A traffic flow perception method and system

Abstract

This application relates to the field of traffic flow sensing technology, and in particular to a traffic flow sensing method and system. The method involves collecting multi-source heterogeneous traffic sensing data from fixed sensing devices, mobile sensing terminals, and environmental sensors within a target sensing area; sending this data to an edge computing node for preliminary processing and feature extraction to generate structured feature data, which is then uploaded to a cloud platform; aligning the data to a unified spatiotemporal reference on the cloud platform and performing multi-source data fusion processing based on dynamic confidence weighting to obtain fused traffic parameters; dynamically adjusting the fusion weights of each data source using a pre-generated environment-weight mapping model; and finally generating traffic situation information including traffic flow, average vehicle speed, lane occupancy, and traffic event types based on the fused traffic parameters. This method can adapt to environmental changes, improve the accuracy and robustness of traffic flow sensing, and solve the problem of perception bias easily occurring with a single data source in specific environments.

Description

Technical Field

[0001] This application relates to the field of traffic flow sensing technology, and in particular to a traffic flow sensing method and system. Background Technology

[0002] With the increasing severity of urban traffic congestion and the rapid development of intelligent transportation systems, traffic management departments need to achieve real-time and accurate perception of traffic flow in various road scenarios. In practical applications, traffic flow sensing is widely deployed in target areas such as urban intersections, highways, tunnels, and bridges, and its perception results directly serve key operations such as traffic light control, congestion warning, route guidance, and emergency dispatch. To cope with complex and ever-changing traffic environments, the sensing system is typically required to have all-weather, highly robust data acquisition capabilities to meet the continuous monitoring needs under different lighting, weather, and road conditions.

[0003] Common traffic flow sensing technologies primarily rely on a single type of data source. For example, fixed detection devices based on inductive loops or geomagnetic sensors count traffic flow by analyzing changes in the electromagnetic field caused by passing vehicles; visual detection systems based on video cameras use image processing and deep learning algorithms to identify vehicle trajectories and lane occupancy; and floating car data based on GPS or mobile communication signaling uses location information from onboard terminals or smartphones to calculate average vehicle speed and travel time for road segments. In addition, some systems incorporate microwave radar, lidar, or weather sensors as auxiliary means. These technologies typically operate independently, uploading the raw data directly to a central cloud platform for further processing and display.

[0004] However, single-data-source solutions are prone to perception bias in specific environments, making it difficult to guarantee overall perception accuracy. Specifically, inductive loop detectors are prone to missed detections when the road surface is damaged or there is strong electromagnetic interference; video cameras experience a significant decrease in detection accuracy at night, in rain, snow, fog, or backlight conditions; and mobile terminal data cannot represent real traffic flow in areas with low penetration or severe signal obstruction (such as tunnels or under overpasses). Therefore, there is an urgent need for a traffic flow perception method and system that can dynamically adjust the weights of multi-source data fusion based on real-time environmental conditions to improve the accuracy and robustness of perception results in complex scenarios. Summary of the Invention

[0005] This application provides a traffic flow sensing method and system to solve the problem that a single data source is prone to sensing bias in specific environments.

[0006] The first aspect of this application provides a traffic flow sensing method, including: Within the target perception area, multi-source heterogeneous traffic perception data from fixed perception devices, mobile perception terminals, and environmental sensors are collected. The multi-source heterogeneous traffic sensing data is sent to edge computing nodes for preliminary processing and feature extraction to generate structured feature data; The structured feature data is uploaded to the cloud platform; In the cloud platform, the structured feature data is uniformly aligned to a spatiotemporal reference. After the structured feature data has been aligned with the unified spatiotemporal reference, a multi-source data fusion process based on dynamic confidence weighting is performed to obtain the fused traffic parameters. The dynamic confidence weighting is calculated based on the perception environment state corresponding to each data source. The parameters of the perception environment state include at least one or more of the following: illumination conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions within the target perception area. The weights of the observations from each data source in the fusion are dynamically adjusted according to the changes in the perception environment state. The dynamic confidence weighting is calculated through a pre-generated environment-weight mapping model, which is used to map the parameter combination of the perceived environment state at the current moment into the confidence weight value of each data source. Based on the fused traffic parameters, traffic situation information is generated that includes at least two of the following: traffic flow, average vehicle speed, lane occupancy, and traffic event type.

[0007] The aforementioned method collects multi-source heterogeneous traffic sensing data from fixed sensing devices, mobile sensing terminals, and environmental sensors. This data is processed into structured feature data at edge computing nodes before being uploaded to a cloud platform. On the cloud platform, the data undergoes spatiotemporal alignment and multi-source data fusion processing based on dynamic confidence weighting. The confidence weights are dynamically calculated based on parameters of the current sensing environment. This method leverages multi-source data fusion to mitigate sensing biases from single data sources in specific environments, improving the accuracy and robustness of traffic parameter estimation and thus obtaining more reliable traffic situation information.

[0008] Optionally, the environmental sensor includes a meteorological sensor for collecting visibility and rain / fog intensity information; If the target perception area is a mountainous road with limited terrain and variable weather, the calculation of the dynamic confidence weighting includes: Based on data collected by meteorological sensors, assess the local visibility or the intensity of rain and fog in the target sensing area; When the evaluation result is lower than the first threshold, the weight of visual sensing devices in data fusion is reduced based on a preset meteorological adjustment coefficient.

[0009] The above scheme can dynamically adjust the weight of the data source according to the variable weather conditions in mountainous areas, alleviate the data deviation of visual sensing devices caused by visibility or rain and fog, and help improve the reliability of multi-source traffic data fusion results in complex mountain road environments.

[0010] Optionally, the method further includes: Acquire historical baseline data related to the current target perception area, collected by visual sensing devices under historical good visibility conditions. When the evaluation result is lower than the first threshold, the Kalman filter algorithm is used to compensate or correct the current observation of the visual perception device based on the residual analysis between historical benchmark data and the current observation. The compensated or corrected observations are combined with the reduced weights and used in the multi-source data fusion process.

[0011] The above method can use data from historical good visibility conditions to correct data, which can alleviate the problems that may arise from directly discarding or completely trusting visual data due to low visibility. This improves the availability of observations from visual sensing devices under low visibility conditions. Furthermore, by combining the compensated or corrected observations with the adjusted weights, the method can participate in multi-source data fusion processing, which helps to improve the accuracy of fused data under adverse weather conditions.

[0012] Optionally, the method further includes: If the target perception area is an area affected by periodic social activities, identify or predict the temporal patterns and intensity of the periodic social activities. Based on the aforementioned time patterns and intensity, an initial confidence adjustment factor is set for the mobile sensing terminal data collected during the periodic social activities. When performing the dynamic confidence weighting, the initial confidence adjustment factor is used as one of the input parameters.

[0013] The above method can take into account the potential impact of periodic social activities on the data quality of mobile sensing terminals. By introducing an initial confidence adjustment factor related to the activity pattern, it can mitigate data fluctuations or anomalies caused by peak periods or specific time periods of social activities. This helps to more reasonably assess the credibility of mobile sensing data in data fusion, thereby improving the adaptability of traffic situation perception in periodic social activity scenarios.

[0014] Optionally, the method further includes: If the periodic social activity is a market activity, a traffic flow prediction model related to the intensity of the market activity is generated based on the historical market activity data of the target sensing area. When a market event is predicted or detected, the traffic flow prediction model is run to generate predicted traffic parameters. The predicted traffic parameters are combined with the fused traffic parameters for traffic incident identification or traffic guidance decision-making.

[0015] The above method can predict traffic conditions during market activities by utilizing historical activity patterns, alleviating the difficulty in interpreting real-time perception data caused by sudden changes in traffic patterns due to market activities. By combining predicted traffic parameters with real-time fused traffic parameters for traffic event identification or traffic guidance decision-making, it helps to improve the comprehensiveness of traffic situation analysis and the effectiveness of decision support in specific social activity scenarios such as market activities.

[0016] Optionally, the fixed sensing device includes radar and geomagnetic coils, and the mobile sensing terminal data includes Bluetooth or Wi-Fi beacon data; If the target sensing area is an underground, tunnel, or high-rise building area with severe multipath effects where signal reception is limited, the calculation of the dynamic confidence weighting includes: Assess the signal strength, number of available satellites, and positioning accuracy fluctuations of GNSS positioning data from mobile sensing terminals; Based on the evaluation results, the fusion weights of GNSS data and other sensing data sources are dynamically adjusted.

[0017] The above scheme can assess the quality of GNSS data based on actual signal reception conditions, mitigate the errors that may be introduced by directly using GNSS data in special scenarios due to poor or unstable GNSS signals, and help improve the reliability of multi-source traffic data fusion results in underground, tunnel, or high-rise areas by dynamically adjusting its fusion weights.

[0018] Optionally, the assessment of the perceived environmental state may be combined with at least one of the historical traffic flow patterns of the target perception area, real-time event information, and a preset sensor performance degradation model as input to the environment-weight mapping model.

[0019] The above solution can integrate multi-dimensional information such as historical operating patterns, real-time external events, and the equipment's own status, alleviating the limitations of relying solely on current physical environment parameters to assess the perceived environmental status. By providing more comprehensive input, it helps improve the adaptability of the environment-weight mapping model to complex real-world traffic scenarios, making the weight allocation of dynamic confidence weighting more reasonable, thereby improving the accuracy of multi-source data fusion.

[0020] A second aspect of this application provides a traffic flow sensing system, applicable to the traffic flow sensing method described in the first aspect, comprising: The data acquisition module, deployed in the target sensing area, is used to collect multi-source heterogeneous traffic sensing data from fixed sensing devices, mobile sensing terminals, and environmental sensors. The data processing and fusion module is used to preprocess the multi-source heterogeneous traffic perception data, unify and align the spatiotemporal references, and perform multi-source data fusion processing based on dynamic confidence weighting. The data processing and fusion module is configured to: calculate based at least on the sensing environment state corresponding to each data source, wherein the parameters of the sensing environment state include at least one or more of the following: illumination conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions within the target sensing area; and dynamically adjust the weight of each data source observation in the fusion according to the changes in the sensing environment state. The data processing and fusion module includes an environment-weight mapping model, which is used to map the parameter combination of the perceived environment state at the current moment into the confidence weight value of each data source. The analysis module is used to generate traffic situation information based on the fused traffic parameters, including at least two of the following: traffic flow, average vehicle speed, lane occupancy, and traffic event type.

[0021] The system provided in the second aspect of this application acquires multi-source heterogeneous traffic sensing data from fixed sensing devices, mobile sensing terminals, and environmental sensors through a data acquisition module. The data processing and fusion module preprocesses the data, aligns it to a unified spatiotemporal reference, and performs multi-source data fusion processing based on dynamic confidence weighting. This fusion dynamically adjusts the weights of each data source in the fusion process based on at least one of the sensing environmental conditions, including illumination conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions. This system can utilize multi-source data fusion to mitigate the sensing bias of a single data source in specific environments, helping to improve the accuracy and robustness of traffic parameter estimation, thereby obtaining more reliable traffic situation information.

[0022] Optionally, the data processing and fusion module includes an environmental status assessment unit, which is used for: Receive data from environmental sensors and basic information about the target perception area; Based on a pre-defined environmental impact model, the impact of the current sensing environment on the performance of each data source is quantitatively assessed. The degree of influence is output as the basis for calculating the dynamic confidence weight.

[0023] This environmental state assessment unit receives data from environmental sensors and basic information about the target sensing area. Based on a pre-defined environmental impact model, it quantifies the impact of the current sensing environment on the performance of each data source. This assessment result serves as the basis for calculating dynamic confidence weights. This unit can transform qualitative environmental information into a quantitative assessment of its impact on the performance of each data source, mitigating the subjectivity of adjusting weights solely based on raw environmental state parameters. By providing a quantified degree of impact, it helps improve the objectivity and accuracy of dynamic confidence weight calculation, making the weight allocation of multi-source data fusion more consistent with the actual impact of the sensing environment on the data sources.

[0024] Optionally, the system may also include a historical database and a prediction module; The historical database is used to store historical traffic data of the target sensing area under different sensing environment conditions and historical performance data of each data source. The prediction module is configured to predict future traffic flow patterns or environmental change trends in the target sensing area based on historical data, and provide the prediction results to the data processing and fusion module for pre-adjusting the dynamic confidence weights.

[0025] The system can use historical data to predict future trends, mitigating the potential lag in adjusting weights based on current data when sudden changes occur in the perceived environment or traffic patterns. By pre-adjusting confidence weights, it helps improve the adaptability of multi-source data fusion processing to future traffic situation changes, making traffic situation perception more forward-looking.

[0026] As can be seen from the above technical solutions, this application provides a traffic flow perception method and system. This involves collecting multi-source heterogeneous traffic perception data from fixed sensing devices, mobile sensing terminals, and environmental sensors within a target perception area; sending the multi-source heterogeneous traffic perception data to an edge computing node for preliminary processing and feature extraction to generate structured feature data; uploading the structured feature data to a cloud platform; performing spatiotemporal benchmark alignment on the structured feature data in the cloud platform; and performing multi-source data fusion processing based on dynamic confidence weighting on the structured feature data after spatiotemporal benchmark alignment to obtain fused traffic parameters. The dynamic confidence weighting is calculated using a pre-generated environment-weight mapping model, which maps the parameter combination of the current perception environment state to the confidence weight values ​​of each data source. Based on the fused traffic parameters, traffic situation information including at least two of traffic flow, average vehicle speed, lane occupancy, and traffic event type is generated to solve the problem of perception bias easily occurring from a single data source in a specific environment. Attached Figure Description

[0027] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart illustrating a traffic flow sensing method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the overall architecture of a traffic flow sensing system provided in an embodiment of this application. Detailed Implementation

[0029] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application.

[0030] To address the issue of perception bias that can easily occur with a single data source in specific environments, see [link to relevant documentation]. Figure 1 This application provides a traffic flow sensing method in some embodiments, including: S100: Within the target perception area, collect multi-source heterogeneous traffic perception data from fixed perception devices, mobile perception terminals, and environmental sensors.

[0031] The fixed sensing equipment includes high-definition cameras, microwave radar, coil detectors, and geomagnetic detectors installed at key locations such as road intersections, road sections, and tunnel entrances and exits. High-definition cameras capture real-time traffic scene videos, extracting visual feature information such as vehicle color, model, and license plate, and calculate traffic flow, vehicle speed, and queue length based on video image analysis technology. Microwave radar utilizes the Doppler effect principle, emitting microwave signals towards the target area and detecting the presence, speed, and direction of travel of vehicles by receiving frequency changes in the reflected signals, offering the advantage of being unaffected by weather conditions. Coil detectors use induction coils buried under the road surface; when a vehicle passes, the change in coil inductance triggers a detection signal, accurately counting the number of vehicles passing per unit time. Geomagnetic detectors detect the presence and driving status of vehicles by sensing changes in the geomagnetic field as vehicles pass. Mobile sensing terminals include in-vehicle GPS (Global Positioning System) positioning devices installed in commercial vehicles (such as taxis and buses) and private cars, as well as smartphone navigation applications. These devices provide supplementary information from a floating car perspective for traffic flow analysis by uploading real-time dynamic data such as vehicle location coordinates, speed, and trajectory. Environmental sensors include temperature and humidity sensors, visibility sensors, and rainfall sensors placed along roadsides. Temperature and humidity sensors collect ambient temperature and humidity data to provide a basis for analyzing the impact of extreme weather on traffic flow. Visibility sensors monitor visibility in real-time by measuring the scattering characteristics of light by particulate matter in the atmosphere; when visibility falls below a threshold, it helps determine the potential impact on traffic flow. Rainfall sensors accurately measure rainfall amounts, providing data support for assessing traffic operation under adverse weather conditions such as rain. These multi-source, heterogeneous traffic sensing data possess different data formats, sampling frequencies, and spatiotemporal resolutions. For example, data collected by fixed sensing devices is mostly structured or semi-structured, with a high sampling frequency and fixed spatial location; data uploaded by mobile sensing terminals is unstructured trajectory data, with a sampling frequency affected by terminal equipment and network conditions, and a wider spatial coverage; environmental sensor data is structured environmental parameter data with a relatively low sampling frequency. By collecting these different types of sensing data, the traffic operation status of the target sensing area can be comprehensively characterized from multiple dimensions, laying the foundation for subsequent data fusion and accurate traffic flow perception.

[0032] S200: Sends multi-source heterogeneous traffic perception data to edge computing nodes for preliminary processing and feature extraction to generate structured feature data.

[0033] Specifically, the edge computing nodes perform preliminary processing and feature extraction on multi-source heterogeneous traffic sensing data. Differentiated strategies are adopted for different types of sensing data to ensure the consistency and usability of the generated structured feature data. For structured or semi-structured data collected by fixed sensing devices, such as traffic flow, speed, and occupancy rates collected by loop detectors, the edge computing nodes first perform data cleaning, removing outliers and missing values ​​caused by equipment failure or interference. Interpolation methods or prediction models based on historical data are used to fill in the missing data. Next, data standardization is performed, unifying parameters of different magnitudes to the same unit, such as unifying the speed unit to kilometers per hour. Finally, key statistical features are extracted, such as average traffic flow per minute, speed variance, and peak hourly traffic flow, and aggregated according to preset time windows (e.g., 5 minutes, 15 minutes) to form a structured feature vector indexed by timestamps and device IDs. For unstructured trajectory data uploaded by mobile sensing terminals, edge computing nodes first perform trajectory parsing and mapping, matching the original latitude and longitude coordinate sequences to the road network of the electronic map to determine the specific road segment and direction of the vehicle's travel. Then, trajectory segmentation and feature extraction are performed, dividing the trajectory into several segments based on time intervals or spatial distances. The average speed, travel time, turning angle, acceleration, and road network-level features such as segment travel time and average speed are calculated for each segment. Simultaneously, the trajectory data is deduplicated and abnormal trajectory filtered, removing duplicate uploads or trajectories that significantly deviate from normal driving patterns (such as prolonged stillness or abnormal speed). Finally, the extracted features are associated with the corresponding road segment ID and time information to form structured trajectory feature data. For structured environmental parameter data collected by environmental sensors, such as temperature, humidity, visibility, and rainfall, edge computing nodes primarily perform data verification and time alignment to ensure data accuracy and integrity. This data is synchronized with the time windows of fixed sensing devices and mobile sensing terminals, extracting statistical features such as the average, maximum, and minimum values ​​of environmental parameters within each time window to construct structured feature data containing environmental factors. Through the above processing, the original multi-source heterogeneous data with different formats, frequencies, and dimensions are transformed into standardized structured feature data with unified time granularity, including equipment / road segment identification, traffic flow parameters, vehicle motion characteristics, and environmental influencing factors. This provides high-quality input for subsequent higher-level data fusion and precise traffic flow perception and analysis on the cloud platform.

[0034] S300: Uploads structured feature data to the cloud platform.

[0035] Specifically, structured feature data is encapsulated according to a pre-defined unified data protocol and interface specification. Structured feature data typically contains multiple layers of information structure: First, there is the basic identification layer, which contains unique identifiers of the data source, such as fixed sensing device IDs (Identity Documents), mobile sensing terminal IDs (such as the unique code of vehicle terminals), road segment IDs, and precise timestamp information, ensuring that each piece of data can be accurately traced and located to a specific spatiotemporal scenario; second, there is the core feature layer, which is subdivided into traffic flow feature sub-layers (such as traffic volume, average vehicle speed, and headway distribution), vehicle motion feature sub-layers (such as instantaneous vehicle speed, acceleration, and turning angle), and environmental feature sub-layers (such as average temperature and humidity, maximum rainfall, and visibility level within a time window), each of which contains specific values, units, and data collection frequency information; finally, there is the data verification layer, which records the verification results, outlier markers, and data integrity indicators during the data processing at edge computing nodes, facilitating secondary verification and quality assessment by the cloud platform after receiving the data. Through this multi-level, structured data organization method, the cloud platform can efficiently parse, store, and index uploaded data, laying a solid foundation for subsequent data fusion and analysis.

[0036] S400: In the cloud platform, structured feature data is uniformly aligned to a spatiotemporal reference.

[0037] Specifically, this spatiotemporal benchmark alignment process mainly includes two key stages: time benchmark alignment and spatial benchmark alignment. Regarding time benchmark alignment, the cloud platform uses its own high-precision time synchronization system (such as a standard time source based on GPS or BeiDou) as a reference to uniformly calibrate the timestamp information in all accessed structured feature data. For potential clock drift or time synchronization errors from different sensing devices, the system uses a timestamp mapping algorithm to accurately convert the collection time of each data point to the cloud platform's standard time axis, ensuring data consistency in the time dimension. For example, traffic flow data reported by different devices within the same minute are accurately merged into the same time window. Regarding spatial benchmark alignment, the cloud platform uses a unified geographic coordinate system (such as the WGS-84 coordinate system or the National Geodetic Coordinate System 2000) as the spatial benchmark. For the local coordinates or relative position descriptions that different sensing devices (such as fixed cameras, floating car GPS, and roadside radar) may use, the platform uses coordinate transformation models (such as affine transformation and projection transformation) to uniformly transform their spatial position information to the standard coordinate system. For regional identifiers such as road segment IDs, the system will establish a road segment spatial topology database and associate it with specific geographical coordinate ranges to ensure that the data can be accurately mapped to the actual road network spatial location. This enables precise matching and unification of structured feature data from different sources and of different types in the spatiotemporal dimension, providing a consistent data foundation for subsequent cross-device and cross-regional traffic flow fusion analysis.

[0038] S500: Performs multi-source data fusion processing based on dynamic confidence weighting on the structured feature data after the spatiotemporal reference has been unified and aligned, in order to obtain the fused traffic parameters.

[0039] The dynamic confidence weighting is calculated based on the perceived environmental state of each data source. The parameters of the perceived environmental state include at least one or more of the following: lighting conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions within the target perception area. The weights of the observations from each data source in the fusion are dynamically adjusted according to changes in the perceived environmental state.

[0040] Specifically, the dynamic confidence-weighted multi-source data fusion processing module includes the following core structure and process: First, the system constructs a perception environment state assessment submodule for each data source (such as cameras, radar, and floating car GPS). This submodule receives or queries real-time data from light sensors (such as light intensity and backlighting), weather stations (such as sunny, rainy, snowy, and foggy weather conditions and corresponding visibility values), electromagnetic environment monitoring equipment (such as electromagnetic interference intensity levels), and road surface monitoring data (such as dry, humid, waterlogged, and snowy conditions) within the target perception area. These parameters are quantified into specific numerical indicators. For example, light conditions can be divided into continuous values ​​from 0 to 100, where 0 represents extreme darkness and 100 represents strong light. Weather conditions can be assigned different weight scores based on their impact on perception; for example, heavy fog may correspond to a lower score, while sunny weather may correspond to a higher score.

[0041] Next, the system incorporates a dynamic confidence calculation model. This model has pre-defined confidence mapping rules or functions for different data sources under different sensing environment conditions. For example, for camera data that relies on image recognition, its confidence level will be significantly lowered in extremely poor lighting conditions (such as nighttime without streetlights and heavy rain) or extremely low visibility (such as dense fog). For millimeter-wave radar, its confidence level is less affected by lighting and visibility, but it will decrease accordingly in environments with strong electromagnetic interference. The model dynamically calculates the real-time confidence value of each data source based on the various sensing environment state parameters collected in real time, using weighted summation, fuzzy logic reasoning, or machine learning algorithms (such as regression models trained on historical data).

[0042] Subsequently, in the data fusion phase, the system employs fusion algorithms such as weighted average or weighted least squares to correlate the traffic parameter observations (such as traffic flow, average vehicle speed, and occupancy rate) output from each data source with their corresponding real-time confidence values. Data sources with higher confidence values ​​have a greater weight in the fusion result; conversely, data sources with low confidence values ​​have reduced weights, and in extreme cases (such as confidence values ​​below a set threshold), their observations may be temporarily excluded from the fusion calculation. For example, if a camera's confidence value drops to 0.3 due to heavy rain, while the radar's confidence value is 0.8, the fused traffic flow parameters will primarily be based on radar data, supplemented by a small amount of camera data for correction.

[0043] In addition, the system incorporates a dynamic adjustment trigger mechanism. When the perceived environmental conditions change significantly (e.g., weather changes from sunny to rainy, or light intensity drops by more than 50%), the system immediately triggers a recalculation of the confidence score and a real-time update of the weights. This ensures the fusion process can quickly adapt to environmental changes and maintain high fusion accuracy. Simultaneously, the system records the confidence score weight adjustments before and after each environmental change, along with the corresponding fusion result errors. This data is used to continuously optimize the parameters of the dynamic confidence score calculation model, enhancing its adaptability to complex and changing perception environments. Through this dynamic confidence score-weighted fusion processing, the impact of measurement errors from a single data source in harsh environments on the overall accuracy of traffic parameters can be effectively reduced, resulting in more robust and reliable fused traffic parameters.

[0044] Dynamic confidence weighting is calculated using a pre-generated environment-weight mapping model, which maps the parameter combination of the perceived environment state at the current moment to the confidence weight values ​​of each data source.

[0045] It should be understood that this environment-weight mapping model can employ a multi-layer neural network structure, specifically including an input layer, hidden layers, and an output layer. The input layer receives a combination of multi-dimensional perceived environmental state parameters at the current moment. These parameters encompass weather conditions (e.g., sunny, cloudy, rainy, snowy, fog, encoded into numerical vectors), light intensity (in lux), visibility (in meters), road conditions (e.g., dry, wet, flooded, icy, also encoded into numerical vectors), and time period information (e.g., peak hours, off-peak hours, nighttime). The hidden layer contains at least two layers of neurons. The first layer has 64 neurons, employing the ReLU (Rectified Linear Unit) activation function to extract the non-linear feature interactions between environmental parameters. The second layer has 32 neurons, also employing the ReLU activation function, to further reduce the dimensionality and abstract the features. The output layer corresponds to various data sources (e.g., radar, cameras). The output value of each neuron is the confidence weight of that data source, and the sum of the weights of all output neurons is normalized to 1 to ensure the rationality of weight allocation. The training process of the model uses historical environmental state data, measured error data from corresponding data sources, and manually labeled empirical weight values ​​as training samples. The network parameters are continuously adjusted through the backpropagation algorithm so that the weight values ​​output by the model can minimize the error of the fusion result, thereby establishing an accurate mapping relationship between environmental state and confidence weight.

[0046] S600: Based on the fused traffic parameters, generate traffic situation information including at least two of the following: traffic flow, average vehicle speed, lane occupancy, and traffic event type.

[0047] Traffic flow specifically refers to the number of vehicles passing through a specific monitoring section per unit time, usually measured in "vehicles / hour". Its calculation requires combining vehicle detection data and time window information from the fused traffic parameters, and is derived by accumulating the effective vehicle identification results within a unit time. Average vehicle speed is obtained by performing an arithmetic or weighted average on the fused vehicle speed data. When using a weighted average, different weights can be assigned based on vehicle type (e.g., small cars, large cars) to better reflect actual traffic flow characteristics. Lane occupancy is calculated using the fused vehicle trajectory data and lane space parameters, specifically the cumulative number of vehicles in the target lane per unit time. The ratio of occupancy time to the time window, or the ratio of vehicle projection area to lane effective area, directly reflects the congestion level of the lane. The determination of traffic event type requires pattern matching of the fused abnormal traffic parameters. For example, when a sudden drop in vehicle speed, abnormally narrowed vehicle distance, and significantly lower traffic flow than the historical average are detected on a certain road segment, the system determines it as a "traffic congestion event." If a vehicle stops for more than a set threshold and shows no signs of normal driving, it is classified as a "vehicle malfunction event." At the same time, the event type can be further subdivided by combining image features collected by cameras (such as vehicle collision marks and abnormal object spillage) to ensure the accuracy and richness of traffic situation information.

[0048] The aforementioned method collects multi-source heterogeneous traffic sensing data from fixed sensing devices, mobile sensing terminals, and environmental sensors. This data is processed into structured feature data at edge computing nodes before being uploaded to a cloud platform. On the cloud platform, the data undergoes spatiotemporal alignment and multi-source data fusion processing based on dynamic confidence weighting. The confidence weights are dynamically calculated based on parameters of the current sensing environment. This method leverages multi-source data fusion to mitigate sensing biases from single data sources in specific environments, improving the accuracy and robustness of traffic parameter estimation and thus obtaining more reliable traffic situation information.

[0049] In some embodiments, the environmental sensor includes a meteorological sensor for collecting visibility and rain / fog intensity information; if the target sensing area is a mountainous road with limited terrain and variable weather, the calculation of dynamic confidence weighting includes: assessing the local visibility or rain / fog impact intensity of the target sensing area based on the data collected by the meteorological sensor; when the assessment result is lower than a first threshold, reducing the weight of visual sensing devices in data fusion based on a preset meteorological adjustment coefficient.

[0050] Specifically, in mountainous road scenarios with limited terrain and variable weather, the core of dynamic confidence weighting calculation lies in constructing a dynamic adjustment mechanism based on real-time meteorological data. First, the system pre-defines a complete meteorological impact assessment model. The model's inputs are visibility values ​​(in meters) and rain / fog intensity values ​​(which can be quantified using levels such as 1-10) collected by meteorological sensors. The output is a meteorological impact factor between 0 and 1. For example, when visibility is greater than 500 meters and the rain / fog intensity level is below 2, the meteorological impact factor approaches 0, indicating that meteorological conditions have minimal impact on visual sensing devices. When visibility drops below 100 meters or the rain / fog intensity level reaches 8, the meteorological impact factor approaches 1, representing a significant impact.

[0051] The first threshold corresponds to a critical value for the meteorological impact factor, for example, set to 0.6. When the meteorological impact factor calculated from meteorological sensor data is lower than 0.6 (i.e., the assessment result is lower than the first threshold), the system will activate a preset meteorological adjustment coefficient to reduce the weight of visual sensing devices (such as cameras, lidar, and other devices that rely on light or image recognition). This meteorological adjustment coefficient is not a fixed value, but is dynamically determined based on the specific value of the meteorological impact factor. For example, it can use a linear function relationship: adjustment coefficient = 1 - meteorological impact factor. In this case, the original weight of visual sensing devices needs to be multiplied by this adjustment coefficient. If the initial weight of visual sensing devices is 0.7, when the meteorological impact factor is 0.5 (lower than the first threshold of 0.6), the adjusted weight will be 0.7 × (1 - 0.5) = 0.35. At the same time, the weight of non-visual sensing devices (such as microwave radar, geomagnetic coil detectors, and other devices less affected by weather) will be increased accordingly to ensure that the overall weight sum is 1 during the multi-source data fusion process, thereby minimizing the impact of visual data distortion caused by severe weather on the traffic situation assessment results.

[0052] In some embodiments, the method further includes: acquiring historical baseline data related to the current target perception area collected by the visual sensing device under historical good visibility conditions; when the evaluation result is lower than a first threshold, using a Kalman filter algorithm to compensate or correct the current observation value of the visual sensing device based on the residual analysis between the historical baseline data and the current observation value; and combining the compensated or corrected observation value with the reduced weight to participate in multi-source data fusion processing.

[0053] Specifically, the first step is the construction and storage of historical benchmark data. The system needs to pre-define quantitative standards for "good visibility conditions," such as visibility greater than or equal to 10 kilometers and the absence of weather phenomena affecting light propagation, such as rain, snow, fog, or dust storms. Under these conditions, visual sensing devices (such as high-definition cameras at specific intersections) continuously collect data from the target sensing area (such as a specific lane at the intersection or the entire intersection area). The collection period can be set according to actual needs, for example, continuous collection for 7 days, collecting 24 hours of video stream or image frame data each day. The collected raw data undergoes preprocessing, including but not limited to image denoising, region of interest extraction, and target detection and tracking (such as identifying and tracking the contours, number, and trajectories of passing vehicles), thereby generating a historical benchmark dataset containing key traffic parameters such as vehicle flow, average speed, and headway. This historical benchmark dataset is categorized and stored according to time dimensions (such as time periods, peak hours, off-peak hours) and spatial dimensions (such as different lanes, different directions), forming a structured database that serves as a reference benchmark for subsequent residual analysis.

[0054] Secondly, the residual analysis process is as follows. When the meteorological impact factor assessment result is below the first threshold (i.e., the current meteorological conditions are deemed to significantly interfere with visual perception), such as when visibility is below 500 meters, rainfall intensity reaches the moderate rain level (5-15 mm of rainfall in 12 hours), or road surface water depth exceeds 3 cm, the meteorological impact factor assessment result is determined to be below the first threshold. The system calls the current observation values ​​collected by visual perception devices in real time (which undergo similar preprocessing to obtain parameters such as current vehicle flow and speed), and retrieves the historical benchmark data that best matches the current time (e.g., the same weekday / weekend, the same time period) and spatial conditions from the historical benchmark database. Then, the residual between the current observation value and the corresponding historical benchmark data is calculated, i.e., residual = current observation value - historical benchmark data. For example, if the average traffic flow of this lane in a certain time period in the historical benchmark data is 100 vehicles / minute, while the current observation value is only 60 vehicles / minute, then the residual is -40 vehicles / minute, indicating that the current visual data may have missed vehicle detection due to meteorological reasons, resulting in an underestimation of the observed value.

[0055] Secondly, there is a compensation or correction mechanism based on Kalman filtering. The Kalman filtering algorithm establishes the system's state equation and observation equation, using the estimated value from the previous moment and the observed value from the current moment to update the optimal estimate of the system state. In this scenario, the observation values ​​of visual sensing devices are considered as measurements affected by noise (mainly observation noise caused by meteorological interference). The system initializes the Kalman filter's state vector (containing state variables such as traffic flow and vehicle speed), covariance matrix (representing the uncertainty of the initial state estimate), state transition matrix (describing how the system state changes over time, which can be set based on the statistical characteristics of historical data or traffic flow models), and observation matrix. Using the residuals calculated in step two as one of the inputs to the Kalman filter, combined with the current observation value, the filter automatically adjusts its gain to dynamically compensate or correct the current observation value. For example, in the case of a residual of -40 vehicles / minute, the Kalman filter algorithm may correct the current observation of 60 vehicles / minute to 85 vehicles / minute, which is closer to the true value, based on the statistical characteristics of the noise (such as the variance of the residual under historical weather interference). The specific correction value is calculated by the filtering algorithm, thereby reducing the observation bias caused by weather factors.

[0056] Finally, the corrected data is fused and applied. Visual perception device observations, after Kalman filtering compensation or correction, no longer use the original weights directly, but instead adopt weights adjusted by a meteorological adjustment factor (such as 0.35 in the previous example). This corrected value is multiplied by the adjusted weight to obtain the weighted contribution of visual data in multi-source data fusion. Simultaneously, observations from non-visual perception devices (such as microwave radar) are weighted using correspondingly enhanced weights. All weighted data from different sources are integrated according to preset fusion rules (such as weighted average, Bayesian estimation), ultimately outputting a more accurate and robust traffic situation assessment result. This approach, combining historical benchmarks with real-time filtering correction, more effectively utilizes the valuable information still contained in visual data while reducing its negative impact from meteorological interference, further improving the reliability of multi-source data fusion.

[0057] In some embodiments, the method further includes: if the target sensing area is an area affected by periodic social activities, identifying or predicting the temporal patterns and intensity of the periodic social activities; based on the temporal patterns and intensity, setting an initial confidence adjustment factor for the mobile sensing terminal data collected during the periodic social activities; and using the initial confidence adjustment factor as one of the input parameters when performing dynamic confidence weighting.

[0058] For example, the target sensing area is the roads surrounding a large stadium. This stadium hosts two large concerts per month (periodic social events), and the peak travel time for attendees is from 5:00 PM to 10:00 PM on each event day (a time pattern). Historical data shows that the average deviation rate of mobile sensing terminal data in this area is 40% higher during event days than on non-event days (a quantitative indicator of intensity). Based on this, an initial confidence adjustment factor of 0.6 can be set (i.e., the original confidence multiplied by 0.6). During the dynamic confidence weighting process, when the system identifies that the current time is between 5:00 PM and 10:00 PM and is a concert day, it automatically calls the initial confidence adjustment factor of 0.6 and multiplies it by a dynamically calculated weight in real time (such as 0.8 generated based on parameters such as signal quality and terminal density) to obtain the final comprehensive confidence weight of 0.48 applied to the mobile sensing terminal data, which is then used in the multi-source data fusion calculation.

[0059] In some embodiments, the method further includes: if the periodic social activity is a market activity, generating a traffic flow prediction model associated with the intensity of the market activity based on historical market activity data of the target sensing area; when a market activity is predicted or detected, running the traffic flow prediction model to generate predicted traffic parameters; and combining the predicted traffic parameters with the fused traffic parameters for traffic event identification or traffic guidance decision-making.

[0060] It should be understood that the specific structure of the traffic flow prediction model can adopt a neural network model containing an input layer, hidden layers, and an output layer. The feature parameters of the input layer include, but are not limited to: the dates of historical market events (e.g., weekends, holidays), event duration, event scale (which can be represented by quantitative indicators such as historical number of participants and number of stalls), weather conditions during the event (temperature, probability of precipitation, etc.), basic traffic flow data of surrounding roads (e.g., average vehicle and pedestrian flow during the same time period on non-market days), and information on adjustments to the frequency of public transportation departures in the area. The hidden layer can be set to 2-3 layers, each containing 32-128 neurons, using activation functions such as ReLU or Leaky ReLU for nonlinear feature extraction to capture the complex mapping relationship between the input parameters and traffic flow. The output layer outputs predicted traffic parameters according to actual needs, such as the average vehicle speed, vehicle flow, peak pedestrian flow, and queue length for the road segment in the next 15 minutes, 30 minutes, and 1 hour. During the model training phase, historical market activity data and corresponding traffic data of the target perception area over the past 3-5 years are used as training samples. The gradient descent method (such as the Adam optimizer) is used to iteratively optimize the model parameters. The model prediction accuracy is evaluated by loss functions such as mean squared error (MSE) or mean absolute error (MAE) until the model converges and reaches the preset prediction error threshold.

[0061] In combining predicted traffic parameters with fused traffic parameters, a weighted fusion strategy can be adopted. For example, the weights can be dynamically adjusted based on the real-time progress of the market activity: 1-2 hours before the market activity begins (the stage when pedestrian and vehicle traffic gradually gathers), the weight of the predicted traffic parameters can be set to 0.3-0.5, and the weight of the fused real-time traffic parameters is 0.5-0.7; during the peak period of the market activity (such as from 10:00 AM to 3:00 PM), the weight of the predicted traffic parameters can be increased to 0.6-0.8 to highlight the model's ability to predict changes in traffic flow caused by the activity; while 1-2 hours after the market activity ends (the stage when pedestrian and vehicle traffic disperses), the weight of the predicted traffic parameters can be gradually reduced, and the weight of the real-time fused data can be increased to 0.7-0.9. The combined traffic parameters will serve as the core input for the traffic event identification module. For example, when the traffic flow in the combined traffic parameters suddenly drops by more than 30% and lasts for more than 10 minutes, while the average vehicle speed is lower than 50% of the normal level for the same period in history, the system can determine that a traffic congestion event may occur. For traffic guidance decisions, the combined traffic parameters can be used to generate dynamic navigation routes, such as marking predicted high-flow road sections as congestion warning areas and recommending detour routes to users, or adjusting the timing schemes of surrounding traffic lights to optimize traffic flow distribution.

[0062] In some embodiments, the fixed sensing device includes radar and geomagnetic coils, and the mobile sensing terminal data includes Bluetooth or Wi-Fi beacon data; if the target sensing area is an underground, tunnel, or high-rise building area with severe multipath effects where signal reception is limited, the calculation of dynamic confidence weighting includes: evaluating the signal strength, number of available satellites, and positioning accuracy fluctuations of the GNSS positioning data from the mobile sensing terminal; and dynamically adjusting the fusion weights of GNSS data and other sensing data sources based on the evaluation results.

[0063] Specifically, the calculation process of dynamic confidence weighting can be broken down into the following structure: First, a GNSS (Global Navigation Satellite System) data quality assessment index system is constructed. This system includes at least three core dimensions: signal strength (e.g., measured in dBm, with a signal strength below -130dBm considered weak), the number of available satellites (e.g., fewer than 4 available satellites indicate low positioning availability), and positioning accuracy fluctuation (measured by the standard deviation of positioning results over 5 consecutive seconds; a standard deviation greater than 5 meters indicates significant accuracy fluctuation). Next, quantitative scoring standards and weighting coefficients are set for each dimension. For example, signal strength accounts for 40%, the number of available satellites for 30%, and positioning accuracy fluctuation for 30%. A comprehensive quality score for the GNSS data is obtained through weighted summation, with a score range of 0-100. When the overall quality score is above 80, the GNSS data is considered to have high reliability, and its fusion weight is adjusted to 0.7-0.9, while the weights of other sensing data sources (such as radar or geomagnetic coil data from fixed sensing devices) are adjusted to 0.1-0.3. If the overall quality score is between 40 and 80, the GNSS data is considered to have medium reliability, and its fusion weight is set to 0.4-0.6, while the weights of other sensing data sources are adjusted to 0.4-0.6. When the overall quality score is below 40, the GNSS data is considered to have low reliability, and its fusion weight is reduced to 0.1-0.3, while the weights of other sensing data sources are increased to 0.7-0.9. In addition, the system will set a dynamic adjustment cycle of 1 minute, reassessing the GNSS data quality and updating the fusion weight every minute to adapt to real-time changes in the signal environment in underground, tunnel, or high-rise areas, ensuring the accuracy and stability of traffic parameter sensing.

[0064] In some embodiments, the assessment of the perceived environmental state combines at least one of the following: historical traffic flow patterns of the target perception area, real-time event information, and a preset sensor performance degradation model, as input to the environment-weight mapping model.

[0065] Specifically, historical traffic flow patterns can be quantitatively represented by constructing a time-series database. This database contains key parameters such as the average hourly traffic volume, vehicle type distribution ratio (e.g., the proportion of small cars and large cars), speed distribution characteristics (e.g., average speed and speed standard deviation), and traffic fluctuation coefficients (e.g., the ratio of peak-hour traffic volume to off-peak traffic volume) for the target sensing area over the past 12 months. The system will use a sliding window algorithm (e.g., with a 30-day cycle) to perform trend analysis on historical data, identify typical traffic flow characteristics such as weekday morning and evening peak hours, weekend off-peak hours, and special patterns on holidays, and convert these characteristic parameters into standardized vectors (with values ​​ranging from 0 to 1) as basic inputs for environmental assessment.

[0066] Real-time event information is obtained by connecting to the city's traffic management platform, meteorological service system, and road maintenance database. This information includes event type (e.g., traffic accidents, temporary traffic control, large-scale events, severe weather), event impact range (radius distance centered on the target sensing area, in meters), event duration (minutes), and event severity level (1-5, with level 1 indicating minor impact and level 5 indicating severe congestion). The system encodes this information; for example, it maps event type to a one-hot vector, and normalizes the impact range, duration, and severity to values ​​between 0 and 1, forming real-time dynamic input parameters.

[0067] The preset sensor performance degradation model is a mathematical model built based on sensor factory parameters, installation time, and historical operation and maintenance records. For GNSS receivers, the performance degradation factor mainly considers the signal receiving sensitivity attenuation rate (0.5dB per quarter), the positioning drift error growth rate (0.1 meters per month), and the multipath interference resistance degradation coefficient (5% decrease per year). For fixed sensing devices, such as radar or geomagnetic coils, the focus is on the detection accuracy attenuation (0.3% decrease per month), the response delay increase (20 milliseconds per six months), and the device temperature drift coefficient (2% increase in error for every 10°C increase in temperature). The system dynamically calculates the current performance degradation coefficient (value 0-1, where 1 represents no attenuation and 0 represents complete failure) based on the actual operating time of the sensor and uses it as a correction item for environmental assessment.

[0068] The aforementioned historical traffic flow pattern vector, real-time event information encoding vector, and sensor performance degradation coefficient are integrated into a comprehensive environmental state assessment value (range 0-10) through weighted summation. The historical traffic flow pattern accounts for 40% of the weight, the real-time event information for 40%, and the sensor performance degradation model for 20%. This comprehensive assessment value, along with the GNSS data comprehensive quality score, is input into the environment-weight mapping model to achieve fine-grained dynamic adjustment of the fusion weights. For example, when the comprehensive assessment value is greater than 8 (indicating low environmental interference and good sensor performance), even if the GNSS data quality score is in the middle range of 40-80, its fusion weight can be increased by 0.1; conversely, if the comprehensive assessment value is less than 3 (indicating a complex environment and severe sensor degradation), the GNSS data quality score must reach above 85 to obtain a base weight of 0.7.

[0069] In some embodiments, the environment-weight mapping model is based on a deep neural network, and its training process includes: acquiring historical multi-source sensing data of the target sensing area under various known sensing environment states and the corresponding calibrated true values ​​of real traffic parameters; using the parameters of the historical multi-source sensing data and the corresponding sensing environment states as input features, and using the deviation or consistency measure between the observation values ​​of each data source and the true values ​​of real traffic parameters as the optimization objective, training the deep neural network so that the weight allocation output by the model can minimize the error between the fusion result and the true value.

[0070] Specifically, the deep neural network adopts a three-layer fully connected architecture. The input layer contains 128 neurons, which are used to receive historical multi-source sensing data (such as historical traffic flow pattern vectors, real-time event information encoding vectors, sensor performance attenuation coefficients, etc., totaling 87 features) and sensing environment state parameters (such as weather level, time period type, regional road network complexity, etc., totaling 15 features). The hidden layer is designed with a staggered neuron configuration of 64-32, which performs nonlinear feature mapping through ReLU and LeakyReLU (Leaky Rectified Linear Unit) activation functions, and a batch normalization operation is added between each layer to accelerate convergence. The output layer contains 5 neurons, corresponding to the fusion weights of 5 types of sensing data sources (GNSS positioning data, roadside radar data, video surveillance data, floating car data, and loop detector data), and the Softmax function is used to ensure that the sum of the weights is 1. During training, the Adam (Adaptive Moment Estimation) optimizer was used (with an initial learning rate of 0.001, decaying by 10% every 50 epochs). The loss function was a weighted combination of the mean squared error (MSE) of the fused result and the true value, and the cross-entropy loss (weight ratio 3:1). An L2 regularization term (λ=0.0005) was also introduced to prevent overfitting. The model was trained with 1.2 million samples, covering six time periods (morning rush hour, evening rush hour, off-peak, and late night), eight weather conditions (sunny, rainy, foggy, etc.), and five road network types (main roads, secondary roads, and side roads). Five-fold cross-validation was used to ensure the model's generalization ability under different environmental conditions.

[0071] In some embodiments, during the dynamic confidence weighting process, weight fine-tuning based on the real-time health status of the data source itself is also introduced, including: maintaining a real-time health score for each data source, and the health score is calculated based on at least one of the continuity of the data reported by the data source recently, the number of internal verification exceptions, and the self-diagnosis status; after obtaining the basic weight through the environment-weight mapping model, the basic weight is weighted or multiplied by the real-time health score of the corresponding data source to obtain the weight value finally applied to data fusion.

[0072] Specifically, the real-time health score maintained for each data source can be achieved through the following methods: First, quantization rules and weights are set for the three key indicators of continuity, the number of internal verification exceptions, and the self-diagnosis status respectively. For example, for data continuity, a time window (such as the past 5 minutes) can be set, and the ratio of the number of times the data source successfully reported data to the theoretically expected number of reports within this window is calculated and denoted as C. The value range of C is [0, 1], and the higher the ratio, the higher the continuity score. For the number of internal verification exceptions, the number of times N that the data source reported data triggering the internal verification mechanism (such as data exceeding the reasonable range, format error) within this time window is counted, and a benchmark threshold N0 (such as 3 times) is set. When N = 0, the exception score is 1; when 0 < N ≤ N0, the exception score = 1 - (N / N0); when N > N0, the exception score is directly reduced to 0. For the self-diagnosis status, the data source usually has a built-in self-diagnosis module that outputs statuses such as normal, warning, and fault. These statuses can be mapped to specific scores S, such as normal corresponding to 1, warning corresponding to 0.5, and fault corresponding to 0.

[0073] Then, different weight coefficients are assigned to these three indicators (continuity C, internal verification exception score A, self-diagnosis status score S), such as 0.4, 0.3, and 0.3 respectively. The real-time health score H of this data source is calculated by weighted summation, and the formula is H = 0.4×C + 0.3×A + 0.3×S. The value range of H is also normalized to between [0, 1]. 1 indicates that the current health status of the data source is excellent, and 0 indicates that the health status is extremely poor.

[0074] After obtaining the base weight W_base (output by the environment-weight mapping model) and the real-time health score H, the final weight W_final can be calculated by multiplication, i.e., W_final = W_base × H. This method effectively reflects the direct impact of the data source's health status on its weights: when the data source's health is high (H is close to 1), the base weights are basically unaffected; when the data source's health decreases (H < 1), its weight in the fusion will decrease accordingly, thereby reducing the interference of unreliable data on the fusion results. For example, if a roadside radar has a base weight of 0.25, but its self-diagnostic status is now a warning (S=0.5) due to a recent temporary equipment failure, its data continuity has slightly decreased (C=0.8), and it has one internal verification anomaly (N=1, N0=3, A=1-1 / 3≈0.6667), then its health score H=0.4×0.8+0.3×0.6667+0.3×0.5≈0.32+0.2+0.15=0.67, and its final weight W_final=0.25×0.67≈0.1675. This is lower than the base weight and is consistent with the actual data reliability.

[0075] In some embodiments, when performing spatiotemporal benchmark unification alignment of structured feature data from different edge computing nodes in a cloud platform, if significant differences are found in the traffic parameters reported by adjacent edge computing nodes in the same target sensing area, a collaborative verification process is triggered, including: extracting the original observation values ​​of each data source in the overlapping area covered by adjacent edge computing nodes during the difference period; based on the sensing environment state of the overlapping area, calling the environment-weight mapping model to calculate the theoretical confidence distribution of the data of the two nodes in that state; comparing the difference between the theoretical distribution and the actual reported data to assist in locating abnormal sensing devices or processing errors of edge computing nodes, and accordingly marking or correcting the affected data before fusion.

[0076] When extracting raw observations from various data sources in overlapping areas within different time periods, data with timestamp deviations within a preset threshold (such as ±1 second) are precisely filtered to ensure consistency in the time dimension. Spatially, the latitude and longitude bounding boxes of the overlapping areas are used as filtering conditions to extract raw data at all lane levels within that range, including but not limited to detailed observations such as vehicle speed, vehicle length, vehicle type classification, and headway. The core structure of the environment-weight mapping model includes an input layer, a feature fusion layer, and a confidence output layer. The input layer receives real-time environmental parameters of the overlapping area, such as light intensity (divided into three levels: 0-500 lux, 500-2000 lux, and above 2000 lux), weather conditions (enumerated values ​​such as sunny, cloudy, rainy, snowy, and foggy), road surface conditions (dry, wet, waterlogged, and snowy), and traffic density (vehicles / km). The feature fusion layer performs nonlinear mapping of environmental parameters through a multilayer perceptron (MLP), converting discrete features into continuous environmental impact factors. The confidence output layer outputs the theoretical confidence intervals of different data sources (such as cameras, radar, and loop detectors) in the current environment based on a logistic regression model trained on historical labeled data (e.g., the confidence interval of a camera in foggy conditions is [0.3, 0.5]). When comparing the theoretical distribution with the actual reported data, the Kolmogorov-Smirnov test is used to calculate the deviation between the theoretical confidence distribution and the actual statistical distribution. When the deviation exceeds a set threshold (e.g., 0.2), the anomaly type is further identified using a decision tree model: if the proportion of radar data with actual confidence values ​​below the theoretical lower limit in rainy or snowy weather exceeds 30%, it is marked as "radar equipment interfered with by severe weather"; if the error rate of vehicle classification in camera data increases sharply and exceeds 5% under normal lighting conditions, it is judged as "abnormal image recognition algorithm of edge computing node". For the marked abnormal data, a dynamic correction strategy is adopted: random errors caused by temporary equipment failures are smoothed using the sliding window averaging method; for systematic deviations (e.g., radar speed measurement is consistently too high), a deviation compensation model (e.g., linear regression equation y=0.95x+2.3) is established based on historical calibration data for real-time correction to ensure the reliability of the data before fusion.

[0077] See Figure 2 This application also provides a traffic flow sensing system in some embodiments, applicable to the traffic flow sensing method provided in the above embodiments, including: a data acquisition module deployed in the target sensing area, used to collect multi-source heterogeneous traffic sensing data from fixed sensing devices, mobile sensing terminals and environmental sensors.

[0078] The fixed sensing equipment includes high-definition cameras (with built-in AI chips, supporting 1080P real-time video stream acquisition and basic feature recognition such as vehicle type, license plate, and lane occupancy) installed on poles along the road, microwave radar (operating frequency band 24GHz, coverage radius 150 meters, capable of outputting motion parameters such as vehicle speed, distance, and heading angle), and coil detectors (buried in the road surface, capturing pulse signals when vehicles pass through through electromagnetic induction, used to accurately count traffic flow and headway). The mobile sensing terminals include onboard GPS positioning modules (sampling frequency 1Hz, positioning accuracy ±5 meters) installed in taxis, buses, and other operating vehicles, and real-time driving status data (such as instantaneous speed, acceleration, and braking status) uploaded by the onboard OBU (On-Board Unit). The environmental sensors integrate temperature and humidity sensors (measurement range -40℃~85℃, humidity 0~100%RH), light intensity sensors (range 0~200000 lux), and raindrop sensors. The sensor (rain sensor, with a detection accuracy of 0.1 mm / h, capable of identifying light, moderate, and heavy rain) is used to collect environmental parameters that affect the performance of sensing devices. The data acquisition module uploads multi-source data to the edge computing node in real time via industrial Ethernet (transmission rate 100 Mbps) or a 4G / 5G wireless communication module (supporting LTE Cat.1 or 5G NR protocol). The camera video stream is compressed using H.265 encoding to reduce bandwidth consumption, while radar and sensor data are encapsulated in JSON format, including fields such as device ID, acquisition timestamp, data type, and specific values.

[0079] The data processing and fusion module is used to preprocess multi-source heterogeneous traffic perception data, align spatiotemporal references, and perform multi-source data fusion processing based on dynamic confidence weighting.

[0080] The data processing and fusion module is configured to: calculate based on the perception environment status corresponding to each data source at least once. The parameters of the perception environment status include at least one or more of the following: illumination conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions within the target perception area. The module also dynamically adjusts the weight of each data source observation in the fusion process based on changes in the perception environment status.

[0081] It should be understood that the data processing and fusion module specifically consists of a data preprocessing unit, a spatiotemporal alignment unit, and a dynamic weighted fusion unit. The data preprocessing unit first cleans the raw data. For camera video streams, it uses frame difference to remove abnormal frames caused by sudden changes in lighting or lens contamination, and employs an adaptive threshold segmentation algorithm to enhance vehicle contour features. For radar data, it uses Kalman filtering to remove motion noise while retaining the effective target's distance, speed, and azimuth information. Sensor data is screened for outliers using the 3σ criterion and smoothed using a sliding window mean method. The spatiotemporal alignment unit uses the standard timestamp of the edge computing node as a reference to synchronize the data from each device (synchronization accuracy ≤1ms). Spatially, it uses a preset coordinate transformation matrix (based on Gauss-Kruger projection) to uniformly transform GPS latitude and longitude and radar polar coordinates to a Cartesian coordinate system with the monitoring section as the origin, achieving centimeter-level spatial registration. The dynamic weighted fusion unit incorporates an environment-weight mapping model. For example, in strong light environments (light intensity > 150,000 lux), the weight of camera data is reduced from the default 0.6 to 0.3, and the weight of radar data is increased from 0.3 to 0.55, while an environmental compensation weight of 0.15 is introduced. In heavy rain (rainfall > 20 mm / h), the weight of camera data is further reduced to 0.2, the weight of radar data is increased to 0.7, and the weight of sensor data is maintained at 0.1. By calling a preset weight adjustment function in real time (such as a membership function based on fuzzy logic), dynamic weighting and optimal fusion of observations from each data source are achieved.

[0082] The data processing and fusion module includes an environment-weight mapping model, which is used to map the parameter combination of the perceived environment state at the current moment to the confidence weight value of each data source.

[0083] Specifically, the environment-weight mapping model adopts a three-layer structure. The bottom layer is the environmental parameter input layer, which receives real-time perceived environmental state parameters such as light intensity (lux), rainfall (mm / h), and visibility (m). The middle layer is the fuzzy inference layer, which has a built-in fuzzy rule library trained based on expert experience and historical data. For example, "If the light intensity is extremely strong (>150,000 lux) and the rainfall is moderate (5-20 mm / h), then the reliability of the camera data is low." The input parameters are fuzzified into fuzzy sets such as "low," "medium," and "high" through membership functions (such as triangular membership functions), and inference is performed according to fuzzy rules. The top layer is the weight output layer, which converts the fuzzy inference results into specific numerical weights through defuzzification processing (such as the centroid method). Finally, it outputs the confidence weight values ​​of various data sources such as cameras, radar, and various sensors, and the sum of the weights of all data sources is 1 to ensure the standardization and effectiveness of the fusion calculation. The model supports dynamic updates of the fuzzy rule library and membership function parameters through OTA upgrades to adapt to changes in environmental characteristics in different regions and seasons.

[0084] The analysis module is used to generate traffic situation information based on the fused traffic parameters, including at least two of the following: traffic flow, average vehicle speed, lane occupancy, and traffic event type.

[0085] Specifically, the analysis module can adopt a three-level processing architecture of "data parsing - feature extraction - situation generation". In the data parsing stage, the module first standardizes the fused traffic parameters (such as raw traffic flow counts, vehicle speed distribution, lane space occupancy time, etc.) to unify the data format and units. For example, the speed unit of different lanes is unified to km / h, and the occupancy time is converted into duty cycle (0-1 range). In the feature extraction stage, time-dimensional features are calculated through a sliding time window (the window size can be configured to 5 minutes, 15 minutes, etc.), such as peak traffic flow per unit time, speed standard deviation, occupancy fluctuation coefficient, etc. At the same time, it combines spatial-dimensional features, including traffic flow difference between adjacent lanes, turning traffic ratio at intersections, etc., to form a multi-dimensional feature matrix. During the situation generation phase, a built-in traffic situation assessment model is implemented. This model is trained based on historical traffic data and analyzes the feature matrix using multi-feature fusion algorithms (such as random forests and neural networks). When the traffic flow exceeds a preset threshold (such as 80% of the road's designed capacity) and the average vehicle speed is below 20 km / h, it is determined to be a "congested situation." When the occupancy rate of the same lane exceeds 90% for 5 consecutive minutes without speed data updates, a "traffic event" judgment is triggered. This is further combined with radar echo signal characteristics (such as the duration of abnormal stationary targets) to distinguish event types such as "vehicle malfunction" and "road construction." The module supports customizing the weights and thresholds of situation evaluation indicators through configuration files. The output traffic situation information contains specific values ​​(such as "traffic flow: 1200 vehicles / hour" and "average vehicle speed: 35 km / h") and situation level labels (such as "smooth traffic," "slow traffic," and "congested") in a structured data format (such as JSON) and supports embedding abnormal event types (such as "rear-end collision" and "road flooding") and their location, duration, and other detailed information into the output results, providing quantitative basis for traffic management decisions.

[0086] The system provided in this application acquires multi-source heterogeneous traffic perception data from fixed sensing devices, mobile sensing terminals, and environmental sensors through a data acquisition module. The data processing and fusion module preprocesses the data, aligns it to a unified spatiotemporal reference, and performs multi-source data fusion processing based on dynamic confidence weighting. This fusion dynamically adjusts the weights of each data source in the fusion process based on at least one of the following environmental conditions: lighting conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions. This system can mitigate the perception bias of a single data source in specific environments by utilizing multi-source data fusion, helping to improve the accuracy and robustness of traffic parameter estimation, thereby obtaining more reliable traffic situation information.

[0087] In some embodiments, the data processing and fusion module includes an environmental state assessment unit, which is used to: receive data from environmental sensors and basic information of the target sensing area; quantitatively assess the degree of impact of the current sensing environment on the performance of each data source based on a preset environmental impact model; and output the degree of impact as the basis for calculating the dynamic confidence weight.

[0088] It should be understood that the environmental status assessment unit can specifically consist of a data receiving subunit, a model calculation subunit, and a result output subunit. The data receiving subunit is responsible for interfacing with the environmental sensor network to collect environmental parameters in real time, including light intensity (lux), rainfall (mm / hour), visibility distance (meter), electromagnetic interference field strength (dB / mV / m), and road surface temperature, humidity, and water depth (cm). Simultaneously, this subunit retrieves basic information about the target sensing area from the system database. This information covers static attribute data such as road type (e.g., highway, urban arterial road, secondary arterial road), number of lanes, road surface material (asphalt or concrete), obstruction from surrounding buildings, and the installation location and orientation of fixed sensing equipment (e.g., cameras, radar). The model calculation subunit has a built-in preset environmental impact model. This model is constructed through a combination of historical data training and expert experience calibration, establishing mapping relationships between environmental parameters and sensing performance degradation for different types of data sources (e.g., video cameras, microwave radar, floating car GPS). For example, for video cameras, image recognition accuracy drops significantly when light intensity is below 50 lux or above 10,000 lux. The model quantifies the impact coefficient on the confidence level of camera data based on the actual light intensity. For microwave radar, velocity measurement accuracy is affected when rainfall intensity exceeds 50 mm / h or electromagnetic interference field strength exceeds -80 dBmV / m. The model also calculates the corresponding impact value. This sub-unit normalizes the received environmental parameters and inputs them into the environmental impact model to calculate the comprehensive impact of the current environmental state on each type of data source, including video data sources, radar data sources, and floating car data sources. This value is usually represented by a coefficient between 0 and 1; the closer to 0, the greater the impact and the lower the confidence level of the data source; the closer to 1, the smaller the impact and the higher the confidence level of the data source. The output subunit then packages the influence coefficients of each data source obtained by the model calculation subunit into a preset data format (such as a structure containing data source ID, influence coefficient, and evaluation timestamp), and sends it in real time to the dynamic confidence weighting calculation module. This serves as the core basis for dynamically adjusting the weights of each data source during the subsequent multi-source data fusion process.

[0089] In some embodiments, the system further includes a historical database and a prediction module; the historical database is used to store historical traffic data of the target sensing area under different sensing environment states and historical performance data of each data source; the prediction module is configured to: predict the future traffic flow pattern or environmental change trend of the target sensing area based on historical data, and provide the prediction results to the data processing and fusion module for pre-adjusting the dynamic confidence weight.

[0090] It should be understood that the historical database may specifically include multiple sub-databases, such as a traffic flow basic information database, an environmental parameter historical database, and a data source performance index database. The traffic flow basic information database stores historical traffic data such as traffic volume, average vehicle speed, and headway in the target sensing area under different time periods (e.g., weekday peak hours, off-peak hours, holidays) and different weather conditions (sunny, rainy, snowy, foggy, etc.). The environmental parameter historical database records historical monitoring values ​​of environmental parameters such as light intensity, rainfall intensity, electromagnetic interference field strength, temperature, and humidity within the corresponding time period. The data source performance index database stores historical recognition accuracy, speed measurement accuracy, data update frequency, and packet loss rate data from various data sources such as video cameras, microwave radar, and floating cars under different combinations of the aforementioned environmental parameters. This data can be organized according to time series, and an association index can be established between environmental parameters, traffic data, and data source performance data to enable efficient querying and retrieval by the prediction module.

[0091] The prediction module typically includes a data preprocessing unit, a feature extraction unit, a prediction algorithm unit, and a result output unit. The data preprocessing unit first retrieves historical data related to the target sensing area from the historical database, performs data cleaning (e.g., removing outliers, imputing missing values), and standardizes or normalizes the data, converting environmental parameters and traffic data of different dimensions into a form suitable for model input. The feature extraction unit then extracts key influencing features from the preprocessed data, such as time features (hours, weekdays / weekends, months), environmental features (light intensity, rainfall intensity, electromagnetic interference intensity), and historical traffic flow features (traffic volume at the previous moment, traffic flow change rate, etc.). The prediction algorithm unit can employ various prediction models, such as statistical time series models (differential autoregressive moving average model, exponential smoothing), and machine learning models (support vector machines, random forests, neural networks, etc.). This unit uses extracted features and historical data to train and optimize the prediction model. Then, based on current and recent environmental parameters and traffic data, it predicts traffic flow patterns (such as trends in vehicle volume increase / decrease and the probability of congestion) or environmental change trends (such as predicted changes in light intensity and rainfall intensity) in the target sensing area over a future period (e.g., 5 minutes, 15 minutes, or 30 minutes). The output unit provides the predicted traffic flow patterns or environmental change trends to the data processing and fusion module in a structured data format (e.g., a dataset containing fields such as prediction time period, predicted traffic parameter values, predicted environmental parameter values, and prediction confidence levels). For example, if the prediction is that rainfall intensity will increase to 60 mm / hour within the next 10 minutes, the data processing and fusion module can pre-adjust the dynamic confidence weight of the microwave radar data source, reducing its weight in data fusion. This allows for preparation of the data fusion strategy before the actual environmental changes occur, improving the foresight and accuracy of traffic flow perception.

[0092] In some embodiments, the above method or system is applied to a vehicle-road cooperative scenario, and the mobile sensing terminal includes the on-board unit of the connected vehicle; the multi-source heterogeneous traffic sensing data also includes vehicle motion status information and sensor sensing results periodically reported by the connected vehicle; when performing dynamic confidence weighting, the calculation of the confidence weight for connected vehicle data also takes into account the frequency of vehicle reporting information, data packet integrity and the calibration level of the vehicle sensor itself.

[0093] For example, the frequency of vehicle information reporting can be quantified by the number of reports within a preset time window. A higher frequency coefficient is assigned if the number of reports reaches 10 or more within a 5-minute window. Data packet integrity is assessed by checking the field missing rate. If the missing rate of key data fields (such as speed, location, and acceleration) is less than 5%, the integrity coefficient is 1.0; for every additional 5% in the missing rate, the coefficient decreases by 0.1. The vehicle sensor calibration level is determined based on calibration reports issued by third-party certification bodies, and is divided into three levels: A, B, and C, corresponding to level coefficients of 1.2, 1.0, and 0.8, respectively. The final formula for calculating the confidence weight of connected vehicle data can be set as: Weight value = Basic weight × (0.4 × Frequency coefficient + 0.3 × Integrity coefficient + 0.3 × Calibration level coefficient). The basic weight is dynamically adjusted based on the distance between the vehicle and the target perception area; the closer the distance, the higher the basic weight, thus achieving a multi-dimensional and refined confidence assessment of connected vehicle data.

[0094] As can be seen from the above technical solutions, the embodiments of this application provide a traffic flow perception method and system. This involves collecting multi-source heterogeneous traffic perception data from fixed sensing devices, mobile sensing terminals, and environmental sensors within a target perception area; sending the multi-source heterogeneous traffic perception data to an edge computing node for preliminary processing and feature extraction to generate structured feature data; uploading the structured feature data to a cloud platform; performing spatiotemporal benchmark unification alignment on the structured feature data after spatiotemporal benchmark unification alignment; and performing multi-source data fusion processing based on dynamic confidence weighting on the structured feature data to obtain fused traffic parameters. Dynamic confidence weighting is calculated using a pre-generated environment-weight mapping model, which maps the parameter combination of the current perception environment state to the confidence weight values ​​of each data source. Based on the fused traffic parameters, traffic situation information including at least two of traffic flow, average vehicle speed, lane occupancy, and traffic event type is generated to address the problem of perception bias easily occurring from a single data source in specific environments.

[0095] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.

Claims

1. A traffic flow sensing method, characterized in that, include: Within the target perception area, multi-source heterogeneous traffic perception data from fixed perception devices, mobile perception terminals, and environmental sensors are collected. The multi-source heterogeneous traffic sensing data is sent to edge computing nodes for preliminary processing and feature extraction to generate structured feature data; The structured feature data is uploaded to the cloud platform; In the cloud platform, the structured feature data is uniformly aligned to a spatiotemporal reference. After the structured feature data has been aligned with the unified spatiotemporal reference, a multi-source data fusion process based on dynamic confidence weighting is performed to obtain the fused traffic parameters. The dynamic confidence weighting is calculated based on the perception environment state corresponding to each data source. The parameters of the perception environment state include at least one or more of the following: illumination conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions within the target perception area. The weights of the observations from each data source in the fusion are dynamically adjusted according to the changes in the perception environment state. The dynamic confidence weighting is calculated through a pre-generated environment-weight mapping model, which is used to map the parameter combination of the perceived environment state at the current moment into the confidence weight value of each data source. Based on the fused traffic parameters, traffic situation information is generated that includes at least two of the following: traffic flow, average vehicle speed, lane occupancy, and traffic event type.

2. The traffic flow sensing method according to claim 1, characterized in that, The environmental sensors include meteorological sensors for collecting visibility and rain / fog intensity information; If the target perception area is a mountainous road with limited terrain and variable weather, the calculation of the dynamic confidence weighting includes: Based on data collected by meteorological sensors, assess the local visibility or the intensity of rain and fog in the target sensing area; When the evaluation result is lower than the first threshold, the weight of visual sensing devices in data fusion is reduced based on a preset meteorological adjustment coefficient.

3. The traffic flow sensing method according to claim 2, characterized in that, The method further includes: Acquire historical baseline data related to the current target perception area, collected by visual sensing devices under historical good visibility conditions. When the evaluation result is lower than the first threshold, the Kalman filter algorithm is used to compensate or correct the current observation of the visual perception device based on the residual analysis between historical benchmark data and the current observation. The compensated or corrected observations are combined with the reduced weights and used in the multi-source data fusion process.

4. The traffic flow sensing method according to claim 1, characterized in that, The method further includes: If the target perception area is an area affected by periodic social activities, identify or predict the temporal patterns and intensity of the periodic social activities. Based on the aforementioned time patterns and intensity, an initial confidence adjustment factor is set for the mobile sensing terminal data collected during the periodic social activities. When performing the dynamic confidence weighting, the initial confidence adjustment factor is used as one of the input parameters.

5. The traffic flow sensing method according to claim 4, characterized in that, The method further includes: If the periodic social activity is a market activity, a traffic flow prediction model related to the intensity of the market activity is generated based on the historical market activity data of the target sensing area. When a market event is predicted or detected, the traffic flow prediction model is run to generate predicted traffic parameters. The predicted traffic parameters are combined with the fused traffic parameters for traffic incident identification or traffic guidance decision-making.

6. The traffic flow sensing method according to claim 1, characterized in that, The fixed sensing device includes radar and geomagnetic coils, and the mobile sensing terminal data includes Bluetooth or Wi-Fi beacon data. If the target sensing area is an underground, tunnel, or high-rise building area with severe multipath effects where signal reception is limited, the calculation of the dynamic confidence weighting includes: Assess the signal strength, number of available satellites, and positioning accuracy fluctuations of GNSS positioning data from mobile sensing terminals; Based on the evaluation results, the fusion weights of GNSS data and other sensing data sources are dynamically adjusted.

7. The traffic flow sensing method according to claim 1, characterized in that, The assessment of the perceived environment state combines at least one of the following: historical traffic flow patterns of the target perception area, real-time event information, and a preset sensor performance degradation model, as input to the environment-weight mapping model.

8. A traffic flow sensing system, characterized in that, The traffic flow sensing method applicable to any one of claims 1 to 7 includes: The data acquisition module, deployed in the target sensing area, is used to collect multi-source heterogeneous traffic sensing data from fixed sensing devices, mobile sensing terminals, and environmental sensors. The data processing and fusion module is used to preprocess the multi-source heterogeneous traffic perception data, unify and align the spatiotemporal references, and perform multi-source data fusion processing based on dynamic confidence weighting. The data processing and fusion module is configured to: calculate based at least on the sensing environment state corresponding to each data source, wherein the parameters of the sensing environment state include at least one or more of the following: illumination conditions, weather conditions, visibility, electromagnetic interference intensity, and road surface conditions within the target sensing area; and dynamically adjust the weight of each data source observation in the fusion according to the changes in the sensing environment state. The data processing and fusion module includes an environment-weight mapping model, which is used to map the parameter combination of the perceived environment state at the current moment into the confidence weight value of each data source. The analysis module is used to generate traffic situation information based on the fused traffic parameters, including at least two of the following: traffic flow, average vehicle speed, lane occupancy, and traffic event type.

9. The traffic flow sensing system according to claim 8, characterized in that, The data processing and fusion module includes an environmental status assessment unit, which is used for: Receive data from environmental sensors and basic information about the target perception area; Based on a pre-defined environmental impact model, the impact of the current sensing environment on the performance of each data source is quantitatively assessed. The degree of influence is output as the basis for calculating the dynamic confidence weight.

10. The traffic flow sensing system according to claim 8, characterized in that, The system also includes a historical database and a prediction module; The historical database is used to store historical traffic data of the target sensing area under different sensing environment conditions and historical performance data of each data source. The prediction module is configured to predict future traffic flow patterns or environmental change trends in the target sensing area based on historical data, and provide the prediction results to the data processing and fusion module for pre-adjusting the dynamic confidence weights.

Images