based on cloud-edge-device collaborative fog monitoring. Background Technology

[0002] Dense fog is a localized, sudden, spatially uneven, and rapidly evolving low-visibility weather phenomenon that often occurs on highways. It is characterized by its sudden onset, irregular distribution, and unstable duration. Because visibility drops drastically in dense fog, drivers struggle to obtain timely traffic information, increasing the risk of rear-end collisions, chain-reaction collisions, and other traffic accidents, thus threatening road safety.

[0003] In existing technologies, the monitoring and lane guidance of fog patches rely on sensor threshold monitoring. This includes acquiring road visibility values ​​using laser visibility detection equipment and triggering warnings based on preset thresholds, or identifying low-visibility areas through video surveillance and issuing speed limits or warning messages. These methods typically rely on observations from a single time segment or simply overlay multi-source data, lacking a unified modeling of the temporal correlations and spatial distribution characteristics between different data points. This results in the system only responding after fog patches have formed or significantly worsened, failing to effectively identify the early stages of fog diffusion. Regarding lane guidance, existing technologies issue uniform speed limits or warning messages based on the overall road condition, lacking assessment of the varying degrees of fog impact on different lanes. This leads to coarse-grained guidance strategies, failing to achieve targeted control and resulting in low traffic utilization. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a lane guidance method and system based on cloud-edge-device collaborative fog monitoring.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A lane guidance method based on cloud-edge-device collaborative fog monitoring includes the following steps: S1. Acquire laser visibility data, millimeter-wave radar data, video data, meteorological data and traffic flow data on the side of the highway through edge sensors, combine them with cloud-based weather forecast data and road label data, and perform time alignment and spatial registration processing through edge computing nodes to obtain a spatiotemporal dataset; S2. Extract visibility change features, video texture attenuation features, millimeter wave echo attenuation features, and temperature and humidity wind field change features in the cloud based on the spatiotemporal dataset, and perform temporal segmentation and fusion encoding to obtain the spatial evolution data of fog. S3. Based on the spatial evolution data of the fog, combined with road lane topology data, road alignment data, traffic flow status data and historical accident data, risk mapping is performed on each lane to obtain a lane risk map. S4. Perform risk ranking and guidance parameter solving based on the lane risk map to obtain the guidance parameter set; construct a regional guidance constraint set based on the lane risk map splicing results of adjacent road segments and the ramp diversion capacity; perform reinforcement learning decision-making based on the guidance parameter set and the regional guidance constraint set to generate lane guidance control instructions and target operation response parameters; S5. Control the roadside guidance equipment and vehicle network publishing terminal to perform lane guidance according to the lane guidance control command, and collect vehicle operation feedback data after guidance; calculate the guidance deviation result based on the vehicle operation feedback data and the target operation response parameters; S6. Correct the fog spatial evolution data according to the induced deviation result, and update the lane guidance control command according to the corrected fog spatial evolution data.

[0006] Further, S1 includes the following steps: Laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data are collected from the roadside of highways using laser visibility detection equipment, millimeter-wave radar, image acquisition equipment, meteorological API, and traffic platforms, and a unified timestamp is added. Time alignment processing is performed on various types of data based on the timestamp identifier to obtain multi-source time-series data arranged according to a uniform sampling interval; Spatial mapping is performed on the multi-source time-series data based on the road label data, and associated with the corresponding road location and lane unit to obtain spatial association data; Data cleaning and normalization are performed on the spatially correlated data to obtain a spatiotemporal dataset.

[0007] Further, S2 includes the following steps: Based on the spatiotemporal dataset, laser visibility data, video data, millimeter-wave radar data, and meteorological data are segmented according to time windows to obtain multi-source data segments; A multi-channel input feature tensor is constructed from the video data and millimeter-wave radar data in the multi-source data segment. The input feature tensor is then input into a first convolutional neural network to perform convolution operations and nonlinear activation, extracting video texture attenuation features and echo attenuation features to obtain a spatial feature vector. The laser visibility data and meteorological data in the multi-source data segment are processed by a second convolutional neural network to extract sequence features, obtain visibility change features and temperature, humidity and wind field change features, and then perform feature concatenation with the spatial feature vector to obtain a fused feature vector. Encoding mapping is performed based on the fused feature vector to obtain the spatial evolution data of the fog.

[0008] Further, S3 includes the following steps: Based on the spatial evolution data of the fog, the boundary position of the fog is divided into intervals along the road mileage direction, and the coverage of each interval within a preset time window is calculated according to the diffusion direction and migration speed. The coverage is then mapped to the corresponding lane unit to obtain the lane fog influence range data. Based on the lane fog impact range data, the current vehicle speed and headway of each lane, the difference between the visible distance and braking distance of the vehicle within the fog impact range is calculated to obtain the visibility safety margin. The rate of change of speed difference between adjacent vehicles is calculated based on the visible safety margin and the vehicle speed distribution gradient, and the rear-end collision risk weight is calculated based on the rate of change of speed difference. Based on the lane fog impact range data and the traffic flow difference between adjacent lanes, the lane change demand intensity per unit time is obtained by weighting, and the lane change conflict probability is calculated based on the lane change demand intensity and lane density. At the same time, the traffic stability is calculated based on the vehicle speed fluctuation variance. The visible safety margin, rear-end collision risk weight, lane change conflict probability, and traffic stability are normalized and weighted to generate a lane risk map.

[0009] Further, S4 includes the following steps: Based on the visible safety margin, rear-end collision risk weight, lane change conflict probability and traffic stability of each lane in the lane risk map, a lane risk state vector is constructed. Combined with the lane risk map splicing results of adjacent road segments and the ramp diversion capacity, a regional collaborative constraint state vector is constructed. Based on the lane risk state vector and the regional collaborative constraint state vector, a reinforcement learning state space is constructed, and the speed limit value, lane opening status and guidance release intensity in the guidance parameter set are used as action sets to construct a reinforcement learning action space. Based on the state space and action space, the reward value corresponding to each action is calculated. The reward value is calculated by weighting the increase in average vehicle speed, the change in queue length, the reduction in abnormal braking, and the change in traffic stability. The strategy for each action is evaluated and updated based on the reward value, and the optimal action is selected as the lane guidance control command. Target running response parameters are then generated based on the optimal action.

[0010] Furthermore, the formula for updating the strategy is as follows: ; in, In the current state Next action The strategy evaluation value; The learning rate; This is the reward value; For discount factor; Execute action The new state afterwards; For the new state Candidate actions from the set of available actions; In the new state The maximum policy evaluation value is obtained by comparing the policy evaluation values ​​corresponding to each candidate action.

[0011] Furthermore, the reward value is calculated using the following reward function: ; in, This is the reward value; The increase in the vehicle's average speed relative to its initial speed before induction; This represents the change in queue length. This represents the change in abnormal braking events; This represents the change in traffic stability. , , and These are the weighting coefficients.

[0012] Furthermore, the vehicle operation feedback data is obtained through the following steps: The lane guidance control command controls the roadside guidance equipment and the vehicle network publishing terminal to perform lane guidance, and collects vehicle operation data after guidance based on traffic flow detection equipment and vehicle terminal; Based on the vehicle operation data, time window statistical processing is performed on vehicle speed, deceleration, headway, and lane change behavior to obtain statistical feature data. Based on the statistical characteristic data, the average vehicle speed, queue length, frequency of abnormal braking events, and degree of vehicle speed fluctuation are calculated to obtain vehicle operation feedback data.

[0013] Further, S6 includes the following steps: Based on the induced bias results, an observation set including vehicle average speed deviation, queue length deviation, abnormal braking event frequency deviation, and vehicle speed fluctuation deviation is constructed. Based on the spatial evolution data of the fog, diffusion range parameters, migration speed parameters, and concentration distribution parameters are extracted to form a set of parameters to be estimated. The time recursive prediction calculation is performed on the set of parameters to be estimated based on the Kalman filter method, and the observation residual is calculated based on the difference between the observation and the prediction result. The Kalman gain is calculated based on the observed residuals, and the set of parameters to be estimated is corrected and updated based on the Kalman gain to obtain the corrected fog spatial evolution data. The lane risk map and lane guidance control instructions are then updated based on the corrected fog spatial evolution data.

[0014] A lane guidance system based on cloud-edge-device collaborative fog monitoring, applied to any of the aforementioned lane guidance methods based on cloud-edge-device collaborative fog monitoring, includes: The end-side sensor module is used to acquire laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data from the roadside of the highway. Edge computing nodes are used to perform time alignment and spatial registration processing on laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data from the roadside of highways to obtain spatiotemporal datasets; The cloud computing module is used to perform the following steps: Based on the spatiotemporal dataset, visibility variation features, video texture attenuation features, millimeter wave echo attenuation features, and temperature and humidity wind field variation features are extracted, and temporal segmentation and fusion encoding are performed to obtain the spatial evolution data of fog. Based on the spatial evolution data of the fog, combined with road lane topology data, road alignment data, traffic flow status data and historical accident data, risk mapping is performed on each lane to obtain a lane risk map; Based on the lane risk map, risk ranking and guidance parameter solving are performed to obtain the guidance parameter set; based on the lane risk map splicing results of adjacent road segments and the ramp diversion capacity, a regional guidance constraint set is constructed; based on the guidance parameter set and the regional guidance constraint set, reinforcement learning decision-making is performed to generate lane guidance control instructions and target operation response parameters; The lane guidance control command controls the roadside guidance equipment and the vehicle-to-everything (V2X) terminal to perform lane guidance, and collects vehicle operation feedback data after guidance; the guidance deviation result is calculated based on the vehicle operation feedback data and the target operation response parameters. The fog spatial evolution data is corrected based on the induced deviation results, and the lane guidance control command is updated based on the corrected fog spatial evolution data.

[0015] The beneficial effects of this invention are as follows: This invention acquires laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data through edge sensors, and combines them with cloud-based weather forecast data and road tag data. Temporal alignment and spatial registration are performed at the edge computing node to obtain a unified spatiotemporal dataset. Furthermore, visibility change features, video texture attenuation features, millimeter-wave echo attenuation features, and temperature, humidity, and wind field change features are extracted from the spatiotemporal dataset in the cloud. Temporal segmentation and fusion encoding are then used to obtain spatial evolution data of fog patches. This transforms the original fog patch judgment method based on a single time section and static threshold into a continuous characterization of the fog patch diffusion process, migration trend, and concentration changes. This solves the problem of existing technologies lacking unified modeling of the temporal correlation and spatial distribution characteristics of multi-source data, and making it difficult to identify the early diffusion stage of fog patches. By combining spatial evolution data of fog with road lane topology data, road alignment data, traffic flow status data, and historical accident data, risk mapping is performed on each lane to generate a lane risk map. Furthermore, based on the lane risk map, risk ranking, guidance parameter solving, and regional collaborative constraint calculation are performed. Through reinforcement learning, lane guidance control instructions for different lanes are generated. This transforms the coarse-grained guidance method of the existing technology, which uniformly issues speed limits or warning information based on the overall road segment status, into a refined guidance method that considers the different degrees of fog impact on different lanes. This solves the problems of the existing technology, such as lack of judgment on the risk differences of different lanes, coarse-grained guidance strategies, and low traffic utilization. Furthermore, by collecting vehicle operation feedback data after guidance and comparing the vehicle operation feedback data with the target operation response parameters to obtain the guidance deviation result, the fog spatial evolution data is corrected based on the guidance deviation result, and the lane guidance control command is updated based on the corrected fog spatial evolution data, forming a closed-loop processing mechanism of fog monitoring, risk assessment, guidance execution and feedback correction. This enables the lane guidance strategy to be dynamically adjusted with the evolution of fog and changes in vehicle operation status, further improving the accuracy, timeliness and adaptive control capability of lane guidance in fog scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of a lane guidance method based on cloud-edge-device collaborative fog monitoring in this invention.

[0017] Figure 2 This is a flowchart of step S4 in this invention. Detailed Implementation

[0018] Please see Figures 1-2 As shown, this invention relates to a lane guidance method based on cloud-edge-device collaborative fog monitoring, comprising the following steps: S1. Acquire laser visibility data, millimeter-wave radar data, video data, meteorological data and traffic flow data on the side of the highway through edge sensors, combine them with cloud-based weather forecast data and road label data, and perform time alignment and spatial registration processing through edge computing nodes to obtain a spatiotemporal dataset; S2. Extract visibility change features, video texture attenuation features, millimeter wave echo attenuation features, and temperature and humidity wind field change features in the cloud based on the spatiotemporal dataset, and perform temporal segmentation and fusion encoding to obtain the spatial evolution data of fog. S3. Based on the spatial evolution data of the fog, combined with road lane topology data, road alignment data, traffic flow status data and historical accident data, risk mapping is performed on each lane to obtain a lane risk map. S4. Perform risk ranking and guidance parameter solving based on the lane risk map to obtain the guidance parameter set; construct a regional guidance constraint set based on the lane risk map splicing results of adjacent road segments and the ramp diversion capacity; perform reinforcement learning decision-making based on the guidance parameter set and the regional guidance constraint set to generate lane guidance control instructions and target operation response parameters; S5. Control the roadside guidance equipment and vehicle network publishing terminal to perform lane guidance according to the lane guidance control command, and collect vehicle operation feedback data after guidance; calculate the guidance deviation result based on the vehicle operation feedback data and the target operation response parameters; S6. Correct the fog spatial evolution data according to the induced deviation result, and update the lane guidance control command according to the corrected fog spatial evolution data.

[0019] In some embodiments, this solution is deployed on a continuous three-lane highway section in a mountainous area prone to fog, including bridges, gentle curves, and low-lying terrain. Along this section, laser visibility detection equipment, millimeter-wave radar, high-definition video acquisition equipment, temperature, humidity, wind speed and direction sensors, and traffic flow detection equipment are deployed. Simultaneously, the cloud accesses regional weather forecast interfaces, road label databases, and historical accident databases. The road label data includes not only station numbers, slopes, radii of curvature, bridge and tunnel locations, ramp connections, and lane numbers, but also accident-prone area labels, historical fog frequency labels, and lane function labels. This transforms fog monitoring from a single-moment local visibility value to a continuous temporal process of fog evolution and its differentiated impact on different lanes. In step S2, the cloud performs fog spatial evolution analysis using the aforementioned spatiotemporal dataset as input. In practice, the spatiotemporal dataset can be segmented according to fixed-length time windows to form continuous multi-source data segments. For example, overlapping segments are constructed using a 30-second basic analysis window and a 5-second sliding step size to preserve subtle changes in the early stages of fog formation. For video and millimeter-wave radar data, multi-channel input feature tensors corresponding to the time windows are constructed and input into a first convolutional neural network for convolution and nonlinear activation processing to extract spatially relevant features such as image texture blurring, contrast attenuation, and radar echo attenuation trends, resulting in spatial feature vectors. For laser visibility and meteorological data, a second convolutional neural network is used to extract sequence features to characterize the continuous decreasing trend of visibility, the rapid increasing trend of humidity, and the switching trend of wind fields, obtaining visibility change features and temperature, humidity, and wind field change features. The spatial feature vectors are then concatenated with the sequence feature vectors and encoded to obtain fog spatial evolution data. This fog spatial evolution data can specifically represent the joint state of the fog boundary position, boundary expansion direction, migration speed, and concentration distribution within a certain time window. In step S3, the fog spatial evolution data is mapped to the lane granularity to generate a lane risk map. Specifically, the boundary of the fog patch is divided into intervals along the road mileage direction, and the coverage area of ​​each interval within a preset time window is calculated based on the diffusion direction and migration speed. Then, combined with the lane topology, the coverage area is mapped to the corresponding lane unit to obtain the lane fog impact range. Taking a three-lane road section as an example, when the fog patch boundary spreads outward from the median strip, the leftmost lane may enter the moderate low visibility area first, while the middle lane and the rightmost lane are still in a slightly affected state at the same time. If a uniform speed limit is still applied to the entire road section at this time, it will cause excessive compression of traffic capacity. Further calculations are made based on the current vehicle speed, headway, traffic flow difference, and historical accident data of each lane, including visibility safety margin, rear-end collision risk weight, lane change conflict probability, and traffic stability.For example, within the fog-affected area, the visibility safety margin is obtained based on the difference between visibility distance and braking distance; the rate of change of speed difference between adjacent vehicles is calculated based on the vehicle speed distribution gradient, thus obtaining the rear-end collision risk weight; the lane change demand intensity per unit time is obtained by weighting the traffic flow difference between adjacent lanes, and the lane change conflict probability is obtained by combining it with lane density; and the traffic stability is obtained based on the vehicle speed fluctuation variance. Finally, the above parameters are normalized and weighted to generate a lane risk map. Step S4 comprehensively sorts the visibility safety margin, rear-end collision risk weight, lane change conflict probability, and traffic stability of each lane to obtain a risk level sequence; then, the guidance parameter set is calculated by combining the preset guidance rule base and equipment capacity constraints. The guidance parameter set may include the speed limit value, opening status, guidance light display scheme, and information release intensity corresponding to each lane. Furthermore, it is not limited to local decision-making on a single road segment, but further reads the lane risk map splicing results and ramp diversion capacity of adjacent road segments to construct a regional guidance constraint set. For example, when the fog spreads rapidly in the leftmost lane of the target road segment and the upstream ramp continues to receive high-density traffic, the system will restrict left lane traffic locally while simultaneously limiting the upstream inflow intensity based on the capacity of adjacent road segments and the diversion capacity of ramps, to avoid localized induction causing new congestion waves to shift forward. Based on this, the set of induction parameters and the set of regional induction constraints are input into the reinforcement learning decision-making process. The lane risk state vector and the regional collaborative constraint state vector constitute the state space, and the speed limit, lane opening status, and induction release intensity constitute the action space. The reward value is calculated based on the increase in average vehicle speed, the change in queue length, the reduction in abnormal braking, and the change in traffic stability. The optimal action is selected through strategy evaluation and strategy update, outputting lane guidance control commands and target operating response parameters. Unlike existing methods that rely on preset rules to directly trigger uniform speed limits, this scheme's induction decision-making is based on a lane risk map, enabling differentiated control for different lanes. The induction decision-making also considers regional collaborative constraints, avoiding regional imbalance caused by single-point optimization. Through reinforcement learning, the benefits of induction actions are evaluated, allowing the system to continuously approach a better control strategy based on historical and real-time feedback. Step S5: The roadside guidance equipment may include a lane controller, variable speed limit signs, variable information boards, and guidance light strips. The vehicle-to-everything (V2X) terminal then issues speed limit, lane change suggestions, and risk warnings to vehicles with vehicle-road cooperative capabilities. After guidance is executed, traffic flow detection equipment and on-board terminals continuously collect vehicle operation data and statistically analyze vehicle speed, deceleration, headway, and lane change behavior according to time windows to obtain statistical characteristic data. Further calculations are made of average vehicle speed, queue length, frequency of abnormal braking events, and speed fluctuation to form vehicle operation feedback data. This vehicle operation feedback data is then compared with the target operation response parameters generated in step S4 to calculate the guidance deviation result.In S6, observations are constructed based on vehicle average speed deviation, queue length deviation, abnormal braking event frequency deviation, and vehicle speed fluctuation deviation. Diffusion range parameters, migration speed parameters, and concentration distribution parameters are extracted from the current fog spatial evolution data as a set of parameters to be estimated. Subsequently, a time-recursive prediction calculation is performed on the set of parameters to be estimated using the Kalman filter method. The observation residual is calculated based on the difference between the observations and the prediction results, and the Kalman gain is further calculated. The Kalman gain is then used to perform a correction update on the set of parameters to be estimated, resulting in corrected fog spatial evolution data. For example, if the system initially judges that the fog is mainly concentrated in the middle and left lanes and issues a strong restriction on the leftmost lane, but after the guidance is implemented, a sudden increase in the frequency of abnormal braking events and a significant increase in vehicle speed fluctuation are observed in the rightmost lane, it indicates that the impact of the fog on the rightmost lane has been underestimated. In this case, the Kalman filter process can re-estimate the fog diffusion range and concentration distribution based on the new observations, appropriately correcting the fog boundary towards the rightmost lane, and further updating the lane risk map and lane guidance control commands. This embodiment achieves spatiotemporal alignment and spatial registration at the edge, completes fog spatial evolution modeling and lane risk mapping in the cloud, completes regional collaborative guidance decision-making through reinforcement learning, and completes online correction through feedback analysis and Kalman filtering. In fog scenarios, it simultaneously ensures timely warning, precise guidance, and traffic operation efficiency.

[0020] Further, S1 includes the following steps: Laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data are collected from the roadside of highways using laser visibility detection equipment, millimeter-wave radar, image acquisition equipment, meteorological API, and traffic platforms, and a unified timestamp is added. Time alignment processing is performed on various types of data based on the timestamp identifier to obtain multi-source time-series data arranged according to a uniform sampling interval; Spatial mapping is performed on the multi-source time-series data based on the road label data, and associated with the corresponding road location and lane unit to obtain spatial association data; Data cleaning and normalization are performed on the spatially correlated data to obtain a spatiotemporal dataset.

[0021] It should be noted that in a typical fog-prone section of a highway, laser visibility detection equipment, millimeter-wave radar, and high-definition video acquisition equipment are deployed along the route. Simultaneously, regional weather forecast data is obtained via a meteorological API, and traffic flow data such as vehicle speed, traffic volume, occupancy rate, and headway are acquired through a traffic management platform. These various data types have different sampling frequencies and data formats during acquisition. For example, laser visibility data is typically output in second-level increments, millimeter-wave radar data is output in continuous echo format, and video data is acquired in frame sequence format. To eliminate temporal scale differences between multi-source data, a unified timestamp is added to each data record during the data acquisition phase. An edge computing node serves as the unified processing entry point, and time alignment processing is performed on various data types based on the timestamps. Resampling and interpolation methods are used to map data of different frequencies to a unified sampling interval, resulting in multi-source time-series data arranged chronologically. Spatial mapping processing is then performed on the multi-source time-series data based on pre-constructed road label data. This road label data includes spatial attribute information such as road mileage markers, lane division information, sensor installation locations, camera field of view coverage, and radar detection areas. Edge computing nodes map each record in multi-source time-series data to a specific road location and corresponding lane unit based on the geographic coordinates of sensors and their corresponding coverage areas. For example, they associate video frame data at a certain timestamp with the lane range it covers, and locate target echo data detected by millimeter-wave radar to a specific lane and distance range, thus obtaining spatially correlated data with clear spatial orientation. Subsequently, data cleaning and normalization processing are performed on the spatially correlated data. Specifically, this includes removing abrupt data caused by equipment failure or communication anomalies through outlier detection, completing missing data using temporal neighborhood interpolation, and performing normalization mapping processing on data of different dimensions. For example, visibility values, radar echo intensity, image grayscale features, and meteorological parameters are uniformly mapped to a preset numerical range, making various features comparable in subsequent processing. After the above processing, a spatiotemporal dataset with a unified time scale, unified spatial coordinates, and unified numerical range is obtained.

[0022] Further, S2 includes the following steps: Based on the spatiotemporal dataset, laser visibility data, video data, millimeter-wave radar data, and meteorological data are segmented according to time windows to obtain multi-source data segments; A multi-channel input feature tensor is constructed from the video data and millimeter-wave radar data in the multi-source data segment. The input feature tensor is then input into a first convolutional neural network to perform convolution operations and nonlinear activation, extracting video texture attenuation features and echo attenuation features to obtain a spatial feature vector. The laser visibility data and meteorological data in the multi-source data segment are processed by a second convolutional neural network to extract sequence features, obtain visibility change features and temperature, humidity and wind field change features, and then perform feature concatenation with the spatial feature vector to obtain a fused feature vector. Encoding mapping is performed based on the fused feature vector to obtain the spatial evolution data of the fog.

[0023] In some embodiments, the spatiotemporal dataset is first segmented into segments according to a fixed-length time window based on a uniform sampling interval. For example, overlapping time segments are constructed using a 30-second window length and a 5-second sliding step, thereby enhancing the ability to capture subtle changes in the early stages of fog formation while maintaining temporal continuity. For the data within each time window, laser visibility sequences, video frame sequences, millimeter-wave radar echo sequences, and meteorological parameter sequences are simultaneously extracted to obtain a set of multi-source data segments. During spatial feature extraction, the video data and millimeter-wave radar data in the multi-source data segments are jointly modeled. Specifically, continuous video frames are stacked in chronological order to form a three-dimensional tensor, while the millimeter-wave radar echo intensity distribution within the corresponding time window is mapped to a two-dimensional matrix consistent with the video spatial resolution and stitched together as additional channels to construct a multi-channel input feature tensor. After the input feature tensor is fed into the first convolutional neural network, sliding convolution operations are performed on local regions through multiple convolutional kernels, and the feature responses are mapped by a nonlinear activation function. This extracts features such as changes in texture sharpness, edge blurring, and contrast attenuation in the video image. Simultaneously, the attenuation trend and spatial distribution characteristics of radar echoes are fused to obtain a spatial feature vector that characterizes the comprehensive impact of fog on spatial visibility. Unlike existing technologies that analyze only video or radar separately, this embodiment constructs a unified multi-channel input, enabling feature-level fusion of the two types of data during convolution. This allows for the identification of complex transitional states such as blurred images with no significant echo attenuation or significant echo attenuation without complete image distortion, improving the ability to identify the early stages of fog. Visibility, temperature, humidity, and wind speed / direction sequences are combined chronologically into a multi-dimensional time series input and fed into the second convolutional neural network. A one-dimensional convolutional kernel performs convolution operations along the time axis to extract temporal dynamic information such as the rate of change in visibility, the rapid increase in humidity, and abrupt changes in wind field characteristics, yielding visibility change features and temperature / humidity / wind field change features. This processing method differs from thresholding methods for a single parameter. It automatically learns local change patterns in the time series through convolutional operations, enabling the identification of potential evolutionary signs before fog formation. Subsequently, the spatial feature vector and the sequence feature vector are concatenated according to a preset feature arrangement order to construct a fused feature vector. To ensure consistency of features from different sources during the fusion process, normalization mapping is performed on each feature vector, and a linear transformation is used to project them to a unified dimensional space. Further, the fused feature vector undergoes encoding mapping processing, for example, through a fully connected layer or an embedded mapping function, compressing the high-dimensional fused features into a low-dimensional representation vector to characterize the spatial distribution and evolution trend of the fog within the current time window, obtaining fog spatial evolution data, including fog boundary locations, diffusion directions, migration speeds, and concentration distribution parameters.

[0024] Further, S3 includes the following steps: Based on the spatial evolution data of the fog, the boundary position of the fog is divided into intervals along the road mileage direction, and the coverage of each interval within a preset time window is calculated according to the diffusion direction and migration speed. The coverage is then mapped to the corresponding lane unit to obtain the lane fog influence range data. Based on the lane fog impact range data, the current vehicle speed and headway of each lane, the difference between the visible distance and braking distance of the vehicle within the fog impact range is calculated to obtain the visibility safety margin. The rate of change of speed difference between adjacent vehicles is calculated based on the visible safety margin and the vehicle speed distribution gradient, and the rear-end collision risk weight is calculated based on the rate of change of speed difference. Based on the lane fog impact range data and the traffic flow difference between adjacent lanes, the lane change demand intensity per unit time is obtained by weighting, and the lane change conflict probability is calculated based on the lane change demand intensity and lane density. At the same time, the traffic stability is calculated based on the vehicle speed fluctuation variance. The visible safety margin, rear-end collision risk weight, lane change conflict probability, and traffic stability are normalized and weighted to generate a lane risk map.

[0025] In some embodiments, the boundary position, diffusion direction, and migration velocity parameters contained in the spatial evolution data of the fog patch are first read, and the fog patch boundary is divided into intervals based on road mileage coordinates, for example, by discretizing the road using fixed-length intervals. Subsequently, based on the fog patch migration velocity vector and diffusion direction, a forward outward calculation is performed on the fog patch boundary within a preset time window to obtain the coverage range of each interval in the future time range. Further, combined with road lane topology information, the coverage range is mapped to specific lane units, thereby obtaining lane fog influence range data. During the visual safety margin calculation process, for each lane unit, the difference between the visible distance and braking distance is calculated for vehicles within the fog patch influence range, based on their current vehicle speed and headway information. The visible distance is obtained by mapping the visibility level and road lighting conditions, while the braking distance is calculated based on the vehicle's current speed and road adhesion coefficient. By performing the difference operation on the two, a visual safety margin reflecting the current driving safety margin within the lane can be obtained. When the visual safety margin is negative, it indicates that the driver cannot guarantee safe braking under the current conditions, and the risk level is significantly increased. Compared to traditional methods that rely solely on visibility thresholds, this indicator couples environmental conditions with vehicle operating conditions in a modeling manner, providing a more realistic reflection of actual driving risks. It calculates the vehicle speed distribution gradient based on the vehicle speed sequences within each lane and estimates the rate of change of speed differences between adjacent vehicles by combining this with the visibility safety margin. For example, by performing a difference operation on the speed sequences of adjacent vehicles and combining it with a time window to calculate the rate of change, the dynamic trend of speed difference can be obtained. The rear-end collision risk weight is then calculated based on the correlation between this trend and the visibility safety margin. When the rate of change of speed difference is large and the visibility safety margin is low, the rear-end collision risk weight is significantly increased. First, based on the lane fog impact range data and the traffic flow difference between adjacent lanes, the lane change demand per unit time is estimated using weighted estimation. For example, when a lane is heavily affected by fog while adjacent lanes have better traffic conditions, significant lateral movement demand will arise. The lane change demand intensity can be obtained by weighting the traffic flow difference and the fog impact intensity. Subsequently, combined with the current lane density, a probability mapping is performed on the lane change demand intensity to obtain the probability of lane change conflict. Simultaneously, by calculating the variance of the vehicle speed time series, the degree of vehicle speed fluctuation is obtained and used as a quantitative indicator of traffic stability. Greater vehicle speed fluctuation indicates more unstable traffic flow and higher potential risks. Preferably, in the calculation of the comprehensive risk value, the weight corresponding to the visible safety margin can be set to 0.35, the weight corresponding to the rear-end collision risk can be set to 0.30, the weight corresponding to the lane change conflict probability can be set to 0.20, and the weight corresponding to traffic stability can be set to 0.15.The reason for adopting the above weighting method is that in fog scenarios, drivers are first directly affected by the rapid compression of visibility, so the visibility safety margin has the most significant impact on risk formation; rear-end collision risk is directly related to the lag in front vehicle recognition and the amplification of speed difference in fog conditions, and belongs to the secondary core risk factors; although the probability of lane change conflict will significantly affect local safety, its occurrence is usually based on the existing fog impact and longitudinal risk disturbance, so its weight is lower than the former two; traffic stability mainly reflects the propagation characteristics of traffic flow disturbances and has a corrective effect on the overall risk assessment.

[0026] Further, S4 includes the following steps: Based on the visible safety margin, rear-end collision risk weight, lane change conflict probability and traffic stability of each lane in the lane risk map, a lane risk state vector is constructed. Combined with the lane risk map splicing results of adjacent road segments and the ramp diversion capacity, a regional collaborative constraint state vector is constructed. Based on the lane risk state vector and the regional collaborative constraint state vector, a reinforcement learning state space is constructed, and the speed limit value, lane opening status and guidance release intensity in the guidance parameter set are used as action sets to construct a reinforcement learning action space. Based on the state space and action space, the reward value corresponding to each action is calculated. The reward value is calculated by weighting the increase in average vehicle speed, the change in queue length, the reduction in abnormal braking, and the change in traffic stability. The strategy for each action is evaluated and updated based on the reward value, and the optimal action is selected as the lane guidance control command. Target running response parameters are then generated based on the optimal action.

[0027] It should be noted that reinforcement learning training first constructs a simulated road environment corresponding to the target highway segment in the cloud, completes policy training in this simulated road environment, and then deploys the converged policy parameters to the actual lane guidance system for online inference. Therefore, this solution constructs a digital simulated road environment, mapping the fog evolution process, traffic flow propagation process, and guidance and control process to an iteratively trainable closed-loop environment, completing policy learning before on-site deployment. First, a road topology simulation base map is constructed based on the road label data of the target road segment. This base map includes structural parameters such as the number of mainline lanes, lane widths, merging and diverging relationships, bridge sections, curve radii, longitudinal slope intervals, ramp access locations, and speed limit sign distribution locations. Fixed-length road segment units are used as discrete simulation units. For example, the target highway segment is divided into several continuous unit segments along the mileage direction, and further divided into grids by lane in the horizontal direction, resulting in a two-dimensional road network unit structure of "mileage interval - lane number". Each two-dimensional road network unit is bound to corresponding geometric attributes, capacity parameters, and prior accident risk parameters. Furthermore, the initial traffic flow state of each road network unit is calibrated based on historical traffic flow data, including free-flow velocity distribution, saturation flow rate, headway statistics for following vehicles, proportion of different vehicle types, and temporal variation patterns of input flow at each ramp. Time-varying fog impact values ​​are assigned to each road network unit based on fog boundary location, diffusion direction, migration speed, and concentration distribution parameters. For each training time, the fog boundary location is updated along the road mileage direction based on fog migration speed, the lateral influence lane range is updated based on diffusion direction, and the visibility attenuation level within each unit is updated based on concentration distribution. The vehicle perception and driving response capabilities within the corresponding road network unit are corrected based on the fog impact values. For example, reduced visibility is mapped to a shortened driver visibility distance, increased concentration is mapped to increased target detection latency, and wind field changes are mapped to increased disturbances in vehicle lateral lane-changing intentions. This allows the impact of the fog environment on traffic behavior to be incorporated into the simulation process through specific calculable parameters. Furthermore, vehicle state update rules are constructed for each lane. Each vehicle in the simulation environment includes at least the following variables: current position, instantaneous speed, acceleration, target lane, desired speed, and current following target. Within each simulation step, the longitudinal speed update of the vehicle is first calculated based on the position and speed of the vehicle in front, the current headway, and the impact value of the fog patch. Then, the lateral lane change demand intensity is calculated based on the traffic flow difference between adjacent lanes, the degree of fog impact in the target lane, and the traffic pressure in the current lane. When a lane is heavily affected by fog while the risk in adjacent lanes is low, the vehicle's lane-changing intention parameter is increased; when the density of adjacent lanes is high or the fog coverage is simultaneously enhanced, lateral lane-changing actions are suppressed.In the reinforcement learning training input construction process, instead of directly inputting all the original simulation data into the strategy, a lane risk state vector is first constructed based on the visible safety margin, rear-end collision risk weight, lane change conflict probability, and traffic stability of each lane at each training step. Then, a regional collaborative constraint state vector is constructed by combining the lane risk map splicing results of adjacent road segments, ramp diversion capacity, and upstream input flow. Further, the risk state vectors of each lane in this road segment are spliced ​​in lane order, and the regional collaborative constraint state vector is appended to it to form a complete training state input. The purpose of this approach is to enable the reinforcement learning strategy to perceive not only the local risk of the current lane at each decision moment, but also the overall constraints of adjacent road segments and ramps, thereby avoiding the trained strategy from only optimizing local single-lane traffic. In the action space, several speed limit levels can be defined for each lane, and the open state can be defined as remain open, restrict traffic, or close. The intensity of the guidance release can be defined as low-frequency, medium-frequency, and high-frequency prompts. Subsequently, the action parameters corresponding to each lane are combined in a preset order to form a set of candidate control actions at a certain moment. During training, the policy selects an action combination from the candidate action set and inputs it into the simulation environment. The simulation environment adjusts the lane passage rules and guidance information output rules of the corresponding road network unit based on this action. For example, when an action lowers the speed limit of the leftmost lane and increases the frequency of lane prompts, the simulation environment simultaneously lowers the expected speed limit of vehicles in that lane and increases the probability of vehicles responding to lane change prompts. In the training implementation, abnormal braking events are not simply judged based on the absolute value of deceleration, but are identified based on the composite condition that the deceleration is lower than a preset threshold and the headway shortens synchronously within a unit time window, to avoid misjudging normal deceleration as risky braking. Queue length can be obtained by accumulating the length of time that low-speed vehicles continuously occupy road network units, making the reward value calculation closer to the meaning of real traffic operation. Through this reward definition method, the policy automatically learns during training that it cannot simply pursue the maximization of average vehicle speed, but needs to simultaneously suppress queue diffusion and the increase of abnormal braking. During the policy update process, a value function-based reinforcement learning training method is used to recursively estimate the rewards of different state-action pairs. After each environmental interaction, the current state, selected action, immediate reward, and new state after the action are recorded. The policy evaluation value of the current action in the current state is then updated according to the policy update formula. In the early stages of training, the impact of different inducement combinations on traffic flow can be explored through a certain proportion of random actions. As the number of training rounds increases, the proportion of random exploration is gradually reduced, and the probability of action selection based on the current optimal estimate is increased, allowing the policy to transition from extensive trial and error to stable convergence. To improve training stability, state transition samples collected over multiple consecutive simulation cycles can be cached and input into the update process in batches. This ensures that the policy evaluation value update does not depend on random fluctuations of a single step length, but is corrected based on the average reward trend over a period of time.After multiple rounds of iterative training under fog scenarios, the strategy will gradually learn to select appropriate combinations of speed limits, lane openings, and warning intensities under different fog boundary locations, diffusion rates, and traffic load levels. During the simulated road training phase, instead of a fixed single fog scenario, multiple training sample scenarios are constructed in batches. For example, scenarios are constructed where localized fog from low-lying areas of bridges spreads laterally towards the middle lane of the main road; scenarios are constructed where fog rapidly advances along the mileage direction on mountain curves; scenarios are constructed where fog and high-density traffic flow overlap in ramp merging areas; and scenarios are constructed where light fog gradually intensifies into high-concentration fog. Within each scenario, traffic flow intensity, vehicle type ratio, and upstream input flow are randomly perturbed, allowing the training strategy to learn stable induced decision-making patterns under multiple scenarios and perturbation conditions, rather than relying on a single fixed condition.

[0028] Furthermore, the formula for updating the strategy is as follows: ; in, In the current state Next action The strategy evaluation value; The learning rate; This is the reward value; For discount factor; Execute action The new state afterwards; For the new state Candidate actions from the set of available actions; In the new state The maximum policy evaluation value is obtained by comparing the policy evaluation values ​​corresponding to each candidate action.

[0029] Specifically, the strategy update employs an iterative recursive approach. Within each time window, the strategy evaluation value of each action is updated based on the currently observed traffic conditions and the effectiveness of the guidance implementation. The updated result serves as the initial evaluation value for the next decision cycle. For example, at a certain moment, if a combination of speed limit reduction and enhanced guidance information is implemented on a lane, resulting in a significant reduction in abnormal braking events and improved traffic stability, the immediate reward value of the corresponding action is high. Through the aforementioned update formula, its strategy evaluation value will be adjusted towards a higher level. Conversely, if an action leads to an increase in queue length or aggravated speed fluctuations, its strategy evaluation value will be weakened. Through continuous iteration, the optimal guidance strategy can be gradually approximated under different fog evolution states and traffic flow conditions. Unlike guidance strategies based on fixed rules or static parameters, this strategy update introduces a recursive update mechanism that includes both immediate rewards and future benefits. This allows the strategy evaluation to not only rely on current observations but also consider the impact of actions on future traffic conditions, thereby maximizing long-term benefits in a dynamic environment.

[0030] Furthermore, the reward value is calculated using the following reward function: ; in, This is the reward value; The increase in the vehicle's average speed relative to its initial speed before induction; This represents the change in queue length. This represents the change in abnormal braking events; This represents the change in traffic stability. , , and These are the weighting coefficients.

[0031] It should be noted that within each decision cycle, the differences in key operational indicators before and after the induction implementation are first calculated based on the vehicle operation feedback data collected in step S5. For example, the increase in average vehicle speed is calculated by comparing the average vehicle speed within the time window before and after the induction; the change in queue length is calculated by comparing the number of queued vehicles or queue length; the change in abnormal braking events is obtained by statistically analyzing the change in the number of sudden deceleration behaviors per unit time; and the change in traffic stability is obtained by calculating the change in the variance of the vehicle speed time series. These indicators can be normalized before calculation to map them to a uniform numerical range, thereby avoiding the influence of different dimensions on the reward calculation results. Different indicators are treated differently in the reward function according to their direction of action. For example, the increase in average vehicle speed and the change in traffic stability are positive benefit indicators; the larger their values, the better the traffic operation. Therefore, they participate in the calculation with positive weights in the reward function. On the other hand, the change in queue length and the change in abnormal braking events are negative cost indicators; when their values ​​increase, it indicates a deterioration in traffic conditions. Therefore, during the calculation process, they can be made to suppress the reward value through sign conversion or weight adjustment. By performing a weighted combination calculation on each indicator, a reward value reflecting the overall effect of the current guidance action is obtained. The construction of this reward function incorporates traffic efficiency, safety, and stability into the evaluation, enabling the reinforcement learning decision-making process to weigh multiple objectives. For example, in some cases, simply increasing vehicle speed may lead to an increase in abnormal braking events. However, by introducing constraints on changes in abnormal braking and stability, the strategy can avoid being overly biased towards efficiency and neglecting safety. By reasonably setting the weight coefficients, the importance of each indicator can be dynamically adjusted according to the characteristics of different road sections or management needs. Preferably, the weight coefficient corresponding to the increase in average vehicle speed can be set to 0.30, the weight coefficient corresponding to the change in queue length to 0.25, the weight coefficient corresponding to the change in abnormal braking events to 0.30, and the weight coefficient corresponding to the change in traffic stability to 0.15. The reason for using this set of weights is that the core objective of lane guidance in fog scenarios is not simply to increase traffic speed, but to maintain controllable traffic flow while ensuring safety. Therefore, the increase in average vehicle speed and the change in abnormal braking events respectively represent the degree of efficiency improvement and the degree of change in safety risk, and should be given relatively high weights. The change in queue length directly reflects the effect of guidance measures on suppressing the spread of local congestion and is a secondary key indicator. The change in traffic stability mainly represents the propagation of traffic flow disturbances and has a corrective effect on the overall strategy.The weights can be adjusted in different levels according to the attributes of different road sections. For bridges, long downhill sections, or curved road sections, the weight coefficient corresponding to the change in abnormal braking events can be appropriately increased to 0.35, while the weight coefficient corresponding to the increase in average vehicle speed can be reduced to 0.25 to strengthen safety constraints. For main road sections with high traffic density but relatively straight alignment, the weight coefficient corresponding to the change in queue length can be increased to 0.30 to enhance the ability to suppress the spread of congestion.

[0032] Furthermore, the vehicle operation feedback data is obtained through the following steps: The lane guidance control command controls the roadside guidance equipment and the vehicle network publishing terminal to perform lane guidance, and collects vehicle operation data after guidance based on traffic flow detection equipment and vehicle terminal; Based on the vehicle operation data, time window statistical processing is performed on vehicle speed, deceleration, headway, and lane change behavior to obtain statistical feature data. Based on the statistical characteristic data, the average vehicle speed, queue length, frequency of abnormal braking events, and degree of vehicle speed fluctuation are calculated to obtain vehicle operation feedback data.

[0033] Specifically, after generating lane guidance control commands in step S4, the roadside guidance equipment performs speed limit adjustments, lane opening control, and guidance information dissemination for each lane. Simultaneously, the vehicle-to-everything (V2X) dissemination terminal sends corresponding guidance information to connected vehicles. After guidance execution, the traffic flow detection equipment continuously collects data on the speed, acceleration, headway, and lane-changing behavior of vehicles in each lane. The onboard terminal further supplements this with individual vehicle-level operational status information, such as instantaneous speed changes and sudden deceleration event markers, thus forming a vehicle operation data sequence containing multi-granularity information. During data processing, statistical processing is first performed on the vehicle operation data within a fixed time window. For example, the mean of the vehicle speed sequence is calculated using a 30-second statistical window; the deceleration sequence is threshold-filtered to identify sudden deceleration events; the headway sequence is statistically analyzed; and lane-changing behaviors are counted and categorized by direction, thereby obtaining statistical feature data within the corresponding time window. Based on this, parameter calculations are performed on the statistical characteristic data: the average vehicle speed is obtained by averaging the speed statistics; the queue length is obtained by statistically analyzing the length of the continuous distribution interval of low-speed vehicles or the cumulative number of queued vehicles; the frequency of abnormal braking events is obtained by counting events where deceleration exceeds a preset threshold; and the degree of vehicle speed fluctuation is obtained by calculating the variance or standard deviation of the speed time series. During the calculation process, each of these indicators can be divided into lane units, so that each lane corresponds to a set of independent operational feedback parameters.

[0034] Further, S6 includes the following steps: Based on the induced bias results, an observation set including vehicle average speed deviation, queue length deviation, abnormal braking event frequency deviation, and vehicle speed fluctuation deviation is constructed. Based on the spatial evolution data of the fog, diffusion range parameters, migration speed parameters, and concentration distribution parameters are extracted to form a set of parameters to be estimated. The time recursive prediction calculation is performed on the set of parameters to be estimated based on the Kalman filter method, and the observation residual is calculated based on the difference between the observation and the prediction result. The Kalman gain is calculated based on the observed residuals, and the set of parameters to be estimated is corrected and updated based on the Kalman gain to obtain the corrected fog spatial evolution data. The lane risk map and lane guidance control instructions are then updated based on the corrected fog spatial evolution data.

[0035] In some embodiments, the deviations of average vehicle speed, queue length, frequency of abnormal braking events, and speed fluctuation are first structured to construct an observation vector, which reflects the differences between the current guidance strategy and the target operational response. Simultaneously, diffusion range parameters, migration velocity parameters, and concentration distribution parameters are extracted from the current fog spatial evolution data and arranged in a preset order to form a set of parameters to be estimated, used to characterize the dynamic distribution of fog in the road space. Then, a time-recursive prediction calculation is performed on the set of parameters to be estimated based on the Kalman filter method. Next, based on the fog parameters estimated at the previous time step and their changing trends, a forward prediction is performed on the diffusion range, migration velocity, and concentration distribution at the current time step to obtain the prediction results. Subsequently, the observations and prediction results are differentially analyzed to obtain the observation residuals, which are used to quantify the degree of deviation between the current model prediction and the actual operational feedback. For example, if the prediction results indicate that a certain lane is less affected by fog, while the observations show a significant decrease in vehicle speed and an increase in abnormal braking events in the corresponding lane, this difference will be explicitly characterized by the observation residuals. The Kalman gain is calculated based on the observation residuals. The Kalman gain is used to characterize the weighting relationship between current observation information and historical prediction information in parameter updates. Its value can be adaptively adjusted according to the residual size and the stability of historical estimates. When the observation residual is large, it indicates that the current prediction deviates significantly from the actual situation, so the weight of the observation information is increased. When the observation residual is small, the confidence in historical predictions is maintained, thus ensuring the smoothness of parameter updates. Subsequently, based on the Kalman gain, a correction update process is performed on the set of parameters to be estimated, and the diffusion range parameter, migration velocity parameter, and concentration distribution parameter are weighted and corrected to obtain the corrected fog spatial evolution data. After completing the fog parameter update, the lane risk map is recalculated based on the corrected fog spatial evolution data, and the reinforcement learning state space input is further updated to generate new lane guidance control commands. For example, in a certain implementation scenario, the initial prediction suggests that the fog is mainly concentrated in the middle lane. However, after the guidance is implemented, significant speed fluctuations and abnormal braking events are observed in the rightmost lane. Through the aforementioned Kalman filter update process, the influence range of the fog can be corrected to the right lane, and the risk weight of that lane can be increased in the subsequent risk assessment. This adjusts the guidance strategy and can continuously correct model biases in scenarios with rapid fog evolution and highly unstable traffic flow, thereby improving the accuracy of fog state estimation and the adaptability of the lane guidance strategy.

[0036] The present invention also includes a lane guidance system based on cloud-edge-device collaborative fog monitoring, comprising: The end-side sensor module is used to acquire laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data from the roadside of the highway. Edge computing nodes are used to perform time alignment and spatial registration processing on laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data from the roadside of highways to obtain spatiotemporal datasets; The cloud computing module is used to perform the following steps: Based on the spatiotemporal dataset, visibility variation features, video texture attenuation features, millimeter wave echo attenuation features, and temperature and humidity wind field variation features are extracted, and temporal segmentation and fusion encoding are performed to obtain the spatial evolution data of fog. Based on the spatial evolution data of the fog, combined with road lane topology data, road alignment data, traffic flow status data and historical accident data, risk mapping is performed on each lane to obtain a lane risk map; Based on the lane risk map, risk ranking and guidance parameter solving are performed to obtain the guidance parameter set; based on the lane risk map splicing results of adjacent road segments and the ramp diversion capacity, a regional guidance constraint set is constructed; based on the guidance parameter set and the regional guidance constraint set, reinforcement learning decision-making is performed to generate lane guidance control instructions and target operation response parameters; The lane guidance control command controls the roadside guidance equipment and the vehicle-to-everything (V2X) terminal to perform lane guidance, and collects vehicle operation feedback data after guidance; the guidance deviation result is calculated based on the vehicle operation feedback data and the target operation response parameters. The fog spatial evolution data is corrected based on the induced deviation results, and the lane guidance control command is updated based on the corrected fog spatial evolution data.

[0037] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A lane guidance method based on cloud-edge-device collaborative fog monitoring, characterized in that, Includes the following steps: S1. Acquire laser visibility data, millimeter-wave radar data, video data, meteorological data and traffic flow data on the side of the highway through edge sensors, combine them with cloud-based weather forecast data and road label data, and perform time alignment and spatial registration processing through edge computing nodes to obtain a spatiotemporal dataset; S2. Extract visibility change features, video texture attenuation features, millimeter wave echo attenuation features, and temperature and humidity wind field change features in the cloud based on the spatiotemporal dataset, and perform temporal segmentation and fusion encoding to obtain the spatial evolution data of fog. S3. Based on the spatial evolution data of the fog, combined with road lane topology data, road alignment data, traffic flow status data and historical accident data, risk mapping is performed on each lane to obtain a lane risk map. S4. Perform risk ranking and guidance parameter solving based on the lane risk map to obtain the guidance parameter set; construct a regional guidance constraint set based on the lane risk map splicing results of adjacent road segments and the ramp diversion capacity; perform reinforcement learning decision-making based on the guidance parameter set and the regional guidance constraint set to generate lane guidance control instructions and target operation response parameters; S5. Control the roadside guidance equipment and vehicle network publishing terminal to perform lane guidance according to the lane guidance control command, and collect vehicle operation feedback data after guidance; calculate the guidance deviation result based on the vehicle operation feedback data and the target operation response parameters; S6. Correct the fog spatial evolution data according to the induced deviation result, and update the lane guidance control command according to the corrected fog spatial evolution data.

2. The lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 1, characterized in that, S1 includes the following steps: Laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data are collected from the roadside of highways using laser visibility detection equipment, millimeter-wave radar, image acquisition equipment, meteorological API, and traffic platforms, and a unified timestamp is added. Time alignment processing is performed on various types of data based on the timestamp identifier to obtain multi-source time-series data arranged according to a uniform sampling interval; Spatial mapping is performed on the multi-source time-series data based on the road label data, and associated with the corresponding road location and lane unit to obtain spatial association data; Data cleaning and normalization are performed on the spatially correlated data to obtain a spatiotemporal dataset.

3. The lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 2, characterized in that, S2 includes the following steps: Based on the spatiotemporal dataset, laser visibility data, video data, millimeter-wave radar data, and meteorological data are segmented according to time windows to obtain multi-source data segments; A multi-channel input feature tensor is constructed from the video data and millimeter-wave radar data in the multi-source data segment. The input feature tensor is then input into a first convolutional neural network to perform convolution operations and nonlinear activation, extracting video texture attenuation features and echo attenuation features to obtain a spatial feature vector. The laser visibility data and meteorological data in the multi-source data segment are processed by a second convolutional neural network to extract sequence features, obtain visibility change features and temperature, humidity and wind field change features, and then perform feature concatenation with the spatial feature vector to obtain a fused feature vector. Encoding mapping is performed based on the fused feature vector to obtain the spatial evolution data of the fog.

4. The lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 3, characterized in that, S3 includes the following steps: Based on the spatial evolution data of the fog, the boundary position of the fog is divided into intervals along the road mileage direction, and the coverage of each interval within a preset time window is calculated according to the diffusion direction and migration speed. The coverage is then mapped to the corresponding lane unit to obtain the lane fog influence range data. Based on the lane fog impact range data, the current vehicle speed and headway of each lane, the difference between the visible distance and braking distance of the vehicle within the fog impact range is calculated to obtain the visibility safety margin. The rate of change of speed difference between adjacent vehicles is calculated based on the visible safety margin and the vehicle speed distribution gradient, and the rear-end collision risk weight is calculated based on the rate of change of speed difference. Based on the lane fog impact range data and the traffic flow difference between adjacent lanes, the lane change demand intensity per unit time is obtained by weighting, and the lane change conflict probability is calculated based on the lane change demand intensity and lane density. At the same time, the traffic stability is calculated based on the vehicle speed fluctuation variance. The visible safety margin, rear-end collision risk weight, lane change conflict probability, and traffic stability are normalized and weighted to generate a lane risk map.

5. The lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 1, characterized in that, S4 includes the following steps: Based on the visible safety margin, rear-end collision risk weight, lane change conflict probability and traffic stability of each lane in the lane risk map, a lane risk state vector is constructed. Combined with the lane risk map splicing results of adjacent road segments and the ramp diversion capacity, a regional collaborative constraint state vector is constructed. Based on the lane risk state vector and the regional collaborative constraint state vector, a reinforcement learning state space is constructed, and the speed limit value, lane opening status and guidance release intensity in the guidance parameter set are used as action sets to construct a reinforcement learning action space. Based on the state space and action space, the reward value corresponding to each action is calculated. The reward value is calculated by weighting the increase in average vehicle speed, the change in queue length, the reduction in abnormal braking, and the change in traffic stability. The strategy for each action is evaluated and updated based on the reward value, and the optimal action is selected as the lane guidance control command. Target running response parameters are then generated based on the optimal action.

6. The lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 5, characterized in that, The formula for updating the strategy is as follows: ； in, In the current state Next action The strategy evaluation value; The learning rate; This is the reward value; For discount factor; Execute action The new state afterwards; For the new state Candidate actions from the set of available actions; In the new state The maximum policy evaluation value is obtained by comparing the policy evaluation values ​​corresponding to each candidate action.

7. The lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 6, characterized in that, The reward value is calculated using the following reward function: ； in, This is the reward value; The increase in the vehicle's average speed relative to its initial speed before induction; This represents the change in queue length. This represents the change in abnormal braking events; This represents the change in traffic stability. , , and These are the weighting coefficients.

8. The lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 1, characterized in that, The vehicle operation feedback data is obtained through the following steps: The lane guidance control command controls the roadside guidance equipment and the vehicle network publishing terminal to perform lane guidance, and collects vehicle operation data after guidance based on traffic flow detection equipment and vehicle terminal; Based on the vehicle operation data, time window statistical processing is performed on vehicle speed, deceleration, headway, and lane change behavior to obtain statistical feature data. Based on the statistical characteristic data, the average vehicle speed, queue length, frequency of abnormal braking events, and degree of vehicle speed fluctuation are calculated to obtain vehicle operation feedback data.

9. A lane guidance method based on cloud-edge-device collaborative fog monitoring according to claim 8, characterized in that, S6 includes the following steps: Based on the induced deviation results, an observation set including vehicle average speed deviation, queue length deviation, abnormal braking event frequency deviation, and vehicle speed fluctuation deviation is constructed. Based on the spatial evolution data of the fog, diffusion range parameters, migration speed parameters, and concentration distribution parameters are extracted to form a set of parameters to be estimated. The time recursive prediction calculation is performed on the set of parameters to be estimated based on the Kalman filter method, and the observation residual is calculated based on the difference between the observation and the prediction result. The Kalman gain is calculated based on the observed residuals, and the set of parameters to be estimated is corrected and updated based on the Kalman gain to obtain the corrected fog spatial evolution data. The lane risk map and lane guidance control instructions are then updated based on the corrected fog spatial evolution data.

10. A lane guidance system based on cloud-edge-device collaborative fog monitoring, applied to the lane guidance method based on cloud-edge-device collaborative fog monitoring as described in any one of claims 1-9, characterized in that, include: The end-side sensor module is used to acquire laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data from the roadside of the highway. Edge computing nodes are used to perform time alignment and spatial registration processing on laser visibility data, millimeter-wave radar data, video data, meteorological data, and traffic flow data from the roadside of highways to obtain spatiotemporal datasets; The cloud computing module is used to perform the following steps: Based on the spatiotemporal dataset, visibility variation features, video texture attenuation features, millimeter wave echo attenuation features, and temperature and humidity wind field variation features are extracted, and temporal segmentation and fusion encoding are performed to obtain the spatial evolution data of fog. Based on the spatial evolution data of the fog, combined with road lane topology data, road alignment data, traffic flow status data and historical accident data, risk mapping is performed on each lane to obtain a lane risk map; Based on the lane risk map, risk ranking and guidance parameter solving are performed to obtain the guidance parameter set; based on the lane risk map splicing results of adjacent road segments and the ramp diversion capacity, a regional guidance constraint set is constructed; based on the guidance parameter set and the regional guidance constraint set, reinforcement learning decision-making is performed to generate lane guidance control instructions and target operation response parameters; The lane guidance control command controls the roadside guidance equipment and the vehicle-to-everything (V2X) terminal to perform lane guidance, and collects vehicle operation feedback data after guidance; the guidance deviation result is calculated based on the vehicle operation feedback data and the target operation response parameters. The fog spatial evolution data is corrected based on the induced deviation results, and the lane guidance control command is updated based on the corrected fog spatial evolution data.

Images