Road traffic safety monitoring and early warning method based on multi-source data fusion

Abstract

The application discloses a road traffic safety monitoring and early warning method based on multi-source data fusion, and belongs to the technical field of road traffic safety monitoring. The method comprises the following steps: collecting vehicle motion trajectory data, meteorological environment data and road surface state data, and constructing a heterogeneous data cube with a timestamp as an index; based on the heterogeneous data cube, the cross-correlation entropy matrix between each data source is calculated, and the cross-correlation entropy matrix is reduced to a low-dimensional feature space by using sparse constraint non-negative matrix decomposition to obtain a fusion feature basis matrix; the real-time collected multi-source data is projected by using the fusion feature basis matrix to obtain a fusion feature vector at the current moment, and the shape distance between the fusion feature vector and a historical safety feature template is calculated based on a dynamic time warping algorithm; according to the shape distance and a preset risk probability distribution, the current early warning threshold is updated through a Bayesian factor, and a road danger early warning signal is generated when the shape distance exceeds the updated early warning threshold.

Description

Technical Field

[0001] This invention relates to the field of road traffic safety monitoring technology, specifically a road traffic safety monitoring and early warning method based on multi-source data fusion. Background Technology

[0002] In the field of road traffic safety monitoring and early warning, conventional technical solutions often rely on data collected by a single type of sensor. For example, they might analyze vehicle motion through video images or measure vehicle speed and occupancy rates using roadside radar. The extracted indicators are then compared with fixed thresholds to determine whether an alarm should be triggered. These solutions only capture some vehicle motion attributes and fail to simultaneously perceive changes in weather and road surface physical conditions within the traffic environment. This results in severely delayed or incomplete perception of dangerous situations in complex scenarios such as rain, fog, and icy roads. Some improved solutions attempt to integrate vehicle trajectory data, meteorological data, and road surface data into a single framework. However, due to the significant heterogeneity of multi-source data in sampling frequency, dimensions, and probability distribution, fusion methods often remain at the level of feature vector concatenation or linear weighting. This makes it difficult to effectively express the deep nonlinear coupling relationships between various factors. The fused feature dimension is high and contains a large amount of redundant information, directly affecting the ability to distinguish changes in road safety conditions. Furthermore, existing early warning systems widely employ fixed threshold values. When weather and road conditions undergo temporary and drastic changes, the characteristic quantities of normal driving conditions will systematically shift, and the fixed thresholds cannot be dynamically adjusted accordingly, easily leading to frequent false alarms or missed detections of real dangers. Historical template matching used in early warning decisions is typically based on point-to-point metrics such as Euclidean distance, ignoring the local scaling and morphological changes of traffic operation states over time, making it difficult to accurately identify progressively accumulating danger trends. Therefore, there is an urgent need for a method that can extract low-dimensional fusion features reflecting the essence of road safety from multi-source heterogeneous data and adjust the early warning trigger boundaries in real time based on risk changes, in order to solve the problems of insufficient multi-source data fusion and fixed thresholds being unsuitable for dynamic environments. Summary of the Invention

[0003] This invention provides a road traffic safety monitoring and early warning method based on multi-source data fusion, aiming to solve the technical problems in the existing technology of insufficient extraction of deep nonlinear correlation features of multi-source heterogeneous data and fixed early warning thresholds that cannot adapt to environmental changes, resulting in false alarms and missed alarms.

[0004] To achieve the above objectives, this invention provides the following technical solution: This invention provides a road traffic safety monitoring and early warning method based on multi-source data fusion. By unifying vehicle trajectory data, meteorological environment data, and road surface condition data into a heterogeneous data cube indexed by timestamps, it achieves spatiotemporal alignment and fusion of multimodal sensing information. During data collection, the method utilizes roadside millimeter-wave radar and vehicle-mounted global navigation satellite system terminals to jointly acquire vehicle trajectory data including vehicle position, instantaneous speed, acceleration, and heading angle. Simultaneously, meteorological environment data consisting of temperature, rainfall intensity, wind speed, and visibility is acquired from meteorological stations deployed on both sides of the road. Furthermore, road surface condition data including road surface temperature, water accumulation thickness, icing probability, and friction coefficient is acquired from piezoelectric sensors and laser road surface condition sensors buried beneath the road surface. The aforementioned multi-source data are stored according to a time-aligned tensor structure. Each time slice retains the feature vectors of all data sources at the corresponding moment, forming a heterogeneous data cube available for subsequent calculations.

[0005] Based on the heterogeneous data cube, a cross-association entropy matrix is ​​calculated between the various data sources. Vehicle trajectory data, meteorological data, and road condition data at the same time stamp are discretized into equally spaced probability mass functions. For any two data sources, the expected value of the logarithmic ratio between their joint probability density and the product of their respective marginal probability densities is calculated as the cross-association entropy between the two data sources. All combinations of data sources are iterated, and the resulting cross-association entropies are arranged into a square matrix to obtain the cross-association entropy matrix. Preferably, to improve the representation of dynamic changes, the continuous time series of the two data sources can be divided into sliding window blocks, with the average cross-association entropy between blocks used as the corresponding element of the matrix. This cross-association entropy matrix quantitatively characterizes the nonlinear statistical dependencies between different data sources, providing structured information for subsequent feature fusion.

[0006] To extract compact and discriminative fusion features from the cross-association entropy matrix, a sparse-constrained nonnegative matrix factorization is used to reduce the matrix's dimensionality. Nonnegative basis and coefficient matrices are initialized, with the number of columns in the basis matrix equal to the preset feature dimension. An objective function is constructed, including a reconstruction error term and a sparse regularization term. The reconstruction error term measures the squared Frobenius norm of the difference between the cross-association entropy matrix and the product of the basis and coefficient matrices, while the sparse regularization term is the sum of the absolute values ​​of all elements in the coefficient matrix. This ensures that only a few components in the coefficient vector are active, enhancing the sparse interpretability of the fusion features. The basis and coefficient matrices are alternately updated until the objective function converges. The converged basis matrix is ​​then used as the fusion feature basis matrix. This fusion feature basis matrix preserves the core patterns in the cross-association of multi-source data while eliminating redundancy and noise interference.

[0007] During the real-time monitoring phase, the multi-source data collected at the current moment is projected using the fusion feature basis matrix to obtain the fusion feature vector for the current moment. As a preferred embodiment of this invention, the real-time collected vehicle trajectory data, meteorological environment data, and road condition data are organized into a real-time data vector according to the same format as constructing the heterogeneous data cube. Then, using the fusion feature basis matrix as the projection basis, the active set gradient projection algorithm iteratively solves for the non-negative coefficient vector that minimizes the projection error. This algorithm dynamically updates the active set in each iteration, moving coefficient components that violate the non-negativity constraint into the active set and performing projection gradient descent, thereby efficiently obtaining the fusion feature vector that satisfies the requirements of non-negativity and sparsity. This vector extracts a compressed representation of the current road traffic safety status from high-dimensional sensor information.

[0008] To measure the deviation of the current traffic state from the safe mode, a dynamic time warping algorithm is used to calculate the morphological distance between the fused feature vector and the historical safety feature template. Multiple fused feature vectors are extracted from multi-source data collected during historical accident-free periods and arranged chronologically. The resulting sequences are clustered and centered to obtain historical safety feature templates representing typical safe states. Using the fused feature vector at the current moment as the query sequence and the historical safety feature template as the reference sequence, a distance matrix is ​​constructed for all corresponding point pairs between the two sequences. Under the premise of satisfying boundary constraints, continuity, and monotonicity, the optimal warped path from the starting point to the ending point is searched, and the minimum cumulative distance of this path is taken as the morphological distance. This morphological distance is robust to local scaling and misalignment on the time axis and can accurately reflect the overall deviation between the traffic state pattern and the safe mode.

[0009] Based on morphological distance and a preset risk probability distribution, the current warning threshold is adaptively adjusted using a Bayesian factor. Probability density functions for morphological distance under normal driving conditions and accident-prone conditions are pre-established from historical data. As a preferred construction method, multi-source data within a preset time period before an accident is extracted, and the morphological distance at each sampling moment relative to the corresponding historical safety feature template is calculated and fitted to a first gamma distribution as the probability density function for the accident-prone state. Simultaneously, continuous multi-source data during accident-free and unobstructed road periods are extracted, and the morphological distance at each sampling moment relative to the same historical safety feature template is calculated and fitted to a second gamma distribution as the probability density function for the normal driving condition. The morphological distances calculated over multiple consecutive time windows prior to the current moment are used as observation samples, and the likelihood ratio of these observation samples under the two probability density functions is calculated to obtain the Bayesian factor. The Bayesian factor is compared with a preset decision threshold. When the Bayesian factor is greater than the decision threshold, it indicates that the current situation is closer to an accident-prone state, and the current warning threshold is lowered to improve sensitivity; when the Bayesian factor is less than the reciprocal of the decision threshold, it indicates that the current situation is more like a normal state, and the current warning threshold is raised to suppress false alarms. This mechanism allows the warning threshold to be dynamically optimized based on the real-time evolution of risk.

[0010] When the morphological distance exceeds the updated warning threshold, a road hazard warning signal is generated. To further improve the reliability and early warning capability, the difference between the updated warning threshold and the morphological distance is used as the hazard margin, and the hazard margins of several adjacent timestamps before and after the current moment are combined into a hazard margin sequence. A sign function transformation is performed on this hazard margin sequence to obtain a sign sequence, and the cumulative sum of the products of adjacent elements in the sign sequence is calculated as a trend strength index. When this trend strength index is positive three times consecutively and the current hazard margin is negative, it indicates that the hazard level has steadily increased and exceeded the threshold. At this point, a road hazard warning signal is generated, and its risk level is marked as high risk. This trend-confirmation-based warning logic effectively eliminates false alarms caused by instantaneous disturbances.

[0011] Preferably, after generating a road hazard warning signal, the dominant data source causing the morphological distance exceedance is located based on the last peak position of the hazard margin sequence. This location process calculates the absolute values ​​of the partial derivatives of vehicle trajectory data, meteorological environment data, and road surface condition data with respect to morphological distance, and identifies the data source with the largest absolute value of the partial derivative as the dominant data source. According to the type of the dominant data source, the corresponding warning information dissemination channel is selected, such as through roadside variable message signs, vehicle-mounted terminal broadcasts, or traffic management platforms, outputting a warning message containing the dominant data source identifier and hazard margin. Through precise source tracing and targeted distribution, the warning information becomes more targeted, assisting drivers, vehicles, and roadside facilities in responding quickly, thereby improving the overall effectiveness of road traffic safety control.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By constructing a heterogeneous data cube of vehicle trajectory data, meteorological environment data, and road surface condition data, and calculating the cross-correlation entropy matrix between different data sources, each data source is discretized into a probability quality function. The logarithmic ratio of the product of the joint probability density and the marginal probability density is used to measure the nonlinear dependence between pairs of data sources. This allows for the acquisition of statistical correlation information of arbitrary forms within the framework of entropy, overcoming the limitation of traditional linear correlation coefficients that can only capture linear relationships. Furthermore, sparse-constrained nonnegative matrix decomposition is used to project the high-dimensional cross-correlation entropy matrix onto a low-dimensional feature space. During the decomposition process, a sparse regularization term based on the sum of the absolute values ​​of the elements is applied to the coefficient matrix, forcing the basis matrix to lock only a few dominant correlation patterns and removing interfering micro-coupling components. The resulting fused feature basis matrix possesses both nonnegativity and interpretability, and can compactly represent the essential correlation structure between multi-source data. When real-time multi-source data is organized in the same format and the non-negative coefficient projection on the basis matrix is ​​solved using the active set gradient projection algorithm, the resulting fused feature vector has a much lower dimension than the original data and contains the most critical coupling information between different data sources. This significantly enhances the sensitivity and noise resistance of subsequent safety state deviation detection, avoiding high-dimensional redundancy and decreased discrimination ability caused by direct feature splicing. The dynamic time warping algorithm is used to calculate the morphological distance between the current fused feature vector and the historical safety feature template. By constructing a distance matrix between the two sequences and searching for the optimal warping path that satisfies boundary, continuity, and monotonicity constraints, adaptive matching of local scaling and velocity changes on the time axis is achieved. Compared to point-by-point Euclidean distance, this morphological distance can more realistically reflect the degree of deviation between the current traffic situation and historical safety patterns in terms of evolutionary morphology, effectively improving the ability to identify progressive danger trends. Based on this, gamma probability density functions for morphological distances under normal driving conditions and near-accident conditions are pre-established using historical data. During real-time monitoring, the morphological distances calculated within multiple consecutive time windows are used as observation samples. The likelihood ratio under the two hypotheses is calculated to obtain the Bayesian factor. When the Bayesian factor exceeds the decision threshold, the warning threshold is automatically lowered, making the system more sensitive when the probability of an accident increases. When the Bayesian factor is lower than the reciprocal of the decision threshold, the warning threshold is raised to suppress false alarms. This real-time data-driven Bayesian factor adaptive threshold adjustment mechanism allows the warning boundary to automatically adjust with dynamic changes in weather conditions, road conditions, and traffic flow patterns. This overcomes the inherent defect of fixed thresholds being unable to adapt to dynamic environments, ensuring timely warnings for high-risk events while significantly reducing the probability of false alarms under normal conditions. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a flowchart of a road traffic safety monitoring and early warning method based on multi-source data fusion; Figure 2 This is a flowchart of the heterogeneous data cube construction and cross-correlation entropy matrix calculation process; Figure 3 This is a flowchart of the process for solving the fusion feature basis matrix and projecting the fusion feature vector based on sparse constrained nonnegative matrix decomposition; Figure 4 This is a flowchart of morphological distance calculation based on dynamic time warping; Figure 5 It is the curve showing the change of cross-correlation entropy over time between vehicle trajectory data, meteorological environment data, and road condition data sources; Figure 6 It is the curve of the fused feature vector changing over time; Figure 7 This is a schematic diagram of the dynamic time warping distance matrix and the optimal warping path that integrates feature vectors and historical security feature templates; Figure 8 This is a comparison chart of the morphological distance distribution and gamma distribution fitting under normal driving and near-accident conditions; Figure 9 This is a schematic diagram of the generation of road hazard warning signals and a curve showing the change in hazard margin. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0016] See Figure 1This invention provides a road traffic safety monitoring and early warning method based on multi-source data fusion, comprising the following steps: collecting vehicle trajectory data, meteorological environment data, and road surface condition data in a road traffic scenario, and constructing a heterogeneous data cube indexed by timestamps; based on the heterogeneous data cube, calculating the cross-correlation entropy matrix between various data sources, and using sparse-constrained non-negative matrix decomposition to reduce the cross-correlation entropy matrix to a low-dimensional feature space to obtain a fusion feature basis matrix; projecting the real-time collected multi-source data using the fusion feature basis matrix to obtain the fusion feature vector at the current moment, and calculating the morphological distance between the fusion feature vector and the historical safety feature template based on a dynamic time warping algorithm; updating the current early warning threshold through a Bayesian factor according to the morphological distance and a preset risk probability distribution, and generating a road hazard early warning signal when the morphological distance exceeds the updated early warning threshold.

[0017] Example 1:

[0018] In specific implementation, please refer to Figure 2 Collecting vehicle trajectory data, meteorological data, and road surface condition data in road traffic scenarios can be achieved through the following methods. For vehicle trajectory data, a roadside millimeter-wave radar transmits linear frequency modulated continuous waves at a fixed frequency and receives the echo signals reflected from vehicles. After mixing and fast Fourier transform, the distance, radial velocity, and azimuth angle of the vehicle are obtained. The vehicle-mounted global navigation satellite system terminal calculates and outputs the latitude, longitude, elevation, and velocity vector of the vehicle position through satellite signal processing. Simultaneously, the heading angle is provided by the terminal's built-in dual-antenna attitude measurement module or inertial measurement unit. During data fusion, the target trajectory detected by the roadside millimeter-wave radar and the trajectory reported by the vehicle-mounted global navigation satellite system terminal are time-stamped and spatially coordinate unified. The nearest neighbor association or global nearest neighbor association method is used to complete the trajectory matching of the same vehicle, generating a vehicle trajectory data sequence containing vehicle position, instantaneous velocity, acceleration, and heading angle.

[0019] For meteorological environmental data, automatic weather stations are deployed at predetermined intervals along both sides of the road. These automatic weather stations include platinum resistance temperature sensors, tipping bucket rain gauges, ultrasonic anemometers, and forward-scattering visibility meters. The resistance value output by the temperature sensor is converted from analog to digital and then into a temperature value according to a calibration table. The tipping bucket rain gauge obtains the rainfall intensity by recording the number of times the bucket tipps, measured in millimeters per minute. The ultrasonic anemometer calculates wind speed by measuring the time difference of ultrasonic wave propagation in the downwind and upwind directions. The forward-scattering visibility meter analyzes the intensity of received infrared light after it has been scattered by atmospheric particles to obtain visibility. Data from all sensors is aggregated by a data acquisition unit and timestamped to form meteorological environmental data including temperature, rainfall intensity, wind speed, and visibility.

[0020] For road surface condition data, piezoelectric sensors buried beneath the road surface convert the mechanical stress generated by vehicle rolling or road vibration into electrical signals. After conditioning by a charge amplifier, the peak value and frequency components of the signal are analyzed, and combined with the road surface temperature calibration curve, the road surface temperature and friction coefficient are calculated. A laser road surface condition sensor emits multi-wavelength laser light downwards, receives reflected light from the road surface, and analyzes the differences in reflectivity and scattering characteristics at different wavelengths to determine the thickness of water accumulation and the probability of icing. The piezoelectric sensor and the laser road surface condition sensor are installed together on the same cross-section, and the output data is packaged after time synchronization into road surface condition data including road surface temperature, water accumulation thickness, icing probability, and friction coefficient.

[0021] A heterogeneous data cube indexed by timestamps is constructed. In the specific implementation, vehicle trajectory data, meteorological environment data, and road condition data from different data sources are time-aligned according to a unified time base, with timestamp precision uniformly set to milliseconds. For time deviations caused by asynchronous sampling, linear interpolation or nearest neighbor interpolation is used to resample the data sequences from each data source onto a common time grid. After time alignment, a third-order tensor is established. The first order of the tensor represents the time dimension, with each time slice corresponding to a sampling moment; the second order represents the data source dimension, corresponding to the three modalities of vehicle trajectory data, meteorological environment data, and road condition data; the third order represents the feature dimension within each data source. The feature dimension of vehicle trajectory data includes vehicle position, instantaneous speed, acceleration, and heading angle; the feature dimension of meteorological environment data includes temperature, rainfall intensity, wind speed, and visibility; and the feature dimension of road condition data includes road surface temperature, water accumulation thickness, icing probability, and friction coefficient. Therefore, each element in the heterogeneous data cube can be uniquely identified by a ternary index (time_stamp, source_index, feature_index), where time_stamp is the timestamp index, source_index is the data source type index, and feature_index is the feature component index of the corresponding data source. In terms of storage implementation, the heterogeneous data cube can be represented as a multidimensional array or tensor object, facilitating subsequent batch extraction and computation.

[0022] Based on a heterogeneous data cube, the cross-correlation entropy matrix between various data sources is calculated. In specific implementation, for a heterogeneous data cube within a given timestamp range, vehicle trajectory data, meteorological environment data, and road condition data at the same timestamp are discretized to obtain their respective probability mass functions. The discretization method is as follows: for each feature dimension of each data source, after determining the dynamic range, it is divided into equally spaced intervals, with the number of intervals set to 100. The frequency of samples falling into each interval is normalized to obtain the probability mass function vector of that data source at that time. For any two data sources, the probability mass function vector of the vehicle trajectory data is denoted as... The probability quality function vector of meteorological and environmental data is The joint probability density of the two vectors can be estimated using the joint occurrence frequency of the corresponding components of the two vectors at the same time. Specifically, the joint probability mass function matrix is ​​obtained by calculating the frequency of samples that simultaneously fall into both interval pairs. The cross-association entropy is defined as the expected value of the logarithm ratio between the product of the joint probability density and the products of their respective marginal probability densities. Its calculation formula can be expressed as:

[0023] in, This represents the cross-correlation entropy between vehicle trajectory data sources and meteorological environment data sources. This is the discrete interval index for the probability mass function of vehicle trajectory data, and the traversal range is from 1 to the total number of discrete intervals of vehicle trajectory data. This is the discrete interval index for the probability quality function of meteorological and environmental data, and the traversal range is from 1 to the total number of discrete intervals of meteorological and environmental data. This indicates that the vehicle's trajectory data is at the [number]th [position]. The discrete intervals and the meteorological environment data are in the first... The joint probability mass function value of each discrete interval is obtained by performing frequency statistics and normalization on all timestamp samples in the heterogeneous data cube. The mass function representing the marginal probability of vehicle trajectory data at the th... The values ​​of each discrete interval are obtained by summing the joint probability mass function over the dimensions of the meteorological and environmental data. The marginal probability mass function of meteorological and environmental data is represented in the first... The values ​​for each discrete interval are obtained by summing the joint probability mass function over the dimensions of the vehicle trajectory data. Iterating through all data source combinations, the cross-correlation entropy is calculated pairwise for each of the vehicle trajectory data, meteorological environment data, and road condition data. The results are arranged in a square matrix: the rows and columns of the matrix correspond to the data source type, and the diagonal elements are defined as the self-correlation entropy of each data source, obtained by calculating the entropy of the data source's own probability mass function in a similar manner. Therefore, the cross-correlation entropy matrix is ​​a third-order square matrix, where each element quantifies the degree of information redundancy and statistical dependence between two data sources.

[0024] In practice, each element in the cross-association entropy matrix is ​​obtained by calculating the average cross-association entropy between blocks after dividing the continuous time series of the two data sources into sliding window blocks. The length of the sliding window can be set to a 10-minute sampling period, and the sliding step size can be set to 5 minutes. For each pair of data sources, all continuous time slices with a time length equal to the sliding window length are extracted from the heterogeneous data cube. The data within this continuous time slice is divided into several sub-blocks, each containing a fixed number of time sampling points, and the sub-block length can be set to 1 minute. The cross-association entropy between the first and second sub-blocks is calculated, the cross-association entropy between the second and third sub-blocks is calculated, and so on, until the last sub-block. Then, the average of the cross-association entropies of all adjacent sub-blocks is taken as the average cross-association entropy between the pair of data sources under this sliding window. The sliding window is moved along the complete time axis, and the above calculation is repeated to obtain the cross-association entropy sequence that changes over time. Finally, the average cross-association entropy matrix calculated by the sliding window at the most recent moment is used as the cross-association entropy matrix participating in subsequent fusion at the current moment.

[0025] See Figure 5 The figure illustrates the changing trends of three sets of cross-correlation entropies calculated based on heterogeneous data cubes in Example 1 over time. The horizontal axis represents time in minutes, ranging from 0 to 500 minutes, and the vertical axis represents the cross-correlation entropy value, ranging from 0.2 to 1.0. The figure contains three curves, representing the cross-correlation entropy between vehicle trajectory data and meteorological environment data. (Solid line) Cross-correlation entropy between meteorological environment data and road surface condition data (Dotted lines) and the cross-correlation entropy between vehicle trajectory data and road surface condition data. (short dash).

[0026] Judging from the trend of the curves, the three cross-correlation entropies remained relatively stable during the period from 0 to approximately 250 minutes, and their values ​​were concentrated in different intervals: The value fluctuated between 0.5 and 0.65, significantly higher than the other two groups; It fluctuates between 0.4 and 0.55, placing it in the middle. The most stable value was the one that remained in the range of 0.3 to 0.4. Furthermore, after approximately 250 minutes, all three sets of cross-correlation entropies showed a significant upward trend, especially the one that... It quickly climbed to a peak close to 0.8, and then stabilized in the 0.7 to 0.8 range. Meanwhile, and They also showed significant growth, reaching the ranges of 0.6 to 0.65 and 0.5 to 0.6 respectively, indicating that the information redundancy and statistical dependence among the three data sources of vehicle trajectory, meteorological environment and road condition have all increased.

[0027] This change reflects the dynamic correlation characteristics of the multi-source data in Example 1 over time. Especially after 250 minutes, the significant increase in cross-correlation entropy indicates enhanced interaction between different data sources, providing effective time-series signal support for subsequent dimensionality reduction fusion and safety monitoring based on the cross-correlation entropy matrix. The overall curve change conforms to the block-based averaging calculation of cross-correlation entropy using a sliding window in Example 1, demonstrating the dynamic evolution of cross-correlation entropy over time and facilitating the capture of potential safety risks in the road traffic environment.

[0028] Example 2:

[0029] In specific implementation, please refer to Figure 3 The cross-association entropy matrix is ​​reduced to a low-dimensional feature space by using sparse-constrained non-negative matrix decomposition to obtain a fused feature basis matrix. The object of processing is the cross-association entropy matrix calculated in Example 1. The cross-association entropy matrix is ​​a third-order square matrix, and the number of rows and columns are the number of data source types. The rows correspond to the first data source, the columns correspond to the second data source, and the matrix elements are the cross-association entropy between the two corresponding data sources.

[0030] When performing nonnegative matrix decomposition with sparse constraints, the cross-association entropy matrix is ​​first denoted as... ,matrix The dimension is ,in This represents the number of data source types, with a value of 3, corresponding to vehicle trajectory data source, meteorological environment data source, and road condition data source, respectively. It also sets the preset feature dimensions. , Value less than The number of components is determined based on the cumulative contribution rate of principal component analysis on the correlation between data sources in the heterogeneous data cube. When the cumulative contribution rate reaches 90% or higher, the corresponding number of components is [number missing]. Values ​​in a scenario with three data sources It can be set to 2. Initialize the non-negative basis matrix. sum coefficient matrix , where the basis matrix The dimension is coefficient matrix The dimension is basis matrix All elements and coefficient matrix All elements are initialized with uniformly distributed random numbers, ranging from 0 to 1, and each element is guaranteed to be non-negative.

[0031] Construct an objective function, the configuration of which includes a reconstruction error term and a sparse regularization term. The reconstruction error term measures the cross-correlation entropy matrix. With basis matrix sum coefficient matrix product The approximation between them is calculated using the square of the Frobenius norm; the sparse regularization term is applied to the coefficient matrix. Apply sparsity constraints and use a coefficient matrix The sum of the absolute values ​​of all elements is calculated, since the coefficient matrix... The elements of the expression are non-negative, and the sum of their absolute values ​​is equal to the sum of their elements. The objective function is expressed as:

[0032] in, Represents the basis matrix sum coefficient matrix The objective function value; This is the cross-association entropy matrix with dimension 1. ; Let be the non-negative basis matrix to be solved, with dimension . ; Let be the non-negative coefficient matrix to be solved, with dimension . ; The Frobenius norm of a matrix is ​​represented by the square root of the sum of the squares of all its elements. is the regularization strength parameter, with values ​​greater than 0. Five-fold cross-validation is used to select the value from the candidate set {0.01, 0.05, 0.1, 0.5, 1.0} that minimizes the reconstruction error on the validation set. When the reconstruction error on the validation set corresponding to 0.1 in the candidate set is minimized, Set to 0.1; The preset feature dimensions; Number of data source types; Represents the coefficient matrix The Middle Line 1 Column elements, The traversal range is 1 to , The traversal range is 1 to .

[0033] In practice, the basis matrix is ​​updated alternately. sum coefficient matrix The update process continues until the objective function converges, employing a multiplicative update rule. In each iteration, the coefficient matrix is ​​preserved. The basis matrix remains unchanged and is updated according to the following formula. :

[0034] Preserving the basis matrix The coefficient matrix remains unchanged and is updated according to the following formula. :

[0035] Updated basis matrix sum coefficient matrix Negative elements that may appear in the expression are truncated to zero to ensure non-negativity. Iteration continues, and the objective function is considered convergent when the relative change in the objective function value between two adjacent iterations is less than a preset convergence threshold. The preset convergence threshold is set to... The converged basis matrix As a fusion feature basis matrix, the fusion feature basis matrix The dimension is It contains a low-dimensional representation of the cross-linking structure of different data sources.

[0036] In practice, a fusion feature basis matrix is ​​used to project the real-time acquired multi-source data to obtain the fusion feature vector at the current moment. First, the real-time acquired vehicle trajectory data, meteorological environment data, and road condition data are organized into a real-time data vector according to the same format as the heterogeneous data cube. Real-time data vector It is a column vector formed by concatenating the feature components from three data sources in a fixed order, with dimensions equal to the total number of feature components within a time slice of the corresponding heterogeneous data cube. This is based on the fused feature basis matrix. Given a projection basis, find the non-negative coefficient vector that minimizes the projection error. The projection error is represented as a real-time data vector. With fusion feature basis matrix Multiply by nonnegative coefficient vector The squared Euclidean distance between vectors. (Non-negative coefficient vectors) The dimension is Consistent with the preset feature dimensions, all components must satisfy the non-negativity constraint.

[0037] Solving for the nonnegative coefficient vector The active set gradient projection algorithm is employed. During initialization, the non-negative coefficient vector is... The initial values ​​are set to random non-negative vectors, and the initial active set is empty. In each iteration, the current non-negative coefficient vector is calculated. The gradient vector, which is the gradient of the non-negative coefficient vector of the projection error. The partial derivatives of the components form the structure; for the current nonnegative coefficient vector Components with a median of zero and a corresponding gradient component greater than zero are considered to violate the nonnegativity constraint, and their corresponding indices are moved into the active set. For indices in the active set, their corresponding components are fixed to zero. For indices outside the active set, projective gradient descent is performed along the negative gradient direction; if a component becomes negative after the update, it is set to zero, and its corresponding index is moved into the active set. This iteration continues until a nonnegative coefficient vector is found. The update magnitude is less than a preset threshold, and the preset threshold is set to... At this point, the converged non-negative coefficient vector This is the output of the fused feature vector at the current moment. The dimension of the fused feature vector is... Each component represents the activation intensity of the fused feature on the corresponding low-dimensional feature dimension.

[0038] See Figure 6 The figure shows the trend of the eigenvector components of the two fused feature dimensions as a function of time after projecting the fused feature basis matrix obtained by the sparse constrained nonnegative matrix factorization method described in Example 2 onto real-time acquired multi-source data. The horizontal axis represents time in minutes, ranging from 0 to 500 minutes; the vertical axis represents the component values ​​of the fused feature vector, ranging from approximately 0.1 to 0.5.

[0039] In the curves, the solid line represents the component changes of fusion feature dimension 1, and the dashed line represents the component changes of fusion feature dimension 2. Overall, both fusion feature dimensions showed a relatively stable trend with slight fluctuations in the first 300 minutes. The component values ​​of fusion feature dimension 1 remained between approximately 0.22 and 0.28, and the component values ​​of fusion feature dimension 2 remained between approximately 0.13 and 0.18. This indicates that the cross-correlation structure between the data sources was relatively stable during this stage, and the activation intensity of the low-dimensional features extracted from the fusion feature basis matrix was at a low level.

[0040] After about 300 minutes, the components of the two fusion feature dimensions began to show a significant upward trend, especially the component of fusion feature dimension 1, which gradually increased from 0.28 to above 0.4, and the component of fusion feature dimension 2 also increased from about 0.18 to close to 0.27. This indicates that the statistical correlation and information redundancy between the multi-source data have changed significantly, and the activation intensity of the fusion features has increased, which may reflect abnormal changes in road traffic conditions or environmental conditions.

[0041] This figure illustrates the dynamic changes over time of the fused feature vector generated by projecting real-time multi-source data after dimensionality reduction of the cross-correlation entropy matrix through non-negative matrix factorization to obtain the fused feature basis matrix in Example 2. The increasing trend of the fused feature dimension provides crucial input data support for safety monitoring and hazard warning based on morphological distance and warning threshold updates in subsequent examples.

[0042] Example 3:

[0043] In specific implementation, please refer to Figure 4 To calculate the morphological distance between the fused feature vector and the historical safety feature template based on the dynamic time warping algorithm, it is necessary to first establish the historical safety feature template. Multiple fused feature vectors are extracted from multi-source data collected during historical accident-free periods. The selection method for historical accident-free periods is as follows: by querying the road traffic accident record database, periods in which no traffic accidents occurred within 90 consecutive natural days and the average daily traffic flow is between 50% and 150% of the annual average daily traffic flow of that road segment are selected. For each sampling time within these periods, the fused feature vector is generated according to the method obtained in Example 2, and the dimension is extracted as follows. The fused feature vector, It is consistent with the preset feature dimensions in Example 2.

[0044] The extracted fused feature vectors are arranged in chronological order to form a historical security feature sequence. Each element in the historical security feature sequence is a... A dimensional vector, where the vector indices correspond to the order of sampling times. The historical security feature sequence is clustered to obtain a historical security feature template. The clustering algorithm uses the K-medoids algorithm based on dynamic time warping distance, with K set to 1, meaning it seeks a template vector sequence representing the central trend of the entire sequence. Specifically, in the historical security feature sequence, each position is considered a candidate center. The sum of the dynamic time warping distances from all subsequences within the time window surrounding the candidate center to the candidate center is calculated. The candidate subsequence with the smallest distance sum is selected as the medoid and used as the historical security feature template. The historical security feature template is a vector of length 1. The vector sequence, denoted as the reference sequence. ,in Indicates the first in the reference sequence indivual 3D fusion feature vector, This is the time index in the reference sequence, with values ​​ranging from 1 to... .

[0045] After obtaining the fused feature vector at the current time step, the continuous vectors before the current time step are... The fused feature vectors from each sampling time point are arranged in chronological order to form a query sequence. ,in Indicates the first in the query sequence indivual 3D fusion feature vector, To query the time index in the sequence, the value range is from 1 to... , Set to with The same numerical value is used to ensure that the two sequences have the same length, which is determined by the historical security feature template.

[0046] Constructing a query sequence and reference sequence Distance matrix of all corresponding point pairs between Distance matrix The dimension is Distance matrix The Middle Line number Column elements Defined as the first in the query sequence vectors With reference sequence vectors The Euclidean distance between them is calculated as follows: ,in The L2 norm of a vector. The value range is 1 to , The value range is 1 to .

[0047] In the distance matrix Search from the starting point To the finish line The optimal regularization path is defined as a sequence of index pairs. Each path point is , The path point index, with values ​​ranging from 1 to... , Let be the total number of steps in the optimal regularization path. The optimal regularization path satisfies three constraints. The boundary constraint is: the starting point of the optimal regularization path is fixed at . The endpoint is fixed at The continuity constraint is: the index change between adjacent path points does not exceed 1, i.e. and The monotonicity constraint is: pathpoint indices cannot be backtracked, i.e. and .

[0048] The search for the optimal regular path is achieved using a dynamic programming algorithm, defining a cumulative distance matrix. Cumulative distance matrix Dimensions and distance matrix Same, cumulative distance matrix The Middle Line number Column elements The recursive formula is:

[0049] in, Indicates starting from the origin Reaching the matrix position The minimum cumulative distance; Position in the distance matrix Local distance at; To perform the minimum value operation, the parentheses contain three candidate values: This represents the cumulative distance value recursively from the top left front position. This represents the cumulative distance value recursively from the upper front position. This represents the cumulative distance recursively calculated from the left-hand forward position. The initial conditions for dynamic programming are set as follows: .when and hour, ;when and hour, Complete the cumulative distance matrix. After filling, the last element of the cumulative distance matrix This is the minimum cumulative distance of the optimal regular path, and this minimum value is output as the morphological distance.

[0050] See Figure 7 The figure illustrates the distance matrix and optimal warping path calculated using the dynamic time warping algorithm in Example 3, between the fused feature vector and the historical security feature template. The horizontal axis represents the index of the query sequence, and the vertical axis represents the index of the reference sequence, both ranging from 0 to 50, reflecting the setting that both sequences are 50 units long. The grayscale values ​​in the figure represent the element values ​​of the distance matrix D; lighter grayscale values ​​indicate a smaller Euclidean distance between the two fused feature vectors at the corresponding positions, while darker grayscale values ​​indicate a larger distance. Overall, the distance matrix shows a trend of lighter grayscale values ​​along the diagonal, indicating a high similarity between the query sequence and the reference sequence when they are time-synchronized, which aligns with the core assumption of sequence morphological similarity in dynamic time warping.

[0051] The optimal regularized path, indicated by dashed lines in the figure, starts from the lower left corner (0,0) and extends along the lighter gray area in the distance matrix to the upper right corner (50,50). The path maintains boundary constraints, continuity constraints, and monotonicity constraints, demonstrating the matching path with the minimum local and cumulative distances selected by the dynamic programming algorithm. This path reflects the optimal time axis alignment between the query sequence and the historical security feature template, effectively measuring the morphological distance between the two sequences.

[0052] The distance matrix and optimal regularization path shown in the figure intuitively reflect the process of calculating morphological distance in Example 3, verifying the rationality and effectiveness of dynamic time regularization of the fused feature vector sequence, and thus providing basic data support for subsequent safety monitoring and early warning based on morphological distance.

[0053] Example 4:

[0054] In practice, based on the morphological distance and a preset risk probability distribution, the current warning threshold is updated using a Bayesian factor, requiring the pre-establishment of two probability density functions: a probability density function for morphological distance under normal driving conditions and a probability density function for morphological distance under accident-prone conditions.

[0055] For the probability density function of morphological distance in the near-accident state, multi-source data within a preset time period before the accident is extracted from historical data. The preset time period is set to 120 seconds, and the extraction method is as follows: for each recorded accident in the road traffic accident record database, a continuous time window of 120 seconds is traced back from the time of occurrence, and multi-source data of all sampling times within this time window are extracted. The sampling frequency is consistent with the time slice frequency of the heterogeneous data cube in Example 1. For each extracted sampling time, the morphological distance of the fused feature vector of the sampling time relative to the historical safety feature template associated with the corresponding accident is calculated according to the method described in Example 3. The historical safety feature template associated with the corresponding accident refers to the historical safety feature template extracted from the most recent continuous accident-free period before the accident on the same road segment and in the same direction of travel.

[0056] The sample set of all calculated morphological distances under the near-accident state is denoted as... ,Will The morphological distance values ​​are fitted to a first gamma distribution. The probability density function of the first gamma distribution is as follows:

[0057] in, This represents the morphological distance variable, and its value is a non-negative real number. The morphological distance value indicates the state of imminent accident. The probability density at that time; This is the set of parameters for the gamma distribution in the near-accident state, including shape parameters. and inverse scaling parameters ; Represents the gamma function; is the shape parameter of the first gamma distribution, and its value is a positive real number; Let be the inverse scaling parameter of the first gamma distribution, taking positive real values. It is determined using the maximum likelihood estimation method. and The numerical values ​​are obtained by solving the log-likelihood equations of the gamma distribution parameters using the Newton-Raphson iterative method, and the converged parameter values ​​are obtained.

[0058] For the probability density function of morphological distance under normal driving conditions, continuous multi-source data of accident-free and unobstructed road periods are extracted from historical data. The selection criteria for accident-free and unobstructed road periods are: no traffic accidents have occurred within 30 consecutive natural days, the average traffic flow speed during the period is not less than 70% of the design speed of the road segment, and the average headway is not less than 2 seconds. For each sampling time within the extracted period, the morphological distance of each sampling time relative to the corresponding historical safety feature template of the same road segment is calculated according to the method described in Example 3. The selection method of the historical safety feature template corresponding to the same road segment is consistent with the selection method of the historical safety feature template used in the morphological distance calculation under the near-accident state, and both are taken from the accident-free period of the same road segment and the same direction of travel.

[0059] Let the sample set of all calculated morphological distances under normal driving conditions be denoted as... ,Will The morphological distance values ​​are fitted to a second gamma distribution. The probability density function of the second gamma distribution is as follows: ,in This is the set of parameters for the gamma distribution under normal driving conditions. The shape parameter of the second gamma distribution. Let be the inverse scaling parameter of the second gamma distribution. The parameter estimation method is the same as that of the first gamma distribution, which uses maximum likelihood estimation combined with Newton-Raphson iteration to solve the problem.

[0060] After establishing two probability density functions, the morphological distances calculated from multiple consecutive time windows prior to the current moment are used as observation samples. The number of time windows is set to 10, and the span of each time window is the same as the query sequence in Example 3. The time spans are consistent, and adjacent windows slide with a step size of one sampling time. The 10 morphological distance values ​​corresponding to the 10 time windows are used to construct the observation sample vector. ,in Indicates the first The morphological distance observations were calculated using a time window. The value range is from 1 to 10.

[0061] Calculate the observed samples The Bayes factor is obtained by calculating the likelihood ratio under the two probability density functions. The likelihood ratio is calculated as follows: under the probability density function of the accident proximity state, the joint likelihood value of the observed samples is... Under the probability density function of normal driving conditions, the joint likelihood value of the observed samples is Bayesian factor The value is the ratio of the two joint likelihood values, i.e. When the Bayes factor is greater than 1, it indicates that the observed sample supports the state of imminent accident; when the Bayes factor is less than 1, it indicates that the observed sample supports the state of normal driving.

[0062] The Bayesian factor is compared with a preset decision threshold, which is set to 3. The threshold is chosen based on the fact that, within the sequential probability ratio test framework, a decision threshold of 3 corresponds to a Type I error probability (false alarm probability) of approximately 0.03, achieving a balance between false alarm rate and detection delay. When the Bayesian factor is greater than the decision threshold of 3, it indicates that the current road traffic condition is approaching an accident, and the current warning threshold is lowered by multiplying it by a decay factor of 0.8. When the Bayesian factor is less than the reciprocal of the decision threshold (1 / 3), it indicates that the current road traffic condition is approaching normal, and the current warning threshold is increased by multiplying it by an amplification factor of 1.2. When the Bayesian factor falls between 1 / 3 and 3, the current warning threshold remains unchanged. The initial value of the current warning threshold is set as the expected value of the morphological distance under normal driving conditions plus twice the standard deviation, where the expected value and standard deviation are determined by the parameters of the second gamma distribution. and Calculated.

[0063] See Figure 8 In the figure, the horizontal axis represents morphological distance, ranging from 0 to 20, and the vertical axis represents probability density, reflecting the likelihood of different morphological distance values ​​occurring. The figure uses two types of filled bar histograms and corresponding probability density curves to represent the morphological distance distribution under normal driving conditions and near-accident conditions, respectively.

[0064] The histogram filled with a left slash and its corresponding solid probability density curve represent the distribution characteristics of morphological distance under normal driving conditions. The curve reaches a peak at a morphological distance of approximately 2, and then gradually decreases, indicating that the morphological distance under normal driving conditions is mostly concentrated between 2 and 6, with a low mean and a relatively concentrated distribution. The probability density rapidly approaches zero, indicating that the morphological distance in this range is more representative and matches the safety feature template of normal driving conditions to a high degree.

[0065] The histogram filled with right slashes and its corresponding dashed probability density curve represent the morphological distance distribution in the near-accident state. The peak of this curve is located at a morphological distance of approximately 4, which is shifted to the right compared to the distribution in the normal driving state, and the tail extends further. This indicates that the morphological distance is generally larger and more discrete in the near-accident state, reflecting a lower matching degree between the fused feature vector and the historical safety feature template, and significant morphological differences.

[0066] Both states exhibit typical gamma distribution characteristics in their probability density functions of morphological distance. However, the distribution location and shape differ significantly in the accident-prone state, demonstrating the distinguishing ability of morphological distance as a risk indicator. The two types of probability density functions and their corresponding histograms clearly marked in the figure provide intuitive verification for the gamma distribution models of morphological distance in normal driving and accident-prone states based on historical data fitting described in Example 4. This supports subsequent risk discrimination and dynamic adjustment of warning thresholds based on Bayesian factors.

[0067] In summary, this figure visually illustrates the difference in probability distribution of morphological distance between normal driving conditions and near-accident conditions. This provides a data foundation and theoretical support for constructing a risk probability model using the morphological distance probability density function in Example 4 and for dynamically updating the early warning threshold by combining real-time observation of morphological distance.

[0068] Example 5:

[0069] In practical implementation, a road hazard warning signal is generated when the morphological distance exceeds the updated warning threshold. The difference between the updated warning threshold in Example 4 and the morphological distance calculated in Example 3 is used as the hazard margin. A negative hazard margin indicates that the morphological distance has exceeded the updated warning threshold; the smaller the value, the more severe the exceedance. The hazard margins of several adjacent timestamps before and after the current time are combined into a hazard margin sequence. The number of adjacent timestamps is set to 3 before and 3 after, plus the current time, for a total of 7 timestamps. The hazard margin sequence is represented as follows: ,in Indicates the current risk margin. This indicates the danger margin of the third timestamp before the current moment. This indicates the danger margin at the third timestamp after the current moment; the meanings of the other subscripts follow the same pattern.

[0070] For hazard margin series Perform a sign function transformation to obtain a sign sequence. The sign function transformation rule is: when the element in the hazard margin sequence When greater than zero, the corresponding element of the symbol sequence The value is +1; when the element in the risk margin sequence When equal to zero, the corresponding element of the symbol sequence The value is 0; when the element in the risk margin sequence When less than zero, the corresponding element in the symbol sequence The value is -1, where This is the timestamp offset index, with a value range of {-3,-2,-1,0,+1,+2,+3}.

[0071] Calculate symbol sequence The cumulative sum of the products of adjacent elements is used as the trend strength index. The formula for calculating the trend strength index is:

[0072] in, This represents a trend strength indicator, with values ​​that are integers. Indicates the first symbol in the sequence The sign value at each position; Indicates the first symbol in the sequence The sign value at each position; Indicates multiplication operation; To accumulate the index variable, it is accumulated from -2 to +3, for a total of 6 adjacent product terms.

[0073] When the trend strength indicator Three consecutive positive values ​​and current risk margin When the value is negative, a road hazard warning signal is generated. Three consecutive positive values ​​mean that at three consecutive sampling times (including the current time), the trend intensity index is calculated according to the above steps at each sampling time, and all three calculated trend intensity index values ​​are positive. Current hazard margin. A negative value means that the current hazard margin is less than zero. When both of the above conditions are met simultaneously, a road hazard warning signal is generated, and the risk level of the road hazard warning signal is marked as high risk. If only the trend intensity index is positive three times consecutively while the current hazard margin is not negative, or if only the current hazard margin is negative while the trend intensity index is not positive three times consecutively, then a high-risk road hazard warning signal is not generated; instead, the risk level is marked as low risk, and monitoring continues.

[0074] After generating a road hazard warning signal, the dominant data source causing the morphological distance exceedance is located based on the last peak position of the hazard margin sequence. The last peak position of the hazard margin sequence refers to the position within the hazard margin sequence... In this process, the search proceeds backward from the end of the sequence to find the index of the first local maximum. A local maximum is defined as an element whose value is simultaneously greater than both its preceding and following neighbors. The process of locating the dominant data source is achieved by calculating the absolute values ​​of the partial derivatives of the vehicle trajectory data, meteorological environment data, and road surface condition data with respect to morphological distance. The partial derivatives are calculated using numerical differentiation: [The text then describes the calculation of the partial derivatives using the real-time data vector at the current moment.] In this process, a small positive perturbation is applied to the components corresponding to the vehicle trajectory data, with the perturbation amplitude being one-thousandth of the corresponding component value. The projected fused feature vector is then recalculated, and the morphological distance is recalculated. The ratio of the change in morphological distance to the amount of perturbation is used as the absolute value of the partial derivative of the vehicle trajectory data with respect to the morphological distance. The absolute value of the partial derivative of the meteorological environment data with respect to the morphological distance is calculated in the same way for the components corresponding to the meteorological environment data. The absolute value of the partial derivative of the road surface condition data with respect to the morphological distance is calculated in the same way for the components corresponding to the road surface condition data. The three absolute values ​​of the partial derivatives are compared, and the data source with the largest absolute value of the partial derivative is determined as the dominant data source.

[0075] After determining the primary data source, the corresponding warning information dissemination channel is selected according to the type of the primary data source. When the primary data source is vehicle trajectory data, the vehicle-mounted terminal broadcast is selected as the warning information dissemination channel; when the primary data source is meteorological environmental data, the roadside variable message sign is selected as the warning information dissemination channel; when the primary data source is road condition data, the traffic management platform push is selected as the warning information dissemination channel. Warning messages are output through the selected warning information dissemination channel. The warning message includes the primary data source identifier and a hazard margin. The primary data source identifier uses an enumerated type value, corresponding to vehicle trajectory data, meteorological environmental data, and road condition data respectively. The hazard margin is the current time value. Numerical value.

[0076] See Figure 9 In the graph, the horizontal axis represents time in seconds, ranging from 0 to 500 seconds; the vertical axis represents the hazard margin, ranging from approximately -3.5 to 2.5. The graph depicts the hazard margin changing over time with a solid black line, and the dashed black line represents the warning threshold, which is always at zero, corresponding to the zero limit of the hazard margin. The star symbol in the legend indicates the specific time a warning signal is generated, but it is not shown in this graph, indicating that a warning signal has not yet been generated.

[0077] The curve shows that the hazard margin was positive in the initial stage (approximately 0 to 150 seconds) and generally fluctuated upwards, reaching a high of nearly 2.5, indicating a relatively high safety margin in road traffic during this phase, not exceeding the warning threshold. Subsequently, starting around 150 seconds, the hazard margin gradually decreased, crossing the zero line of the warning threshold at approximately 225 seconds and entering the negative range, indicating that the morphological distance exceeded the updated warning threshold, posing a potential hazard. After entering the negative range, the hazard margin continued to decrease, reaching a low of approximately -3.3, indicating a gradual increase in the degree of danger. The overall decline was relatively smooth, with occasional small fluctuations, reflecting a gradual deterioration in real-time traffic safety.

[0078] This graph illustrates the dynamic changes in hazard margin over time in Example 5. The moment the hazard margin crosses the zero line of the warning threshold represents a potential hazard point. If the trend intensity index is positive three times consecutively and the current hazard margin is negative, a high-risk road hazard warning signal will be triggered. Although the specific warning signal generation time is not indicated in the graph, the hazard margin remains negative and continues to decline in the latter half of the phase, meeting the prerequisites for warning signal generation. This curve effectively demonstrates the sensitive response capability of the hazard margin index, calculated based on the difference between morphological distance and the warning threshold, to potential risks in monitoring road safety status.

[0079] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A road traffic safety monitoring and early warning method based on multi-source data fusion, characterized in that, include: Collect vehicle trajectory data, meteorological environment data, and road surface condition data in road traffic scenarios, and construct a heterogeneous data cube indexed by timestamps; Based on the heterogeneous data cube, the cross-association entropy matrix between each data source is calculated, and the cross-association entropy matrix is ​​reduced to a low-dimensional feature space by sparse-constrained non-negative matrix decomposition to obtain the fused feature basis matrix. The fusion feature basis matrix is ​​used to project the real-time acquired multi-source data to obtain the fusion feature vector at the current moment, and the morphological distance between the fusion feature vector and the historical security feature template is calculated based on the dynamic time warping algorithm. Based on the morphological distance and the preset risk probability distribution, the current warning threshold is updated using a Bayesian factor, and a road hazard warning signal is generated when the morphological distance exceeds the updated warning threshold.

2. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, The steps for collecting vehicle trajectory data, meteorological environment data, and road surface condition data in road traffic scenarios specifically include: Vehicle trajectory data is acquired jointly by roadside millimeter-wave radar and vehicle-mounted global navigation satellite system terminal. The vehicle trajectory data includes vehicle position, instantaneous speed, acceleration, and heading angle. Meteorological environmental data is obtained by setting up weather stations on both sides of the road. The meteorological environmental data includes temperature, rainfall intensity, wind speed and visibility. Road surface condition data is acquired by piezoelectric sensors and laser road surface condition sensors buried under the road surface. The road surface condition data includes road surface temperature, water accumulation thickness, icing probability, and friction coefficient.

3. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, Based on the heterogeneous data cube, the steps for calculating the cross-correlation entropy matrix between the various data sources specifically include: The vehicle trajectory data, meteorological environment data, and road condition data at the same time stamp in the heterogeneous data cube are discretized into equally spaced probability mass functions. For any two data sources, calculate the log ratio between their joint probability density and the product of their respective marginal probability densities, and use the expected value of this log ratio as the cross-association entropy. By iterating through all data source combinations, the calculated cross-association entropy is arranged into a square matrix to obtain the cross-association entropy matrix.

4. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, The heterogeneous data cube uses a time-aligned tensor structure to store vehicle trajectory data, meteorological environment data, and road surface condition data, where each time slice contains feature vectors from all data sources at the corresponding time.

5. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, Each element in the cross-association entropy matrix is ​​obtained by calculating the average cross-association entropy between blocks after dividing the continuous time series of two data sources into sliding window blocks.

6. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 3, characterized in that, The step of reducing the cross-association entropy matrix to a low-dimensional feature space using sparse-constrained nonnegative matrix decomposition to obtain the fused feature basis matrix specifically includes: Initialize a non-negative basis matrix and coefficient matrix, wherein the number of columns in the basis matrix is ​​equal to the preset feature dimension; Construct an objective function that includes a reconstruction error term and a sparse regularization term. The reconstruction error term is the square of the Frobenius norm of the difference between the cross-association entropy matrix and the basis matrix multiplied by the coefficient matrix. The sparse regularization term is the sum of the absolute values ​​of all elements of the coefficient matrix. The basis matrix and the coefficient matrix are alternately updated until the objective function converges, and the converged basis matrix is ​​used as the fused feature basis matrix.

7. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 6, characterized in that, The steps of projecting the real-time acquired multi-source data onto the fused feature basis matrix to obtain the fused feature vector at the current moment specifically include: The vehicle trajectory data, meteorological environment data, and road condition data collected in real time at the current moment are organized into a real-time data vector in the same format as the heterogeneous data cube. Using the fused feature basis matrix as the projection basis, the non-negative coefficient vector that minimizes the projection error is solved. This solution process uses the active set gradient projection algorithm to iteratively update the non-negative coefficient vector. The converged non-negative coefficient vector is output as the fused feature vector at the current time step.

8. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 7, characterized in that, The active set gradient projection algorithm dynamically updates the active set in each iteration, moves coefficient components that violate the non-negativity constraint into the active set, and performs projective gradient descent.

9. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, The step of calculating the morphological distance between the fused feature vector and the historical security feature template based on the dynamic time warping algorithm specifically includes: Multiple fusion feature vectors are extracted from multi-source data collected during historical accident-free periods. These multiple fusion feature vectors are arranged in chronological order to form a historical safety feature sequence. The sequence is then clustered to obtain a historical safety feature template. The current fused feature vector and the historical security feature template are used as the query sequence and reference sequence, respectively, to construct the distance matrix of all corresponding point pairs between the two sequences; The optimal regular path from the starting point to the ending point is searched on the distance matrix. The optimal regular path satisfies the boundary constraints, continuity and monotonicity, and the minimum cumulative distance of the optimal regular path is taken as the morphological distance.

10. The road traffic safety monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, The steps for updating the current early warning threshold using a Bayesian factor based on the morphological distance and the preset risk probability distribution specifically include: The probability density functions of morphological distance under normal driving conditions and morphological distance under near-accident conditions are pre-established; The morphological distances calculated from multiple consecutive time windows prior to the current moment are used as observation samples. The likelihood ratios of the observation samples under two probability density functions are calculated to obtain the Bayes factor. The Bayesian factor is compared with a preset decision threshold. When the Bayesian factor is greater than the decision threshold, the current warning threshold is lowered. When the Bayesian factor is less than the reciprocal of the decision threshold, the current warning threshold is raised.

Images