A fusion calling method and system based on multi-source traffic data

Abstract

The application relates to the field of data fusion analysis, in particular to a fusion calling method and system based on multi-source traffic data, which acquires a geomagnetic original signal and video structured data of a target vehicle and performs space-time alignment, extracts a reasonable fluctuation range of a vehicle length and color distribution probability from a pre-constructed vehicle model-scene adaptive baseline library, and if license plate information cannot be recognized, obtains a nominal vehicle length of a candidate model from a vehicle model knowledge graph; the deviation of the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between an instantaneous speed and a historical speed curve are weighted and fused to generate a contradiction index of the target vehicle characteristics, and if the contradiction index exceeds a preset contradiction index threshold, an abnormality alarm is triggered. The problems that geomagnetic signals are susceptible to electromagnetic interference in complex traffic flow and that video data and geomagnetic data have poor space-time alignment accuracy and lead to inaccurate vehicle recognition are solved.

Description

Technical Field

[0001] This invention relates to the field of data fusion analysis, specifically a method and system for fusing and calling multi-source traffic data. Background Technology

[0002] Against the backdrop of the rapid development of intelligent transportation systems, multi-source traffic data fusion has become a core technical means to improve traffic management efficiency. However, the instantaneous fluctuations in vehicle physical characteristics and the conflict between multi-sensor data in dynamic traffic scenarios remain a technical bottleneck restricting the accurate identification of traffic incidents. In scenarios such as smart toll stations and key freight corridors, geomagnetic detectors are susceptible to electromagnetic interference from adjacent lanes in complex traffic flows, leading to instantaneous distortion in single-vehicle length measurements. Simultaneously, semantic differences exist between vehicle brand, color, and other information in structured video data and vehicle length, speed, and other information in geomagnetic features, further exacerbating the complexity of cross-modal data fusion. While geomagnetic detectors are increasingly used, their data quality is highly dependent on anti-interference algorithms, and the accuracy of video data recognition significantly decreases in harsh environments.

[0003] Therefore, there is an urgent need for a multi-source data fusion and invocation method based on dynamic quality assessment to achieve fast and high-precision vehicle feature verification, providing reliable technical support for traffic inspection and congestion management. Summary of the Invention

[0004] (1) Technical problems to be solved

[0005] The purpose of this invention is to provide a method and system for fusing and calling multi-source traffic data, in order to solve the problems of vehicle length measurement distortion caused by the susceptibility of geomagnetic signals to electromagnetic interference from adjacent lanes, poor spatiotemporal alignment accuracy between video data and geomagnetic data, and inaccurate vehicle identification due to insufficient cross-modal data fusion.

[0006] (2) Technical solution

[0007] To achieve the above objectives, on the one hand, the present invention provides a method for fusing and invoking multi-source traffic data, the method comprising:

[0008] S1. Real-time acquisition of the target vehicle's original geomagnetic signal and video structured data; magnetic field gradient decomposition processing of the original geomagnetic signal to obtain an anti-interference geomagnetic signal; spatiotemporal alignment of the anti-interference geomagnetic signal and video structured data; extraction of the measured vehicle length and instantaneous speed from the spatiotemporally aligned anti-interference geomagnetic signal; and extraction of vehicle brand, color features, and license plate information from the video structured data.

[0009] S2. Based on the target vehicle type and scene context, extract the reasonable fluctuation range of vehicle length and color distribution probability from the pre-built vehicle model-scene adaptive baseline library; if the license plate information in the video structured data is identifiable, query the vehicle management office database based on the license plate information to obtain the nominal vehicle length of the target vehicle registration model; if the license plate information is not identifiable, based on the matching result of the measured vehicle length and the reasonable fluctuation range, and combined with the vehicle brand similarity threshold, filter candidate models from the vehicle model knowledge graph and obtain the nominal vehicle length of the candidate models.

[0010] S3. The deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scenario adaptive baseline library are weighted and fused to generate a contradiction index of the target vehicle features. If the contradiction index exceeds the preset contradiction index threshold, an abnormal alarm is triggered.

[0011] Furthermore, the method for obtaining an anti-interference geomagnetic signal by performing magnetic field gradient decomposition processing on the original geomagnetic signal includes:

[0012] The original geomagnetic signal is decomposed into low-frequency approximate components and high-frequency detail components of multiple scales through wavelet transform.

[0013] KL divergence analysis is used to analyze the noise distribution of high-frequency detail components at each scale, and an adaptive threshold related to the electromagnetic interference intensity of adjacent channels is generated. The adaptive threshold is compared with the amplitude of the high-frequency detail components, and the interference components of the high-frequency detail components that exceed the adaptive threshold are truncated to obtain the first high-frequency detail component.

[0014] After denoising the first high-frequency detail component, the second high-frequency detail component is obtained. Wavelet reconstruction is performed on the second high-frequency detail component and the low-frequency approximation component, and the reconstructed anti-interference geomagnetic signal is smoothed by sliding window mean filtering to generate an anti-interference geomagnetic signal.

[0015] Furthermore, the method for generating an adaptive threshold related to the intensity of adjacent channel electromagnetic interference by analyzing the noise distribution of high-frequency detail components at each scale using KL divergence analysis includes:

[0016] The noise signals of high-frequency detail components at each scale are extracted and their probability density distributions are calculated to obtain the noise distribution of high-frequency detail components; a preset reference noise distribution is obtained, which is statistically generated based on geomagnetic signals in a scenario without adjacent channel interference.

[0017] The deviation between the high-frequency detail component noise distribution and the preset reference noise distribution is quantified by KL divergence, and the adaptive threshold is dynamically adjusted according to the deviation.

[0018] Furthermore, the method for spatiotemporal alignment of the anti-interference geomagnetic signal and the video structured data includes:

[0019] The spatiotemporal alignment includes spatial calibration and time synchronization. The spatial calibration is to calculate the transformation matrix between the geomagnetic coordinate system and the video image coordinate system based on a preset calibration board, and to correct the spatial position deviation between the geomagnetic sensor and the video acquisition device in real time through a dual Kalman filter to ensure that the spatial position deviation is less than a preset position deviation threshold.

[0020] The time synchronization is achieved by aligning the timestamps of the geomagnetic sensor and the video acquisition device according to a precise clock synchronization protocol, so that the data acquisition time deviation is less than a preset time deviation threshold.

[0021] The spatially calibrated anti-interference geomagnetic signal and the time-synchronized video structured data are correlated and matched according to the time window to obtain a spatiotemporally consistent fused data stream.

[0022] Furthermore, the method for correlating and matching the spatially calibrated anti-interference geomagnetic signal with the time-synchronized video structured data according to a time window to obtain a spatiotemporally consistent fused data stream includes:

[0023] The time window is divided based on the frame timestamp of the video structured data, and the video frames in each time window and their corresponding vehicle brand and color features are extracted.

[0024] Within the same time window, the corresponding geomagnetic waveform data is extracted from the anti-interference geomagnetic signal, and the start and end times of the vehicle passing the geomagnetic sensor are located by the peak detection method; based on the start and end times, the average speed of the geomagnetic signal and the vehicle length measurement value within the time window are calculated.

[0025] The vehicle brand and color features of the video frame are aligned with the average speed and vehicle length measurements by timestamp to generate spatiotemporally correlated fused data units; the fused data units of consecutive time windows are serialized and spliced ​​to form a spatiotemporally consistent fused data stream.

[0026] Furthermore, the method for extracting the reasonable fluctuation range of vehicle length and the probability of color distribution from the vehicle model-scene adaptive baseline library based on the target vehicle type and scene context includes:

[0027] Based on the vehicle model-scenario adaptive baseline library, the benchmark scenario set corresponding to the target vehicle type is matched. The benchmark scenario set includes the vehicle length distribution characteristics under different traffic flow levels, weather conditions and time period combinations within the historical statistical period.

[0028] Based on the real-time traffic flow level, weather conditions, and time period parameters in the current scenario context, extract the N historical scenario clusters with the highest correlation from the benchmark scenario set, calculate the weighted quantile interval of the vehicle length measurement value of each historical scenario cluster, and generate a reasonable fluctuation range of vehicle length, where N≥2.

[0029] Based on the scene color feature map stored in the vehicle-scene adaptive baseline library, spatiotemporal similarity matching is performed on the current scene context to extract scene color distribution patterns that meet the preset similarity threshold. Combined with real-time collected vehicle color data, the color distribution probability is obtained through Bayesian probability analysis.

[0030] Furthermore, the method for selecting candidate models and obtaining their nominal vehicle lengths from the vehicle model knowledge graph based on the matching results of the measured vehicle length and the reasonable fluctuation range, combined with a vehicle brand similarity threshold, includes:

[0031] The nominal vehicle length range and associated brand characteristics of all candidate models are obtained from the vehicle model knowledge graph; candidate models whose nominal vehicle length range intersects with the reasonable fluctuation range of the measured vehicle length are selected to obtain a preliminary candidate set.

[0032] The similarity between the vehicle brand features extracted from the video structured data and the associated brand features of each candidate model in the preliminary candidate set is quantitatively calculated using a semantic matching model.

[0033] The final candidate model list is obtained by selecting candidate models from the preliminary candidate set whose similarity to related brand features is greater than the vehicle brand similarity threshold.

[0034] Based on the matching degree between the nominal vehicle length range and the actual vehicle length of each model in the final candidate model list, the nominal vehicle length of the candidate model with the highest matching degree is selected as the output result.

[0035] Furthermore, the method for quantifying the similarity between the vehicle brand features extracted from the video structured data and the associated brand features of each candidate model in the preliminary candidate set through a semantic matching model includes:

[0036] Semantic description text of the brand features associated with candidate models is extracted from the vehicle model knowledge graph. The semantic description includes the brand name, brand logo, and vehicle series information.

[0037] The vehicle brand features extracted from the video structured data are converted into semantic vectors; the cosine similarity between the semantic vectors and the embedding vectors of the candidate model semantic description text is calculated according to the semantic matching model and recorded as the associated brand feature similarity.

[0038] Furthermore, the method for generating a contradiction index of target vehicle features by weighted fusion of the deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scene adaptive baseline library includes:

[0039] The deviation between the measured vehicle length and the nominal vehicle length is characterized by the ratio of the absolute difference between the measured vehicle length and the nominal vehicle length to the nominal vehicle length.

[0040] The color distribution probability deviation is quantified by the difference between the frequency of occurrence of the current vehicle color in the corresponding scenario of the model-scenario adaptive baseline library and the historical statistical frequency.

[0041] The difference between the instantaneous velocity and the historical velocity curve is characterized by the degree of deviation of the instantaneous velocity from the probability distribution of the historical velocity curve.

[0042] The fusion weights of the deviation, color distribution probability deviation, and speed difference are dynamically adjusted according to the current scene context; the deviation, color distribution probability deviation, and speed difference are linearly weighted and fused according to the adjusted fusion weights to generate a contradiction index of the target vehicle features.

[0043] On the other hand, based on the same inventive concept, the present invention also provides a fusion and retrieval system based on multi-source traffic data, the system comprising: a data processing module, a nominal vehicle length acquisition module, and a vehicle analysis module, wherein each module is sequentially connected in communication.

[0044] The data processing module is used to acquire the raw geomagnetic signal and video structured data of the target vehicle in real time, perform magnetic field gradient decomposition on the raw geomagnetic signal to obtain an anti-interference geomagnetic signal, and perform spatiotemporal alignment on the anti-interference geomagnetic signal and the video structured data; extract the measured vehicle length and instantaneous speed from the spatiotemporally aligned anti-interference geomagnetic signal, and extract the vehicle brand, color features and license plate information from the video structured data.

[0045] The nominal vehicle length acquisition module is used to extract the reasonable fluctuation range of vehicle length and color distribution probability from a pre-built vehicle model-scene adaptive baseline library based on the target vehicle type and scene context. If the license plate information in the video structured data is identifiable, the module queries the vehicle management office database based on the license plate information to obtain the nominal vehicle length of the target vehicle registration model. If the license plate information is not identifiable, the module selects candidate models from the vehicle model knowledge graph and obtains the nominal vehicle length of the candidate models based on the matching result of the measured vehicle length and the reasonable fluctuation range, combined with the vehicle brand similarity threshold.

[0046] The vehicle analysis module is used to generate a contradiction index of the target vehicle features by weighted fusion of the deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scenario adaptive baseline library. If the contradiction index exceeds the preset contradiction index threshold, an abnormal alarm is triggered.

[0047] (3) Beneficial effects

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. By performing magnetic field gradient decomposition on the raw geomagnetic signal, an anti-interference geomagnetic signal is obtained, which effectively reduces the impact of adjacent channel electromagnetic interference and ensures the quality of geomagnetic data. Combined with video structured data, multi-source data fusion is achieved through spatiotemporal alignment. Vehicle feature information is extracted from different data sources, and mutual verification and supplementation are carried out to improve the accuracy and reliability of vehicle feature recognition.

[0050] 2. Based on a vehicle-scenario adaptive baseline library, a contradiction index is generated through comprehensive analysis of multi-dimensional features and weighted fusion to fully assess the consistency of target vehicle features. When the contradiction index exceeds a preset threshold, an anomaly alarm is triggered, enabling timely detection of potentially abnormal vehicles and improving anomaly detection capabilities in traffic management.

[0051] 3. When license plate information is unrecognizable, based on the matching results of the measured vehicle length and a reasonable fluctuation range, combined with the vehicle brand similarity threshold, candidate models are screened from the vehicle model knowledge graph, and the nominal vehicle length is obtained. Through comprehensive judgment of multiple factors, the problem of vehicle model recognition in the case of missing license plate information is effectively solved, improving the accuracy and intelligence level of vehicle model recognition. Attached Figure Description

[0052] Figure 1 This is a flowchart of a method for fusing and invoking multi-source traffic data according to Embodiment 1 of the present invention.

[0053] Figure 2 This is a schematic diagram of the module composition of a multi-source traffic data fusion and invocation system according to Embodiment 2 of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Before giving examples, it is necessary to elaborate on the application scenario of the inventive concept. The present invention is a method and system for fusion and invocation of multi-source traffic data, which is applied to perform high-precision vehicle feature verification in a complex traffic environment, providing reliable technical support for traffic inspection and congestion management.

[0056] Example 1: As Figure 1 shown, this example provides a method for fusion and invocation of multi-source traffic data, and the method includes:

[0057] S1. Real-time obtain the geomagnetic raw signal and video structured data of the target vehicle, perform magnetic field gradient decomposition processing on the geomagnetic raw signal to obtain an anti-interference geomagnetic signal, and perform spatio-temporal alignment on the anti-interference geomagnetic signal and the video structured data; extract the measured vehicle length and instantaneous speed from the spatio-temporally aligned anti-interference geomagnetic signal, and extract the vehicle brand, color feature and license plate information from the video structured data.

[0058] Exemplarily, at the exit of a certain highway toll station, when a blue A-type sedan of a certain brand passes through the detection area, the traffic monitoring device simultaneously activates the geomagnetic sensor and the high-definition camera. The geomagnetic sensor captures the magnetic field change signal generated when the vehicle passes by, but due to a large truck passing through the adjacent lane at the same time, the geomagnetic raw signal is significantly interfered. The measured vehicle length extracted from the anti-interference geomagnetic signal is 4.78 meters, and the instantaneous speed is 26.4 km / h. The blue A-type sedan of a certain brand is identified from the video data, and the license plate number is "Ji A12***".

[0059] S2. According to the target vehicle type and scene context, extract the reasonable fluctuation range of the vehicle length and the color distribution probability from the pre-constructed vehicle type-scene adaptive baseline library; if the license plate information in the video structured data is recognizable, query the vehicle management office database according to the license plate information to obtain the nominal vehicle length of the target vehicle's registered model; if the license plate information is not recognizable, according to the matching result of the measured vehicle length and the reasonable fluctuation range, and in combination with the vehicle brand similarity threshold, screen candidate models from the vehicle type knowledge graph and obtain the nominal vehicle length of the candidate models.

[0060] Exemplarily, query the vehicle management office database according to the license plate number "Ji A12***", and obtain that the registration information of this vehicle is an A3-type sedan of a certain brand, and the nominal vehicle length is 4.82 meters. At the same time, extract the reasonable fluctuation range of the vehicle length of the A-type sedan of a certain brand from the pre-constructed vehicle type-scene adaptive baseline library as 4.75 - 4.85 meters.

[0061] S3. The deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scenario adaptive baseline library are weighted and fused to generate a contradiction index of the target vehicle features. If the contradiction index exceeds the preset contradiction index threshold, an abnormal alarm is triggered.

[0062] Furthermore, the method for obtaining an anti-interference geomagnetic signal by performing magnetic field gradient decomposition processing on the original geomagnetic signal includes:

[0063] The original geomagnetic signal is decomposed into low-frequency approximate components and high-frequency detail components of multiple scales through wavelet transform;

[0064] KL divergence analysis is used to analyze the noise distribution of high-frequency detail components at each scale, and an adaptive threshold related to the electromagnetic interference intensity of adjacent channels is generated. The adaptive threshold is compared with the amplitude of the high-frequency detail components, and the interference components of the high-frequency detail components that exceed the adaptive threshold are truncated to obtain the first high-frequency detail component.

[0065] After denoising the first high-frequency detail component, the second high-frequency detail component is obtained. Wavelet reconstruction is performed on the second high-frequency detail component and the low-frequency approximation component, and the reconstructed anti-interference geomagnetic signal is smoothed by sliding window mean filtering to generate an anti-interference geomagnetic signal.

[0066] For example, the original geomagnetic signal generated by the blue sedan of a certain brand was decomposed into five components at different scales using wavelet transform: one low-frequency approximation component and four high-frequency detail components. The low-frequency approximation component retains the main characteristics of the vehicle's magnetic field changes, representing the overall trend of the vehicle's passage; while the high-frequency detail components contain edge features and potential noise interference. Analysis shows that the high-frequency components at the second and third scales (corresponding to the 20-60Hz and 60-120Hz frequency bands) contain significant electromagnetic interference, which is consistent with the frequency characteristics generated when a large truck passes through an adjacent lane.

[0067] The noise distribution characteristics of high-frequency detail components at various scales were analyzed using KL divergence. The portion of the high-frequency detail components exceeding an adaptive threshold was identified as interference components and truncated to obtain the first high-frequency detail component.

[0068] The first high-frequency detail component was further denoised using a soft-threshold shrinkage method. A shrinkage coefficient of 0.15 was applied to the second scale and 0.22 to the third scale, effectively reducing noise caused by adjacent channel electromagnetic interference and obtaining the second high-frequency detail component. Finally, the processed second high-frequency detail component and the retained low-frequency approximation component were reconstructed using wavelet analysis, and the reconstructed signal was smoothed using a sliding window mean filter of length 7 to generate the final anti-interference geomagnetic signal. The processed anti-interference geomagnetic signal reduced vehicle length distortion caused by adjacent channel interference from the initial 19.7% to 2.3%, and improved the signal-to-noise ratio from 7.2dB to 15.4dB, providing a reliable data foundation for subsequent vehicle feature extraction.

[0069] Furthermore, the method for generating an adaptive threshold related to the intensity of adjacent channel electromagnetic interference by analyzing the noise distribution of high-frequency detail components at each scale using KL divergence analysis includes:

[0070] The noise signals of high-frequency detail components at each scale are extracted and their probability density distributions are calculated to obtain the noise distribution of high-frequency detail components; a preset reference noise distribution is obtained, which is statistically generated based on geomagnetic signals in a scenario without adjacent channel interference.

[0071] The deviation between the high-frequency detail component noise distribution and the preset reference noise distribution is quantified by KL divergence, and the adaptive threshold is dynamically adjusted according to the deviation.

[0072] For example, to accurately assess the impact of adjacent channel electromagnetic interference on geomagnetic signals, noise signals of high-frequency detail components at each scale are first extracted and probability density analysis is performed. For the geomagnetic signal when a target vehicle passes, the noise component is separated from the high-frequency components at the second and third scales, and the probability density functions of its amplitude distribution are calculated respectively. The second-scale noise exhibits a skewed distribution, mainly concentrated in the positive offset region, with a peak value of 0.36 and a skewness coefficient reaching 0.78. The third-scale noise exhibits a more complex bimodal distribution characteristic, with a main peak value of 0.42, a secondary peak value of 0.27, and a distance of 0.15 between the two peaks. This distribution characteristic is highly consistent with the standard electromagnetic interference mode.

[0073] A standard noise distribution, obtained from a pre-collected, interference-free environment, was used as a reference noise distribution and compared with the current noise distribution of each high-frequency component. This reference noise distribution was generated statistically from 720 hours of continuous geomagnetic signal collection from the same toll station under conditions of no adjacent lane interference. The reference noise distribution exhibits a typical Gaussian distribution, with a mean close to 0, a standard deviation of 0.12, and a kurtosis of 2.98. The noise distributions of the currently collected high-frequency detail components at each scale were compared with the reference noise distribution, and the degree of deviation between the two was quantified using the KL divergence algorithm. The KL divergence value was 0.86 at the second scale and reached 1.73 at the third scale, while the first and fourth scales were only 0.28 and 0.33, respectively, indicating that the second and third scales were indeed significantly affected by external interference.

[0074] Based on the calculated KL divergence values, a nonlinear mapping method is used to dynamically adjust the adaptive threshold. For the second scale, the initial adaptive threshold of 10% is adjusted to 18% based on a KL divergence value of 0.86; for the third scale, the initial adaptive threshold of 12% is adjusted to 24% based on a KL divergence value of 1.73. This dynamic threshold adjustment mechanism ensures that optimal noise suppression is maintained under different interference intensities, avoiding excessive loss of effective signals while filtering out interference components to the greatest extent possible.

[0075] Furthermore, the method for spatiotemporal alignment of the anti-interference geomagnetic signal and the video structured data includes:

[0076] The spatiotemporal alignment includes spatial calibration and time synchronization. The spatial calibration is to calculate the transformation matrix between the geomagnetic coordinate system and the video image coordinate system based on a preset calibration board, and to correct the spatial position deviation between the geomagnetic sensor and the video acquisition device in real time through a dual Kalman filter to ensure that the spatial position deviation is less than a preset position deviation threshold.

[0077] The time synchronization is achieved by aligning the timestamps of the geomagnetic sensor and the video acquisition device according to a precise clock synchronization protocol, so that the data acquisition time deviation is less than a preset time deviation threshold.

[0078] The spatially calibrated anti-interference geomagnetic signal and the time-synchronized video structured data are correlated and matched according to the time window to obtain a spatiotemporally consistent fused data stream.

[0079] For example, spatiotemporal alignment preprocessing of the geomagnetic sensor and video equipment was completed before the vehicle passed through the detection area. For spatial calibration, managers used a standard 1.5m × 1.5m calibration board to place 11 preset calibration points within the detection area, establishing a transformation relationship between the geomagnetic coordinate system and the video image coordinate system. The calculated transformation matrix had an accuracy of 0.05m in the X-axis direction, 0.08m in the Y-axis direction, and 0.12m in the Z-axis direction. To address minor changes in the positional relationship between sensors (such as ground subsidence, camera attitude changes, etc.), a dual Kalman filter was used to correct spatial positional deviations in real time. For instance, if a horizontal offset of 0.07m and a vertical offset of 0.03m were detected between the geomagnetic sensor and its installation reference position, the spatial positional deviation was controlled within 0.04m after correction using the dual Kalman filter, far below the preset 0.15m positional deviation threshold.

[0080] For time synchronization, a precise clock synchronization protocol is used to align the timestamps of the geomagnetic sensor and the video acquisition equipment. This protocol is based on an improved version of the Network Time Protocol (NTP) and adds hardware timestamp functionality. A time synchronization server is deployed locally at the toll station. The geomagnetic sensor and video equipment synchronize with the server every 15 seconds, ultimately controlling the time deviation between the two types of equipment to within 2.3 milliseconds, far below the preset 10-millisecond time deviation threshold. For example, when a blue A-type sedan of a certain brand passes through the detection area, the geomagnetic sensor records the start time of passage as 16:30:25.762, while the corresponding time recorded by the video equipment is 16:30:25.764, a time difference of only 2 milliseconds, ensuring high-precision time alignment of the data.

[0081] Furthermore, the method for correlating and matching the spatially calibrated anti-interference geomagnetic signal with the time-synchronized video structured data according to a time window to obtain a spatiotemporally consistent fused data stream includes:

[0082] The time window is divided based on the frame timestamp of the video structured data, and the video frames in each time window and their corresponding vehicle brand and color features are extracted.

[0083] Within the same time window, the corresponding geomagnetic waveform data is extracted from the anti-interference geomagnetic signal, and the start and end times of the vehicle passing the geomagnetic sensor are located by the peak detection method; based on the start and end times, the average speed of the geomagnetic signal and the vehicle length measurement value within the time window are calculated.

[0084] The vehicle brand and color features of the video frame are aligned with the average speed and vehicle length measurements by timestamp to generate spatiotemporally correlated fused data units; the fused data units of consecutive time windows are serialized and spliced ​​to form a spatiotemporally consistent fused data stream.

[0085] Exemplarily, in the specific association matching process, time windows are first divided based on the frame timestamps of the video structured data. For example, within 1.8 seconds when a certain blue brand A-type sedan passes through the detection area, the camera captures 54 images at a rate of 30 frames per second. These 54 frames are divided into 18 time windows in groups of 3 frames each, with each window having a width of 100 milliseconds. In the first window, the video structured algorithm extracts the front bumper feature of the vehicle, preliminarily identifies a certain brand logo, and captures the blue feature with a hue value of 215 - 225 and a saturation of 75% - 85%. At the same time, the license plate area is detected but not fully in view. As the vehicle moves forward, the 6th window captures the complete front part of the vehicle. At this time, the brand is identified as "a certain brand", the vehicle type is initially judged as "type A" with a confidence level of 87%, and the license plate number "Ji A12***" is fully identified with a confidence level of 95%.

[0086] For the same time windows, corresponding geomagnetic waveform data is extracted from the anti-interference geomagnetic signals. Taking the 6th window as an example, the sampling rate of the geomagnetic sensor is 1000Hz, and this window contains 100 sampling points. The geomagnetic signal changes significantly when the vehicle passes through, forming peaks. By detecting these peaks, the specific time when the vehicle passes through the geomagnetic sensor can be determined. Through the peak detection method, the start time when the vehicle passes through the geomagnetic sensor is determined to be 16:30:25.762 (front wheel trigger), the end time is 16:30:27.581 (rear wheel passes), and the total duration is 1.819 seconds. Based on this time range and signal characteristics, the instantaneous speed of the vehicle within the 6th window is calculated to be 27.2 km / h, slightly higher than its average speed of 26.4 km / h when passing through the entire detection area, which is consistent with the behavior of the vehicle slightly decelerating within the detection area. At the same time, as the vehicle continuously passes through and accumulates body length information, by the end of the 18th window, the measured vehicle length is calculated to be 4.78 meters.

[0087] For each time window, the features extracted from the video frames are precisely aligned with the features extracted from the geomagnetic signals according to the timestamps to form a fused data unit. For example, in the 12th window, at this time the middle section of the vehicle is exactly in the center of the detection area, and the video captures the complete side profile, confirming it as a four-door sedan, and the confidence level of the A-type vehicle increases to 94%. At the same time, the geomagnetic signal intensity reaches a peak of 79.5 microteslas, and the vehicle feature matching degree is the highest. A unique identifier is assigned to each fused data unit. The fused data units of these 18 time windows are spliced in sequence to generate a continuous fused data stream, which completely records the whole process of the certain blue brand A-type sedan passing through the detection area, reflects the continuous movement trajectory of the vehicle within the entire monitoring area, and provides a high-quality data basis for subsequent calculation of the contradiction index. The missing data is filled by linear interpolation or data compensation of adjacent windows.

[0088] Furthermore, the method for extracting the reasonable fluctuation range of vehicle length and the probability of color distribution from the vehicle model-scene adaptive baseline library based on the target vehicle type and scene context includes:

[0089] Based on the vehicle model-scenario adaptive baseline library, the benchmark scenario set corresponding to the target vehicle type is matched. The benchmark scenario set includes the vehicle length distribution characteristics under different traffic flow levels, weather conditions and time period combinations within the historical statistical period.

[0090] Based on the real-time traffic flow level, weather conditions, and time period parameters in the current scenario context, extract the N historical scenario clusters with the highest correlation from the benchmark scenario set, calculate the weighted quantile interval of the vehicle length measurement value of each historical scenario cluster, and generate a reasonable fluctuation range of vehicle length, where N≥2.

[0091] Based on the scene color feature maps stored in the vehicle model-scene adaptive baseline library, spatiotemporal similarity matching is performed on the current scene context to extract scene color distribution patterns that meet a preset similarity threshold. Combined with real-time collected vehicle color data, Bayesian probability analysis is used to obtain color distribution probabilities. The vehicle length distribution features, reasonable fluctuation range of vehicle length, and color distribution probabilities of the current scene context are written back to the vehicle model-scene adaptive baseline library as incremental datasets, and the scene feature weight coefficients are updated through a sliding time window mechanism. The vehicle model-scene adaptive baseline library contains three core data structures: vehicle model-scene mapping relationships, scene color feature maps, and historical speed curve sets. The scene color feature maps store the vehicle color probability distribution under different spatiotemporal scenarios.

[0092] For example, to determine whether the characteristics of a blue model A sedan of a certain brand meet the expected range, the reasonable fluctuation range of vehicle length and the probability of color distribution are extracted from the vehicle type-scenario adaptive baseline library. First, based on the vehicle recognition result "model A sedan of a certain brand", the corresponding benchmark scene set is matched from the baseline library. This benchmark scene set contains 2784 passage records of model A sedans of a certain brand collected at the toll station exit in the past 180 days. These records are classified according to traffic flow level (low, medium, high), weather conditions (sunny, cloudy, rainy, snowy), and time period (morning peak, daytime off-peak, evening peak, night), forming a vehicle length distribution feature library of 48 different scene combinations.

[0093] Based on the current scenario context—16:30 (off-peak daytime), moderate traffic flow (350 vehicles / hour), and cloudy weather—the five historical scenario clusters with the highest similarity were selected from the baseline scenario set: ① Off-peak daytime - moderate traffic flow - cloudy (98.3% similarity); ② Off-peak daytime - moderate traffic flow - sunny (92.1% similarity); ③ Off-peak daytime - high traffic flow - cloudy (87.6% similarity); ④ Evening peak daytime - moderate traffic flow - cloudy (82.4% similarity); ⑤ Off-peak daytime - low traffic flow - cloudy (79.2% similarity). Weights were assigned to these five scenario clusters: 0.35, 0.25, 0.18, 0.12, and 0.10, respectively. Then, the weighted quantile intervals for the length of a certain brand of Type A vehicle within each scenario cluster were calculated. In the scene cluster ① with the highest similarity, the 25th percentile of the length of a certain brand's A-type vehicle is 4.76 meters and the 75th percentile is 4.84 meters. After applying the weighted calculation of all scene clusters, the reasonable fluctuation range obtained is 4.75-4.85 meters, which indicates that the measured vehicle length of 4.78 meters falls within the reasonable range.

[0094] Simultaneously, it is necessary to determine whether the vehicle color features conform to the expected distribution. From the scene color feature map of the vehicle-scene adaptive baseline library, color distribution patterns with high similarity (threshold set to 80%) to the current scene are extracted. In the matched scene data, the color distribution of a certain brand's Type A sedan is: black 38.6%, white 27.5%, silver 15.8%, blue 13.2%, red 3.2%, and other colors 1.7%. Combining the color data of 462 Type A sedans of a certain brand collected in the last 30 days, a Bayesian probability update model is applied to generate the latest color distribution probability of a certain brand's Type A sedan in the current scene: black 37.9%, white 28.1%, silver 16.3%, blue 12.7%, red 3.4%, and other colors 1.6%. As can be seen from the comparison, the probability of the blue color of a certain brand type A appearing in the current scenario is 12.7%, which is very close to the historical statistical value of 13.2%, with a deviation of only 0.5 percentage points, far below the deviation threshold of 2 percentage points. This means that the color feature of the blue color of a certain brand type A is consistent with the expected distribution and is not an abnormal situation.

[0095] Furthermore, the method for selecting candidate models and obtaining their nominal vehicle lengths from the vehicle model knowledge graph based on the matching results of the measured vehicle length and the reasonable fluctuation range, combined with a vehicle brand similarity threshold, includes:

[0096] The nominal vehicle length range and associated brand characteristics of all candidate models are obtained from the vehicle model knowledge graph. Candidate models whose nominal vehicle length range intersects with the reasonable fluctuation range of the measured vehicle length are selected to obtain a preliminary candidate set. The vehicle model knowledge graph and the vehicle model-scenario adaptive baseline library are associated and mapped through vehicle model coding.

[0097] The similarity between the vehicle brand features extracted from the video structured data and the associated brand features of each candidate model in the preliminary candidate set is quantitatively calculated using a semantic matching model.

[0098] The final candidate model list is obtained by selecting candidate models from the preliminary candidate set whose similarity to related brand features is greater than the vehicle brand similarity threshold.

[0099] Based on the matching degree between the nominal vehicle length range and the actual vehicle length of each model in the final candidate model list, the nominal vehicle length of the candidate model with the highest matching degree is selected as the output result.

[0100] For example, in a real-world scenario, if the license plate information cannot be identified, candidate models need to be selected from the vehicle model knowledge graph based on the measured vehicle length and the matching results within a reasonable fluctuation range, combined with vehicle brand similarity. For instance, when a blue sedan passes through the detection area, the license plate number cannot be identified due to severe damage, but the measured vehicle length is 4.78 meters according to geomagnetic measurements. The video identifies the brand characteristics of "a certain brand," with a confidence level of 89%. First, the nominal vehicle length range of all models of that brand is retrieved from the vehicle model knowledge graph: Type D (4.42-4.46 meters), Type C (4.56-4.62 meters), Type A (4.75-4.85 meters), Type B (4.92-5.05 meters), etc. Based on the measured vehicle length of 4.78 meters and its reasonable fluctuation range of 4.75-4.85 meters, candidate models with overlapping nominal vehicle length ranges were selected to form a preliminary candidate set: Type A sedan (4.75-4.85 meters), Type E four-door coupe (4.70-4.76 meters), and Type F SUV coupe (4.73-4.77 meters).

[0101] Next, the similarity between the vehicle brand features extracted from the video data and the associated brand features of each candidate model in the preliminary candidate set is quantitatively calculated using a semantic matching model.

[0102] Similarity scores were calculated using a semantic matching model: Type A sedan 0.92, Type E four-door coupe 0.76, and Type F SUV coupe 0.68. These similarity scores were compared to a preset vehicle brand similarity threshold of 0.80, filtering out candidate models with similarities greater than the threshold, resulting in a final candidate model list containing only Type A sedans with a similarity of 0.92. The matching degree between the nominal length range (4.75-4.85 meters) of Type A sedan and the measured length (4.78 meters) was then calculated. The formula is: 1 - |Measured length - Midpoint of nominal length| / (Width of nominal length range / 2), yielding a matching degree of 0.96. Since Type A sedans are the only model in the final candidate model list and have a high matching degree of 0.96, the nominal length of Type A sedans (midpoint of the range) of 4.80 meters was selected as the output result for subsequent contradiction index calculation. Even when license plate information is unavailable, the vehicle model and nominal length can be accurately inferred through multi-source data fusion.

[0103] Furthermore, the method for quantifying the similarity between the vehicle brand features extracted from the video structured data and the associated brand features of each candidate model in the preliminary candidate set through a semantic matching model includes:

[0104] Semantic description text of the brand features associated with candidate models is extracted from the vehicle model knowledge graph. The semantic description includes the brand name, brand logo, and vehicle series information.

[0105] Vehicle brand features extracted from structured video data are converted into semantic vectors. The cosine similarity between these semantic vectors and the embedding vectors of candidate model semantic description texts is calculated using the semantic matching model and recorded as the associated brand feature similarity. The conversion into semantic vectors is achieved by encoding the vehicle brand name and visual feature description using a pre-trained natural language processing model.

[0106] For example, to accurately quantify the similarity of vehicle brand features, a deep semantic matching model is used for calculation. Continuing with the example of a blue sedan from a certain brand, when the license plate cannot be identified, it is necessary to extract the associated brand features of candidate models from the vehicle model knowledge graph for semantic description. For the three models in the initial candidate set, their complete semantic description text is extracted respectively: The semantic description of the A-type sedan is "A-type of a certain brand, mid-to-high-end luxury sedan, distinctive grille design, brand logo in the center, streamlined body outline, horizontally extended distinctive taillights, standard platform"; the semantic description of the E-type four-door coupe is "E-type of a certain brand, entry-level four-door coupe, diamond grille, fastback roofline, sharp headlight design, aerodynamic design, trapezoidal dual exhaust layout"; the semantic description of the F-type SUV coupe is "F-type of a certain brand, mid-size luxury SUV coupe, large-size grille, high ground clearance design, sliding fastback roof, large-size five-spoke wheels, sporty rear spoiler".

[0107] Simultaneously, vehicle brand features extracted from the video structured data are converted into semantic vectors. These vehicle brand features include: "brand logo front grille, classic sedan styling, smooth body side lines, horizontally arranged taillights, traditional three-box design, four-door layout, and standard wheel design." A pre-trained vehicle semantic model is used to transform these vehicle brand features into 768-dimensional semantic vectors, which represent the key visual elements and structural features of the vehicle's appearance.

[0108] The cosine similarity between the semantic vector and the embedding vector of the semantic description text of the candidate models was calculated. The semantic descriptions of the three candidate models were converted into embedding vectors of the same dimension using a semantic encoder. The cosine similarity between the video extracted features and the semantic description of the Type A sedan was 0.92, the similarity with the Type E four-door coupe was 0.76, and the similarity with the Type F SUV coupe was 0.68. These similarity scores directly reflect the degree of matching between the vehicle in the video and each candidate model in terms of appearance, structure, and brand characteristics. The high similarity (0.92) of the Type A sedan far exceeds the preset threshold of 0.80, while the other two models do not reach the threshold, confirming that the car is most likely a Type A sedan of a certain brand. The deep semantic matching method can accurately identify the specific model of the vehicle even when license plate information is missing, providing a reliable basis for subsequent determination of nominal vehicle length and calculation of the contradiction index.

[0109] Furthermore, the method for generating a contradiction index of target vehicle features by weighted fusion of the deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scene adaptive baseline library includes:

[0110] The deviation between the measured vehicle length and the nominal vehicle length is characterized by the ratio of the absolute difference between the measured vehicle length and the nominal vehicle length to the nominal vehicle length.

[0111] The color distribution probability deviation is quantified by the difference between the frequency of occurrence of the current vehicle color in the corresponding scenario of the model-scenario adaptive baseline library and the historical statistical frequency.

[0112] The difference between the instantaneous velocity and the historical velocity curve is characterized by the degree of deviation of the instantaneous velocity from the probability distribution of the historical velocity curve.

[0113] The fusion weights of deviation, color distribution probability deviation, and speed difference are dynamically adjusted based on the current scene context. A contradiction index of the target vehicle's features is generated by linearly weighting and fusing these factors according to the adjusted fusion weights. The higher the traffic flow, the greater the weight allocation for deviation; the stronger the environmental interference, the smaller the weight allocation for speed difference.

[0114] For example, after all features of a blue A-type sedan of a certain brand are extracted, its feature contradiction index is calculated to determine whether the vehicle has any abnormalities. The deviation between the measured vehicle length and the nominal vehicle length is obtained. The measured vehicle length is 4.78 meters, and the nominal vehicle length is 4.82 meters (according to the vehicle management office database of a certain brand A3 model). The absolute difference between the two is 0.04 meters. The ratio relative to the nominal vehicle length is calculated using the formula: |4.78-4.82| / 4.82×100%, which is 0.83%. According to historical data analysis, under normal circumstances, the deviation of vehicle length measured by geomagnetism is usually within 1.5%, so the deviation of 0.83% is judged to be within the normal range.

[0115] According to the vehicle model-scenario adaptive baseline library, under the current scenario conditions (daytime off-peak, moderate traffic, cloudy weather), the historical frequency of blue color distribution on a certain brand's Type A vehicle is 13.2%, while the real-time frequency over the last 30 days is 12.7%. The difference between the two is 0.5 percentage points, which is far below the preset threshold of 2 percentage points, indicating that the vehicle color characteristics conform to the historical distribution pattern.

[0116] The speed difference between the instantaneous speed and the historical speed curve was obtained. The average speed of the vehicle passing through the detection area was 26.4 km / h, while according to the historical speed curve of this vehicle type for this time period in the model-scenario adaptive baseline library, the average speed of a certain brand's Type A vehicle in this area was 25.8 km / h, with a standard deviation of 1.7 km / h. The speed difference between the two is 2.3%, which is lower than the preset 5% threshold, indicating that the vehicle speed characteristics are within the normal fluctuation range.

[0117] The fusion weights of each indicator are dynamically adjusted based on the current scene context. Considering that traffic flow is moderate and lighting is sufficient in the afternoon, resulting in higher video recognition accuracy, and that the probability of interference from adjacent lanes to the geomagnetic sensor increases, the weights for vehicle length deviation, color distribution probability deviation, and speed difference are automatically set to 0.5, 0.3, and 0.2, respectively. Applying the linear weighted fusion formula, the conflict index of a blue vehicle of a certain brand (Type A) is calculated as: 0.5 × 0.83% / 1.5% + 0.3 × 0.5 / 2 + 0.2 × 2.3% / 5% = 0.44. The final conflict index for this vehicle is 0.44, which is lower than the preset conflict index threshold of 1.5. Therefore, no abnormal alarm is triggered, and the vehicle is determined to be a normally passing vehicle. This reliably filters out normal behavior, effectively reducing the false alarm rate and improving traffic monitoring efficiency.

[0118] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a fusion and invocation system based on multi-source traffic data. The system includes: a data processing module, a nominal vehicle length acquisition module, and a vehicle analysis module, with each module connected in a sequential communication manner.

[0119] The data processing module is used to acquire the raw geomagnetic signal and video structured data of the target vehicle in real time, perform magnetic field gradient decomposition on the raw geomagnetic signal to obtain an anti-interference geomagnetic signal, and perform spatiotemporal alignment on the anti-interference geomagnetic signal and the video structured data; extract the measured vehicle length and instantaneous speed from the spatiotemporally aligned anti-interference geomagnetic signal, and extract the vehicle brand, color features and license plate information from the video structured data.

[0120] The nominal vehicle length acquisition module is used to extract the reasonable fluctuation range of vehicle length and color distribution probability from a pre-built vehicle model-scene adaptive baseline library based on the target vehicle type and scene context. If the license plate information in the video structured data is identifiable, the module queries the vehicle management office database based on the license plate information to obtain the nominal vehicle length of the target vehicle registration model. If the license plate information is not identifiable, the module selects candidate models from the vehicle model knowledge graph and obtains the nominal vehicle length of the candidate models based on the matching result of the measured vehicle length and the reasonable fluctuation range, combined with the vehicle brand similarity threshold.

[0121] The vehicle analysis module is used to generate a contradiction index of the target vehicle features by weighted fusion of the deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scenario adaptive baseline library. If the contradiction index exceeds the preset contradiction index threshold, an abnormal alarm is triggered.

[0122] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0123] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fusing and retrieving multi-source traffic data, characterized in that, The method includes: The system acquires the raw geomagnetic signal and video structured data of the target vehicle in real time. It performs magnetic field gradient decomposition on the raw geomagnetic signal to obtain an anti-interference geomagnetic signal. It then performs spatiotemporal alignment on the anti-interference geomagnetic signal and the video structured data. The system extracts the measured vehicle length and instantaneous speed from the spatiotemporally aligned anti-interference geomagnetic signal and extracts the vehicle brand, color features, and license plate information from the video structured data. Based on the target vehicle type and scene context, the reasonable fluctuation range of vehicle length and color distribution probability are extracted from the pre-built vehicle model-scene adaptive baseline library; if the license plate information in the video structured data is identifiable, the nominal vehicle length of the target vehicle registration model is obtained by querying the vehicle management office database based on the license plate information; if the license plate information is not identifiable, the candidate models are selected from the vehicle model knowledge graph and the nominal vehicle length of the candidate models are obtained based on the matching result of the measured vehicle length and the reasonable fluctuation range, combined with the vehicle brand similarity threshold. The deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scenario adaptive baseline library are weighted and fused to generate a contradiction index of the target vehicle features. If the contradiction index exceeds the preset contradiction index threshold, an abnormal alarm is triggered. The method for extracting the reasonable fluctuation range of vehicle length and the probability of color distribution from the vehicle model-scene adaptive baseline library based on the target vehicle type and scene context includes: Based on the vehicle model-scenario adaptive baseline library, the benchmark scenario set corresponding to the target vehicle type is matched. The benchmark scenario set includes the vehicle length distribution characteristics under different traffic flow levels, weather conditions and time period combinations within the historical statistical period. Based on the real-time traffic flow level, weather conditions and time period parameters in the current scenario context, extract the N historical scenario clusters with the highest correlation from the benchmark scenario set, calculate the weighted quantile interval of the vehicle length measurement value of each historical scenario cluster, and generate a reasonable fluctuation range of vehicle length, N≥2. Based on the scene color feature map stored in the vehicle model-scene adaptive baseline library, spatiotemporal similarity matching is performed on the current scene context to extract scene color distribution patterns that meet the preset similarity threshold. Combined with real-time collected vehicle color data, color distribution probability is obtained through Bayesian probability analysis. The method for generating a contradiction index of target vehicle features by weighted fusion of the deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scene adaptive baseline library includes: The deviation between the measured vehicle length and the nominal vehicle length is characterized by the ratio of the absolute difference between the measured vehicle length and the nominal vehicle length to the nominal vehicle length. The color distribution probability deviation is quantified by the difference between the frequency of occurrence of the current vehicle color in the corresponding scenario of the model-scenario adaptive baseline library and the historical statistical frequency. The difference between the instantaneous velocity and the historical velocity curve is characterized by the degree of deviation of the instantaneous velocity from the probability distribution of the historical velocity curve; The fusion weights of the deviation, color distribution probability deviation, and speed difference are dynamically adjusted according to the current scene context; the deviation, color distribution probability deviation, and speed difference are linearly weighted and fused according to the adjusted fusion weights to generate a contradiction index of the target vehicle features.

2. The method for fusing and invoking multi-source traffic data according to claim 1, characterized in that, The method for obtaining an anti-interference geomagnetic signal by performing magnetic field gradient decomposition processing on the original geomagnetic signal includes: The original geomagnetic signal is decomposed into low-frequency approximate components and high-frequency detail components of multiple scales through wavelet transform; KL divergence analysis is used to analyze the noise distribution of high-frequency detail components at each scale, and an adaptive threshold related to the electromagnetic interference intensity of adjacent channels is generated. The adaptive threshold is compared with the amplitude of the high-frequency detail components, and the interference components of the high-frequency detail components that exceed the adaptive threshold are truncated to obtain the first high-frequency detail component. After denoising the first high-frequency detail component, the second high-frequency detail component is obtained. Wavelet reconstruction is performed on the second high-frequency detail component and the low-frequency approximation component, and the reconstructed anti-interference geomagnetic signal is smoothed by sliding window mean filtering to generate an anti-interference geomagnetic signal.

3. The method for fusing and invoking multi-source traffic data according to claim 2, characterized in that, The method for generating an adaptive threshold related to the intensity of adjacent channel electromagnetic interference by analyzing the noise distribution of high-frequency detail components at various scales using KL divergence analysis includes: The noise signals of high-frequency detail components at each scale are extracted and their probability density distributions are calculated to obtain the noise distribution of high-frequency detail components; a preset reference noise distribution is obtained, which is statistically generated based on the geomagnetic signal in a scenario without adjacent channel interference. The deviation between the high-frequency detail component noise distribution and the preset reference noise distribution is quantified by KL divergence, and the adaptive threshold is dynamically adjusted according to the deviation.

4. The method for fusing and invoking multi-source traffic data according to claim 1, characterized in that, The method for spatiotemporal alignment of the anti-interference geomagnetic signal and video structured data includes: The spatiotemporal alignment includes spatial calibration and time synchronization. The spatial calibration is to calculate the transformation matrix between the geomagnetic coordinate system and the video image coordinate system based on a preset calibration board, and to correct the spatial position deviation between the geomagnetic sensor and the video acquisition device in real time through a dual Kalman filter to ensure that the spatial position deviation is less than a preset position deviation threshold. The time synchronization is to align the timestamps of the geomagnetic sensor and the video acquisition device according to a precise clock synchronization protocol, so that the data acquisition time deviation is less than a preset time deviation threshold. The spatially calibrated anti-interference geomagnetic signal and the time-synchronized video structured data are correlated and matched according to the time window to obtain a spatiotemporally consistent fused data stream.

5. The method according to claim 4, characterized in that, The method for correlating and matching spatially calibrated anti-interference geomagnetic signals with time-synchronized video structured data according to a time window to obtain a spatiotemporally consistent fused data stream includes: The time window is divided based on the frame timestamp of the video structured data, and the video frames in each time window and their corresponding vehicle brand and color features are extracted. Within the same time window, the corresponding geomagnetic waveform data is extracted from the anti-interference geomagnetic signal, and the start and end times of the vehicle passing the geomagnetic sensor are located by the peak detection method; based on the start and end times, the average speed of the geomagnetic signal and the vehicle length measurement value within the time window are calculated; The vehicle brand and color features of the video frame are aligned with the average speed and vehicle length measurements by timestamp to generate spatiotemporally correlated fused data units; the fused data units of consecutive time windows are serialized and spliced ​​to form a spatiotemporally consistent fused data stream.

6. The method for fusing and invoking multi-source traffic data according to claim 1, characterized in that, The method for selecting candidate models and obtaining their nominal vehicle lengths from the vehicle model knowledge graph based on the matching results of the measured vehicle length and the reasonable fluctuation range, combined with a vehicle brand similarity threshold, includes: Obtain the nominal vehicle length range and associated brand characteristics of all candidate models from the vehicle knowledge graph; filter out candidate models whose nominal vehicle length range intersects with the reasonable fluctuation range of the measured vehicle length to obtain a preliminary candidate set; The similarity between vehicle brand features extracted from video structured data and associated brand features of each candidate model in the preliminary candidate set is quantitatively calculated using a semantic matching model. The final candidate model list is obtained by selecting candidate models from the preliminary candidate set whose similarity to related brand features is greater than the vehicle brand similarity threshold. Based on the matching degree between the nominal vehicle length range and the actual vehicle length of each model in the final candidate model list, the nominal vehicle length of the candidate model with the highest matching degree is selected as the output result.

7. The method for fusing and invoking multi-source traffic data according to claim 6, characterized in that, The method for quantifying the similarity between the vehicle brand features extracted from the video structured data and the associated brand features of each candidate model in the preliminary candidate set through a semantic matching model includes: Semantic description text of brand features associated with candidate models is extracted from the vehicle model knowledge graph. The semantic description includes brand name, brand logo and vehicle series information. The vehicle brand features extracted from the video structured data are converted into semantic vectors; the cosine similarity between the semantic vectors and the embedding vectors of the candidate model semantic description text is calculated according to the semantic matching model and recorded as the associated brand feature similarity.

8. A fusion and invocation system based on multi-source traffic data, characterized in that, The system includes: a data processing module, a nominal vehicle length acquisition module, and a vehicle analysis module, with each module communicating with the others in sequence. The data processing module is used to acquire the raw geomagnetic signal and video structured data of the target vehicle in real time, perform magnetic field gradient decomposition on the raw geomagnetic signal to obtain an anti-interference geomagnetic signal, and perform spatiotemporal alignment on the anti-interference geomagnetic signal and the video structured data; extract the measured vehicle length and instantaneous speed from the spatiotemporally aligned anti-interference geomagnetic signal, and extract the vehicle brand, color features and license plate information from the video structured data; The nominal vehicle length acquisition module is used to extract the reasonable fluctuation range of vehicle length and color distribution probability from a pre-built vehicle model-scene adaptive baseline library based on the target vehicle type and scene context; if the license plate information in the video structured data is identifiable, the module queries the vehicle management office database based on the license plate information to obtain the nominal vehicle length of the target vehicle registration model; if the license plate information is not identifiable, the module selects candidate models from the vehicle model knowledge graph and obtains the nominal vehicle length of the candidate models based on the matching result of the measured vehicle length and the reasonable fluctuation range, combined with the vehicle brand similarity threshold. The vehicle analysis module is used to generate a contradiction index of the target vehicle features by weighted fusion of the deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scenario adaptive baseline library. If the contradiction index exceeds the preset contradiction index threshold, an abnormal alarm is triggered. The method for extracting the reasonable fluctuation range of vehicle length and the probability of color distribution from the vehicle model-scene adaptive baseline library based on the target vehicle type and scene context includes: Based on the vehicle model-scenario adaptive baseline library, the benchmark scenario set corresponding to the target vehicle type is matched. The benchmark scenario set includes the vehicle length distribution characteristics under different traffic flow levels, weather conditions and time period combinations within the historical statistical period. Based on the real-time traffic flow level, weather conditions and time period parameters in the current scenario context, extract the N historical scenario clusters with the highest correlation from the benchmark scenario set, calculate the weighted quantile interval of the vehicle length measurement value of each historical scenario cluster, and generate a reasonable fluctuation range of vehicle length, N≥2. Based on the scene color feature map stored in the vehicle model-scene adaptive baseline library, spatiotemporal similarity matching is performed on the current scene context to extract scene color distribution patterns that meet the preset similarity threshold. Combined with real-time collected vehicle color data, color distribution probability is obtained through Bayesian probability analysis. The method for generating a contradiction index of target vehicle features by weighted fusion of the deviation between the measured vehicle length and the nominal vehicle length, the color distribution probability deviation, and the speed difference between the instantaneous speed and the historical speed curve in the vehicle model-scene adaptive baseline library includes: The deviation between the measured vehicle length and the nominal vehicle length is characterized by the ratio of the absolute difference between the measured vehicle length and the nominal vehicle length to the nominal vehicle length. The color distribution probability deviation is quantified by the difference between the frequency of occurrence of the current vehicle color in the corresponding scenario of the model-scenario adaptive baseline library and the historical statistical frequency. The difference between the instantaneous velocity and the historical velocity curve is characterized by the degree of deviation of the instantaneous velocity from the probability distribution of the historical velocity curve; The fusion weights of the deviation, color distribution probability deviation, and speed difference are dynamically adjusted according to the current scene context; the deviation, color distribution probability deviation, and speed difference are linearly weighted and fused according to the adjusted fusion weights to generate a contradiction index of the target vehicle features.

Images