A vehicle positioning method, a vehicle positioning device, and a computer storage medium

Abstract

The application provides a vehicle positioning method, a vehicle positioning device and a computer storage medium. The method comprises the following steps: obtaining a field end vehicle detection result positioned by current frame field end point cloud data; obtaining an uncertainty estimation value of each vehicle detection parameter value; obtaining a time deviation between a current frame vehicle end positioning time and a current frame field end true value time; determining a comprehensive uncertainty measure of the field end vehicle detection result based on the uncertainty estimation value of each vehicle detection parameter value; determining a current frame positioning state classification result according to the comprehensive uncertainty measure and the time deviation; compensating the field end vehicle detection result according to the positioning state classification result; and outputting a vehicle positioning result according to the compensated field end vehicle detection result. Through the above vehicle positioning method, the uncertainty decomposition is introduced for field end laser radar true value quality evaluation, so that the accuracy of the field end true value is improved, and the accurate positioning of the vehicle is realized.

Description

Technical Field

[0001] This application relates to the field of vehicle positioning technology, and in particular to a vehicle positioning method, a vehicle positioning device, and a computer storage medium. Background Technology

[0002] With the rapid development of intelligent connected vehicles and autonomous driving technologies, accurate vehicle positioning in indoor GNSS-denied scenarios such as underground parking garages, enclosed factories, and logistics sorting centers has become a focus of industry attention. Currently, such scenarios mainly rely on a technical solution combining field-based LiDAR perception and vehicle-side autonomous positioning. Existing solutions typically collect point cloud data using LiDAR deployed at the field site, and use target detection algorithms to extract parameters such as the vehicle's position, size, and heading angle, using this as a global ground truth benchmark. Simultaneously, the vehicle performs autonomous positioning calculations based on sensors such as LiDAR SLAM, UWB, or IMU. The system compares the field-based detection results with the vehicle-side positioning results to calculate the spatiotemporal deviation, and then verifies and compensates for the vehicle-side positioning to maintain driving safety and path tracking accuracy in complex indoor environments.

[0003] However, in existing technologies, the lack of explicit quantification of the uncertainty in the ground truth quality of the field and the ability to adaptively predict the nonlinear motion patterns of vehicles pose challenges to the accuracy and robustness of positioning verification in scenarios with intense vehicle maneuvering, severe environmental occlusion, or minor deviations in spatiotemporal synchronization. It is difficult to effectively balance verification sensitivity and system stability, and misjudgments or missed detections of positioning anomalies are prone to occur. Summary of the Invention

[0004] To address the aforementioned technical problems, this application proposes a vehicle positioning method, a vehicle positioning device, and a computer storage medium.

[0005] To address the aforementioned technical problems, this application proposes a vehicle positioning method, which includes: Obtain the field-end vehicle detection result obtained from the current frame field endpoint cloud data positioning; wherein, the field-end vehicle detection result includes several vehicle detection parameter values; Obtain an uncertainty estimate for each vehicle's detection parameter value; Obtain the time deviation between the current frame vehicle-end positioning time and the current frame field-end true value time; Based on the uncertainty estimate of each vehicle detection parameter value, a comprehensive uncertainty measure of the field vehicle detection results is determined; The current frame positioning status classification result is determined based on the comprehensive uncertainty metric and the time deviation. The vehicle detection results at the vehicle end are compensated based on the positioning status classification results; The vehicle positioning result is output based on the compensated vehicle detection results.

[0006] The uncertainty estimates include cognitive uncertainty estimates and accidental uncertainty estimates.

[0007] The uncertainty estimate for obtaining the detection parameter value for each vehicle includes: The current frame field endpoint cloud data is input into the vehicle detection network, and several forward propagations are performed to obtain several sets of vehicle detection parameter values. Obtain the estimated mean of several vehicle detection parameter values ​​for the target class of vehicles; The estimated variance measure is determined based on the vehicle detection parameter values ​​of the target class vehicle detection parameters and the estimated mean. The estimated variance metric is used as an estimate of the uncertainty of the target class vehicle detection parameters.

[0008] The uncertainty estimate for obtaining the detection parameter value for each vehicle includes: Input the current frame field endpoint cloud data into the vehicle detection network to obtain the vehicle detection parameter values ​​and their logarithmic variance; The random uncertainty estimate of the vehicle detection parameter value is determined based on the logarithmic variance.

[0009] The vehicle positioning method further includes: Obtain a training dataset, which contains multiple frames of LiDAR point cloud data and corresponding ground truth labels for target bounding boxes; A target vehicle detection network is constructed, which includes a feature extraction backbone network, a bounding box regression branch, and an uncertainty prediction branch. The lidar point cloud data is input into the target vehicle detection network. The bounding box regression branch outputs the predicted value, and the uncertainty prediction branch simultaneously outputs the corresponding logarithmic variance. The uncertainty-weighted loss value is determined based on the predicted value, the true value label, and the log-variance. The target vehicle detection network is trained using the uncertainty-weighted loss value to obtain the vehicle detection network.

[0010] The vehicle localization method further includes, after obtaining the uncertainty estimate of each vehicle detection parameter value: When the estimated cognitive uncertainty of all vehicle detection parameter values ​​is less than the cognitive uncertainty threshold and the estimated random uncertainty is less than the random uncertainty threshold, the field vehicle detection result is marked as a high-confidence true value. When there is a cognitive uncertainty estimate of at least one vehicle detection parameter value that is greater than the cognitive uncertainty threshold, and the cognitive uncertainty estimate of all vehicle detection parameter values ​​is less than a preset multiple of the cognitive uncertainty threshold, or when there is a random uncertainty estimate of at least one vehicle detection parameter value that is greater than the random uncertainty threshold, and the cognitive uncertainty estimate of all vehicle detection parameter values ​​is less than a preset multiple of the random uncertainty threshold, the field vehicle detection result is marked as a medium confidence true value. When the estimated cognitive uncertainty of at least one vehicle detection parameter value is greater than a preset multiple of the cognitive uncertainty threshold, or when the estimated random uncertainty of at least one vehicle detection parameter value is greater than a preset multiple of the random uncertainty threshold, the field vehicle detection result is marked as a low-confidence true value.

[0011] The vehicle positioning method further includes: When the time deviation is greater than the time deviation threshold, or when the vehicle detection result at the field is a low-confidence true value, the vehicle detection result at the field is updated by extrapolation prediction based on a learning model.

[0012] The step of updating the on-site vehicle detection results using extrapolation prediction based on a learning model includes: The true state sequence of the field end within the historical time window is input into the first extrapolation prediction model. The input features of the state sequence include lateral position error, longitudinal position error, heading angle error, yaw rate, longitudinal velocity, lateral velocity, and longitudinal acceleration. Obtain the predicted state sequence within the future time window output by the first extrapolation prediction model, the predicted state sequence including predicted lateral position, predicted longitudinal position, predicted heading angle and predicted speed; Kinematic constraints are applied to the predicted state sequence to obtain the corrected vehicle detection results at the field end.

[0013] The kinematic constraint correction includes: When the predicted acceleration at adjacent moments in the predicted state sequence exceeds the preset acceleration range, the predicted acceleration is truncated to the boundary value of the preset acceleration range. When the predicted curvature exceeds the preset curvature threshold, the yaw rate is truncated.

[0014] The vehicle localization method further includes, after applying kinematic constraint correction to the predicted state sequence to obtain the corrected field vehicle detection results: The prediction deviation is calculated based on the corrected field-end vehicle detection results and the vehicle-end vehicle detection results. Based on the relationship between the predicted deviation and the preset deviation threshold, a tiered prediction is performed: When the prediction deviation is less than the first deviation threshold, it is determined to be in a normal state; When the prediction deviation is between the first deviation threshold and the second deviation threshold, it is determined to be a potential degradation state and the first level of compensation is triggered. When the prediction deviation is greater than or equal to the second deviation threshold, it is determined that the deviation is about to occur and the second level of compensation is triggered. The first deviation threshold and the second deviation threshold are determined based on the statistical distribution of historical data.

[0015] The vehicle localization method further includes, after applying kinematic constraint correction to the predicted state sequence to obtain the corrected field vehicle detection results: The spatial deviation is calculated based on the corrected predicted trajectory and the vehicle-end positioning data. Obtain the extrapolation uncertainty index corresponding to the extrapolation prediction; The spatial deviation threshold is dynamically adjusted based on the extrapolation uncertainty index, wherein the spatial deviation threshold increases as the extrapolation uncertainty index increases.

[0016] The vehicle positioning method further includes: Construct multiple first extrapolation prediction models with identical structures but different initialization parameters; Obtain the independent prediction results of each of the first extrapolation prediction models for the same input; Based on the dispersion of each independent prediction result and the preset model inherent variance, the extrapolation uncertainty index is determined.

[0017] The comprehensive uncertainty measure is generated by weighted summation of the following items: normalized cognitive uncertainty index, normalized accidental uncertainty index, normalized extrapolation uncertainty index, and comprehensive verification compensation term; The normalization indices are obtained by standardizing the historical statistics of the corresponding uncertainty components.

[0018] Wherein, when the extrapolation prediction based on the learning model is not executed, the normalized extrapolation uncertainty index takes the value of zero.

[0019] The comprehensive verification confidence index corresponding to the comprehensive verification compensation item is generated by weighted summation of the following items: The confidence terms for time deviation, space deviation, velocity deviation, heading angle deviation, and true value uncertainty penalty are included. The truth uncertainty penalty term is determined based on the proportion of parameters with different confidence levels in the vehicle detection results at the field. The higher the proportion of parameters with high confidence levels, the closer the truth uncertainty penalty term is to the upper limit.

[0020] The positioning status classification results include: Level 1, Level 2, Level 3, and Level 4, wherein: The triggering conditions corresponding to the first level include: the time deviation meets the first condition, the spatial deviation meets the second condition, and the comprehensive uncertainty measure meets the third condition; The triggering conditions for the second level include: the spatial deviation meets the fourth condition or the comprehensive uncertainty measure meets the fifth condition; The triggering conditions for the third level include: the spatial deviation meets the sixth condition, the comprehensive uncertainty measurement meets the seventh condition, or the number of cycles in which the prediction deviation continues to exceed the warning threshold reaches a preset value. The triggering conditions corresponding to the fourth level include: the spatial deviation meets the eighth condition, the comprehensive verification confidence is lower than the preset confidence lower limit, the comprehensive uncertainty measure meets the ninth condition, the number of consecutive verification failure frames reaches a preset value, or the communication interruption duration reaches a preset value.

[0021] The compensation of vehicle detection results based on positioning status classification results includes: The compensation strategy corresponding to the first level is to use the vehicle detection results at the vehicle end; The compensation strategy corresponding to the second level is to calculate the correction amount based on the corrected predicted trajectory and the vehicle detection results at the vehicle end, and correct the vehicle detection results at the vehicle end in a weighted manner. The weighting coefficient is dynamically adjusted according to the comprehensive uncertainty measure. The compensation strategy corresponding to the third level is to use a filtering fusion algorithm to fuse the field vehicle detection results and the vehicle detection results, and adjust the observation noise parameters in the filtering fusion algorithm according to the field truth confidence level. The compensation strategy corresponding to the fourth level is to trigger a security downgrade operation.

[0022] The step of adjusting the observation noise parameters in the filtering and fusion algorithm based on the ground truth confidence level includes: When the confidence level of the true value at the field end is the first confidence level, the first observation noise covariance is used; When the confidence level of the true value at the field end is the second confidence level, the second observation noise covariance is used; Wherein, the second confidence level is lower than the first confidence level, and the second observation noise covariance is greater than the first observation noise covariance.

[0023] The vehicle positioning method further includes, after outputting the vehicle positioning result based on the compensated vehicle detection result, the vehicle positioning result after that: The actual position and pose data of the vehicle and the vehicle detection results of the field are obtained, and a multi-dimensional residual sequence is calculated. The multi-dimensional residual includes position residual, heading residual and velocity residual. Perform time-series statistical analysis on the multidimensional residual sequence, including mean drift detection, variance change detection, and autocorrelation characteristic analysis. Based on the results of the time-series statistical analysis, a graded early warning information is generated, and differentiated response actions are executed according to the graded early warning information. The response actions include adjusting calibration parameters to optimize the frequency, triggering online recalibration, or performing security degradation. Based on the residual sequence analysis results, the parameter set is optimized online. The parameter set includes at least one of the following: time delay parameter, coordinate transformation matrix, fusion weight parameter, and hierarchical threshold parameter.

[0024] The mean drift detection uses a cumulative sum control chart algorithm to determine the warning level by the ratio of the cumulative statistic to a preset decision threshold; the variance change detection uses the F-test method to determine the warning level by the variance change rate of the residual sequence in the first and second halves; and the autocorrelation characteristic analysis determines the warning level by the number of autocorrelation coefficients exceeding the confidence interval under multiple lag orders.

[0025] The tiered early warning system includes three levels: when the first level early warning is triggered, the frequency of calibration parameter optimization is increased; when the second level early warning is triggered, online recalibration is performed using static environmental characteristics to update coordinate transformation parameters; when the third level early warning is triggered, the verification service is suspended and the recalibration process is started.

[0026] The online optimization parameter set includes: model parameters adjusted based on residual feedback, and hierarchical threshold parameters dynamically adjusted based on the historical statistical distribution of comprehensive uncertainty measurement; an outlier removal mechanism is introduced during the optimization process to ensure long-term robustness.

[0027] To solve the above-mentioned technical problems, this application also proposes a vehicle positioning device, which includes a memory and a processor coupled to the memory; The memory is used to store program data, and the processor is used to execute the program data to implement the vehicle positioning method described above.

[0028] To address the aforementioned technical problems, this application also proposes a computer storage medium for storing program data, which, when executed by a computer, is used to implement the vehicle positioning method described above.

[0029] Compared with existing technologies, the beneficial effects of this application are as follows: This application constructs a complete closed-loop system of uncertainty quantification, predictive compensation, and adaptive hierarchical classification. By acquiring field endpoint cloud data and decomposing its cognitive and accidental uncertainties, fluctuations in the quality of the true value can be explicitly perceived, avoiding the risk of using low-quality data as a benchmark. Combined with the accurate measurement of time deviation, the verification error problem caused by vehicle-road spatiotemporal asynchrony is effectively solved. Based on the positioning state classification jointly driven by comprehensive uncertainty measurement and time deviation, the compensation strategy can be dynamically adjusted according to the complexity of the scene, preventing overcompensation in open scenes and ensuring timely and safe degradation in complex occluded scenes. Finally, the differentiated compensation mechanism uses high-confidence true values ​​to correct vehicle drift or triggers fusion and degradation at low confidence levels, significantly reducing the probability of falsely triggering safe degradation and the risk of missed positioning failure, achieving high reliability and high safety operation of indoor dynamic vehicle positioning. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the first embodiment of the vehicle positioning method provided in this application; Figure 2 This is a flowchart illustrating the second embodiment of the vehicle positioning method provided in this application; Figure 3 This is a flowchart illustrating the third embodiment of the vehicle positioning method provided in this application; Figure 4 This is a flowchart illustrating the fourth embodiment of the vehicle positioning method provided in this application; Figure 5 This is a flowchart illustrating the fifth embodiment of the vehicle positioning method provided in this application; Figure 6 This is a flowchart illustrating the sixth embodiment of the vehicle positioning method provided in this application; Figure 7 This is a schematic diagram of an embodiment of the vehicle positioning device provided in this application; Figure 8 This is a schematic diagram of the structure of an embodiment of the computer storage medium provided in this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0032] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such orders can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] In GNSS-denied scenarios such as indoor underground parking garages and enclosed factories, the positioning systems of autonomous vehicles face severe challenges. Existing technical solutions typically treat the detection results of field-end LiDAR directly as a deterministic ground truth benchmark, ignoring factors such as incomplete point clouds caused by vehicle occlusion, observation noise introduced by distant low-reflectivity targets, and insufficient model generalization ability in complex dynamic scenarios. This results in significant cognitive and accidental uncertainties in the ground truth quality.

[0034] To address the aforementioned issues, this application provides an indoor dynamic vehicle positioning verification and compensation method. Please refer to [link / reference needed] for details. Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the vehicle positioning method provided in this application.

[0035] like Figure 1 As shown, the vehicle positioning method provided in this application includes the following steps: Step S11: Obtain the field-end vehicle detection result obtained from the current frame field endpoint cloud data positioning; wherein, the field-end vehicle detection result includes several vehicle detection parameter values.

[0036] In this embodiment, the point cloud data at the site can refer to a set of three-dimensional spatial point clouds collected by fixed LiDAR sensors deployed in indoor environments (such as parking lots and logistics warehouses). The site vehicle detection result is structured data generated after feature extraction and regression processing of the original point cloud data by a target detection network running on edge computing nodes (MECs) or cloud servers. Several vehicle detection parameter values ​​specifically include a seven-dimensional vector describing the vehicle's geometry and motion state, namely, the three-dimensional center coordinates (x, y, z), the three-dimensional dimensions (length l, width w, height h), and the heading angle (theta).

[0037] These parameter values ​​are obtained by inputting the current frame point cloud into a pre-trained deep neural network (such as PointPillars or CenterPoint architecture), where the backbone network extracts voxel features, the region proposal network generates candidate boxes, and the regression head outputs the final bounding box parameters. For example, in a certain detection, the network outputs the center coordinates of a vehicle as (28.52, 16.78, 0.15) meters, its size as (4.5, 1.8, 1.5) meters, and its heading angle as 86.3 degrees. This step aims to extract the benchmark object and its key attributes for subsequent verification from the raw perception data, providing a data foundation for uncertainty quantification.

[0038] Step S12: Obtain the uncertainty estimate of the detection parameter value for each vehicle.

[0039] In the embodiments of this application, the uncertainty estimate includes, but is not limited to, two components: cognitive uncertainty estimate and accidental uncertainty estimate.

[0040] The estimation of cognitive uncertainty is mainly obtained through the Monte Carlo Dropout (MC-Dropout) method. Please refer to the following for details. Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the vehicle positioning method provided in this application.

[0041] like Figure 2 As shown, the vehicle positioning method provided in this application includes the following steps: Step S21: Input the current frame field endpoint cloud data into the vehicle detection network, perform several forward propagations, and obtain several sets of vehicle detection parameter values.

[0042] In this embodiment, the vehicle detection network can refer to a target detection model that keeps the Dropout layer active during the inference phase, used for feature extraction and bounding box regression of the input LiDAR point cloud data.

[0043] Several forward propagations can refer to using the Monte Carlo Dropout (MC-Dropout) method to repeatedly perform the forward inference process N times on the same frame field endpoint cloud data, where N is a preset positive integer, typically ranging from 30 to 50 times.

[0044] Because the Dropout layer randomly discards some neuron connections during each forward propagation, it causes subtle differences in the network structure at the microscopic level. This results in the same input data outputting N slightly different sets of vehicle detection parameter values ​​after N propagations. These N sets of parameter values ​​constitute the sample distribution of the target class vehicle detection parameters, and their dispersion directly reflects the detection model's understanding and grasp of the current input scene.

[0045] For example, when the vehicle is in a severely occluded or sparse point cloud region, the center coordinates of the bounding box output from N forward propagations fluctuate significantly; conversely, in scenarios with a wide field of view and distinct features, the N sets of output values ​​are highly concentrated. This multiple sampling mechanism transforms the knowledge blind spots within the model into quantifiable data distribution features, allowing it to be embedded into existing deep learning detection frameworks without additional modifications to the network's main structure.

[0046] Step S22: Obtain the estimated mean of several vehicle detection parameter values ​​for the target class of vehicles.

[0047] In this embodiment, the vehicle detection parameters can refer to specific physical quantities whose uncertainty needs to be evaluated, including the center point coordinates (x, y, z) of the three-dimensional bounding box, size parameters (length l, width w, height h), and heading angle (theta). The estimated mean can refer to the statistical expectation value obtained by arithmetically averaging the N sets of vehicle detection parameter values ​​obtained above. Specifically, for the i-th vehicle detection parameter... Its estimated mean The calculation formula is:

[0048] in This represents the estimated value of the i-th parameter obtained during the j-th forward propagation.

[0049] This estimated mean represents the most likely prediction of the vehicle's state given the current level of model understanding, serving as a baseline for subsequent calculations of dispersion. For example, if 30 forward propagations are performed on the longitudinal position parameters of a vehicle, resulting in 30 different longitudinal coordinate values, summing these 30 values ​​and dividing by 30 yields the estimated mean of the vehicle's longitudinal position. Obtaining this mean eliminates random noise interference from single inferences, providing a central reference for measuring the stability of the model's output.

[0050] Step S23: Determine the estimated variance measure based on the vehicle detection parameter values ​​of the target class vehicle detection parameters and the estimated mean.

[0051] Among them, the estimated variance measure can be a statistical indicator used to quantify the fluctuation range of N groups of vehicle detection parameter values ​​around their estimated mean, specifically calculated using sample variance.

[0052] This metric is obtained by calculating the sum of squares of the differences between the parameter values ​​obtained in each forward propagation and the estimated mean, and then dividing by the number of propagations N (or N-1). Mathematically, it is a measure of the estimated variance of the i-th parameter. Calculated as .

[0053] This variance metric directly characterizes the magnitude of cognitive uncertainty: the smaller the variance value, the more consistent the N prediction results are, and the more certain the model's perception of the current data is; the larger the variance value, the more significant the differences in the model's output under different random masks, indicating that the current input data may deviate from the training data distribution, and the model is in an uncertain state.

[0054] For example, for the heading angle parameter, if the distribution range of the 30 predicted values ​​is only within ±0.5 degrees, the calculated variance is extremely small; if the distribution range is as high as ±10 degrees, the variance increases significantly. This process transforms the abstract problem of model confidence into a specific numerical calculation problem, making the uncertainties between different parameters comparable.

[0055] Step S24: Use the estimated variance measure as an estimate of the uncertainty of the target class vehicle detection parameters.

[0056] In this embodiment, the uncertainty estimate specifically refers to the cognitive uncertainty estimate, whose value is directly equal to the estimated variance measure calculated in the previous step. By directly assigning the estimated variance measure to the uncertainty estimate, the sensitivity of the model to differences in the distribution of input data is explicitly quantified. This uncertainty estimate will then be used for subsequent true value quality grading, serving as one of the core criteria for distinguishing between high-confidence, medium-confidence, or low-confidence true values.

[0057] For example, when the estimated variance of a calculated vehicle length parameter is less than a preset cognitive uncertainty threshold, the detection result corresponding to that parameter is considered highly reliable; conversely, if the variance exceeds a multiple of the threshold, it is marked as low reliable, indicating to the system that the detection result may have a large cognitive bias.

[0058] In this way, this step completes a closed loop from the original point cloud input to the final uncertainty quantification index, providing a reliable truth quality assessment basis for downstream positioning verification and compensation strategies.

[0059] This application also provides a specific implementation method for obtaining uncertainty estimates for each vehicle detection parameter value. This method focuses on addressing the unreliability of observations introduced by physical factors such as lidar ranging noise, target occlusion, and reflectivity variations, and quantifies accidental uncertainties end-to-end through a deep learning network.

[0060] Please continue reading for details. Figure 3 , Figure 3 This is a flowchart illustrating the third embodiment of the vehicle positioning method provided in this application.

[0061] like Figure 3 As shown, the vehicle positioning method provided in this application includes the following steps: Step S31: Input the current frame field endpoint cloud data into the vehicle detection network to obtain the vehicle detection parameter values ​​and their logarithmic variance.

[0062] In this embodiment, the vehicle detection network is an improvement on the basic target detection architecture (such as PointPillars or CenterPoint), with an additional uncertainty prediction branch added to its network head.

[0063] The uncertainty prediction branch is set up in parallel with the original bounding box regression branch, sharing the high-dimensional feature map output by the underlying feature extraction backbone network. After the current frame field endpoint cloud data is encoded by the feature extraction backbone network, it is sent to the two branches for decoding: the bounding box regression branch outputs the determined vehicle detection parameter values, including physical quantities in seven dimensions such as three-dimensional center coordinates, length, width, height, and heading angle; at the same time, the uncertainty prediction branch synchronously outputs the logarithmic variance corresponding to the above seven dimensions.

[0064] This application uses logarithmic variance instead of directly outputting variance in order to maintain numerical stability during training and prevent the variance prediction value from approaching zero, causing gradient vanishing or diverging due to excessively large values.

[0065] For example, for the longitudinal position parameters of a certain vehicle being detected, the network might output a predicted value of 28.52 meters, with a log-variance of -4.8. This log-variance characterizes the network's confidence assessment of the degree to which the specific observation is affected by sensor noise. Through this dual-branch structure, the network can simultaneously complete target localization and reliability assessment in a single forward propagation, achieving joint learning of predicted values ​​and uncertainty indicators.

[0066] Step S32: Determine the estimated random uncertainty of the vehicle detection parameter value based on the logarithmic variance.

[0067] In this embodiment of the application, the estimate of random uncertainty is obtained by performing an exponential transformation operation on the logarithmic variance obtained above.

[0068] Specifically, the logarithmic variance of the network output is used as the input to an exponential function to calculate its natural exponent, thereby restoring the probability distribution parameter in the logarithmic domain to its variance value in physical dimensions, i.e., the estimate of random uncertainty. This estimate of random uncertainty reflects the inherent random noise in the observation process, and its magnitude directly depends on the physical characteristics of the lidar (such as angular resolution and ranging accuracy) and environmental factors (such as rain and fog attenuation, dynamic occlusion). For example, if the logarithmic variance of a certain parameter obtained above is -4.8, then by calculating exp(-4.8), the estimated random uncertainty of this parameter can be approximately 0.0082 square meters (or the square of the corresponding unit). The smaller this value, the more reliable the observation; the larger it is, the more severe the noise interference.

[0069] This quantification result serves directly as the basis for subsequent true value quality grading and weighted loss calculation, avoiding the subjectivity and lag of manually setting a fixed noise covariance in traditional methods. This allows the system to dynamically adjust the degree of trust in the field detection results based on the real-time scenario.

[0070] Furthermore, the vehicle detection network used for calculating the aforementioned accidental uncertainty estimate can be trained as follows: A training dataset is obtained, comprising multiple frames of LiDAR point cloud data and corresponding ground truth labels for target bounding boxes; a target vehicle detection network is constructed, including a feature extraction backbone network, a bounding box regression branch, and an uncertainty prediction branch; the LiDAR point cloud data is input into the target vehicle detection network, the bounding box regression branch outputs a predicted value, and the uncertainty prediction branch simultaneously outputs the corresponding logarithmic variance; an uncertainty-weighted loss value is determined based on the predicted value, the ground truth label, and the logarithmic variance; the uncertainty-weighted loss value is used to train the target vehicle detection network to obtain the vehicle detection network.

[0071] Specifically, the target vehicle detection network is an end-to-end deep learning architecture with a specially designed dual-output structure. The feature extraction backbone, typically employing a sparse convolutional neural network (such as the backbone of VoxelNet or PointPillars), transforms unstructured LiDAR point clouds into high-dimensional feature maps to capture the vehicle's geometry and spatial context. The bounding box regression branch, connected after the backbone, is responsible for regressing the vehicle's deterministic position and pose parameters based on the extracted features. Crucially, a new uncertainty prediction branch is added. This branch runs in parallel with the bounding box regression branch, sharing underlying features but possessing an independent output layer, specifically designed to predict the log-variance of each bounding box parameter.

[0072] For example, if the bounding box regression branch outputs a 7-dimensional vector representing the vehicle state, the uncertainty prediction branch simultaneously outputs another 7-dimensional vector, corresponding to the logarithmic variance of the aforementioned 7 parameters. This dual-branch structure enables the network to simultaneously infer what the vehicle is (detection value) and how accurate it is (uncertainty) in a single forward propagation, avoiding the inefficiency of traditional methods that require multiple samplings or post-processing to estimate uncertainty.

[0073] The uncertainty-weighted loss value can be a core indicator guiding network parameter updates, and its construction logic deeply integrates prediction error and uncertainty estimation. The calculation formula of the loss function usually contains two terms: the first term is the squared residual term divided by the variance weighting term, and the second term is the regularization term of the log-variance.

[0074] Specifically, when the predicted value of a parameter deviates significantly from the true label, if the network's predicted log-variance is also large (i.e., acknowledging its uncertainty), the denominator in the first term increases, automatically reducing the loss weight of that sample on that parameter and preventing noisy samples from dominating the gradient update direction. Conversely, if the network shows overconfidence in incorrect predictions (small variance), it will suffer a huge loss penalty. The second term, log-variance, prevents the network from predicting the variance of all parameters as infinite in order to minimize the total loss, forcing the network to find a balance between acknowledging uncertainty and pursuing accuracy.

[0075] This application embodiment performs iterative parameter optimization using the backpropagation algorithm. The calculated uncertainty-weighted loss value is used to calculate the gradient of all learnable parameters in the network, and the optimizer (such as Adam or SGD) adjusts the weights of the feature extraction backbone, bounding box regression branch, and uncertainty prediction branch according to the gradient direction.

[0076] This application achieves a technical upgrade from deterministic truth assumptions to uncertainty measurement plus confidence grading by introducing a dual decomposition mechanism of cognitive uncertainty estimation and random uncertainty estimation. With the help of cognitive uncertainty estimation, the system can identify the decrease in model confidence caused by scene distribution shifts, preventing blind trust in detection results when the model is unaware of the issue. Simultaneously, using random uncertainty estimation, the system can quantify the inherent observation noise level of the sensor, distinguishing between high-quality low-noise data and high-noise data.

[0077] Based on the aforementioned dual uncertainty measure, this application classifies the quality of each frame's detection result into the following three levels: When the estimated cognitive uncertainty of all vehicle detection parameter values ​​is less than the cognitive uncertainty threshold, and the estimated random uncertainty is less than the random uncertainty threshold, the field vehicle detection result is marked as a high-confidence true value.

[0078] When there is a cognitive uncertainty estimate of at least one vehicle detection parameter value that is greater than the cognitive uncertainty threshold, and the cognitive uncertainty estimate of all vehicle detection parameter values ​​is less than a preset multiple of the cognitive uncertainty threshold, or when there is a random uncertainty estimate of at least one vehicle detection parameter value that is greater than the random uncertainty threshold, and the cognitive uncertainty estimate of all vehicle detection parameter values ​​is less than a preset multiple of the random uncertainty threshold, the field vehicle detection result is marked as a medium confidence true value.

[0079] When the estimated cognitive uncertainty of at least one vehicle detection parameter value is greater than a preset multiple of the cognitive uncertainty threshold, or when the estimated random uncertainty of at least one vehicle detection parameter value is greater than a preset multiple of the random uncertainty threshold, the field vehicle detection result is marked as a low-confidence true value.

[0080] This application constructs a truth value quality grading mechanism based on double uncertainty decomposition, realizing the mapping from abstract uncertain numerical values ​​to specific confidence levels (high / medium / low). A rigorous logical complementarity is formed among high-confidence, medium-confidence, and low-confidence truth values: high-confidence truth values ​​provide an unbiased benchmark reference, ensuring the system's positioning accuracy under ideal conditions; medium-confidence truth values, through the introduction of weight decay and penalty terms, enable the system to still utilize effective information for appropriate compensation when some parameters are disturbed, avoiding the complete discarding of useful data due to local noise; the identification of low-confidence truth values ​​acts as a safety valve, forcing the system to switch to backup schemes such as extrapolation prediction when truth value quality deteriorates, preventing a chain of errors caused by poor-quality truth values.

[0081] Based on this, the two uncertainty components work together to construct the comprehensive uncertainty metric, enabling the subsequent positioning compensation strategy to adaptively adjust according to the source of uncertainty (whether it is a model problem or a sensor problem) and its magnitude: when cognitive uncertainty dominates, it tends to trigger model-level verification or downgrading; when accidental uncertainty dominates, it tends to adjust the observation noise parameters in the filtering algorithm.

[0082] This synergy effectively solves the misjudgment problem caused by ignoring the differences in the sources of uncertainty in existing technologies. It reduces the false alarm rate of misjudging normal positioning as abnormal and the false detection rate of missing real drift as normal, significantly improving the robustness and security of vehicle positioning systems in indoor GNSS denial scenarios.

[0083] Step S13: Obtain the time deviation between the current frame vehicle-end positioning time and the current frame field-end true value time.

[0084] In this embodiment, the time deviation can refer to the time difference between the time when the vehicle-side positioning calculation is completed and the time when the field-side lidar data acquisition or target calculation is completed. Since the field-side equipment relies on the NTP protocol or local crystal oscillator, while the vehicle-side positioning module is based on GNSS timing or local IMU clock, there is an inherent clock asynchrony problem between the two.

[0085] This time deviation is calculated by timestamping each key node using a high-precision time synchronization protocol (such as IEEE 1588PTP), specifically including the field-end lidar acquisition time, the field-end target calculation time, and the vehicle-end positioning calculation time. The system calculates Delta_t = |T_veh_loc - T_lidar_proc| by comparing the vehicle-end positioning time T_veh_loc with the field-end true value time T_lidar_proc.

[0086] For example, if the vehicle-side positioning time is 10:00:00.150s and the field-side true value time is 10:00:00.132s, then the time deviation is 18ms. This step aims to quantify the degree of spatiotemporal asynchrony, because a large time deviation means that directly using field-side data to verify vehicle-side data will introduce position errors caused by vehicle movement, and this deviation must be compensated or extrapolated.

[0087] Step S14: Based on the uncertainty estimate of each vehicle detection parameter value, determine the comprehensive uncertainty measure of the field vehicle detection result.

[0088] In this embodiment, the comprehensive uncertainty metric is a scalar value used to uniformly characterize the overall reliability level of vehicle detection results in the current frame. This metric is generated by normalizing and weighting the estimated cognitive uncertainty and accidental uncertainty of each vehicle detection parameter value obtained above.

[0089] Specifically, the comprehensive uncertainty measure is generated by weighted summation of the following items: normalized cognitive uncertainty index, normalized accidental uncertainty index, normalized extrapolation uncertainty index, and comprehensive verification compensation term; wherein, each of the normalized indices is obtained by standardizing the historical statistics of the corresponding uncertainty component.

[0090] The normalized cognitive uncertainty index is a dimensionless value obtained by standardizing the estimated cognitive uncertainty of each vehicle detection parameter value obtained above.

[0091] Specifically, the system first calculates the mean and standard deviation of the cognitive uncertainty estimates for all frames within the historical time window. Then, it subtracts the mean from the current frame's cognitive uncertainty estimate and divides it by the standard deviation to obtain the normalized cognitive uncertainty index. This index characterizes the deviation of the detection model's knowledge blind spot regarding the data distribution in the current scenario from the historical average level.

[0092] For example, if the current frame has sparse point clouds due to severe vehicle occlusion, its original cognitive uncertainty variance is 0.05m², while the historical statistical mean is 0.01m² and the standard deviation is 0.008m². The calculated normalized index is approximately 6.25, which is significantly higher than the normal level, indicating that the model does not have sufficient understanding of the detection result.

[0093] The normalized random uncertainty index is obtained by applying the same standardized logic to the estimated random uncertainty of each vehicle detection parameter value. This index reflects the relative magnitude of inherent random noise (such as sensor ranging noise and changes in environmental reflectivity) during the observation process.

[0094] The normalized extrapolation uncertainty index is based on the uncertainty estimate generated during the extrapolation prediction process (e.g., obtained by integrating the output dispersion of multiple prediction models), and is standardized using historical statistics. It is used to quantify the reliability of trajectory extrapolation results. The value of this index increases accordingly when the extrapolation time span increases or when vehicle movement is drastic.

[0095] The three normalization indices mentioned above eliminate the differences in numerical scale between different physical dimensions (such as position variance m² and angle variance deg²) and uncertainties from different sources, enabling them to be uniformly incorporated into the subsequent calculation framework.

[0096] The comprehensive verification compensation term is a correction factor used to adjust the comprehensive uncertainty measure, and its value is negatively correlated with the consistency of the multidimensional verification. Specifically, this term is calculated by weighting the confidence levels of time deviation, spatial deviation, velocity deviation, and heading angle deviation, as well as the true value uncertainty penalty term. When the true value quality level is high and the deviations in each dimension are small, the comprehensive verification compensation term tends to be close to zero or a small positive value; conversely, when the true value is marked as low confidence or the deviations in each dimension are large, the value of this term increases, thereby improving the final comprehensive uncertainty measure.

[0097] Specifically, when the extrapolation prediction based on the learning model is not executed, the normalized extrapolation uncertainty index is zero. The non-execution of the extrapolation prediction based on the learning model can mean that the system determines, within the current frame positioning verification period, that there is no need to activate the trajectory extrapolation compensation mechanism. Specifically, when the time deviation between the vehicle-side positioning time and the field-side ground truth time is less than or equal to a preset time deviation threshold, and the field-side vehicle detection result is marked as a high-confidence ground truth, the system directly uses the field-side detection result as the verification benchmark, and at this time, the first extrapolation prediction model is not activated for state sequence prediction.

[0098] In this step, the index value before extrapolation is performed is forcibly set to zero, meaning that this contribution is completely eliminated in the weighted summation calculation of the overall uncertainty measure. For example, in an open area of ​​an underground parking garage, if a vehicle is traveling in a straight line at a constant speed, the time synchronization accuracy is better than 10ms, and the field radar detects the vehicle without obstruction, the system determines that no extrapolation is needed. In this case, regardless of the value of the extrapolation uncertainty statistic in the historical data, the normalized extrapolation uncertainty index of the current frame is assigned a value of 0.0. Through this zero-value filling mechanism, it is possible to avoid a false increase in the overall uncertainty measure due to the reference of invalid data or historical residual data when no extrapolation result is generated, thereby preventing the system from mistakenly triggering high-level security degradation strategies.

[0099] The comprehensive verification confidence index corresponding to the above comprehensive verification compensation item is generated by weighted summation of the following items: time deviation confidence item, spatial deviation confidence item, speed deviation confidence item, heading angle deviation confidence item, and true value uncertainty penalty item; among which, the true value uncertainty penalty item is determined according to the proportion of parameters with different confidence levels in the field vehicle detection results. The higher the proportion of high confidence level parameters, the closer the true value uncertainty penalty item is to the upper limit value.

[0100] The time deviation confidence term is determined based on the time deviation between the current frame vehicle-end positioning time and the field-end ground truth time. Specifically, the system calculates the absolute value of the time difference between the two and maps it to a confidence score. When the time deviation is less than a preset time synchronization threshold (e.g., 20ms), the value of this term is close to the upper limit of 1.0, indicating good spatiotemporal alignment. As the time deviation increases, the value of this term decreases non-linearly to reflect the decrease in verification reliability caused by clock asynchrony.

[0101] The spatial deviation confidence term is determined based on the Euclidean distance between the corrected field-end predicted trajectory and the actual vehicle-end positioning result. For example, a high confidence score is assigned when the spatial deviation is less than 0.1 meters; if the deviation exceeds 0.5 meters, the term approaches zero.

[0102] The confidence terms for speed deviation and heading angle deviation are quantified based on the degree of difference between the speed vector and heading angle reported by the vehicle and the values ​​calculated by the field. Their function is to evaluate the consistency of vehicle positioning from the perspective of motion state.

[0103] The above four confidence items together constitute the basic score for verification consistency, reflecting the degree of agreement between the vehicle-side data and the field-side benchmark in terms of time series, location, and kinematic characteristics.

[0104] The truth value uncertainty penalty term is not a fixed constant, but dynamically depends on the confidence level statistics of the field vehicle detection results generated in the above steps.

[0105] Specifically, the system first counts the number of parameters that are high-confidence ground truth values ​​(i.e., Grade A, where both cognitive uncertainty and random uncertainty are below the threshold) among all vehicle detection parameters in the current frame (including three-dimensional coordinates x, y, z, length, width, height l, w, h, and heading angle theta), and calculates the proportion of these parameters to the total number of parameters. The truth value uncertainty penalty term Configured to be with Positive correlation.

[0106] When the time deviation is greater than the time deviation threshold, or when the vehicle detection result at the field is a low-confidence true value, the vehicle detection result at the field is updated by extrapolation prediction based on a learning model.

[0107] For example, when a radar frame is marked as a low-confidence true value due to obstruction by a pillar, the system automatically switches to extrapolation mode. It then uses a high-quality true value sequence from the past two seconds to predict the vehicle's current pose, thus filling in the gaps caused by missing or low-quality true values. This dynamic switching mechanism effectively avoids misjudgments caused by forced comparisons in cases of time asynchrony or unreliable true values, significantly improving the robustness of the positioning verification system under typical conditions such as communication jitter and sensor transient failures.

[0108] Specifically, this application provides a flowchart of extrapolation prediction and kinematic constraint correction based on a learning model. Please refer to the following for details. Figure 4 , Figure 4 This is a flowchart illustrating the fourth embodiment of the vehicle positioning method provided in this application.

[0109] like Figure 4 As shown, the vehicle positioning method provided in this application includes the following steps: Step S41: Input the field end true state sequence within the historical time window into the first extrapolation prediction model. The input features of the state sequence include lateral position error, longitudinal position error, heading angle error, yaw rate, longitudinal velocity, lateral velocity, and longitudinal acceleration.

[0110] In this embodiment, the first extrapolation prediction model can refer to a temporal prediction network built based on a deep learning architecture, such as a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU), used to capture the nonlinear dynamic features of vehicle motion. This state sequence comes from the historical records of the field vehicle detection results marked as low-confidence ground truth or having time deviations. Specifically, the length of the historical time window can be set according to the scenario requirements; for example, if it is 2 seconds and the sampling frequency is 10Hz, the input sequence contains state vectors for 20 time steps.

[0111] The lateral and longitudinal position errors in the input features reflect the degree of vehicle deviation from the reference trajectory, the heading angle error characterizes attitude deviation, while the yaw rate, longitudinal velocity, lateral velocity, and longitudinal acceleration comprehensively describe the vehicle's instantaneous motion state. These multidimensional features collectively form the basis for the model to infer the vehicle's future motion trend. For example, when a vehicle is driving on a spiral ramp in an underground parking garage, its longitudinal and lateral velocities exhibit coupled changes, accompanied by significant fluctuations in yaw rate. By inputting these seven features sequentially into the first extrapolation prediction model, the model can learn the motion patterns under such complex maneuvers, rather than simply relying on linear assumptions. This multidimensional state sequence input method provides rich contextual information for subsequent nonlinear trajectory prediction, significantly improving the model's adaptability to complex indoor scenarios.

[0112] Step S42: Obtain the predicted state sequence within the future time window output by the first extrapolation prediction model, the predicted state sequence including predicted lateral position, predicted longitudinal position, predicted heading angle and predicted speed.

[0113] In this application embodiment, the future time window can refer to a preset duration extending backward from the current moment, such as 0.5 seconds to 2 seconds, to cover potential positioning blind spots within the verification period.

[0114] The predicted state sequence is calculated through forward propagation of the first extrapolation prediction network, and it contains vehicle state estimates for each future time step. Specifically, the predicted lateral and longitudinal positions constitute the vehicle's expected trajectory points in the planar coordinate system, the predicted heading angle describes the vehicle's future orientation change, and the predicted speed reflects the vehicle's future travel speed. These outputs are derived from historical motion patterns and are intended to replace real-time ground truth values ​​that are unavailable due to time skewness or low confidence.

[0115] Taking an automated parking scenario as an example, if the current frame's field data is missing, the model can output the vehicle's expected position coordinates and heading angle within the next second based on the acceleration, deceleration, and turning data from the past 2 seconds, thus filling the ground truth gap. The predicted state sequence generated in this step provides a basic data source for subsequent physical consistency verification, ensuring that the extrapolation process not only considers data-driven trends but also retains room for further correction.

[0116] Step S43: Apply kinematic constraint correction to the predicted state sequence to obtain the corrected field vehicle detection results.

[0117] In the embodiments of this application, kinematic constraint correction can refer to the process of post-processing and correcting the prediction results of pure data-driven predictions using the physical limits of vehicle dynamics, with the aim of eliminating abnormal prediction values ​​that do not conform to physical laws.

[0118] This correction process is based on prior knowledge of the vehicle's maximum acceleration, maximum deceleration, and maximum curvature. Specifically, firstly, the velocity difference between adjacent moments in the predicted state sequence is calculated to obtain the predicted acceleration. When the predicted acceleration exceeds a preset acceleration range (e.g., maximum acceleration 2.0 m / s², maximum deceleration -4.0 m / s²), the predicted velocity is truncated to the boundary value, and the position is recalculated. Secondly, the predicted curvature is calculated based on the predicted velocity and yaw rate. When the predicted curvature exceeds a preset curvature threshold (e.g., 0.3 m), the correction is applied. - ¹ When the minimum turning radius is approximately 3.3 meters, the yaw rate is truncated to ensure that the steering action meets the vehicle's mechanical limits.

[0119] For example, if an LSTM model predicts that a vehicle accelerates from a standstill to 20 m / s (an acceleration of 200 m / s²) in 0.1 seconds due to noise in the training data, kinematic constraint correction will forcibly limit this speed to a physically permissible increment, avoiding absurd trajectory jumps. By combining the first extrapolation prediction model with kinematic constraint correction, the powerful fitting ability of deep learning to nonlinear motion patterns is utilized, while unreasonable prediction results are filtered out through physical rules. This results in highly reliable corrected on-site vehicle detection results, effectively solving the problem of rapid error accumulation in traditional uniform-speed extrapolation models under severe maneuvering scenarios.

[0120] Furthermore, this application also provides a scheme for performing hierarchical prediction and compensation triggering based on prediction deviation. After obtaining the corrected on-site vehicle detection results, predictive anomaly detection and hierarchical compensation triggering steps are performed.

[0121] Specifically, this application calculates the prediction deviation based on the corrected on-site vehicle detection results and the vehicle-end vehicle detection results. The prediction deviation can refer to the spatial difference between the corrected on-site predicted trajectory and the real-time vehicle-end positioning results, and is used to characterize the degree of consistency between the vehicle-end positioning and the on-site predicted trend.

[0122] This application performs hierarchical prediction based on the relationship between the prediction deviation and the preset deviation threshold: when the prediction deviation is less than the first deviation threshold, it is determined to be in a normal state; when the prediction deviation is between the first deviation threshold and the second deviation threshold, it is determined to be in a potential degradation state and triggers the first level of compensation; when the prediction deviation is greater than or equal to the second deviation threshold, it is determined to be in a state of impending deviation and triggers the second level of compensation; wherein, the first deviation threshold and the second deviation threshold are determined based on the statistical distribution of historical data.

[0123] Specifically, tiered prediction is an adaptive decision-making mechanism based on the magnitude of prediction deviation, aiming to shift safety intervention from post-event remediation to in-event early warning. The first and second deviation thresholds are not fixed engineering experience values, but are dynamically determined based on the statistical distribution of historical operating data. For example, the 70th and 95th percentiles of the prediction deviation distribution on the historical validation set can be used to adapt to noise levels in different scenarios.

[0124] When the prediction deviation is less than the first deviation threshold, it indicates that the vehicle-end positioning and the field-end prediction trend are highly consistent. The system determines this as a normal state, maintains the current positioning trust weight, and does not trigger additional compensation actions. When the prediction deviation is between the first and second deviation thresholds, it indicates that the vehicle-end positioning may experience slight drift or enter a feature degradation region (such as the initial stage of pillar occlusion). The system determines this as a potential degradation state and triggers the first level of compensation. This compensation strategy typically includes reducing the trust weight of the vehicle-end positioning, enabling lightweight coordinate system fine-tuning, or increasing the state monitoring frequency to prevent the deviation from further expanding.

[0125] When the prediction deviation is greater than or equal to the second deviation threshold, it indicates that the vehicle positioning is about to deviate significantly and there is a risk of loss of control. The system determines that the deviation is imminent and immediately triggers the second level of compensation. This compensation strategy usually involves using a more aggressive fusion algorithm (such as extended Kalman filter to fuse the field end true value) or directly switching to a backup positioning source to avoid safety degradation caused by the accumulation of deviation.

[0126] Furthermore, this application also provides a scheme for dynamically adjusting the spatial deviation threshold based on an extrapolation uncertainty index.

[0127] In this application, the spatial deviation is calculated based on the corrected predicted trajectory and the vehicle-side positioning data. The spatial deviation can refer to the Euclidean distance or vector difference between the corrected field-side predicted trajectory and the real-time vehicle-side positioning data in a unified coordinate system, used to quantify the degree of inconsistency between the two in spatial position. The corrected predicted trajectory is obtained through the aforementioned kinematic constraint correction process, eliminating abnormal predicted points that do not conform to the vehicle's physical motion characteristics, ensuring the rationality of the trajectory used as a comparison benchmark in terms of acceleration and curvature. The vehicle-side positioning data originates from the vehicle's pose calculated at the current moment by the onboard positioning module (such as a laser SLAM, UWB, or IMU integrated navigation system).

[0128] For example, when the corrected predicted trajectory has a longitudinal position of 28.64 meters and a lateral position of 16.79 meters, while the vehicle-side positioning data shows a longitudinal position of 28.67 meters and a lateral position of 16.92 meters, the calculated spatial deviation is approximately 0.13 meters. This spatial deviation directly reflects the deviation of the vehicle-side positioning from the high-confidence field prediction and is a core input variable for subsequently determining whether the positioning status is normal.

[0129] Then, this application obtains the extrapolation uncertainty index corresponding to the extrapolation prediction.

[0130] The extrapolation uncertainty index is a quantitative value used to characterize the credibility of extrapolation prediction results. It is derived from the dispersion statistics of the output results of multiple independent extrapolation prediction models based on the ensemble learning strategy.

[0131] Specifically, this application constructs multiple first extrapolation prediction models (such as LSTM networks) with identical structures but different initialization parameters. The ground truth state sequences within the same historical time window are input into each model to obtain the independently output predicted state sequences for future time windows. The extrapolation uncertainty index is obtained by weighted summation of the variance (or standard deviation) of these independent prediction results with a pre-set model intrinsic variance. This index reflects the range of prediction fluctuations caused by the randomness of model parameter initialization and differences in training data distribution. For example, if five independent LSTM models are used to predict the position of the same vehicle 0.5 seconds in the future, the predicted longitudinal positions are 28.60m, 28.65m, 28.63m, 28.70m, and 28.62m, respectively. Their sample variance is an important component of the extrapolation uncertainty index. If the scenario is complex, leading to significant differences in the prediction results among the models (e.g., a variance as high as 0.05m²), it indicates a high extrapolation uncertainty index, meaning the reliability of the predicted trajectory at the current moment is low. The purpose of this indicator is to provide a basis for the dynamic adjustment of the spatial deviation threshold, so as to avoid making incorrect positioning anomaly judgments when the prediction itself is unreliable.

[0132] Finally, this application dynamically adjusts the spatial deviation threshold based on the extrapolation uncertainty index, wherein the spatial deviation threshold increases as the extrapolation uncertainty index increases.

[0133] Dynamic adjustment can refer to establishing a positive correlation between extrapolation uncertainty indicators and spatial deviation thresholds, so that the tolerance for determining whether vehicle positioning is abnormal can adaptively fluctuate with changes in prediction reliability.

[0134] When the extrapolation uncertainty index increases, it indicates that the extrapolated predicted trajectory is greatly affected by noise or model blind spots. At this time, the spatial deviation threshold is automatically increased to relax the tolerance limit for vehicle-end positioning deviation and prevent false alarms of vehicle-end positioning failure due to inaccurate reference. Conversely, when the extrapolation uncertainty index decreases, it indicates that the predicted trajectory is highly reliable. At this time, the spatial deviation threshold is tightened to improve the detection sensitivity of small drifts in vehicle-end positioning.

[0135] Furthermore, this application also provides a scheme for determining an extrapolation uncertainty index. The method further includes constructing multiple first extrapolation prediction models with the same structure but different initialization parameters, obtaining independent prediction results of each first extrapolation prediction model for the same input, and determining the extrapolation uncertainty index based on the dispersion of each independent prediction result and the preset model inherent variance.

[0136] Please continue reading for details. Figure 5 , Figure 5 This is a flowchart illustrating the fifth embodiment of the vehicle positioning method provided in this application.

[0137] like Figure 5 As shown, the vehicle positioning method provided in this application includes the following steps: Step S51: Construct multiple first extrapolation prediction models with the same structure but different initialization parameters.

[0138] In this embodiment, M identical first extrapolation prediction models are constructed. For example, each model contains two hidden layers with 64 hidden units per layer. The input feature dimension is 7 (including lateral position error, longitudinal position error, heading angle error, yaw rate, longitudinal velocity, lateral velocity, and longitudinal acceleration), and the output feature dimension is 4 (predicted lateral position, predicted longitudinal position, predicted heading angle, and predicted velocity). The difference between these M models lies in the initialization parameters of their network weights. For example, different random seeds are used for Xavier or He initialization, or different data augmentation perturbations are introduced during training, enabling each model to learn different aspects of the data distribution when faced with the same input.

[0139] Step S52: Obtain the independent prediction results of each of the first extrapolation prediction models for the same input.

[0140] In this embodiment, the independent prediction result can refer to the predicted state sequence within the future time window output by each of the M first extrapolation prediction models after simultaneously inputting the field-end true state sequence within the same historical time window into the aforementioned M models. This step is executed by the inference engine in a cloud server or edge computing node.

[0141] Step S53: Determine the extrapolation uncertainty index based on the dispersion of each independent prediction result and the preset model inherent variance.

[0142] In this embodiment, the extrapolation uncertainty index is a numerical measure that quantifies the reliability of the extrapolation prediction results. It is generated by fusing the statistical dispersion of the prediction results with the model's prior intrinsic variance. The dispersion reflects the consistency of the model set under the current input scenario. The larger the dispersion, the greater the divergence between models and the more unstable the prediction. The model's intrinsic variance is the baseline noise level obtained through the validation set statistics, which characterizes the lower bound of the average uncertainty of the model architecture itself.

[0143] For example, if the predicted values ​​of five models for a certain time position are [28.64, 28.66, 28.63, 28.65, 28.70], and their mean is 28.656, the calculated variance dispersion is 0.0006m²; if the preset model inherent variance is 0.0004m², then the determined extrapolation uncertainty index is 0.0010m².

[0144] Step S15: Determine the current frame positioning status classification result based on the comprehensive uncertainty metric and the time deviation.

[0145] In this embodiment, the positioning status classification result divides the current positioning verification environment into several discrete status levels, each corresponding to a different risk level and processing strategy. This classification result is determined jointly based on the comprehensive uncertainty measure and the time deviation obtained above, combined with preset classification threshold conditions.

[0146] The grading logic includes, but is not limited to: when the time deviation is less than the first time threshold and the comprehensive uncertainty measure is less than the first uncertainty threshold, it is judged as Level 1 (normal state); when the time deviation meets the second condition or the comprehensive uncertainty measure is in the second range, it is judged as Level 2 (slight deviation state); when the comprehensive uncertainty measure exceeds the third threshold or the time deviation continues to exceed the limit, it is judged as Level 3 (moderate drift state); when the comprehensive uncertainty measure is extremely high, continuous verification fails, or communication is interrupted, it is judged as Level 4 (positioning failure state).

[0147] For example, if the calculated U_total is 0.35 and the time deviation is 18ms, both are below the set slight deviation threshold, then the current level is determined to be the second level. This step aims to establish a mapping relationship between uncertainty measurement and hierarchical triggering, enabling the hierarchical strategy to dynamically and adaptively adjust according to the uncertainty level of the scene and the spatiotemporal synchronization state, avoiding misjudgments or omissions caused by fixed thresholds.

[0148] The positioning status classification results provided in this application include: Level 1, Level 2, Level 3, and Level 4.

[0149] The triggering conditions for the first level include: the time deviation meets the first condition, the spatial deviation meets the second condition, and the comprehensive uncertainty measure meets the third condition.

[0150] The first level represents the positioning system in a normal, high-availability state. Its triggering logic uses a combination of logic and requires that the time, space, and uncertainty dimensions simultaneously meet optimal constraints. Time deviation can refer to the time difference between the current frame's vehicle-end positioning time and the field-end true value time. The first condition is that the time deviation is less than or equal to a preset time synchronization threshold (e.g., 20ms) to ensure the effectiveness of spatiotemporal alignment. Spatial deviation can refer to the Euclidean distance between the field-end predicted trajectory after kinematic constraint correction and the vehicle-end positioning result. The second condition is that the spatial deviation is less than or equal to the first-level spatial deviation threshold (e.g., 0.1m), indicating that the vehicle-end positioning has not experienced significant drift. The comprehensive uncertainty measure is a normalized index generated by weighting cognitive uncertainty, accidental uncertainty, and extrapolation uncertainty. The third condition is that this measure value is lower than the first-level uncertainty threshold (e.g., 0.2), indicating that the field-end true value is reliable and the extrapolation process has high confidence.

[0151] The system will only enter the first level when all three conditions are met. At this time, no compensation operation is required, and the vehicle positioning result will be directly accepted.

[0152] The triggering conditions for the second level include: the spatial deviation meets the fourth condition or the comprehensive uncertainty measure meets the fifth condition.

[0153] The second level represents a slight deviation or increased local uncertainty in the positioning system. Its triggering logic uses an OR logic to identify potential risks in advance. The fourth condition is set as a spatial deviation between the first and second level spatial deviation thresholds (e.g., 0.1m to 0.3m), indicating a slight drift in vehicle-end positioning that is not yet out of control. The fifth condition is set as a comprehensive uncertainty measure between the first and second level uncertainty thresholds (e.g., 0.2 to 0.5), indicating that although the spatial deviation may still be within acceptable limits, the cognitive uncertainty of the field-end true value or the random uncertainty caused by long-term extrapolation has significantly increased. If either the spatial deviation exceeds the limit or the uncertainty increases, the system is determined to enter the second level. At this point, the system will activate a lightly weighted compensation strategy based on LSTM-predicted trajectories to suppress the drift trend.

[0154] The triggering conditions for the third level include: the spatial deviation meets the sixth condition, the comprehensive uncertainty measure meets the seventh condition, or the prediction deviation exceeds the warning threshold for a number of periods that reach a preset value.

[0155] The third level indicates that the positioning system is in a state of moderate drift or instability, requiring stronger fusion intervention. The sixth condition is set as spatial deviation between the second and third level spatial deviation thresholds (e.g., 0.3m to 0.5m), indicating that the vehicle-end positioning error has affected the driving safety boundary. The seventh condition is set as comprehensive uncertainty measurement between the second and third level uncertainty thresholds (e.g., 0.5 to 0.8), indicating a decline in true value quality or a decrease in the reliability of the extrapolation model. Furthermore, a persistence criterion in the time dimension is introduced: the prediction deviation (the difference between the vehicle-end positioning and the extrapolated prediction) exceeds the dynamic warning threshold for N consecutive cycles (e.g., 3-5 cycles). This is used to filter occasional measurement noise interference and prevent false triggering of high-level compensation due to single-frame anomalies. When any of the above spatial, uncertainty, or persistence deviation conditions is met, the system is determined to enter the third level, and will switch to the Extended Kalman Filter (EKF) fusion mode, dynamically adjusting the observation noise covariance based on the true value confidence.

[0156] The triggering conditions for the fourth level include: spatial deviation meeting the eighth condition, comprehensive verification confidence level being lower than the preset confidence level lower limit, comprehensive uncertainty measurement meeting the ninth condition, and the number of consecutive verification failure frames reaching a preset value or the communication interruption duration reaching a preset value.

[0157] The fourth level represents a positioning system failure or a serious security risk, requiring immediate security downgrade. The eighth condition is set as a spatial deviation greater than the third-level spatial deviation threshold (e.g., >0.5m), indicating that the vehicle-end positioning has completely deviated from the true trajectory. The comprehensive verification confidence is a comprehensive index generated by fusing time, space, speed, and truth value uncertainty penalties. When it falls below a preset confidence lower limit (e.g., 0.3), it indicates that the overall verification result is unreliable. The ninth condition is set as a comprehensive uncertainty metric higher than the third-level uncertainty threshold (e.g., >0.8), indicating that the system is in an extremely uncertain state. Simultaneously, a fault counting mechanism is set. When the number of consecutive verification failure frames reaches a preset value (e.g., 5 frames) or the communication interruption duration between the field station and the vehicle reaches a preset value (e.g., 2 seconds), regardless of other parameters, the system is forcibly judged to enter the fourth level. At this level, the system will trigger the minimum risk strategy (MRC), cutting off the cloud control channel and performing an emergency stop or slow docking.

[0158] Furthermore, this application also provides a scheme for adjusting the observation noise covariance based on the confidence level.

[0159] Specifically, when the confidence level of the field-end true value is the first confidence level, the first observation noise covariance is used.

[0160] The first confidence level refers to the state where the estimated cognitive uncertainty and random uncertainty of all bounding box parameters in the field vehicle detection results are less than the corresponding threshold, typically corresponding to the high confidence level (Grade A) defined in [the standard]. This level indicates that the target position, size, and heading angle detected by the current field-end lidar have extremely high reliability, the data distribution is highly consistent with the model training set, and the sensor observation noise is at a low level. The first observation noise covariance is a matrix parameter used in the Extended Kalman Filter (EKF) or factor graph optimization algorithm to characterize the statistical properties of measurement noise; its magnitude directly determines the filter's confidence weight for the field observation data.

[0161] Specifically, when the true value at the field end is determined to be at the first confidence level, the system sets the first observation noise covariance to a smaller value (denoted as ). For example, a position variance of 0.01 m² and an angle variance of 0.001 rad² can be assigned to the diagonal. Since the smaller the observation noise covariance, the larger the Kalman gain, the filter will rely more on the field observations to correct the vehicle-end predicted trajectory in the state update step, thereby making full use of high-quality global ground truth information to eliminate vehicle-end accumulated drift.

[0162] When the confidence level of the true value at the field end is the second confidence level, the second observation noise covariance is used.

[0163] The second confidence level refers to the true value state in the field vehicle detection results where the cognitive or accidental uncertainty of some bounding box parameters exceeds the basic threshold but does not reach twice the threshold, or where the quality of local parameters deteriorates due to occlusion or changes in reflectivity. This typically corresponds to the medium confidence true value (Grade B) defined in [the standard framework]. This level indicates that while the true value is usable, its accuracy is subject to fluctuations due to environmental dynamics or limitations in model generalization ability. The second observation noise covariance is a measurement noise parameter set for this type of medium-quality true value, and its value is significantly greater than the first observation noise covariance (denoted as [missing information]). ,and For example, the location variance can be adjusted to 0.05. Up to 0.1 The angle variance was adjusted to 0.005. Up to 0.01 .

[0164] This application increases the observation noise covariance and automatically reduces the Kalman gain of the filtering algorithm, thereby reducing the correction amplitude of field observations during state updates. The filter then relies more on the kinematic model derivation results of the vehicle itself. This mechanism effectively avoids forcibly fusing field data containing moderate errors into the vehicle positioning solution, preventing low-quality ground truth from contaminating or diverging high-precision vehicle pose.

[0165] Step S16: Compensate the vehicle detection results at the vehicle end based on the positioning status classification results.

[0166] In the embodiments of this application, compensation may refer to adopting differentiated data processing strategies to correct errors in vehicle detection results at the vehicle end, based on different positioning status classification results.

[0167] Specific compensation methods include: if the location is at level one, the vehicle detection results at the vehicle end are used directly without additional compensation; if the location is at level two, the correction amount is calculated based on the deviation between the predicted trajectory and the vehicle end results, and the vehicle end results are weighted and corrected using dynamically adjusted weighting coefficients, with higher uncertainty resulting in lower compensation gain; if the location is at level three, a filtering fusion algorithm (such as Extended Kalman Filter EKF) is used to fuse the field vehicle detection results and the vehicle end vehicle detection results, and the observation noise covariance parameter in the filtering algorithm is dynamically adjusted according to the confidence level of the field ground truth, with lower confidence resulting in higher observation noise; if the location is at level four, a safety degradation operation is triggered, such as cutting off the cloud control channel and controlling the vehicle to enter the minimum risk state.

[0168] For example, when the location is classified as Level 2, the system calculates a correction of (-0.03, -0.13, 0.01) meters, determines a compensation gain of 0.325 based on the current comprehensive uncertainty metric, and finally generates the compensated position coordinates. This step, through an adaptive compensation mechanism, ensures both positioning accuracy and system stability and security.

[0169] Step S17: Output the vehicle positioning result based on the compensated vehicle detection result.

[0170] In this embodiment, the vehicle positioning result is the final pose data after the aforementioned compensation processing, including optimized three-dimensional coordinates, heading angle, and velocity information, which can be directly used for vehicle path planning and control execution. This result is directly output based on the compensated vehicle detection results generated above, or it is output after smoothing the original observation data from the vehicle sensors during the compensation process. The output format can be a CAN bus message sent to the vehicle controller or a data packet sent to the cloud monitoring platform.

[0171] For example, the system encapsulates the compensated coordinates (28.66, 16.88, 0.14) meters and the updated heading angle into a standard positioning message and publishes it. This step aims to deliver high-quality positioning data, which has undergone uncertainty quantification verification and differential compensation, to downstream application modules, thereby significantly improving the positioning accuracy, robustness, and long-term operational stability of the autonomous driving system in GNSS-denied scenarios.

[0172] Based on the above embodiments, this application, after completing the indoor vehicle positioning and the vehicle's autonomous driving control, continues to implement a scheme of multi-dimensional residual time series analysis and online parameter optimization.

[0173] Please continue reading for details. Figure 6 , Figure 6 This is a flowchart illustrating the sixth embodiment of the vehicle positioning method provided in this application.

[0174] like Figure 6 As shown, the vehicle positioning method provided in this application includes the following steps: Step S61: Obtain the actual position and pose data of the vehicle and the vehicle detection results of the field, and calculate the multi-dimensional residual sequence, wherein the multi-dimensional residual includes position residual, heading residual and velocity residual.

[0175] In this embodiment, the actual vehicle-side posture data refers to the physical state data achieved by the vehicle through the actual motion executed by the chassis control system after receiving the positioning compensation command. This data originates from real-time feedback from the vehicle's CAN bus or chassis domain controller and specifically includes the three-dimensional coordinates, heading angle, and linear velocity vector at the execution moment. The field-side vehicle detection result refers to the field-side perception data, corrected by the aforementioned steps and used as a global reference.

[0176] The process of calculating the multidimensional residual sequence involves performing a difference operation between the vehicle-side execution data and the field-side reference data at the same time alignment point.

[0177] Specifically, the position residual is obtained by calculating the Euclidean distance between the position vector executed by the vehicle end and the position vector detected by the field end, and is used to characterize the absolute deviation of the system in spatial positioning; the heading residual is obtained by calculating the absolute value of the difference between the heading angle executed by the vehicle end and the heading angle detected by the field end, and is used to characterize the deviation of the system in attitude pointing; the velocity residual is obtained by calculating the magnitude difference or vector difference between the velocity vector executed by the vehicle end and the velocity vector detected by the field end, and is used to characterize the deviation of the system in dynamic response.

[0178] For example, in a certain calibration cycle, the field end detects the vehicle at coordinates (28.5, 16.8) meters, with a heading of 86.5 degrees and a speed of 1.2 m / s. However, the actual position feedback from the vehicle end is (28.6, 16.9) meters, with a heading of 87.0 degrees and a speed of 1.15 m / s. The calculated position residual is approximately 0.14 meters, the heading residual is 0.5 degrees, and the speed residual is 0.05 m / s. These residual data are stored in a sliding time window in chronological order, forming a multi-dimensional residual sequence that changes over time. This sequence constitutes the basic input data for subsequent time-series statistical analysis. By constructing a multi-dimensional residual sequence including position, heading, and speed, the error characteristics of the system in various dimensions of static positioning, dynamic tracking, and attitude control can be comprehensively captured, avoiding missed detections caused by single-dimensional analysis.

[0179] Step S62: Perform time series statistical analysis on the multidimensional residual sequence, including mean drift detection, variance change detection, and autocorrelation characteristic analysis.

[0180] In this embodiment of the application, time-series statistical analysis aims to extract statistical characteristics of system performance degradation from historical residual data.

[0181] Mean drift detection uses the Cumulative Sum Control Chart (CUSUM) algorithm to identify small, persistent drifts by calculating the cumulative deviation of the residual sequence from the target mean. When the cumulative statistic exceeds a preset decision threshold, the system is determined to have mean drift, which usually corresponds to a slow change in calibration parameters (such as translation components in the coordinate transformation matrix).

[0182] Variance change detection uses the F-test method, which divides the residual sequence within the current sliding window into two sub-intervals, calculates the sample variance of each sub-interval, and constructs the F-statistic. By comparing the F-statistic with the critical value, it is determined whether the fluctuation range of the residual has changed abruptly. A significant increase in variance usually means that the sensor noise characteristics have changed or environmental interference has intensified.

[0183] Autocorrelation analysis involves calculating the autocorrelation coefficients of the residual sequence at different lag orders to detect whether there are non-random systematic patterns in the residuals. If the autocorrelation coefficients at certain lag orders exceed the confidence interval, it indicates that the residual sequence no longer satisfies the independent and identically distributed assumption, suggesting the existence of unmodeled systematic errors or model mismatch.

[0184] For example, in a residual sequence of 1000 consecutive frames, if the CUSUM statistic increases linearly from 0 to 3.5 (threshold 5) within 50 frames, it indicates a positive drift trend; if the variance of the first half is 0.01 and the variance of the second half suddenly increases to 0.05 and the F test is significant, it indicates a sudden decrease in system stability; if the autocorrelation coefficient of the lag 3 frames is as high as 0.2 while the rest are close to 0, it indicates the existence of periodic control delay or oscillation.

[0185] These three analytical methods analyze the residual sequence from three orthogonal dimensions: deviation trend, fluctuation amplitude, and sequence correlation, respectively, and together constitute a three-dimensional assessment of the system's health status.

[0186] Step S63: Generate graded early warning information based on the results of the time series statistical analysis, and execute differentiated response actions according to the graded early warning information. The response actions include adjusting calibration parameters to optimize the frequency, triggering online recalibration, or performing security degradation.

[0187] In this application embodiment, the graded early warning information is a risk level divided according to the quantitative results of mean drift detection, variance change detection and autocorrelation characteristic analysis. It typically includes a first-level early warning (minor anomaly), a second-level early warning (moderate anomaly) and a third-level early warning (severe anomaly).

[0188] When a slight mean drift trend or a single autocorrelation coefficient exceeding the limit is detected, a first-level warning message is generated. The differentiated response action executed at this time is to adjust the calibration parameter optimization frequency, for example, shortening the parameter optimization trigger interval from every 100 frames to every 20 frames, in order to speed up the correction of minor drifts.

[0189] When a significant change in variance is detected, or multiple lag order autocorrelation coefficients exceed the limit simultaneously, or when the first-level warning persists for a certain period of time without being eliminated, a second-level warning is generated. The differentiated response action executed at this time is to trigger online recalibration, using static features in the environment (such as walls and columns) as natural calibration points, and using an incremental scanning matching algorithm to update the coordinate transformation matrix in real time without manual intervention.

[0190] When a sharp mean shift, an explosive increase in variance, or a continuous deterioration of autocorrelation characteristics reaching a critical value is detected, a Level 3 warning message is generated. The differentiated response action at this time is to perform a security downgrade, immediately suspend the current verification service, cut off the cloud control channel, and force the vehicle to enter a minimum risk state (such as parking on the side of the road), while simultaneously initiating a comprehensive manual recalibration process.

[0191] This tiered response mechanism ensures that the system can take appropriate intervention measures when faced with signs of degradation of varying severity, avoiding both service interruptions caused by overreaction and security incidents caused by underreaction.

[0192] Step S64: Based on the residual sequence analysis results, optimize the parameter set online. The parameter set includes at least one of the following: time delay parameter, coordinate transformation matrix, fusion weight parameter, and hierarchical threshold parameter.

[0193] In this embodiment of the application, the online optimization parameter set can refer to the key system configuration variables that are dynamically adjusted based on the error root causes revealed by residual sequence analysis.

[0194] The optimization of the delay parameter is based on the characteristic that the velocity residual is abnormal but the position residual is normal. It is inferred that there is a deviation in time synchronization. Then, the network transmission delay is fine-tuned or the delay parameter is processed by gradient descent to eliminate the error caused by spatiotemporal asynchrony.

[0195] The optimization of the coordinate transformation matrix is ​​based on the systematic offset characteristics of the position residual or heading residual. The loss function is constructed using residual feedback, and the rotation matrix and translation vector from the vehicle-end coordinate system to the field-end global coordinate system are iteratively updated to compensate for installation errors caused by equipment vibration or temperature deformation.

[0196] The optimization of the fusion weight parameters involves dynamically adjusting the observation noise covariance in Kalman filtering or factor graph optimization based on the variance changes of the residual sequence. When the residual variance increases, the trust weight of the field data is automatically reduced to improve the robustness of the system.

[0197] The optimization of the tiered threshold parameters is based on the historical statistical distribution of the comprehensive uncertainty measure. The thresholds that trigger early warning and compensation strategies at each level are updated regularly to adapt to changes in scenario uncertainty during long-term operation.

[0198] For example, if autocorrelation analysis shows a fixed hysteresis pattern, the system can automatically increase the estimated delay parameter value by 5ms; if mean drift detection shows a constant deviation in the X-axis direction, the system can automatically correct the X-axis translation component in the coordinate transformation matrix. Through this closed-loop feedback mechanism, the system can adaptively maintain high-precision positioning performance during long-term operation, realizing the transformation from reactive repair to predictive maintenance.

[0199] The online optimization parameter set specifically includes: adjusting the model parameters of the prediction model based on residual feedback, and dynamically adjusting the hierarchical threshold parameters based on the historical statistical distribution of the comprehensive uncertainty measure; an outlier removal mechanism is introduced during the optimization process to ensure long-term robustness.

[0200] Specifically, adjusting the model parameters of the prediction model based on residual feedback can refer to using the multi-dimensional residual sequence calculated in the aforementioned steps as a supervision signal to iteratively update the internal weights or coefficients of the prediction model involved in the vehicle localization method (e.g., the state transition matrix of an LSTM network or Kalman filter used for trajectory extrapolation). These model parameters are obtained by minimizing the mean squared error or negative log-likelihood function of the residual sequence using gradient descent or Bayesian optimization methods. Its purpose is to enable the prediction model to adaptively learn the vehicle's motion characteristics and environmental interference patterns in the current scene, thereby correcting model mismatch problems caused by environmental changes or sensor aging.

[0201] Dynamically adjusting the grading threshold parameter based on the historical statistical distribution of comprehensive uncertainty measurement can refer to collecting comprehensive uncertainty measurement data over a long period (such as the past hour or 1000 frames), constructing its probability density distribution, and recalculating the trigger threshold required for positioning status grading in real time based on the statistical characteristics of this distribution (such as mean, standard deviation, or specific quantiles). This grading threshold parameter is set according to the 3σ principle of comprehensive uncertainty measurement or a pre-set confidence interval (such as the 95th percentile). Its purpose is to solve the problem that fixed thresholds cannot adapt to changes in uncertainty levels in different scenarios, achieving scenario-adaptive grading compensation strategies.

[0202] Introducing an outlier removal mechanism during optimization involves cleaning the input residual sequence or uncertainty measure data before executing the optimization loop of adjusting model parameters and updating threshold parameters. This process identifies and removes outliers that deviate from the normal distribution range. This outlier removal mechanism can be implemented using the Random Sample Consensus (RANSAC) algorithm or the M-estimator. Its purpose is to prevent dirty data caused by occasional sensor failures, communication packet loss, or extreme environmental interference from dominating the optimization direction, leading to model parameter divergence or threshold setting failure.

[0203] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0204] To implement the above vehicle positioning method, this application also proposes a vehicle positioning device, please refer to the details below. Figure 7 , Figure 7 This is a schematic diagram of an embodiment of the vehicle positioning device provided in this application.

[0205] The vehicle positioning device 400 in this embodiment includes a processor 41, a memory 42, an input / output device 43, and a bus 44.

[0206] The processor 41, memory 42, and input / output device 43 are respectively connected to the bus 44. The memory 42 stores program data, and the processor 41 is used to execute the program data to implement the vehicle positioning method described in the above embodiments.

[0207] In this embodiment, processor 41 can also be referred to as CPU (Central Processing Unit). Processor 41 may be an integrated circuit chip with signal processing capabilities. Processor 41 can also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component. The general-purpose processor can be a microprocessor, or processor 41 can be any conventional processor.

[0208] This application also provides a computer storage medium; please refer to the following: Figure 8 , Figure 8 This is a schematic diagram of a computer storage medium according to an embodiment of the present application. The computer storage medium 600 stores a computer program 61, which, when executed by a processor, is used to implement the vehicle positioning method of the above embodiment.

[0209] When the embodiments of this application are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0210] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A vehicle positioning method, characterized in that, The vehicle positioning method includes: Obtain the field-end vehicle detection result obtained from the current frame field endpoint cloud data positioning; wherein, the field-end vehicle detection result includes several vehicle detection parameter values; Obtain an uncertainty estimate for each vehicle's detection parameter value; Obtain the time deviation between the current frame vehicle-end positioning time and the current frame field-end true value time; Based on the uncertainty estimate of each vehicle detection parameter value, a comprehensive uncertainty measure of the field vehicle detection results is determined; The current frame positioning status classification result is determined based on the comprehensive uncertainty metric and the time deviation. The vehicle detection results at the vehicle end are compensated based on the positioning status classification results; The vehicle positioning result is output based on the compensated vehicle detection results.

2. The vehicle positioning method according to claim 1, characterized in that, The uncertainty estimates include cognitive uncertainty estimates and accidental uncertainty estimates.

3. The vehicle positioning method according to claim 2, characterized in that, The uncertainty estimate for obtaining the detection parameter value for each vehicle includes: The current frame field endpoint cloud data is input into the vehicle detection network, and several forward propagations are performed to obtain several sets of vehicle detection parameter values. Obtain the estimated mean of several vehicle detection parameter values ​​for the target class of vehicles; The estimated variance measure is determined based on the vehicle detection parameter values ​​of the target class vehicle detection parameters and the estimated mean. The estimated variance metric is used as an estimate of the cognitive uncertainty of the target class vehicle detection parameters.

4. The vehicle positioning method according to claim 2, characterized in that, The uncertainty estimate for obtaining the detection parameter value for each vehicle includes: Input the current frame field endpoint cloud data into the vehicle detection network to obtain the vehicle detection parameter values ​​and their logarithmic variance; The random uncertainty estimate of the vehicle detection parameter value is determined based on the logarithmic variance.

5. The vehicle positioning method according to claim 4, characterized in that, The vehicle positioning method further includes: Obtain a training dataset, which contains multiple frames of LiDAR point cloud data and corresponding ground truth labels for target bounding boxes; A target vehicle detection network is constructed, which includes a feature extraction backbone network, a bounding box regression branch, and an uncertainty prediction branch. The lidar point cloud data is input into the target vehicle detection network. The bounding box regression branch outputs the predicted value, and the uncertainty prediction branch simultaneously outputs the corresponding logarithmic variance. The uncertainty-weighted loss value is determined based on the predicted value, the true value label, and the log-variance. The target vehicle detection network is trained using the uncertainty-weighted loss value to obtain the vehicle detection network.

6. The vehicle positioning method according to claim 2, characterized in that, After obtaining the uncertainty estimate of each vehicle detection parameter value, the vehicle localization method further includes: When the estimated cognitive uncertainty of all vehicle detection parameter values ​​is less than the cognitive uncertainty threshold and the estimated random uncertainty is less than the random uncertainty threshold, the field vehicle detection result is marked as a high-confidence true value. When there is a cognitive uncertainty estimate of at least one vehicle detection parameter value that is greater than the cognitive uncertainty threshold, and the cognitive uncertainty estimate of all vehicle detection parameter values ​​is less than a preset multiple of the cognitive uncertainty threshold, or when there is a random uncertainty estimate of at least one vehicle detection parameter value that is greater than the random uncertainty threshold, and the cognitive uncertainty estimate of all vehicle detection parameter values ​​is less than a preset multiple of the random uncertainty threshold, the field vehicle detection result is marked as a medium confidence true value. When the estimated cognitive uncertainty of at least one vehicle detection parameter value is greater than a preset multiple of the cognitive uncertainty threshold, or when the estimated random uncertainty of at least one vehicle detection parameter value is greater than a preset multiple of the random uncertainty threshold, the field vehicle detection result is marked as a low-confidence true value.

7. The vehicle positioning method according to claim 6, characterized in that, The vehicle positioning method further includes: When the time deviation is greater than the time deviation threshold, or when the vehicle detection result at the field is a low-confidence true value, the vehicle detection result at the field is updated by extrapolation prediction based on a learning model.

8. The vehicle positioning method according to claim 7, characterized in that, The step of updating the on-site vehicle detection results using extrapolation prediction based on a learning model includes: The true state sequence of the field end within the historical time window is input into the first extrapolation prediction model. The input features of the state sequence include lateral position error, longitudinal position error, heading angle error, yaw rate, longitudinal velocity, lateral velocity, and longitudinal acceleration. Obtain the predicted state sequence within the future time window output by the first extrapolation prediction model, the predicted state sequence including predicted lateral position, predicted longitudinal position, predicted heading angle and predicted speed; Kinematic constraints are applied to the predicted state sequence to obtain the corrected vehicle detection results at the field end.

9. The vehicle positioning method according to claim 8, characterized in that, The kinematic constraint correction includes: When the predicted acceleration at adjacent moments in the predicted state sequence exceeds the preset acceleration range, the predicted acceleration is truncated to the boundary value of the preset acceleration range. When the predicted curvature exceeds the preset curvature threshold, the yaw rate is truncated.

10. The vehicle positioning method according to claim 8 or 9, characterized in that, After applying kinematic constraint correction to the predicted state sequence to obtain the corrected field vehicle detection results, the vehicle localization method further includes: The prediction deviation is calculated based on the corrected field-end vehicle detection results and the vehicle-end vehicle detection results. Based on the relationship between the predicted deviation and the preset deviation threshold, a tiered prediction is performed: When the prediction deviation is less than the first deviation threshold, it is determined to be in a normal state; When the prediction deviation is between the first deviation threshold and the second deviation threshold, it is determined to be a potential degradation state and the first level of compensation is triggered. When the prediction deviation is greater than or equal to the second deviation threshold, it is determined that the deviation is about to occur and the second level of compensation is triggered. The first deviation threshold and the second deviation threshold are determined based on the statistical distribution of historical data.

11. The vehicle positioning method according to claim 8 or 9, characterized in that, After applying kinematic constraint correction to the predicted state sequence to obtain the corrected field vehicle detection results, the vehicle localization method further includes: The spatial deviation is calculated based on the corrected predicted trajectory and the vehicle-end positioning data. Obtain the extrapolation uncertainty index corresponding to the extrapolation prediction; The spatial deviation threshold is dynamically adjusted based on the extrapolation uncertainty index, wherein the spatial deviation threshold increases as the extrapolation uncertainty index increases.

12. The vehicle positioning method according to claim 8, characterized in that, The vehicle positioning method further includes: Construct multiple first extrapolation prediction models with identical structures but different initialization parameters; Obtain the independent prediction results of each of the first extrapolation prediction models for the same input; Based on the dispersion of each independent prediction result and the preset model inherent variance, the extrapolation uncertainty index is determined.

13. The vehicle positioning method according to claim 6, characterized in that, The comprehensive uncertainty measure is generated by weighted summation of the following items: normalized cognitive uncertainty index, normalized accidental uncertainty index, normalized extrapolation uncertainty index, and comprehensive verification compensation term; The normalization indices are obtained by standardizing the historical statistics of the corresponding uncertainty components.

14. The vehicle positioning method according to claim 13, characterized in that, When the extrapolation prediction based on the learning model is not executed, the normalized extrapolation uncertainty index takes the value of zero.

15. The vehicle positioning method according to claim 13, characterized in that, The comprehensive verification confidence index corresponding to the comprehensive verification compensation item is generated by weighted summation of the following items: The confidence terms for time deviation, space deviation, velocity deviation, heading angle deviation, and true value uncertainty penalty are included. The truth uncertainty penalty term is determined based on the proportion of parameters with different confidence levels in the vehicle detection results at the field. The higher the proportion of parameters with high confidence levels, the closer the truth uncertainty penalty term is to the upper limit.

16. The vehicle positioning method according to claim 1, characterized in that, The positioning status classification results include: Level 1, Level 2, Level 3, and Level 4, wherein: The triggering conditions corresponding to the first level include: the time deviation meets the first condition, the spatial deviation meets the second condition, and the comprehensive uncertainty measure meets the third condition; The triggering conditions for the second level include: the spatial deviation meets the fourth condition or the comprehensive uncertainty measure meets the fifth condition; The triggering conditions for the third level include: the spatial deviation meets the sixth condition, the comprehensive uncertainty measurement meets the seventh condition, or the number of cycles in which the prediction deviation continues to exceed the warning threshold reaches a preset value. The triggering conditions corresponding to the fourth level include: the spatial deviation meets the eighth condition, the comprehensive verification confidence is lower than the preset confidence lower limit, the comprehensive uncertainty measure meets the ninth condition, the number of consecutive verification failure frames reaches a preset value, or the communication interruption duration reaches a preset value.

17. The vehicle positioning method according to claim 16, characterized in that, The compensation of vehicle detection results based on positioning status classification results includes: The compensation strategy corresponding to the first level is to use the vehicle detection results at the vehicle end; The compensation strategy corresponding to the second level is to calculate the correction amount based on the corrected predicted trajectory and the vehicle detection results at the vehicle end, and correct the vehicle detection results at the vehicle end in a weighted manner. The weighting coefficient is dynamically adjusted according to the comprehensive uncertainty measure. The compensation strategy corresponding to the third level is to use a filtering fusion algorithm to fuse the field vehicle detection results and the vehicle detection results, and adjust the observation noise parameters in the filtering fusion algorithm according to the field truth confidence level. The compensation strategy corresponding to the fourth level is to trigger a security downgrade operation.

18. The vehicle positioning method according to claim 17, characterized in that, The step of adjusting the observation noise parameters in the filtering and fusion algorithm based on the ground truth confidence level includes: When the confidence level of the true value at the field end is the first confidence level, the first observation noise covariance is used; When the confidence level of the true value at the field end is the second confidence level, the second observation noise covariance is used; Wherein, the second confidence level is lower than the first confidence level, and the second observation noise covariance is greater than the first observation noise covariance.

19. The vehicle positioning method according to claim 1, characterized in that, After outputting the vehicle positioning result based on the compensated vehicle detection result, the vehicle positioning method further includes: The actual position and pose data of the vehicle and the vehicle detection results of the field are obtained, and a multi-dimensional residual sequence is calculated. The multi-dimensional residual includes position residual, heading residual and velocity residual. Perform time-series statistical analysis on the multidimensional residual sequence, including mean drift detection, variance change detection, and autocorrelation characteristic analysis. Based on the results of the time-series statistical analysis, a graded early warning information is generated, and differentiated response actions are executed according to the graded early warning information. The response actions include adjusting calibration parameters to optimize the frequency, triggering online recalibration, or performing security degradation. Based on the residual sequence analysis results, the parameter set is optimized online. The parameter set includes at least one of the following: time delay parameter, coordinate transformation matrix, fusion weight parameter, and hierarchical threshold parameter.

20. The vehicle positioning method according to claim 19, characterized in that, The mean drift detection uses a cumulative sum control chart algorithm to determine the warning level by the ratio of the cumulative statistic to a preset decision threshold; the variance change detection uses the F-test method to determine the warning level by the variance change rate of the residual series in the first and second halves; the autocorrelation characteristic analysis determines the warning level by the number of autocorrelation coefficients exceeding the confidence interval under multiple lag orders.

21. The vehicle positioning method according to claim 19, characterized in that, The tiered early warning system includes three levels: when the first level early warning is triggered, the frequency of calibration parameter optimization is increased; when the second level early warning is triggered, online recalibration is performed using static environmental characteristics to update coordinate transformation parameters; when the third level early warning is triggered, the verification service is suspended and the recalibration process is started.

22. The vehicle positioning method according to claim 19, characterized in that, The online optimization parameter set includes: model parameters adjusted based on residual feedback, and hierarchical threshold parameters dynamically adjusted based on the historical statistical distribution of comprehensive uncertainty measure; an outlier removal mechanism is introduced during the optimization process to ensure long-term robustness.

23. A vehicle positioning device, characterized in that, The vehicle positioning device includes a memory and a processor coupled to the memory; The memory is used to store program data, and the processor is used to execute the program data to implement the vehicle positioning method as described in any one of claims 1 to 22.

24. A computer storage medium, characterized in that, The computer storage medium is used to store program data, which, when executed by the computer, is used to implement the vehicle positioning method as described in any one of claims 1 to 22.

Images