Queue length estimation method, apparatus, device, storage medium, and program product

Abstract

The application discloses a queuing length estimation method, device, equipment, storage medium and program product, and relates to the technical field of intelligent traffic control. The method comprises the following steps: obtaining trajectory data according to the distance measurement between the forward prediction state of first radar detection data and the backward smoothing state of satellite positioning data of a first floating car in a road side radar detection range; obtaining parking related data from the trajectory data, and classifying the parking behavior of the first floating car; if the first parking behavior type is a queuing parking type or a pick-up parking type, performing kinematic backstepping on a second floating car outside the road side radar detection range according to a preset deceleration parameter to obtain the parking position and the uncertainty of the second floating car; taking the parking position of the target first floating car as an initial queuing length, taking the parking position of the second floating car as an observation point, and performing Bayesian iterative updating on the queuing length according to the prior distribution parameter corresponding to a current period to obtain the queuing length.

Description

Technical Field

[0001] This application relates to the field of intelligent traffic control technology, specifically to a queue length estimation method, device, equipment, storage medium, and program product. Background Technology

[0002] Accurately assessing queue length at intersections is a key indicator for optimizing signal timing and coordinated control. Existing methods rely on fixed detectors (such as coils, geomagnetic sensors, and microwave radar), which suffer from low upload frequencies, high failure rates, and high maintenance costs. In recent years, methods based on connected vehicle trajectory data have faced problems such as low penetration rates, long sampling periods, and unreliable single-vehicle judgments.

[0003] Especially in areas more than 150 meters upstream of the stop line, the limited detection range of roadside radar makes it impossible to obtain detailed queue status in that area; while the sparse satellite positioning data from floating car rovers makes it difficult to accurately pinpoint parking positions. Therefore, how to organically integrate high-frequency but short-range radar detection data with low-frequency but wide-area satellite positioning data from floating cars to accurately assess the queue length at intersections is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0004] This application provides a queue length estimation method, apparatus, device, storage medium, and program product, which solves the problem of how to accurately assess the queue length at intersections in the prior art.

[0005] In a first aspect, embodiments of this application provide a queue length estimation method, including: Within the current signal period, for each of the multiple first floating vehicles within the detection range of the roadside radar, trajectory data of the first floating vehicle is obtained based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle; wherein, the first radar detection data is the radar detection data corresponding to the first floating vehicle in the radar detection dataset of the current signal period; the radar detection dataset and the satellite positioning data of the first floating vehicle are aligned to the same coordinate system and synchronized in time; Parking-related data is obtained from the trajectory data, and the parking behavior of the first floating car is classified according to the parking-related data to obtain the first parking behavior type; If the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, then according to the preset deceleration parameters corresponding to the first parking behavior type, kinematic back-inference is performed on multiple second floating cars outside the detection range of the roadside radar to obtain the parking position and uncertainty of each second floating car; wherein, the satellite positioning data of the second floating car indicates that the second floating car stops between two adjacent samplings, and the parking behavior type of the second floating car is queuing parking. The initial queue length is determined by taking the parking position of the target first floating car as the initial queue length and the parking position of each second floating car as the observation point. The queue length is then updated using Bayesian iteration based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met, thus obtaining the queue length. The target first floating car is the first floating car that is furthest from the stop line among multiple first floating cars. The prior distribution parameters are obtained by updating historical queue data according to the corresponding time period.

[0006] Optionally, the queue length estimation method, wherein obtaining the trajectory data of the first floating vehicle based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle, includes: Based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle, and a preset fusion parameter, a first fusion weight corresponding to the first radar detection data and a second fusion weight corresponding to the satellite positioning data of the first floating vehicle are obtained; wherein, the sum of the first fusion weight and the second fusion weight is equal to 1; the preset fusion parameter includes at least one of the following: a reference threshold; a scale parameter of the hyperbolic tangent function; a basic radar weight; and an upper limit for the weight adjustment range; Based on the first fusion weight and the second fusion weight, the first radar detection data and the satellite positioning data of the first floating vehicle are weighted and fused to obtain the trajectory data of the first floating vehicle.

[0007] Optionally, the queue length estimation method, wherein classifying the parking behavior of the first floating car based on the parking-related data to obtain a first parking behavior type includes: Using the causal effect estimation method, historical radar detection datasets, and historical satellite positioning datasets, we obtain the initial weights corresponding to strong causal features, spatial constraint features, and motion pattern features, respectively. A first score is obtained based on the parking duration in the parking-related data, which is related to the strong causal feature. A second score is obtained based on the distance from the stop line in the parking-related data, which is related to the spatial constraint feature. A third score is obtained based on the trajectory curvature in the parking-related data, which is related to the motion pattern feature. The first floating car's parking behavior is obtained by weighted fusion based on the first score, the second score, the third score, and the initial weights corresponding to the strong causal feature, the spatial constraint feature, and the motion pattern feature. The parking behavior of the first floating car is classified based on the comprehensive score and the posterior probability calculated through the conditional probability table to obtain the first parking behavior type.

[0008] Optionally, in the queue length estimation method, if the first parking behavior type is a queuing parking type or a passenger pick-up / drop-off parking type, kinematic back-calculation is performed on multiple second floating cars outside the detection range of the roadside radar based on the preset deceleration parameters corresponding to the first parking behavior type to obtain the parking position and uncertainty of each second floating car, including: If the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, then according to the preset deceleration parameter corresponding to the first parking behavior type, the probability distribution of the parking position of each of the multiple second floating cars outside the detection range of the roadside radar is obtained. Based on the probability distribution, multiple samplings are performed to obtain the parking position and uncertainty of the second floating car.

[0009] Optionally, the queue length estimation method further includes: Perform domain-based collaborative verification on the first parking behavior type to obtain the verification results; The first floating car whose verification result is "verification failed" is excluded from the plurality of first floating cars to obtain a plurality of excluded first floating cars; The target first floating car is obtained based on the parking position of each of the excluded first floating cars; For multiple first floating cars whose verification results are passed and which are located in the same parking feature area, the spatiotemporal similarity between the multiple first floating cars is obtained. The target first floating car is obtained based on the parking positions of multiple first floating cars whose spatiotemporal similarity is greater than a preset similarity.

[0010] Optionally, the queue length estimation method, wherein performing domain-based collaborative verification on the first parking behavior type to obtain verification results includes: Within the collaboration window, a spatial search window is established with the floating car as the center to obtain the domain floating cars located within the spatial search window; The cooperative confidence between the domain floating car and the first floating car is obtained based on the longitudinal and lateral distances between the domain floating car and the first floating car, the parking indication function of the domain floating car, and the number of domain floating cars in the spatial search window. If the collaboration confidence level is greater than the collaboration upper limit threshold, then the second parking behavior type of the first floating car is determined to be the queuing parking type. If the collaboration confidence level is less than or equal to the collaboration upper limit threshold and greater than or equal to the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be the passenger pick-up and drop-off parking type. If the collaboration confidence level is less than the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be either faulty parking or illegal parking. The second parking behavior type is compared with the first parking behavior type to obtain the verification result.

[0011] Optionally, in the queue length estimation method, for multiple first floating cars whose verification results are successful and located in the same parking feature area, obtaining the spatiotemporal similarity between the multiple first floating cars includes: For multiple floating cars that have passed the verification and are located in the same parking feature area, a kernel function for joint Gaussian process regression is obtained based on the preset time feature length and the spatial feature length corresponding to the parking feature area. The joint Gaussian process regression is constructed based on the kernel function and the preset mean function; Based on the joint Gaussian process regression, the spatiotemporal similarity between multiple first floating cars is obtained.

[0012] Optionally, the queue length estimation method further includes, after performing Bayesian iterative updates on the queue length based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met: The particle filtering method is used to estimate the growth rate of the parking position of the second floating car; The parking position of the second floating car is updated according to the growth rate to obtain the updated parking position of the second floating car; Obtaining the queue length includes: The queue length is obtained based on the updated parking position of the second floating car.

[0013] Optionally, in the queue length estimation method, after obtaining the queue length, the method further includes: When the current signal period ends, the queue length is checked for consistency based on the radar detection dataset to obtain the check result; If the verification result is unsuccessful, the first fusion weight and the second fusion weight are updated to obtain the updated first fusion weight and the updated second fusion weight. In the next signal cycle, the trajectory data of the first floating car is obtained based on the updated first fusion weight and the updated second fusion weight.

[0014] Secondly, embodiments of this application also provide a queue length estimation device, comprising: The fusion module is used to obtain trajectory data of each of the multiple first floating vehicles within the detection range of the roadside radar during the current signal cycle, based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle; wherein, the first radar detection data is the radar detection data corresponding to the first floating vehicle in the radar detection dataset of the current signal cycle; the radar detection dataset and the satellite positioning data of the first floating vehicle are aligned to the same coordinate system and synchronized in time; The classification module is used to obtain parking-related data from the trajectory data and classify the parking behavior of the first floating car according to the parking-related data to obtain a first parking behavior type; The estimation module is used to perform kinematic back-calculation on multiple second floating cars outside the detection range of the roadside radar based on the preset deceleration parameters corresponding to the first parking behavior type if the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, so as to obtain the parking position and uncertainty of each second floating car; wherein, the satellite positioning data of the second floating car indicates that the second floating car stops between two adjacent samplings, and the parking behavior type of the second floating car is queuing parking; An iterative module is used to take the parking position of the target first floating car as the initial queue length, the parking position of each second floating car as an observation point, and perform Bayesian iterative updates based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met to obtain the queue length; wherein, the target first floating car is the first floating car that is farthest from the stop line among multiple first floating cars; the prior distribution parameters are obtained by updating according to the corresponding time period based on historical queue data.

[0015] Thirdly, embodiments of this application also provide a queue length estimation device, including: a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the processor executes the program or instructions to implement the queue length estimation method as described in the first aspect.

[0016] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the queue length estimation method as described in the first aspect.

[0017] Fifthly, embodiments of this application also provide a computer program product, including computer instructions that, when executed by a processor, implement the queue length estimation method as described in the first aspect.

[0018] Compared with the prior art, this application provides a queue length estimation method, apparatus, device, storage medium, and program product. Within the current signal cycle, for each of a plurality of first floating vehicles within the detection range of roadside radar, trajectory data of the first floating vehicle is obtained based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle. The first radar detection data is the radar detection data corresponding to the first floating vehicle in the radar detection dataset of the current signal cycle; the radar detection dataset and the satellite positioning data of the first floating vehicle are aligned to the same coordinate system and synchronized in time. Parking-related data is obtained from the trajectory data, and the parking behavior of the first floating vehicle is classified according to the parking-related data to obtain a first parking behavior type. If the first parking behavior type is a queue parking type or a passenger pick-up / drop-off parking type... Based on the preset deceleration parameters corresponding to the first parking behavior type, kinematic back-calculation is performed on multiple second floating cars outside the detection range of the roadside radar to obtain the parking position and uncertainty of each second floating car. The satellite positioning data of the second floating car indicates that the second floating car has stopped between two adjacent samplings, and the parking behavior type of the second floating car is a queuing parking type. The parking position of the target first floating car is used as the initial queue length, and the parking position of each second floating car is used as an observation point. The queue length is then updated using Bayesian iteration based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met, thus obtaining the queue length. The target first floating car is the first floating car furthest from the stop line among multiple first floating cars. The prior distribution parameters are obtained by updating historical queuing data according to the corresponding time period. In this way, the queue length at the intersection can be accurately assessed. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the queue length estimation method described in the embodiments of this application; Figure 2 This is a diagram illustrating the completion effect in an embodiment of this application; Figure 3 This is one of the schematic diagrams of the smooth fusion process in the embodiments of this application; Figure 4 This is a second schematic diagram of the smooth fusion process in the embodiments of this application; Figure 5 This is a schematic diagram comparing the ACE estimate in the embodiments of this application with existing empirical values; Figure 6 This is a schematic diagram of the decision-making process for counterfactual verification in the embodiments of this application; Figure 7 This is a schematic diagram of the parking location in an embodiment of this application; Figure 8 This is one of the schematic diagrams of the Bayesian iterative update process in the embodiments of this application; Figure 9 This is the second schematic diagram of the Bayesian iterative update process in the embodiments of this application; Figure 10 This is one of the schematic diagrams of the particle filtering method in the embodiments of this application; Figure 11 This is a second schematic diagram of the particle filtering method in the embodiments of this application; Figure 12 This is a schematic diagram illustrating the beneficial effects in the embodiments of this application; Figure 13 This is a schematic diagram of the queue length estimation device described in the embodiments of this application; Figure 14 This is a hardware block diagram of the queue length estimation device described in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not used to describe a specified order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0022] Please see Figure 1 This application provides a queue length estimation method that can be applied to intersection scenarios or urban expressway ramp merging areas.

[0023] Furthermore, the method includes: Step 101: Within the current signal period, for each of the multiple first floating vehicles within the detection range of the roadside radar, the trajectory data of the first floating vehicle is obtained based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle; wherein, the first radar detection data is the radar detection data corresponding to the first floating vehicle in the radar detection dataset of the current signal period; the radar detection dataset and the satellite positioning data of the first floating vehicle are aligned to the same coordinate system and synchronized in time; Optionally, the distance metric is Mahalanobis distance.

[0024] In one implementation, optionally, the trajectory data of the first floating vehicle is obtained based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle, including: Based on the distance metric between the forward prediction state of the radar detection data of the first floating vehicle and the backward smoothing state of the satellite positioning data of the first floating vehicle, and a preset fusion parameter, a first fusion weight corresponding to the first radar detection data and a second fusion weight corresponding to the satellite positioning data of the first floating vehicle are obtained; wherein, the sum of the first fusion weight and the second fusion weight is equal to 1; the preset fusion parameter includes at least one of the following: a reference threshold; a scale parameter of the hyperbolic tangent function; a basic radar weight; and an upper limit for the weight adjustment amplitude; Based on the first fusion weight and the second fusion weight, the first radar detection data and the satellite positioning data of the first floating vehicle are weighted and fused to obtain the trajectory data of the first floating vehicle.

[0025] The embodiments of this application include a boundary buffer zone at the intersection ( , Within a certain range (m), the first and second fusion weights are dynamically adjusted based on Mahalanobis distance. Compared to existing technologies that use fixed-weight fusion, this embodiment can adaptively adjust the confidence level based on the deviation between the two data sources (including the first radar detection data and the satellite positioning data of the first floating vehicle) at that moment. The Mahalanobis distance between the forward prediction state of the first radar detection data and the backward smoothed state of the satellite positioning data of the first floating vehicle is calculated. As shown in the formula below: ; in, , Let represent the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle, respectively. Here, the forward prediction state of the first radar detection data is the state vector obtained by forward prediction of the first radar detection data, and the backward smoothing state of the satellite positioning data of the first floating vehicle is the state vector obtained by backward smoothing the satellite positioning data of the first floating vehicle. The above state vectors are represented as follows: , For location, For speed; The joint state covariance matrix is ​​derived from the posterior covariance of the first radar detection data. Smooth covariance with satellite positioning data of the first floating vehicle Adding them together, ; Furthermore, the first radar detection data and the satellite positioning data of the first floating vehicle are fused according to the following formula to obtain the trajectory data of the first floating vehicle: ; in, This represents the trajectory data of the first floating car; Indicates the first fusion weight; This represents the first radar detection data; Indicates the second fusion weight; This represents the satellite positioning data of the first floating vehicle.

[0026] The first fusion weight and the second fusion weight are dynamically adjusted according to the Mahalanobis distance, as shown in the following formula: ; Basic radar weights (reflecting high-frequency, high-precision priors of the radar); This is the upper limit of the weight adjustment range; This is a reference threshold, specifically the Mahalanobis distance reference threshold; is the scaling parameter of the tanh function.

[0027] It should be noted that when the Mahalanobis distance... When the data is smaller (consistent). The baseline regression value is 0.8; when the Mahalanobis distance is... When the data conflict increases, the first fusion weight can be increased to a maximum of 0.95, further reducing the impact of satellite positioning data.

[0028] Therefore, the embodiments of this application use Mahalanobis distance as a data conflict metric to dynamically adjust the fusion weights. Compared with the existing fixed weights, this approach can adaptively address boundary data conflicts and achieve spatiotemporal alignment and boundary adaptive fusion.

[0029] Prior to step 101, the method further includes: Data preprocessing: The radar detection dataset (sampling frequency of 10Hz, including vehicle identification, millisecond-level timestamps, and location) is processed. ,speed The satellite positioning dataset of the floating car (sampling period of 15 to 30 seconds, including vehicle identification, second-level timestamp, latitude and longitude, instantaneous speed) is used. The timestamps of the satellite positioning data of the first floating car are uniformly converted to Unix (a general operating system) timestamps (seconds). The latitude and longitude of the floating car are transformed to the same local plane coordinate system as the radar through Gauss-Kruger projection. Abnormal trajectory points (such as trajectory points with speed > 30m / s or adjacent frame position jump > 5m) are removed from the radar detection dataset according to physical constraints.

[0030] Boundary bidirectional Kalman filtering: at the boundary of roadside radar coverage (distance from the stop line) A bidirectional Kalman filter is established at point m. Specifically, on the radar side, for the radar detection dataset, forward prediction is performed using the 10Hz trajectory as the observation, and the state vector... The standard CV (constant velocity) model is adopted, and the diagonal elements of the process noise covariance are taken as... On the floating car side, a Rauch-Tung-Striebel smoother is used to perform backward state estimation based on the satellite positioning data of the first floating car.

[0031] Step 102: Obtain parking-related data from the trajectory data of the first floating car, and classify the parking behavior of the first floating car according to the parking-related data to obtain the first parking behavior type; In one implementation, optionally, the parking behavior of the first floating car is classified according to the parking-related data to obtain a first parking behavior type, including: We use causal effect estimation, historical radar detection datasets, and historical satellite positioning datasets to obtain the initial weights corresponding to strong causal features, spatial constraint features, and motion pattern features, respectively. A first score is obtained based on the parking duration in the parking-related data, which is related to the strong causal feature. A second score is obtained based on the distance from the stop line in the parking-related data, which is related to the spatial constraint feature. A third score is obtained based on the trajectory curvature in the parking-related data, which is related to the motion pattern feature. The first floating car's parking behavior is obtained by weighted fusion based on the first score, the second score, the third score, and the initial weights corresponding to the strong causal feature, the spatial constraint feature, and the motion pattern feature. The parking behavior of the first floating car is classified based on the comprehensive score and the posterior probability calculated through the conditional probability table to obtain the first parking behavior type.

[0032] Prior to step 102, the method further includes: Constructing a spatiotemporal semantic graph Node set The features include five dimensions, as shown in Table 1 below: Table 1: .

[0033] The edge weights between nodes are obtained through conditional mutual information from historical radar detection datasets or historical satellite positioning datasets. The calculation is used to quantify the causal contribution of each feature to the queuing decision.

[0034] Furthermore, in step 102, the causal effect estimation method is introduced into the weight determination process for queuing behavior classification: Based on the do-calculus idea, for each candidate feature (Including strong causal features, spatial constraint features, and motion pattern features), estimate its impact on queuing decisions. The average causal effect (ACE) is shown in the following formula: ; The inverse probability weight (IPW) estimation is achieved by combining historical radar detection datasets and historical satellite positioning datasets, as shown in the following formula: ; in The number of historical samples (including historical radar detection datasets and / or historical satellite positioning datasets). This is the propensity score. The covariate vector (containing other features); , For the first The feature values ​​and labels of each sample.

[0035] The initial weights of the above features are given by normalized causal effects, as shown in the following formula: ; The initial weights of strong causal features are estimated by analyzing historical radar detection datasets and / or historical satellite positioning datasets. The initial weight of the spatial constraint feature is approximately 0.40. The initial weight of motion pattern features is approximately 0.30. The value is approximately 0.30. The advantage of the causal effect estimation method used in this application is that the initial weights are derived from the true causal effect estimation rather than empirical assignments, thus possessing interpretability and falsificationability.

[0036] Specifically, strong causal features (initial weights) ): Used to indicate the remaining time of a red light With parking duration Based on the temporal causality, a first score is obtained that correlates the parking behavior of the first floating car with the strong causal feature, according to the parking duration in the parking-related data. As shown in the formula below: ; Physical meaning: If the first floating car stops during a red light and the remaining red light time is long enough, then the stopping behavior of the first floating car is most likely caused by signal control.

[0037] Spatial constraint features (weights) ): Distance from the stop line m, based on the distance from the stop line in the parking-related data, the second score related to the parking behavior of the first floating car and the spatial constraint characteristics is obtained as shown in the following formula: ; Motion pattern features (weights) ): Trajectory curvature The third score, derived from the trajectory curvature in the parking-related data, relating the parking behavior of the first floating car to the motion pattern features, is shown in the following formula: ; in, This is an indicator function.

[0038] A comprehensive score for the parking behavior of the first floating car is obtained by weighting and fusing the first score, the second score, and the third score, along with the initial weights corresponding to the strong causal features, the spatial constraint features, and the motion pattern features. As shown in the formula below: .

[0039] Calculate the posterior probability using a conditional probability table (CPT). .

[0040] The classification rules are as follows: If the comprehensive score is greater than or equal to the first comprehensive score, and the posterior probability is greater than the first probability, then the first parking behavior type is determined to be the queuing parking type, i.e. and 0.85, classified as a queuing parking type; If the overall score is greater than or equal to the second overall score and less than the first overall score, then the first parking behavior type is determined to be a passenger pick-up / drop-off parking type, i.e., 0.3. 0.7 indicates the parking type is for picking up or dropping off passengers; If the overall score is less than the second overall score, then the first parking behavior type is determined to be either a faulty parking type or an illegal parking type, i.e. If the value is 0.3, it is determined to be an isolated event (including faulty parking or illegal parking).

[0041] Furthermore, this application embodiment employs a counterfactual verification mechanism. For the first floating car determined to be of the queuing parking type, counterfactual query verification is performed, as shown in the following formula: ; like A score of 0.3 indicates that the red light timing condition does not contribute sufficiently to the causal relationship for this determination, thus downgrading it to a suspected queue and allowing for further neighborhood collaborative verification. This counterfactual verification mechanism effectively avoids misclassification caused by accidental spatial-curvature feature matching.

[0042] It should be noted that if the failure rate of counterfactual verification is greater than 20% for three consecutive cycles, then a check on each candidate feature will be triggered. (Including strong causal features, spatial constraint features, and motion pattern features), estimate its impact on queuing decisions. ACE online revaluation.

[0043] Therefore, the embodiments of this application can realize parking behavior classification based on causal inference. By estimating the average causal effect of three types of features on queuing determination through do-calculus and inverse probability weighting, the initial weights that can be explained and falsified are normalized, and a counterfactual verification mechanism is introduced.

[0044] Step 103: If the first parking behavior type is queuing parking or passenger pick-up / drop-off parking, then according to the preset deceleration parameters corresponding to the first parking behavior type, kinematic back-calculation is performed on multiple second floating cars outside the detection range of the roadside radar to obtain the parking position and uncertainty of each second floating car; wherein, the satellite positioning data of the second floating car indicates that the second floating car stops between two adjacent samplings, and the parking behavior type of the second floating car is queuing parking; It should be noted that the radar detection dataset and the satellite positioning data of the second floating car have been aligned to the same coordinate system and synchronized in time.

[0045] The parking behavior of the second floating car is classified to obtain the parking behavior type of the second floating car. The classification method used can be the same as the classification method used to classify the parking behavior of the first floating car.

[0046] In one implementation, optionally, if the first parking behavior type is queuing parking or passenger pick-up / drop-off parking, then based on the preset deceleration parameters corresponding to the first parking behavior type, kinematic back-calculation is performed on multiple second floating cars outside the detection range of the roadside radar to obtain the parking position and uncertainty of each second floating car, including: Step a: If the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, then according to the preset deceleration parameter corresponding to the first parking behavior type, obtain the probability distribution of the parking position of each of the multiple second floating cars outside the detection range of the roadside radar. Step b: Based on the probability distribution, perform multiple samplings to obtain the parking position and uncertainty of the second floating car.

[0047] In this embodiment, step a uses the parking behavior classification result from step 102 as the decision basis for the deceleration parameter, avoiding the systematic bias caused by the use of a uniform deceleration parameter in existing methods. Specifically, different first parking behavior types correspond to different preset deceleration parameters, including the mean deceleration and the standard deviation of deceleration, as shown in Table 2 below: Table 2: .

[0048] In step a above, a further probabilistic processing of deceleration is introduced, whereby the deceleration no longer takes a fixed value, but is modeled as a Gaussian distribution. And through Monte Carlo sampling (number of samplings) The probability distribution of the parking position of the second floating car is obtained instead of a single fixed value, which is more in line with the real uncertainty of driving behavior.

[0049] Specifically, firstly, when the speed of the first floating car is detected to be from the first speed... (satisfies a speed greater than the first speed threshold) The first speed threshold (5m / s) will reach the second speed (satisfying a speed less than the second speed threshold) The second speed threshold At speeds of 0.5 m / s, the kinematic constraint equations are established as shown in the following formula: ; in, This represents the displacement during the deceleration process. This refers to the deceleration duration; The average of the above-mentioned decelerations; Then, determine the parking time using the following formula. : ; The probability distribution of the parking position of the second floating car (derived from Monte Carlo) is calculated as follows: ; in, Indicates the first Deceleration of sub-Monte Carlo sampling; , These represent the intervals between two adjacent sampling times of satellite positioning data. 15 to 30 seconds); This indicates the vehicle's position at the previous sampling time.

[0050] Therefore, in this embodiment, based on the parking behavior type, differentiated preset deceleration parameters are selected, and the probability distribution of parking position is obtained through deceleration Gaussian distribution modeling and Monte Carlo sampling, thereby realizing probabilistic kinematic back-inference driven by classification.

[0051] Next, in step b above, the probability distribution of the parking position of the second floating car obtained in step a is sampled multiple times to obtain the parking position of the second floating car. Specifically, take The mean of the samples. Here, the variance corresponding to this parking location is taken as the uncertainty of that parking location.

[0052] It should be noted that, due to the sampling period of the satellite positioning data of the floating car... The time interval between sampling and sampling is relatively long, ranging from 15 to 30 seconds. When the floating car stops between two samplings, the uncertainty of the stopping position can reach up to about 30m. Therefore, the embodiments of this application can achieve high-precision back-calculation by selecting differentiated deceleration parameters driven by classification and kinematic constraint equations, thus achieving the effect of kinematic back-calculation driven by classification.

[0053] In one embodiment, optionally, the method further includes: Step c: Perform domain-based collaborative verification on the first parking behavior type to obtain the verification result; Step d: The first floating car whose verification result is "verification failed" is excluded from the multiple first floating cars to obtain multiple first floating cars after exclusion; Step e: Obtain the target first floating car based on the parking position of each of the excluded first floating cars; Step f: For multiple first floating cars whose verification results are verified and located in the same parking feature area, obtain the spatiotemporal similarity between the multiple first floating cars; Step g: Obtain the target first floating car based on the parking positions of multiple first floating cars whose spatiotemporal similarity is greater than a preset similarity.

[0054] In one implementation, optionally, in step c above, domain-based collaborative verification is performed on the first parking behavior type to obtain verification results, including: Within the collaboration window, a spatial search window is established with the first floating car as the center to obtain the domain floating cars located within the spatial search window; The cooperative confidence between the domain floating car and the first floating car is obtained based on the longitudinal and lateral distances between the domain floating car and the first floating car, the parking indication function of the domain floating car, and the number of domain floating cars in the spatial search window. If the collaboration confidence level is greater than the collaboration upper limit threshold, then the second parking behavior type of the first floating car is determined to be the queuing parking type. If the collaboration confidence level is less than or equal to the collaboration upper limit threshold and greater than or equal to the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be the passenger pick-up and drop-off parking type. If the collaboration confidence level is less than the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be either faulty parking or illegal parking. The second parking behavior type is compared with the first parking behavior type to obtain the verification result.

[0055] In this embodiment, a cooperative window is first determined. This embodiment adopts a cooperative window mechanism with signal period alignment: existing cooperative methods use a fixed time window, which is prone to mixing in parking data under different signal phases across periods. However, this embodiment uses a cooperative window... With signal period (e.g., 90 seconds) Alignment, and only include the trajectory data of the first floating car within the same red light phase, as shown in the following formula: ; Within the collaboration window, a spatial search window is established centered on the first floating vehicle to obtain the domain floating vehicles located within the spatial search window. The spatial search window satisfies the following conditions: Longitudinal range: ±50m from the parking position of the first floating car; Lateral range: ±3.5m (one lane width) from the lane where the first floating car is located.

[0056] Then, this application embodiment extends the existing one-dimensional vertical Gaussian kernel to a two-dimensional Gaussian kernel (including a vertical kernel and a horizontal kernel), which can more accurately characterize the lane correlation of queuing. Specifically, the cooperative confidence between the floating car in the domain and the first floating car is calculated according to the following formula. : ; in, The weights are two-dimensional Gaussian kernels: ; For the first The longitudinal distance between the floating car in the vehicle area and the first floating car; For the first The lateral distance between the floating car in the vehicle area and the first floating car; 25m represents the longitudinal characteristic length (reflecting the longitudinal extension scale of the queue). 3.5m indicates the lateral characteristic length (corresponding to the lane width, emphasizing coordination within the same lane); For the first Parking instruction function for floating cars in the vehicle domain; The number of floating vehicles in the domain within the space search window; 0.01 is a small value used to prevent division by zero.

[0057] This application introduces an adaptive threshold mechanism. Existing collaborative confidence uses a fixed collaborative threshold, for example... A value of 0.65 is prone to failure in low-penetration scenarios. However, in this embodiment, the number of floating cars within the spatial search window is used as the basis for the solution. Dynamically adjust collaboration thresholds, including collaboration upper limit thresholds. and the lower limit threshold of collaboration : ; in, , For threshold adjustment range, The number of floating cars in the reference domain. When the number of floating cars in the domain within the space search window is small (e.g., ... When the threshold for collaboration is reduced to 0.55, the threshold for collaboration can be reduced to 0.28 to avoid missing judgments under low sample size.

[0058] Here, the embodiments of this application adopt a three-level determination: If the collaboration confidence level is greater than the collaboration upper limit threshold, then the second parking behavior type of the first floating car is determined to be the queuing parking type, i.e. It was determined to be a queuing parking type (multi-vehicle collaborative verification, high confidence). If the collaboration confidence level is less than or equal to the upper collaboration threshold and greater than or equal to the lower collaboration threshold, then the second parking behavior type of the first floating car is determined to be the passenger pick-up / drop-off parking type, i.e. : This is classified as a passenger pick-up and drop-off parking type (medium confidence level); If the collaboration confidence level is less than the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be either a faulty parking type or an illegal parking type, i.e. It is determined to be an isolated event (low confidence level, which may be a faulty parking type or an illegal parking type).

[0059] It should be noted that the second parking behavior type is compared with the first parking behavior type to obtain the verification result. If the two are consistent, the verification result is that the verification is successful.

[0060] Therefore, in this embodiment, the collaborative window is aligned with the signal period, and a two-dimensional Gaussian kernel is incorporated into the lane dimension. The threshold is adaptive with the number of samples, which can realize the collaborative confidence verification of signal period alignment and adaptive threshold.

[0061] In one implementation, optionally, in step f above, for multiple first floating cars whose verification results are successful and located in the same parking feature area, the spatiotemporal similarity between the multiple first floating cars is obtained, including: For multiple first floating cars whose verification results are passed and located in the same parking feature area, the kernel function of joint Gaussian process regression is obtained according to the preset time feature length and the spatial feature length corresponding to the parking feature area; The joint Gaussian process regression is constructed based on the kernel function and the preset mean function; Based on the joint Gaussian process regression, the spatiotemporal similarity between multiple first floating cars is obtained.

[0062] In this embodiment of the application, a radar observation set is first constructed based on the radar detection dataset. This includes high-frequency trajectory points with a sampling frequency of 10Hz within the detection range of roadside radar; and a satellite positioning constraint set constructed based on the satellite positioning dataset. This includes the parking location and the uncertainty of the parking location obtained from the probability distribution mentioned above.

[0063] Then, based on the radar observation set and satellite positioning constraint set constructed above, a joint Gaussian process regression is constructed. , where the mean function The constant is set to zero. This application's embodiment improves the kernel function with a dual-bandwidth design. This kernel function is essentially a dual-bandwidth spatiotemporal kernel function, based on the spatial characteristic length. Distinguish between parking characteristic zones, including dense queue characteristic zones and upstream free-flow characteristic zones. Specifically, the kernel function is shown in the following formula: ; in, The signal variance is estimated by data-driven estimation (typical value approximately...). ); Represents the Euclidean distance in space; Indicates a time interval; To observe the noise variance; The spatial feature lengths corresponding to the different parking feature areas It needs to satisfy the following formula: ; Among them, the preset time feature length Seconds. This dual-bandwidth improvement allows the kernel function to use different spatial feature lengths, enabling it to distinguish between densely queued feature regions (short feature length, requiring fine interpolation) and upstream free-flow feature regions (long feature length, requiring smooth interpolation).

[0064] Step 104: Perform Bayesian iterative updates on the queue length based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met, and obtain the queue length.

[0065] It is understood that step 104 is to obtain the queue length based on the parking position of the first floating car and the parking position of the second floating car at the end of the iteration.

[0066] In this embodiment, the initial queue length is first estimated: the initial queue length is estimated based on the parking position of the last floating car (i.e., the target floating car, which is the floating car farthest from the stop line among multiple floating cars) obtained from the first radar detection data. For example, if the floating vehicle detected by the roadside radar is located 120m from the stop line, then 120m.

[0067] Then, the parking positions of the floating cars that have passed the domain collaboration verification obtained in step 103 are used. Perform a Bayesian iterative update, as shown in the following formula: ; According to the Uncertainty of the parking position of a floating car Determine the variance of the likelihood distribution of the likelihood function. , m is the standard deviation of the satellite positioning data error, as shown in the following formula: .

[0068] Next, this application embodiment adopts a dynamic Gamma prior that combines time period and period: existing Bayesian methods use static priors, which cannot adapt to queuing differences in different time periods (such as morning peak, evening peak, off-peak, and night). However, this application embodiment maintains four types of time period priors online, as shown in the following formula: ; Among them, the prior distribution parameters for each time period ( , The prior distribution parameters can be updated online using the maximum likelihood estimation based on the historical queuing data for this period over the previous 14 days. Typical prior distribution parameters are shown in Table 3 below: Table 3: .

[0069] Assuming the current time period is According to prior knowledge The iterative correction is performed as shown in the following formula: ; in, The actual queue length (a state variable to be estimated, in meters); Gamma function (standard mathematical function) ); , They represent the current time period respectively. The shape and scale parameters of the Gamma distribution; Convergence threshold m.

[0070] Here, the iterative results are obtained through numerical optimization, such as solving for the posterior mode using maximum a posterioriprobability (MAP) estimation. The difference between the two estimates. Convergence is considered to occur when m is reached, and typically 3 to 5 iterations are sufficient.

[0071] Therefore, this application embodiment maintains the four-time period Gamma prior online and uses the initial value detected by roadside radar as a strong constraint to accelerate iteration, thereby realizing the Bayesian iterative update of the dynamic time period prior.

[0072] In one implementation, optionally, after performing Bayesian iterative updates on the queue length based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met, the method further includes: The particle filtering method is used to estimate the growth rate of the parking position of the second floating car; The parking position of the second floating car is updated according to the growth rate to obtain the updated parking position of the second floating car; Obtaining the queue length includes: The queue length is obtained based on the updated parking position of the second floating car.

[0073] In this embodiment, the particle filtering method is used to estimate the parking position of the second floating car at a distance (i.e., outside the detection range of the roadside radar). The specific process is as follows: For areas outside the detection range of the roadside radar (e.g., more than 150m from the stop line), a particle filtering method is used to observe the second floating car; specifically, the number of particles... Process noise m; Resampling threshold: effective particle ratio Resampling is triggered at certain times.

[0074] Here, the embodiments of this application extend the particle state vector: existing particle filtering methods only estimate the parking position of the last second floating car. However, the embodiments of this application extend the state to two dimensions. The system estimates the growth rate of parking positions for the second floating car and also estimates the parking position of the second floating car, thus enabling it to predict queuing dynamics. The state transition equation is shown in the following formula: ; in, The second is the update step size for the particle filter; , m / s is the noise standard deviation of the growth rate at the second parking position.

[0075] Finally, the parking location of the second floating car The actual value is the weighted average of the particles, as shown in the following formula: .

[0076] It should be noted that the growth rate of the parking positions of the second floating car... This is used to update the parking position of the second floating car, obtain the updated parking position of the second floating car, and then obtain the queue length based on the updated parking position, forming a closed-loop system of Bayesian principal estimation, particle filter trend prediction and prior dynamic update.

[0077] Therefore, the state in this application embodiment is extended to parking position and growth rate, providing the ability to dynamically predict queues and realize the remote queueing trend estimation of two-dimensional state particle filtering.

[0078] In one embodiment, optionally, after obtaining the queue length, the method further includes: When the current signal period ends, the queue length is checked for consistency based on the radar detection dataset to obtain the check result; If the verification result is unsuccessful, the first fusion weight and the second fusion weight are updated to obtain the updated first fusion weight and the updated second fusion weight. In the next signal cycle, the trajectory data of the first floating car is obtained based on the updated first fusion weight and the updated second fusion weight.

[0079] In the embodiments of this application, at the end of each signal cycle, a consistency check is performed on the output results of that signal cycle (such as trajectory data and the number of floating cars).

[0080] Specifically, when the current signal cycle ends, the queue length is adjusted based on the first radar detection data. Perform a consistency check and obtain the check result. As shown in the formula below: ; in, It is the queue length obtained based on the first radar detection data; here, it is expected that the verification result is less than the preset verification result (e.g., 3m), that is... If the verification result is greater than or equal to the preset verification result, i.e. the deviation is large, then the second fusion weight corresponding to the satellite positioning data of the first floating car is reduced to obtain the updated second fusion weight. Thus, in the next signal cycle, the trajectory data of the first floating car can be obtained according to the updated first fusion weight and the updated second fusion weight.

[0081] Furthermore, in the next signal cycle, the trajectory data of the first floating car is obtained based on the updated first fusion weight and second fusion weight, including: In the next signal cycle, based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle and the preset fusion parameters, the first fusion weight corresponding to the first radar detection data and the second fusion weight corresponding to the satellite positioning data of the first floating vehicle are obtained. Based on the first fusion weight corresponding to the first radar detection data and the updated first fusion weight, the adjusted first fusion weight is obtained; and based on the second fusion weight corresponding to the second radar detection data and the updated second fusion weight, the adjusted second fusion weight is obtained. Based on the adjusted first fusion weight and the adjusted second fusion weight, the first radar detection data and the satellite positioning data of the first floating vehicle are weighted and fused to obtain the trajectory data of the first floating vehicle.

[0082] Furthermore, if the current signal period ends, the method further includes: Verify the rationality of queue growth based on the inflow and outflow relationship at intersections, and define the flow conservation deviation. (in , These represent the number of floating cars entering and leaving the queue during the current signal period, respectively. (This refers to the actual change in the number of floating cars in the queue); here, it is expected that the flow conservation deviation is less than a preset flow rate, for example... If the flow rate conservation deviation is greater than or equal to the preset flow rate, then step 104 above is repeated.

[0083] Therefore, the embodiments of this application drive parameter self-correction through a three-dimensional verification of failure rate of consistency verification and counterfactual verification, thereby achieving closed-loop feedback and multi-source consistency verification.

[0084] The queue length estimation method described in the embodiments of this application is illustrated below with examples: Example 1: Application scenario: Intersections of main urban roads; Scenario Description: A cross-shaped signalized intersection on a main urban road. The east entrance has three lanes for straight traffic and one lane for left turns. A 77GHz millimeter-wave radar is deployed on the roadside, covering an area 150m upstream of the stop line, with a sampling frequency of 10Hz. Floating car data is sourced from the GPS trajectories of taxis and ride-hailing vehicles, with a sampling period of... 15 to 30 seconds, penetration rate approximately 8 Signal period The traffic light cycle is 90 seconds, with a 35-second green light and a 55-second red light for east-west straight traffic. During the morning rush hour, the queue length at the east entrance often exceeds 150m of radar detection range (reaching up to approximately 220m), creating a radar blind spot of about 70m, which is a key scenario addressed in this implementation plan.

[0085] The overall technical approach of this application follows a closed-loop architecture of causal prior, collaborative verification, uncertainty propagation, Bayesian iteration, and closed-loop correction.

[0086] The physical scene of Example 1 and the completion effect achievable in the embodiments of this application are visualized as follows: Figure 2 As shown, floating cars queuing within the roadside radar detection range are visible, while those queuing outside the roadside radar detection range are not visible and require supplementary data obtained through satellite positioning data of the floating cars (orange F1 to F4). The queue length estimated using the embodiment of this application is... m and truth value The difference in m is only 2m (relative error 0.93%), while the existing baseline method estimates it to be 185m (absolute error 29m, relative error 13.5%).

[0087] The specific steps and parameters are as follows: Step 1, Data Preprocessing and Boundary Fusion: Integrate radar detection data (sampling frequency 10Hz) and satellite positioning data (sampling interval 22 seconds) to complete coordinate unification and outlier removal. Within the boundary buffer zone (140 to 160 m), the Mahalanobis distance was measured. Adjust the first fusion weight Second fusion weight The measured continuity error of the boundary transition zone position is 0.76m (meeting the requirement of <1m).

[0088] like Figure 3As shown, the smooth fusion process of radar detection data (high precision, blue) and satellite positioning data (low precision, orange) within the boundary buffer zone is illustrated. The fusion curve (dark blue solid line) shows a positional continuity error of less than 0.8m within the buffer zone. Figure 4 The diagram illustrates the Mahalanobis distance adaptive weight curve of an embodiment of this application. When the data is highly consistent In the event of a serious conflict, It automatically adjusts to 0.95, demonstrating significantly better adaptability than existing fixed-weight baseline methods.

[0089] Step 2, Causal Inference Classification: Estimate the ACE of three types of causal features based on the historical radar detection dataset and historical satellite positioning dataset of the previous 14 days. (Strong causal characteristics) (Spatial constraint characteristics) (Motion pattern characteristics), normalized to obtain , , The results are consistent with the initial weights, which is sufficient to demonstrate the rationality of the empirical assignment.

[0090] Detection of a floating car sample: s、 s、 m、 ,have to , posterior It was determined to be a queuing parking type.

[0091] Employ counterfactual verification mechanism: Verification passed.

[0092] like Figure 5 As shown, a comparison is given between the ACE estimates (red) of three types of causal features and the existing empirical values ​​(gray). The results show that the ACE estimates (0.41 / 0.30 / 0.29) conform to the empirical values ​​(0.40 / 0.30 / 0.30), verifying the rationality of the empirical values. More importantly, the embodiments of this application provide an interpretable, falsifiable, and online-updable weight determination mechanism; as shown... Figure 6 As shown, this illustrates the decision-making process for counterfactual verification: the probability of fact under the current observation. Counterfactual The probability drops to 0.42, indicating a causal effect. The verification was successful, and the parking type was ultimately determined to be queuing.

[0093] Step 3, Classification-driven reverse inference: Floating car detected from m / s decreased to m / s. According to the type of parking queue. Monte Carlo sampling This yields the average value of the first parking position. m, standard deviation m. Compared to a fixed The uncertainty given by the m / s² method and the Monte Carlo method can be used for the likelihood function in subsequent step 6.

[0094] like Figure 7 As shown, satellite positioning data from two sampling sessions ( s time m / s m, s time m / s The motion of the floating car between m is not directly observable (the gray dashed line represents the actual trajectory and is for reference only). Existing baseline methods (gray triangles) use a fixed deceleration. m / s², only the parking position of 184.5m is given, without any uncertainty information; the embodiment of this application (blue star and scatter cloud) samples the deceleration. conduct The Monte Carlo simulation outputs the mean value of 184.7m and the standard deviation of the first parking position. m. This uncertainty serves as the variance term of the Bayesian likelihood function in subsequent step 6, enabling the fusion system to distinguish between reliable and unreliable observations.

[0095] Step 4, Collaborative Confidence Verification: Within the current red light phase, the target floating car (located in...) There are 4 domain floating vehicles within a 50m longitudinal range around (m), of which 3 domain floating vehicles are stationary. They are -22m, -8m, and 15m respectively. All range from 0 to 3.5 meters), and one floating vehicle moves within the area. m, m, located in the adjacent lane).

[0096] Calculate the collaborative confidence score: .

[0097] Number of floating vehicles in the space search window Collaboration upper limit threshold , The parking situation was determined to be a queuing parking type (domain collaboration verification passed).

[0098] Step 5, Gaussian process regression spatiotemporal fusion: Combine the radar detection data with the parking location obtained in Step 3 (including the uncertainty of the parking location). Joint input Gaussian process regression. Because... m, using spatial characteristic length m; Preset time feature length seconds; signal variance For parking locations outside the detection range of the roadside radar, the predicted mean and variance are output.

[0099] Step 6, Bayesian iterative update: The current time period is the morning rush hour (07:30), dynamic prior. The average value is 81m.

[0100] Initial value: m (roadside radar detects the parking position of the last floating car). First observation: m, m, likelihood MAP m; Second observation m, m, get m; Third observation m, m, get m; Fourth observation m, get m; 5th observation m, get m; Differences between the two times m< m converges to m (actual ground value measured 214m, error 13.5%).

[0101] like Figure 8 and Figure 9 The Bayesian iterative update process is visualized. Figure 8 The embodiment of this application (blue solid line, convergence after 5 iterations) starts from the radar initial value. Starting from m, the method updates via MAP after each observation from the floating car, and converges to 185.1m after 3 to 5 iterations; the existing baseline method (gray dashed line) fails to converge after 11 iterations to 180m due to the lack of strong initial values ​​and dynamic priors. Figure 9 This application demonstrates a unique dynamic time-period Gamma prior mechanism, which automatically selects the corresponding time period based on the current time period (morning peak / evening peak / off-peak / night). The distribution allows the prior mean to dynamically adapt from 16m at night to 96m during the evening rush hour.

[0102] Step 7, Particle Filter Far-End Estimation: in two-dimensional state Running a 500-particle filter yields the following results: m、 m / s (queue is still growing), with an error of only 0.93% from the ground truth of 214m. The state extension of the particle filter improves the estimation accuracy by 6.6% compared to using the Bayesian module alone.

[0103] like Figure 10 As shown, the distribution of 500 particles gradually changes over time from the initial Gamma prior (diffuse distribution) towards the true value. The convergence process of m is completed within 30 seconds. The estimated value in this application's embodiment... m (difference from the true value by only 2m); such as Figure 11 As shown, this demonstrates the key value of two-dimensional state extension, whereas existing 1D particle filters (gray dashed lines) only estimate position. The embodiments of this application are extended to After that, it is possible to pass Predicting the position of the tail of the queue 45 seconds later, an overflow warning (green star) is triggered in advance, giving upstream variable speed limit control valuable response time.

[0104] Step 8, Closed-loop verification: The verification result is 1.8m < 3m; the flow conservation deviation is 1 vehicle / cycle < 2 vehicles / cycle, the verification passes; the feedback loop will... The updated value is written to the history database.

[0105] Thus, after steps 1 to 8 above, and after 7 days of continuous testing at the intersection (including 5 weekdays of morning and evening peak hours and 2 off-peak days), as shown in Table 4 below, the embodiments of this application achieve the following 7 beneficial effects compared to the existing baseline methods. And as... Figure 12 As shown, the beneficial effects are visually presented in the form of a bar chart.

[0106] Table 4: .

[0107] Example 2: Application scenario: Merging area of ​​urban expressway ramps; Scenario Description: A traffic signal control system is installed at the junction of an off-ramp and the surface road on a city expressway. Each lane of the off-ramp is approximately 180m long. Due to structural limitations, roadside radar cannot be installed; only one roadside radar at the ramp entrance can monitor the first 60m. The queue length along the entire ramp is estimated based on fused trajectory data, with particular attention paid to queue completion in the later sections of the ramp (60 to 180m from the merging point).

[0108] Key parameters and results include the following: The detection range of the roadside radar is 60m, and the detection range outside the roadside radar is 120m. Floating car penetration rate is approximately A single ramp per cycle is approximately A floating car sample; Due to the low sample size scenario, an adaptive thresholding mechanism is used to adjust the collaborative upper limit threshold. Lowered to 0.59; During the morning rush hour, the queue length can reach a maximum of 165m. The queue length estimated using this embodiment is 61.8m, with an absolute error of 3.2m and a relative error of 1.9%. Compared to existing baseline methods (absolute error of 27.5m and relative error of 16.7%), this embodiment reduces the estimation error of queue lengths outside the roadside radar detection range by 88.4%. Due to rapid changes in ramp merging pressure, the particle filter method is used to estimate the growth rate of the second parking position. It provides an early warning signal (approximately 45 seconds in advance) for when the queue is about to overflow, which can assist upstream variable speed limit control.

[0109] In summary, the queue length estimation method described in the embodiments of this application provides a three-layer closed-loop architecture of causal prior, multi-source collaboration, and Bayesian iteration: interpretable pre-screening of parking behavior is performed through causal prior, multi-source collaboration is used to verify and eliminate uncertainty of individual vehicles, and Bayesian iteration is used to achieve refined estimation of remote queues. The three layers are mutually causal feedback. Introducing classification-driven differential kinematics backpropagation: Based on the parking behavior type, preset deceleration parameters are dynamically selected to accurately convert low-frequency satellite positioning data into high-precision parking position constraint points. The longitudinal error of the parking position can be reduced from 12m to 3.8m (a reduction of 68.3%). Adaptive weighted fusion of radar detection data and satellite positioning data using two-way Kalman and Mahalanobis distance at the boundary: The fusion weights are dynamically adjusted using Mahalanobis distance as a measure of data conflict, and the position continuity error can be less than 0.8m in the boundary transition zone; By combining true causal effect estimation with Gaussian kernel collaborative confidence, the causal contribution of each feature to queuing decision is quantified using conditional mutual information, and then verified using the collaborative confidence of neighborhood floating cars. The queuing identification accuracy is now improved from 89% to 96.4% compared to the existing baseline method. Bayesian iterative update mechanism for radar initial value and floating car sparse observation: The last parking position of the floating car detected by the roadside radar is used as the strong initial value, and the parking position of the floating car is used as the sparse posterior observation. It can converge in 3 to 5 iterations, and the contribution of the far-end completion error can be less than 1.8%. The cooperative window mechanism for signal period awareness: the cooperative window is aligned with the signal period, the prior is automatically reset, and cross-period data pollution is avoided; The overall solution reduces the average absolute percentage error of queue length estimation across the entire intersection area (including roadside radar blind spots) from 17.6% to 4.7%, meeting the needs of real-time traffic control.

[0110] like Figure 13 As shown in the figure, this application embodiment also provides a queue length estimation device, including: The fusion module 1301 is used to obtain trajectory data of each of the multiple first floating vehicles within the detection range of the roadside radar during the current signal cycle, based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle; wherein, the first radar detection data is the radar detection data corresponding to the first floating vehicle in the radar detection dataset of the current signal cycle; the radar detection dataset and the satellite positioning data of the first floating vehicle are aligned to the same coordinate system and synchronized in time. The classification module 1302 is used to obtain parking-related data from the trajectory data and classify the parking behavior of the floating car according to the parking-related data to obtain a first parking behavior type; The estimation module 1303 is used to perform kinematic back-calculation on multiple second floating cars outside the detection range of the roadside radar based on the preset deceleration parameters corresponding to the first parking behavior type if the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, so as to obtain the parking position and uncertainty of each second floating car; wherein, the satellite positioning data of the second floating car indicates that the second floating car stops between two adjacent samplings, and the parking behavior type of the second floating car is queuing parking; The iteration module 1304 is used to take the parking position of the target first floating car as the initial queue length, the parking position of each second floating car as the observation point, and perform Bayesian iterative updates on the queue length according to the prior distribution parameters corresponding to the current time period until the iteration end condition is met, thereby obtaining the queue length.

[0111] Optionally, in the queue length estimation device, the fusion module 1301 is specifically used for: Based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle, and a preset fusion parameter, a first fusion weight corresponding to the first radar detection data and a second fusion weight corresponding to the satellite positioning data of the first floating vehicle are obtained; wherein, the sum of the first fusion weight and the second fusion weight is equal to 1; the preset fusion parameter includes at least one of the following: a reference threshold; a scale parameter of the hyperbolic tangent function; a basic radar weight; and an upper limit for the weight adjustment range; Based on the first fusion weight and the second fusion weight, the first radar detection data and the satellite positioning data of the first floating vehicle are weighted and fused to obtain the trajectory data of the first floating vehicle.

[0112] Optionally, in the queue length estimation device, the classification module 1302 is specifically used for: Using the causal effect estimation method, historical radar detection datasets, and historical satellite positioning datasets, we obtain the initial weights corresponding to strong causal features, spatial constraint features, and motion pattern features, respectively. A first score is obtained based on the parking duration in the parking-related data, which is related to the strong causal feature. A second score is obtained based on the distance from the stop line in the parking-related data, which is related to the spatial constraint feature. A third score is obtained based on the trajectory curvature in the parking-related data, which is related to the motion pattern feature. The floating car's parking behavior is obtained by weighted fusion based on the first score, the second score, and the third score, as well as the initial weights corresponding to the strong causal feature, the spatial constraint feature, and the motion pattern feature. The parking behavior of the first floating car is classified based on the comprehensive score and the posterior probability calculated through the conditional probability table to obtain the first parking behavior type.

[0113] Optionally, in the queue length estimation device, the estimation module 1303 includes: The obtaining unit is used to obtain the probability distribution of the parking position of the floating car based on the preset deceleration parameter corresponding to the first parking behavior type if the first parking behavior type is a queuing parking type or a passenger pick-up and drop-off parking type. The sampling unit is used to perform multiple samplings based on the probability distribution to obtain the parking position and uncertainty of each of the multiple second floating cars outside the detection range of the roadside radar.

[0114] Optionally, the queue length estimation device further includes: The verification module is used to perform domain-based collaborative verification on the first parking behavior type and obtain the verification result; An exclusion module is used to exclude the first floating car whose verification result is that the verification fails from the plurality of first floating cars, so as to obtain a plurality of first floating cars after exclusion. The first obtaining module is used to obtain the target first floating car based on the parking position of each of the first floating cars in the plurality of excluded first floating cars; The acquisition module is used to acquire the spatiotemporal similarity between multiple first floating cars whose verification results are verified and which are located in the same parking feature area; The second acquisition module is used to obtain the target first floating car based on the parking positions of multiple first floating cars whose spatiotemporal similarity is greater than a preset similarity.

[0115] Optionally, in the queue length estimation device, the verification module is specifically used for: Within the collaboration window, a spatial search window is established with the first floating car as the center to obtain the domain floating cars located within the spatial search window; The cooperative confidence between the domain floating car and the first floating car is obtained based on the longitudinal and lateral distances between the domain floating car and the first floating car, the parking indication function of the domain floating car, and the number of domain floating cars in the spatial search window. If the collaboration confidence level is greater than the collaboration upper limit threshold, then the second parking behavior type of the first floating car is determined to be the queuing parking type. If the collaboration confidence level is less than or equal to the collaboration upper limit threshold and greater than or equal to the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be the passenger pick-up and drop-off parking type. If the collaboration confidence level is less than the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be either faulty parking or illegal parking. The second parking behavior type is compared with the first parking behavior type to obtain the verification result.

[0116] Optionally, in the queue length estimation device, the acquisition module is specifically used for: For multiple first floating cars whose verification results are passed and located in the same parking feature area, a joint Gaussian process regression is obtained based on the preset time feature length and the spatial feature length corresponding to the parking feature area. The joint Gaussian process regression is constructed based on the kernel function and the preset mean function; Based on the joint Gaussian process regression, the spatiotemporal similarity between multiple first floating cars is obtained.

[0117] Optionally, the queue length estimation device further includes: The rate estimation module is used to estimate the growth rate of the parking position of the second floating car using the particle filtering method; The first update module is used to update the parking position of the second floating car according to the growth rate, so as to obtain the updated parking position of the second floating car.

[0118] The iteration module 1304 is specifically used for: The queue length is obtained based on the updated parking position of the second floating car.

[0119] Optionally, the queue length estimation device further includes: The verification module is used to perform a consistency verification on the queue length based on the radar detection dataset when the current signal cycle ends, and obtain the verification result. The second update module is used to update the first fusion weight and the second fusion weight if the verification result is unsuccessful, so as to obtain the updated first fusion weight and the updated second fusion weight. The third acquisition module is used to obtain the trajectory data of the first floating car in the next signal cycle based on the updated first fusion weight and the updated second fusion weight.

[0120] It should be noted that the queue length estimation device provided in this application embodiment can implement all the method steps implemented in the above queue length estimation method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0121] This application also provides a queue length estimation device, such as... Figure 14 As shown, it includes:

[0122] The processor 1401, memory 1402, transceiver 1403, and a program or instructions stored in the memory 1402 and executable on the processor 1401; when the processor 1401 executes the program or instructions, it implements the various processes of the above-described queue length estimation method embodiments and achieves the same technical effect. To avoid repetition, these will not be described again here.

[0123] The transceiver 1403 is used to receive and send data under the control of the processor 1401.

[0124] Among them, Figure 14In this context, the bus architecture can include any number of interconnected buses and bridges, specifically connecting various circuits of one or more processors represented by processor 1401 and memory represented by memory 1402. The bus architecture can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 1403 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. For different user equipment, the user interface 1404 can also be an interface capable of connecting external or internal devices, including but not limited to keypads, displays, speakers, microphones, joysticks, etc.

[0125] Processor 1401 is responsible for managing the bus architecture and general processing, while memory 1402 can store the data used by processor 1401 when performing operations.

[0126] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described queue length estimation method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc.

[0127] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described queue length estimation method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0128] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0130] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for estimating queue length, characterized in that, include: Within the current signal period, for each of the multiple first floating vehicles within the detection range of the roadside radar, trajectory data of the first floating vehicle is obtained based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle; wherein, the first radar detection data is the radar detection data corresponding to the first floating vehicle in the radar detection dataset of the current signal period; the radar detection dataset and the satellite positioning data of the first floating vehicle are aligned to the same coordinate system and synchronized in time; Parking-related data is obtained from the trajectory data, and the parking behavior of the first floating car is classified according to the parking-related data to obtain the first parking behavior type; If the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, then according to the preset deceleration parameters corresponding to the first parking behavior type, kinematic back-inference is performed on multiple second floating cars outside the detection range of the roadside radar to obtain the parking position and uncertainty of each second floating car; wherein, the satellite positioning data of the second floating car indicates that the second floating car stops between two adjacent samplings, and the parking behavior type of the second floating car is queuing parking. The initial queue length is determined by taking the parking position of the target first floating car as the initial queue length and the parking position of each second floating car as the observation point. The queue length is then updated using Bayesian iteration based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met, thus obtaining the queue length. The target first floating car is the first floating car that is furthest from the stop line among multiple first floating cars. The prior distribution parameters are obtained by updating historical queue data according to the corresponding time period.

2. The method according to claim 1, characterized in that, Based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle, the trajectory data of the first floating vehicle is obtained, including: Based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle, and a preset fusion parameter, a first fusion weight corresponding to the first radar detection data and a second fusion weight corresponding to the satellite positioning data of the first floating vehicle are obtained; wherein, the sum of the first fusion weight and the second fusion weight is equal to 1; the preset fusion parameter includes at least one of the following: a reference threshold; a scale parameter of the hyperbolic tangent function; a basic radar weight; and an upper limit for the weight adjustment range; Based on the first fusion weight and the second fusion weight, the first radar detection data and the satellite positioning data of the first floating vehicle are weighted and fused to obtain the trajectory data of the first floating vehicle.

3. The method according to claim 1, characterized in that, Based on the parking-related data, the parking behavior of the first floating car is classified to obtain a first parking behavior type, including: Using the causal effect estimation method, historical radar detection datasets, and historical satellite positioning datasets, we obtain the initial weights corresponding to strong causal features, spatial constraint features, and motion pattern features, respectively. A first score is obtained based on the parking duration in the parking-related data, which is related to the strong causal feature. A second score is obtained based on the distance from the stop line in the parking-related data, which is related to the spatial constraint feature. A third score is obtained based on the trajectory curvature in the parking-related data, which is related to the motion pattern feature. The first floating car's parking behavior is obtained by weighted fusion based on the first score, the second score, the third score, and the initial weights corresponding to the strong causal feature, the spatial constraint feature, and the motion pattern feature. The parking behavior of the first floating car is classified based on the comprehensive score and the posterior probability calculated through the conditional probability table to obtain the first parking behavior type.

4. The method according to claim 1, characterized in that, If the first parking behavior type is queuing parking or passenger pick-up / drop-off parking, then based on the preset deceleration parameters corresponding to the first parking behavior type, kinematic back-calculation is performed on multiple second floating cars outside the detection range of the roadside radar to obtain the parking position and uncertainty of each second floating car, including: If the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, then according to the preset deceleration parameter corresponding to the first parking behavior type, the probability distribution of the parking position of each of the multiple second floating cars outside the detection range of the roadside radar is obtained. Based on the probability distribution, multiple samplings are performed to obtain the parking position and uncertainty of the second floating car.

5. The method according to claim 1, characterized in that, The method further includes: Perform domain-based collaborative verification on the first parking behavior type to obtain the verification results; The first floating car whose verification result is "verification failed" is excluded from the plurality of first floating cars to obtain a plurality of excluded first floating cars; The target first floating car is obtained based on the parking position of each of the excluded first floating cars; For multiple first floating cars whose verification results are passed and which are located in the same parking feature area, the spatiotemporal similarity between the multiple first floating cars is obtained. The target first floating car is obtained based on the parking positions of multiple first floating cars whose spatiotemporal similarity is greater than a preset similarity.

6. The method according to claim 5, characterized in that, Perform domain-based collaborative verification on the first parking behavior type to obtain verification results, including: Within the collaboration window, a spatial search window is established with the first floating car as the center to obtain the domain floating cars located within the spatial search window; The cooperative confidence between the domain floating car and the first floating car is obtained based on the longitudinal and lateral distances between the domain floating car and the first floating car, the parking indication function of the domain floating car, and the number of domain floating cars in the spatial search window. If the collaboration confidence level is greater than the collaboration upper limit threshold, then the second parking behavior type of the first floating car is determined to be the queuing parking type. If the collaboration confidence level is less than or equal to the collaboration upper limit threshold and greater than or equal to the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be the passenger pick-up and drop-off parking type. If the collaboration confidence level is less than the collaboration lower limit threshold, then the second parking behavior type of the first floating car is determined to be either faulty parking or illegal parking. The second parking behavior type is compared with the first parking behavior type to obtain the verification result.

7. The method according to claim 5, characterized in that, For multiple first floating cars whose verification results are successful and located in the same parking feature area, the spatiotemporal similarity between the multiple first floating cars is obtained, including: For multiple first floating cars whose verification results are passed and located in the same parking feature area, the kernel function of joint Gaussian process regression is obtained according to the preset time feature length and the spatial feature length corresponding to the parking feature area; The joint Gaussian process regression is constructed based on the kernel function and the preset mean function; Based on the joint Gaussian process regression, the spatiotemporal similarity between multiple first floating cars is obtained.

8. The method according to claim 1, characterized in that, After performing Bayesian iterative updates on the queue length based on the prior distribution parameters corresponding to the current time period until the iteration termination condition is met, the method further includes: The particle filtering method is used to estimate the growth rate of the parking position of the second floating car; The parking position of the second floating car is updated according to the growth rate to obtain the updated parking position of the second floating car; Obtaining the queue length includes: The queue length is obtained based on the updated parking position of the second floating car.

9. The method according to claim 2, characterized in that, After obtaining the queue length, the method further includes: When the current signal period ends, the queue length is checked for consistency based on the radar detection dataset to obtain the check result; If the verification result is unsuccessful, the first fusion weight and the second fusion weight are updated to obtain the updated first fusion weight and the updated second fusion weight. In the next signal cycle, the trajectory data of the first floating car is obtained based on the updated first fusion weight and the updated second fusion weight.

10. A queue length estimation device, characterized in that, include: The fusion module is used to obtain trajectory data of each of the multiple first floating vehicles within the detection range of the roadside radar during the current signal cycle, based on the distance metric between the forward prediction state of the first radar detection data and the backward smoothing state of the satellite positioning data of the first floating vehicle; wherein, the first radar detection data is the radar detection data corresponding to the first floating vehicle concentrated in the radar detection dataset of the current signal cycle; the radar detection data and the satellite positioning data of the first floating vehicle are aligned to the same coordinate system and synchronized in time; The classification module is used to obtain parking-related data from the trajectory data and classify the parking behavior of the first floating car according to the parking-related data to obtain a first parking behavior type; The estimation module is used to perform kinematic back-calculation on multiple second floating cars outside the detection range of the roadside radar based on the preset deceleration parameters corresponding to the first parking behavior type if the first parking behavior type is queuing parking or passenger pick-up and drop-off parking, so as to obtain the parking position and uncertainty of each second floating car; wherein, the satellite positioning data of the second floating car indicates that the second floating car stops between two adjacent samplings, and the parking behavior type of the second floating car is queuing parking; An iterative module is used to take the parking position of the target first floating car as the initial queue length, the parking position of each second floating car as an observation point, and perform Bayesian iterative updates on the queue length according to the prior distribution parameters corresponding to the current time period until the iteration termination condition is met, thereby obtaining the queue length; wherein, the target first floating car is the first floating car that is farthest from the stop line among multiple first floating cars; the prior distribution parameters are obtained by updating historical queue data according to the corresponding time period.

11. A queue length estimation device, characterized in that, include: A processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the processor, when executing the program or instructions, implements the queue length estimation method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the queue length estimation method as described in any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the queue length estimation method as described in any one of claims 1 to 9.

Images