Fusion analysis and quality evaluation method based on multi-source data elements

Abstract

The invention discloses a fusion analysis and quality evaluation method based on multi-source data elements, and relates to the technical field of traffic data processing. In order to solve the problems that traditional traffic multi-source data fusion is weak in space-time relevance and quality evaluation is not combined with traffic scene characteristics, the method comprises the following steps: firstly, constructing a traffic multi-source data space-time alignment model to realize space-time unification of road condition images, vehicle GPS tracks, traffic signal lamp states and citizen reported data; secondly, proposing a dynamic weight fusion algorithm based on traffic congestion association degree, and dynamically adjusting the weight according to the association strength of data and congestion assessment; and finally, a multi-dimensional quality evaluation system containing traffic scene exclusive dimensions is established, quality dynamic tracking and early warning are realized in combination with a traffic flow time sequence prediction model, and according to the method, in a smart city traffic scene, the traffic jam prediction accuracy and the data quality evaluation and traffic decision adaptation degree are improved, and the method is significantly superior to a traditional method.

Description

Technical Field

[0001] This invention relates to the field of traffic data processing technology, specifically a method for fusion analysis and quality assessment based on multi-source data elements. Background Technology

[0002] In smart city traffic management, multi-source data is the core basis for congestion assessment and decision-making. However, traditional methods have significant drawbacks: First, they do not fully consider the spatiotemporal characteristics of traffic data. For example, GPS trajectories and road condition images have large fusion errors due to timestamp discrepancies, making it impossible to accurately reflect real-time traffic conditions. Second, the fusion weights are fixed and do not take into account the correlation between the data and the congestion assessment. For example, non-congested road segment data reported by citizens has a weak correlation with congestion judgments, but it is weighted equally with camera data, affecting the accuracy of fusion. Third, the quality assessment does not incorporate the characteristics of traffic scenarios, and only general dimensions cannot meet the needs of traffic decision-making.

[0003] The camera images became blurry due to heavy rain, and traditional methods still fused them with fixed weights, resulting in a 35% bias in congestion judgment. At the same time, the existing assessment only detects whether the image is missing, without assessing the impact of blurry images on congestion judgment, which further exacerbates decision-making errors. Summary of the Invention

[0004] This application provides a method for fusion analysis and quality assessment based on multi-source data elements. It solves the problem that traditional traffic multi-source data fusion does not fully consider spatiotemporal characteristics. Due to timestamp deviations and geographic coordinate mismatches, data such as GPS trajectories and road condition images have large fusion errors and cannot accurately reflect real-time road conditions. It realizes the solution of data spatiotemporal deviation problem through a traffic multi-source data spatiotemporal alignment model, so as to achieve spatiotemporal unification of multi-source data.

[0005] This application provides a method for fusion analysis and quality assessment based on multi-source data elements, characterized by the following steps: Heterogeneous data in smart city traffic scenarios are collected. The data sources include traffic camera road condition images, vehicle GPS trajectory data, traffic signal control data, traffic condition report text from citizen APP and traffic flow detection station data. The collected data is processed by denoising, filling missing values ​​and spatiotemporal alignment. The spatiotemporal alignment is achieved by timestamp calibration and geographic coordinate matching to obtain a standardized traffic dataset. A traffic-specific feature mapping model was constructed. MobileNet was used to extract vehicle density and queue length features from road condition images. A trajectory pattern feature extraction algorithm was used to extract average vehicle speed and turning frequency features from GPS data. BERT was used to extract congestion location and cause features from reported text. Statistical analysis was performed to extract time-period distribution features of traffic light duration and traffic flow detection data. All types of data were mapped to a unified traffic feature space. Z-score standardization was applied to numerical features, and L2 normalization was applied to vector features to eliminate scale differences. Data were aligned using a 5-minute time window to form a three-dimensional traffic feature vector matrix. The linear correlation between numerical features such as vehicle density and average speed and congestion assessment indicators is calculated using the Pearson correlation coefficient. The nonlinear correlation between congestion cause vectors, turning frequency, and congestion assessment indicators is calculated using mutual information entropy. Then, the weighted average of the correlation between the core features of each data source and the congestion assessment indicators is obtained. The fusion weights are dynamically adjusted in conjunction with the real-time data quality score. Multi-source traffic data fusion is achieved through weighted summation, and the real-time congestion level and traffic suggestions for road segments are output.

[0006] Furthermore, the traffic camera deployment described in step 1 meets the following requirements: the high-definition camera deployed at each intersection has a resolution of no less than 2 million pixels, a frame rate of 25fps, a road section covered by a single camera with a length of 1.8-2.2 kilometers, and the camera lens angle can be remotely adjusted through the smart city traffic management platform with an adjustment accuracy of ±0.5°.

[0007] Furthermore, the vehicle-mounted intelligent terminal mentioned in step 1 must comply with the "Technical Requirements for Satellite Positioning Systems for Road Transport Vehicles", integrate a Beidou / GPS dual-mode positioning module, have a positioning accuracy of ≤10 meters, use the JT / T808 protocol for data transmission, have a local cache capacity of ≥1000 GPS records in a weak network environment, and complete the retransmission of cached data within 30 seconds after the network is restored. Furthermore, the denoising stage of the spatiotemporal preprocessing described in step 1: the state equation of the Kalman adaptive filter is as follows: The observation equation is ,in For process noise, To observe the noise, the noise covariance matrix was calibrated offline using GPS drift data from the past 7 days; the Retinex image enhancement algorithm uses multi-scale Gaussian filtering with scale parameters set to 0.1, 1.0, and 10.0 to adapt to different severe weather scenarios such as heavy rain and backlighting. Furthermore, the road condition images employ an improved YOLOv8 algorithm, inserting a CBAM attention mechanism after the C3 module in the backbone layer. The input layer integrates an adaptive enhancement module with Retinex and histogram equalization, outputting 64-dimensional vehicle density and queue length feature vectors. The GPS data employs an improved DBSCAN algorithm, dynamically adjusting the cluster radius and minimum sample number based on the average vehicle speed of the road segment, removing drift points with speeds >120km / h or deviations from the road >50 meters, and outputting 32-dimensional average speed and turning frequency feature vectors. Furthermore, the feature value range mapping of the Min-Max normalization algorithm described in step 2: for vehicle density features, the historical minimum value x min Take 5 vehicles / km, historical maximum value x max Take 80 vehicles / km; for average speed characteristics, the historical minimum value x min Take 5 km / h as the historical maximum value x max Take 60 km / h; x for other characteristics min With x max All data were determined based on the same period data from the past 30 days, with the statistical time window being the morning rush hour from 7:00 to 9:00 every day. Furthermore, the Pearson correlation coefficient calculation described in step 3 uses 30 sets of samples from the same period of the past 30 days for the target road segment. Each set of samples contains the normalized mean of four types of features and a manually labeled congestion index; the real-time quality score calculation formula is as follows: ,in: T: The proportion of data transmission without packet loss or delay; E: The average error ratio between the characteristic data and the true value over the past 30 days.

[0008] Furthermore, in step 3, the transmission stability T statistics for calculating the real-time quality score are as follows: T is the percentage of data packets with no packet loss and a delay of ≤200ms during the data collection period, where the delay threshold for camera image data is 300ms, the delay threshold for GPS data and traffic light data is 100ms, and the delay threshold for text data reported by citizens is 500ms.

[0009] Furthermore, the feature normalization in step 2 employs the Min-Max normalization algorithm, mapping all features to the [0,1] interval, as shown in the formula. Where x is the original eigenvalue, x min x max These are the historical minimum and maximum values ​​of this feature, respectively, with the historical minimum value x being the lowest value. min Take 5 vehicles / km, historical maximum value x max Take 80 vehicles / km; for average speed characteristics, the historical minimum value x min Take 5 km / h as the historical maximum value xmax Take 60 km / h; x for other characteristics min With x max All data were determined based on statistics from the same period over the past 30 days.

[0010] Furthermore, when the predicted quality value is ≤88%, a level 2 warning is triggered, and the administrator is notified via pop-up window and SMS through the smart city traffic management platform. The optimization plan includes remotely adjusting the angle of intersection cameras, dispatching traffic police to handle road accidents, and activating Retinex and adaptive noise reduction image enhancement plans to ensure that the traffic quality value recovers by ≥2% within 15 minutes after optimization. Detailed Implementation

[0011] To facilitate understanding of the present invention, a more complete description is provided below; the present invention may be implemented in many different forms and is not limited to the embodiments described herein; rather, these embodiments are provided so that the disclosure of the present invention may be more thorough and complete.

[0012] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0014] Taking the morning rush hour traffic management of a core main road in a first-tier city as an example, the implementation steps of this method are explained in detail: Step 1: Multi-source traffic data acquisition and spatiotemporal preprocessing Traffic multi-source data includes traffic camera images, vehicle GPS trajectory data, traffic light control data, traffic condition reports from citizen apps, and traffic flow monitoring station data. By deploying one high-definition camera at each intersection, with a resolution of no less than 2 million pixels and a frame rate of 25fps, and the camera lens angle being remotely adjustable via the smart city traffic management platform with an adjustment accuracy of ±0.5°, and capturing one frame of road condition image every 2 seconds, with each camera covering a road segment of 1.8-2.2 kilometers, a stable and quantifiable source of raw data is provided for subsequent extraction of vehicle density and queue length features using Mobile Net. This ensures a reliable data input foundation for the image feature extraction process, thereby guaranteeing the feasibility of the entire multi-source data fusion and congestion assessment scheme. GPS data of 100 taxis and 50 ride-hailing vehicles passing through the section of road is collected through in-vehicle intelligent terminals, with one record per second. The records include latitude and longitude, instantaneous speed and driving direction. The collected data is transmitted to the cloud platform through 4G / 5G network, which facilitates the subsequent extraction of core features such as average vehicle speed and turning frequency. By collecting the red, yellow, and green durations of traffic lights at each intersection using cameras, and updating the traffic lights once per cycle, this feature is combined with image features extracted by MobileNet and GPS trajectory features to provide key support for dynamic weight adjustment and congestion assessment in multi-source data fusion, ensuring that congestion assessment takes into account both traffic flow status and traffic light regulation. The city's traffic app receives text reports from citizens, averaging 3-5 reports every 5 minutes. The reports include information such as "congestion at the intersection of XX Road and A Road" and "traffic accident in the middle section of XX Road". This information complements GPS trajectory features and image features, allowing congestion assessment to combine objective data and subjective feedback. One traffic flow monitoring station is set up at the beginning, middle and end of the road segment to collect two-way traffic flow every 5 minutes, in units of vehicles per hour. This provides an accurate and continuous objective data source for subsequent statistical analysis to extract features such as the average traffic flow and fluctuation coefficient over a period of time. At the same time, the three-point distribution comprehensively reflects the differences in traffic flow distribution on the road segment, complementing GPS, image and other data to support the accuracy of congestion assessment. Spatiotemporal preprocessing: Spatiotemporal preprocessing includes denoising, missing value imputation, and spatiotemporal alignment; Noise reduction: Kalman adaptive filtering is used to remove drift points in GPS data, and Retinex image enhancement algorithm is used to improve the clarity of camera images in rainy weather. This can eliminate noise and interference in the original data and ensure the accuracy of extracting GPS average speed and image vehicle density. The state equation for Kalman adaptive filtering is:

[0015] The observation equation is:

[0016] Among them W k−1 For process noise, V k To observe the noise, the noise covariance matrix was calibrated offline using GPS drift data from the past 7 days; the Retinex image enhancement algorithm uses multi-scale Gaussian filtering with scale parameters set to 0.1, 1.0, and 10.0 to adapt to different severe weather scenarios such as heavy rain and backlighting.

[0017] Missing value imputation: For missing segments of GPS trajectory, the missing segments are filled by predicting the average speed and driving direction of the previous 10 seconds; for missing frames of road condition images, interpolated images are generated based on the traffic flow change trends of the previous 5 frames to ensure data continuity and integrity, and to avoid distortion of subsequent feature extraction and multi-source data fusion results due to data breaks. Spatiotemporal alignment: All data timestamps are calibrated to the unified clock of the city traffic management platform, and geographic coordinates are transformed into the city's electronic traffic map through Gaussian projection to ensure that the coverage area of ​​the camera images corresponds completely to the road segments on the GPS track. The vehicle-mounted intelligent terminal must comply with the "Technical Requirements for Satellite Positioning Systems of Road Transport Vehicles", integrate a Beidou / GPS dual-mode positioning module, have a positioning accuracy of ≤10 meters, use the JT / T808 protocol for data transmission, have a local cache capacity of ≥1000 GPS records in weak network environments, and complete the retransmission of cached data within 30 seconds after the network is restored. Step 2: Traffic Feature Mapping and Unified Representation A traffic-specific feature mapping model is constructed. This model achieves unified feature representation of heterogeneous data through source feature extraction, feature normalization, and unified vector matrix generation. Road condition images: Features are extracted using an improved YOLOv8 algorithm, and an attention mechanism is introduced in the backbone layer to enhance feature focusing on small vehicles, thereby improving the accuracy of small target recognition; An adaptive image enhancement module is added to the input layer, which combines the Retinex algorithm with histogram equalization to eliminate image noise in rainstorm and backlight scenes. The number of vehicles detected in a single frame image is counted, and combined with the length of the road segment covered by the camera, it is converted into vehicles / km, and then mapped into a 32-dimensional feature vector through a fully connected layer; Based on the queuing area mask, the vertical pixel span is calculated, converted to "meters" by combining the actual road width, and mapped to a 32-dimensional feature vector through a fully connected layer; By concatenating the 32-dimensional feature vectors of vehicle density and queue length, a 64-dimensional road condition image feature vector is obtained. GPS data: The trajectory is clustered using an improved DBSCAN clustering algorithm. The clustering radius and minimum number of samples are dynamically adjusted based on the average vehicle speed of the road segment. Kalman filtering and map matching modules are added before clustering to remove GPS drift points, such as those with speeds >120km / h or those deviating from the road by more than 50 meters, thereby improving the accuracy of trajectory matching. The GPS data of 150 vehicles were denoised and matched with maps, and then bound to specific road segments; By clustering the preprocessed trajectories using the improved DBSCAN algorithm, if the density of trajectory points in a certain area is 20 per 100 meters and the average speed is <20 km / h, it is determined to be a congested road section. The average instantaneous speed of all vehicles in the congested road section is statistically analyzed and mapped to a 16-dimensional feature vector. The number of times the vehicle's driving direction changed by more than 30° within a 5-minute time window was counted and mapped to a 16-dimensional feature vector. The 16-dimensional feature vectors of average speed and turning frequency were concatenated to obtain a 32-dimensional GPS data feature vector. Citizen-reported text: The pre-trained BERT model is used to extract keyword features, such as "congestion" and "traffic accident". The citizen-reported text (such as "XX Road and A Road intersection are congested due to rear-end collision") is cleaned, segmented and encoded, and converted into a token ID in BERT input format. The encoded text is input into the fine-tuned BERT model, and the hidden layer vectors of the location and reason keywords are output. The dimensionality is reduced to 8 dimensions by PCA. The probability distribution vectors of the three types of text are combined with the historical reporting accuracy of citizens and then mapped to 8 dimensions. The 8-dimensional vectors of keyword features and semantic classification features are concatenated to obtain a 16-dimensional feature vector of citizen-reported text. Traffic light data: The duration of red, yellow and green lights is collected by intersection cameras and updated once per signal cycle. Two core indicators are extracted: "green light duration" and "red light duration stability". Traffic flow data: By using traffic flow detectors at the beginning, middle, and end of the road segment, we collect 5 minutes of two-way traffic flow and calculate two core indicators: "average traffic flow during the morning peak (7:00-9:00)" and "traffic fluctuation coefficient".

[0018] Feature dimension mapping: The stability of green light duration and red light duration are mapped to 4-dimensional feature vectors, for a total of 8 dimensions. The average traffic flow during the morning peak and the traffic flow fluctuation coefficient are mapped to 4-dimensional feature vectors, for a total of 8 dimensions. The above 8-dimensional vectors are concatenated to obtain 16-dimensional traffic light and traffic flow data feature vectors. To eliminate the dimensional differences in features from different data sources (such as vehicle density units "vehicles / km" and average speed units "km / h"), a Min-Max normalization algorithm is used to map all features to the [0,1] interval, as shown in the following formula:

[0019] Where x is the original feature value, x min x max These are the historical minimum and maximum values ​​of this feature, respectively, with the historical minimum value x being the lowest value. min Take 5 vehicles / km, historical maximum value x max Take 80 vehicles / km; for average speed characteristics, the historical minimum value x min Take 5 km / h as the historical maximum value x max Take 60 km / h; x for other characteristics min With x max All data were determined based on the same period data from the past 30 days, with the statistical time window being the morning rush hour from 7:00 to 9:00 every day.

[0020] The normalized 64-dimensional image features, 32-dimensional GPS features, 16-dimensional text features, and 16-dimensional traffic light and flow features are concatenated in the order of "image → GPS → text → traffic light and flow" to obtain a 128-dimensional traffic feature vector. One hundred core road segments in the city were selected, and each road segment was divided into 5-minute time windows, with 288 windows per day. A total of 3,600 samples were generated over 12.5 consecutive days. The 3600 128-dimensional feature vectors are organized in the format of "sample number × feature dimension" to generate a 3600×128-dimensional traffic feature vector matrix. This matrix is ​​stored in a MySQL database and supports subsequent multi-source data fusion and congestion assessment algorithm calls. Step 3: Dynamic weight fusion analysis based on congestion correlation The correlation is used to measure the "influence of a single feature on the congestion index". The value range is [-1, 1]. The larger the value of R, the stronger the correlation. Positive correlation means "the higher the feature value, the higher the congestion index", and negative correlation means "the higher the feature value, the lower the congestion index".

[0021] Calculate the congestion correlation: Historical data (7:30-7:35) from the same period over the past 30 days for the target road section were selected as the historical sample, totaling 30 groups, each containing: Independent variable: Normalized mean of the four types of features (e.g., image feature mean X) 1j GPS feature mean X 2j, j=1,2,...,30); Dependent variable: Manually labeled congestion index Y for the corresponding time period j The value ranges from 0 to 10 and is determined by traffic management personnel based on on-site surveys and multi-source data. The congestion correlation coefficient is calculated using the Pearson correlation coefficient, as shown in the following formula:

[0022] Where X is the feature value of the data source, Y is the road segment congestion index, the higher the correlation, the greater the weight allocation, and the weight is dynamically updated every 5 minutes according to the real-time congestion changes. The correlation between camera image features and road congestion index is positive; the higher the vehicle density, the higher the congestion index. The correlation between GPS data features and congestion index is negative; the lower the average speed, the higher the congestion index. The correlation between citizen-reported text features and congestion index is positive; the higher the frequency of accident keywords, the higher the congestion index. The correlation between traffic light and traffic flow features and congestion index is negative; the longer the green light duration and the lower the traffic flow, the lower the congestion index. Calculate the real-time quality score: Real-time quality score measures the reliability of each feature data at the current moment, with a value range of [0,1]. The higher the score, the more reliable the data. This implementation evaluates the data from two dimensions: "transmission stability (T)" and "historical error rate (E)," using a weighted summation formula, as follows: , in:

[0023] T: The percentage of data transmission without packet loss or delay; E: The average error ratio between the feature data and the true value over the past 30 days, where (1−E) is the confidence level. Weighting and fusion: The weight allocation uses the "quality score normalization" method to ensure that the sum of the weights of the four feature classes is 1. The formula is as follows: Where i = 1, 2, 3, 4, corresponding to four types of features: camera images, GPS data, citizen-reported text, and traffic light and flow data. The real-time congestion index is calculated using a two-step method of "weighted summation + scaling mapping". The formula is as follows: =

[0024] Among them W i The normalized mean of the i-th class of features is the average of all dimensional feature values ​​of that class. Based on the "congestion status of the target road segment + feature vector matrix of surrounding road segments", differentiated suggestions are generated and pushed to users within a 5-kilometer radius of the target road segment through the city traffic APP, and simultaneously displayed on traffic guidance screens. T represents the percentage of data packets with no packet loss and a latency of ≤200ms during the data collection period. The latency threshold for camera image data is 300ms, the latency threshold for GPS data and traffic light data is 100ms, and the latency threshold for text data reported by citizens is 500ms.

[0025] Step 4: Multi-dimensional quality assessment of traffic scenarios Spatiotemporal consistency is used to measure whether the timestamps and spatial locations of different data sources match the same traffic event, avoiding fusion errors caused by "time misalignment" or "spatial offset".

[0026] Data acquisition and calibration: Timestamp collection: Extracting timestamps from four types of data sources for key events—such as GPS data records, camera image records, citizen report text records, and traffic light data records; Spatial coordinate calibration: Based on the electronic road map of the traffic police system, the spatial coordinates of each data source are calibrated—GPS positioning deviation is 2 meters, camera field of view center deviation is 1 meter, citizen-reported location deviation is 3 meters, and traffic light position deviation is zero.

[0027] The congestion correlation assessment measures the "mean strength of the correlation between the characteristics of four types of data sources and the congestion index". The higher the correlation, the stronger the support of the data for congestion assessment and the better the quality. The arithmetic mean of the correlation between the four types of data sources is calculated by taking the absolute value, since the sign of the correlation only indicates the direction of the correlation. The formula is as follows: Scoring criteria: R ≥ 0.85 scores 100%. 90% If you get 80%, and so on, for every 0.05 decrease, the score decreases by 10%. Real-time assessment measures the total delay from data collection to the output of congestion index and traffic recommendations. The lower the delay, the better it can support real-time traffic decisions, such as dynamic updates of traffic guidance screens and real-time push notifications from apps. Credibility assessment is used to measure the "reliability weighting of different data sources". Considering that traffic police system data, such as traffic light and traffic detector data, are collected by professional equipment and have higher credibility than subjective data reported by citizens, the analytic hierarchy process (AHP) is used to determine the weights and then calculate the overall credibility. Overall credibility = Traffic police system data score × 0.6 + Citizen-reported data score × 0.4.

[0028] Step 5: Dynamic Tracking and Early Warning of Traffic Quality Temporal Prediction: A multi-head attention module is added between the hidden and output layers of the LSTM model to strengthen the influence weight of "key historical data" on the prediction results and improve the prediction accuracy of long sequences. Non-temporal discrete data such as weather (W(t)) and events (E(t)) are converted into vectors through the embedding layer and concatenated with temporal quality data (Q(t)) and flow data (F(t)) before being input into the model, thus overcoming the limitation of traditional LSTM in processing only single temporal data. Early warning and optimization: Triggering an alert on the smart city traffic management platform includes pop-up windows and SMS notifications to administrators. The platform homepage displays a prominent red border with the message "XX Road Traffic Quality Level 2 Alert". The pop-up message includes "predicted time period, current quality value, predicted minimum value, and reasons for quality decline, such as traffic accidents causing GPS data fluctuations or light rain affecting image quality". Output data optimization scheme: The angle of the intersection camera; Dispatch traffic police to the relevant road sections to handle traffic accidents; Image enhancement algorithm: The Retinex image enhancement algorithm mentioned above is enabled, and an adaptive noise reduction module is superimposed. It is optimized for light rain scenes to reduce pixel interference caused by raindrops. The algorithm parameters are remotely pushed to the camera terminal through the edge computing server, which can take effect quickly without restarting the device.

[0029] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for fusion analysis and quality assessment based on multi-source data elements, characterized in that: Includes the following steps: Heterogeneous data in smart city traffic scenarios are collected. The data sources include traffic camera road condition images, vehicle GPS trajectory data, traffic signal control data, traffic condition report text from citizen APP and traffic flow detection station data. The collected data is processed by denoising, filling missing values ​​and spatiotemporal alignment. The spatiotemporal alignment is achieved by timestamp calibration and geographic coordinate matching to obtain a standardized traffic dataset. A traffic-specific feature mapping model is constructed. MobileNet is used to extract vehicle density and queue length features from road condition images. A trajectory pattern feature extraction algorithm is used to extract vehicle average speed and turning frequency features from GPS data. BERT is used to extract congestion location and cause features from reported text. Statistical analysis is used to extract time period distribution features of traffic light duration and traffic flow detection data. All types of data are mapped to a unified traffic feature space. Z-Score standardization is performed on numerical features and L2 normalization is performed on vector features to eliminate scale differences. Data are aligned using a 5-minute time window to form a three-dimensional traffic feature vector matrix; The linear correlation between numerical features such as vehicle density and average speed and congestion assessment indicators is calculated using the Pearson correlation coefficient. The nonlinear correlation between congestion cause vectors, turning frequency, and congestion assessment indicators is calculated using mutual information entropy. Then, the weighted average of the correlation between the core features of each data source and the congestion assessment indicators is obtained to obtain the overall correlation between each data source and the congestion assessment indicators. The fusion weights are dynamically adjusted in combination with the real-time data quality score. Multi-source traffic data fusion is achieved through weighted summation to output the real-time congestion level and traffic suggestions for road segments.

2. The traffic multi-source data fusion and congestion assessment method according to claim 1, characterized in that, The traffic camera deployment described in step 1 meets the following requirements: the high-definition camera deployed at each intersection has a resolution of no less than 2 million pixels, a frame rate of 25fps, a road section covered by a single camera with a length of 1.8-2.2 kilometers, and the camera lens angle can be remotely adjusted through the smart city traffic management platform with an adjustment accuracy of ±0.5°.

3. The traffic multi-source data fusion and congestion assessment method according to claim 1, characterized in that, The vehicle-mounted intelligent terminal mentioned in step 1 must comply with the "Technical Requirements for Satellite Positioning Systems of Road Transport Vehicles", integrate a Beidou / GPS dual-mode positioning module, have a positioning accuracy of ≤10 meters, use the JT / T808 protocol for data transmission, have a local cache capacity of ≥1000 GPS records in a weak network environment, and complete the retransmission of cached data within 30 seconds after the network is restored.

4. The traffic multi-source data fusion and congestion assessment method according to claim 1, characterized in that, The denoising stage of the spatiotemporal preprocessing described in step 1: The state equation of the Kalman adaptive filter is as follows: The observation equation is ,in For process noise, To observe the noise, the noise covariance matrix was calibrated offline using GPS drift data from the past 7 days; The Retinex image enhancement algorithm uses multi-scale Gaussian filtering with scale parameters set to 0.1, 1.0, and 10.0 to adapt to different severe weather scenarios such as heavy rain and backlighting. According to the traffic multi-source data fusion and congestion assessment method according to claim 1, the state equation of the spatiotemporal preprocessing denoising step in step 1 is: The observation equation is ,in For process noise, To observe the noise, the noise covariance matrix was calibrated offline using GPS drift data from the past 7 days; The Retinex image enhancement algorithm uses multi-scale Gaussian filtering with scale parameters set to 0.1, 1.0, and 10.0 to adapt to different severe weather scenarios such as heavy rain and backlighting.

5. The traffic multi-source data fusion and congestion assessment method according to claim 1, characterized in that, The road condition images employ an improved YOLOv8 algorithm, inserting a CBAM attention mechanism after the C3 module in the backbone layer. The input layer integrates an adaptive enhancement module with Retinex and histogram equalization, outputting 64-dimensional vehicle density and queue length feature vectors. The GPS data employs an improved DBSCAN algorithm, dynamically adjusting the cluster radius and minimum sample number based on the average vehicle speed of the road segment, removing drift points with speeds >120km / h or deviations from the road >50 meters, and outputting 32-dimensional average speed and turning frequency feature vectors.

6. The traffic multi-source data fusion and congestion assessment method according to claim 1, characterized in that, The feature value range mapping of the Min-Max normalization algorithm described in step 2 is as follows: For vehicle density features, the historical minimum value xmin is 5 vehicles / km and the historical maximum value xmax is 80 vehicles / km; for average speed features, the historical minimum value xmin is 5 km / h and the historical maximum value xmax is 60 km / h; the xmin and xmax values ​​of other features are determined based on the statistical data of the same period in the past 30 days, and the statistical time window is the morning peak from 7:00 to 9:00 every day.

7. The traffic multi-source data fusion and congestion assessment method according to claim 1, characterized in that, The Pearson correlation coefficient calculation described in step 3 uses 30 sets of samples from the same period of the past 30 days for the target road segment. Each set of samples contains the normalized mean of 4 types of features and the manually labeled congestion index. The formula for calculating the real-time quality score is as follows: ,in: T: The percentage of data transmission without packet loss or delay; E: The average error ratio between the feature data and the true value over the past 30 days.

8. The traffic multi-source data fusion and congestion assessment method according to claim 1, characterized in that, The transmission stability T statistics for the real-time quality score calculation in step 3: T is the percentage of data packets with no packet loss and a delay of ≤200ms during the data collection period, where the delay threshold for camera image data is 300ms, the delay threshold for GPS data and traffic light data is 100ms, and the delay threshold for text data reported by citizens is 500ms.

9. The traffic multi-source data fusion and congestion assessment method according to claim 6, characterized in that, The feature normalization described in step 2 uses the Min-Max normalization algorithm to map all features to the [0,1] interval, as shown in the formula. Where x is the original eigenvalue, x min x max These are the historical minimum and maximum values ​​of this feature, respectively, with the historical minimum value x being the lowest value. min Take 5 vehicles / km, historical maximum value x max Take 80 vehicles / km; for average speed characteristics, the historical minimum value x min Take 5 km / h as the historical maximum value x max Take 60 km / h; x for other characteristics min With x max All data were determined based on statistics from the same period over the past 30 days.

10. The traffic multi-source data fusion and congestion assessment method according to any one of claims 1 to 9, characterized in that, When the predicted quality value is ≤88%, a level 2 warning is triggered, and the administrator is notified via pop-up window and SMS through the smart city traffic management platform. The optimization plan includes remotely adjusting the angle of intersection cameras, dispatching traffic police to handle road accidents, and activating Retinex and adaptive noise reduction image enhancement plans to ensure that the traffic quality value recovers by ≥2% within 15 minutes after optimization.

Images