A vehicle-road-cloud collaborative perception multi-source data processing method and system

Abstract

This application relates to the field of intelligent transportation and vehicle-road-cloud collaborative perception technology, and discloses a multi-source data processing method and system for vehicle-road-cloud collaborative perception. This method uses real-time scene parameter vectors as the core driver. After completing the acquisition of multi-source raw data from vehicles, roads, and the cloud, it sequentially performs adaptive data source filtering, scene-based dynamic preprocessing, multi-source fusion and spatiotemporal alignment, followed by data standardization and encapsulation and full-dimensional quality assessment. Finally, it achieves parameter closed-loop self-optimization at the edge and cloud levels through a global loss function. Compared with existing technologies, this method solves the problems of data redundancy, rigid processing strategies, insufficient spatiotemporal alignment accuracy, and lack of self-optimization feedback, improving data processing quality and efficiency, outputting high-quality standardized data with spatiotemporal consistency, enhancing the reliability and intelligent operation and maintenance level of the collaborative perception system, while reducing computing power and operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and vehicle-road-cloud collaborative sensing technology, and more specifically, to a multi-source data processing method and system for vehicle-road-cloud collaborative sensing. Background Technology

[0002] With the large-scale application of intelligent connected vehicles and advanced autonomous driving technologies, integrated vehicle-road-cloud collaborative perception has become a key supporting technology for improving the accuracy of traffic environment perception and ensuring the safety of autonomous driving. By integrating multi-source heterogeneous data from onboard sensors, roadside perception devices, and cloud computing platforms, the vehicle-road-cloud collaborative perception system can overcome the line-of-sight limitations and environmental occlusion constraints of single-vehicle perception, and achieve global perception of traffic conditions across all road sections and in all weather conditions.

[0003] In vehicle-road-cloud collaborative perception systems, the preprocessing quality of multi-source data directly affects the accuracy, real-time performance, and reliability of subsequent data fusion. However, existing technologies still have the following technical shortcomings in practical applications:

[0004] First, the lack of a scenario-adaptive data source selection mechanism leads to low resource utilization efficiency. Existing solutions generally perform indiscriminate full data collection and transmission from all sensors on the vehicle and roadside, without dynamically adjusting the data collection strategy according to the complexity of the real-time traffic scenario. In simple scenarios such as smooth traffic and no obstructions, a large amount of redundant sensor data is still continuously sampled and transmitted at a fixed maximum frequency. However, the computing power and communication bandwidth resources of roadside edge computing nodes are physically constrained, and full data processing leads to increased system latency, making it difficult to meet the millisecond-level response requirements of autonomous driving.

[0005] Second, data preprocessing strategies are rigid and lack adaptability to various scenarios. Existing data cleaning, filtering, and sampling control often employ fixed threshold strategies, such as constant noise filtering parameters and fixed sensor sampling frequencies. These strategies cannot adaptively adjust to complex scenarios such as sudden changes in lighting, severe weather like rain and snow, abrupt changes in traffic flow density, and dynamic building occlusion. Static preprocessing strategies are prone to problems in complex environments, such as thresholds that are too wide, leading to residual noise and false targets, or thresholds that are too strict, causing the erroneous deletion of valid target information, directly affecting the accuracy of subsequent fusion sensing.

[0006] Third, the spatiotemporal alignment accuracy of multi-source data is insufficient, which can easily lead to perception errors in special scenarios. Existing spatiotemporal synchronization solutions mostly employ GNSS single-source positioning or GNSS / IMU dual-source fusion positioning. In scenarios with limited GNSS signals, such as urban canyons, tunnels, and densely populated areas with tall buildings, positioning drift or interruption problems are prone to occur. Regarding time synchronization, existing solutions mostly use network time protocols or trigger-based hardware synchronization methods. The former can only achieve millisecond-level synchronization accuracy, while the latter has poor adaptability to heterogeneous sensors. Both are insufficient to meet the microsecond-level time alignment requirements of multi-source sensor data fusion. Spatiotemporal alignment deviations lead to errors in target position and motion state judgment, posing safety hazards to advanced autonomous driving.

[0007] Fourth, the lack of a closed-loop quality assessment and self-optimization mechanism results in insufficient long-term system stability. The existing preprocessing process uses a one-way open-loop architecture and lacks a mechanism for quality assessment and feedback adjustment of the preprocessing results. When sensor parameter drift or long-term changes in environmental characteristics lead to deterioration in preprocessing quality, the system cannot automatically detect and trigger parameter optimization, relying solely on manual on-site calibration, which results in high maintenance costs and fails to guarantee the long-term stability of the system.

[0008] In summary, existing technologies have not fully solved the technical problems of dynamically adjusting data source selection strategies based on real-time traffic scenarios, adaptively optimizing preprocessing parameters, achieving high-precision spatiotemporal alignment, and closed-loop self-optimization. This poses challenges to the engineering application and long-term stable operation of vehicle-road-cloud collaborative perception technology. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the purpose of this application is to provide a multi-source data processing method and system for vehicle-road-cloud collaborative perception.

[0010] To achieve the above objectives, this application provides the following technical solution:

[0011] A method for processing multi-source data in vehicle-road-cloud collaborative sensing includes the following steps:

[0012] S1, Multi-source data acquisition: Real-time acquisition of multi-source raw data from vehicle-mounted equipment, roadside equipment and cloud, generating multi-source raw datasets and constructing real-time scene parameter vectors;

[0013] S2, Data source filtering: Based on multi-source original datasets and scene parameter vectors, a scene parameter-driven adaptive mapping mechanism for data sources is constructed. Data is filtered through a multi-factor weighted priority scoring function, and a subset of data sources is output.

[0014] S3, Dynamic Preprocessing: Based on a subset of the data source and the scene parameter vector, the sampling frequency of the data source is dynamically adjusted, adaptive filtering and data cleaning are performed to obtain the preprocessed dataset;

[0015] S4, Spatiotemporal Alignment: Based on the preprocessed dataset, spatiotemporal alignment is performed using a multi-source fusion localization and multi-level time synchronization mechanism, and the spatiotemporally aligned dataset is output.

[0016] S5, Data Output and Quality Assessment: The spatiotemporally aligned dataset is packaged into a standard preprocessed data package in a unified data format and output as a quality assessment index. The data efficiency, spatial bias and temporal bias are calculated simultaneously and output as quality assessment indicators.

[0017] S6, Closed-loop self-optimization: Based on the quality assessment index, deviation classification and attribution are performed, a global loss function of comprehensive quality assessment index is constructed, and the screening parameters of S2, the preprocessing parameters of S3 and the spatiotemporal alignment parameters of S4 are iteratively optimized and updated to form a closed-loop self-optimization of the data processing strategy.

[0018] In a preferred embodiment, the scene parameter vector in S1 includes scene type encoding, traffic flow density, light intensity, occlusion factor, and GNSS signal quality factor.

[0019] In a preferred embodiment, the data source adaptive mapping mechanism in S2 includes scene parameter standardization processing and adaptive filtering rules based on a multi-factor weighted priority scoring function.

[0020] Scene parameter standardization is used to unify parameters with different mass dimensions in the scene parameter vector to a preset value range;

[0021] Adaptive filtering rules are used to filter valid data sources based on priority scoring results and dynamic filtering thresholds.

[0022] In a preferred embodiment, the dynamic adjustment of the data source sampling frequency in S3 is performed in segments based on traffic flow density and the priority score of the data source.

[0023] The adaptive filtering and data cleaning dynamically adjust the filtering threshold based on light intensity and occlusion factor to remove noise and false data.

[0024] In a preferred embodiment, the multi-source fusion positioning described in S4 is multi-source fusion spatial alignment, which adopts a GNSS, IMU, and geomagnetic three-source fusion positioning scheme. Through a weighted fusion algorithm with adaptive signal quality allocation, combined with the projection transformation of high-precision electronic maps and intersection-level calibration, the spatial coordinates of multi-source data are unified.

[0025] In a preferred embodiment, the multi-level time synchronization mechanism in S4 is a three-level precise time synchronization architecture, which sequentially completes clock synchronization from the cloud to the roadside, transmission delay correction from the roadside to the vehicle, and performs linear interpolation compensation for residual time deviations to achieve precise unification of timestamps for multi-source data.

[0026] In a preferred embodiment, the standard preprocessed data packet in S5 includes uniform coordinates, uniform timestamps, sensor identifiers, and confidence labels.

[0027] In a preferred embodiment, the deviation classification attribution in S6 refers to comparing the quality assessment indicators with the preset target values, classifying the indicators that exceed the standards, and matching the corresponding parameter optimization paths.

[0028] The global loss function refers to a multi-dimensional quality assessment index that integrates data effectiveness, spatial bias, and temporal bias, used to quantify the overall quality bias of the entire preprocessing process.

[0029] In a preferred embodiment, the iterative optimization update in step S6 includes defining the set of parameters to be optimized, gradient-based parameter update rules, and edge-cloud hierarchical update and version management mechanism.

[0030] The set of parameters to be optimized includes the filtering parameters of S2, the preprocessing parameters of S3, and the spatiotemporal alignment parameters of S4.

[0031] Gradient-based parameter update rules complete parameter iteration by calculating the gradient of the global loss function;

[0032] The edge-cloud hierarchical update and version management mechanism adopts an update strategy that combines rapid local iteration of roadside edge nodes with long-term global optimization in the cloud, and is equipped with parameter version management and anomaly rollback mechanisms.

[0033] A multi-source data processing system for vehicle-road-cloud collaborative perception is characterized by comprising a multi-source data acquisition module, an adaptive data source filtering module, a dynamic preprocessing module, a spatiotemporal alignment module, a data output and quality assessment module, and a closed-loop self-optimization module.

[0034] The multi-source data acquisition module is used to collect multi-source raw data from vehicles, roadsides, and the cloud, generate multi-source raw datasets, and construct real-time scene parameter vectors.

[0035] The adaptive data source filtering module is used to adaptively filter multi-source raw data based on scene parameter vectors and output a subset of valid data sources.

[0036] The dynamic preprocessing module is used to dynamically sample and adjust the effective data source subset based on the scene parameter vector and perform adaptive data cleaning, outputting the preprocessed dataset.

[0037] The spatiotemporal alignment module is used to accurately align the spatial coordinates and timestamps of the preprocessed dataset and output the spatiotemporally aligned dataset.

[0038] The data output and quality assessment module is used to standardize and encapsulate the spatiotemporally aligned dataset for output, and simultaneously calculate and output the quality assessment indicators of data processing.

[0039] The closed-loop self-optimization module is used to iteratively optimize the core operating parameters of the aforementioned modules based on quality assessment indicators, forming a closed-loop self-optimization of the data processing strategy.

[0040] By adopting the above technical solution, the beneficial effects of the present invention are as follows:

[0041] Through a scenario-parameter-driven adaptive data source filtering mechanism, high-value data sources can be dynamically filtered based on real-time traffic scenarios, eliminating invalid and redundant data, significantly reducing the computing power consumption and communication bandwidth usage of roadside edge nodes, and improving the real-time performance of data processing.

[0042] By dynamically adjusting the sampling frequency in conjunction with traffic flow and priority, and optimizing the adaptive filtering threshold in conjunction with illumination and occlusion, this approach replaces the traditional fixed parameter scheme. It can adapt to various dynamic traffic scenarios, significantly improve data efficiency, and accurately remove noise and false target data.

[0043] By adopting a three-source adaptive fusion spatial alignment scheme of GNSS, IMU and geomagnetic three sources, the problem of positioning drift and interruption in occlusion scenarios is solved; by adopting a three-level time synchronization architecture, the time synchronization accuracy is improved from the millisecond level of the existing technology to the microsecond level, providing high-quality data with high spatiotemporal consistency for downstream fusion perception.

[0044] By using multi-dimensional quality assessment, deviation classification and attribution, and parameter iterative optimization driven by a global loss function, the traditional open-loop processing architecture can be replaced. This allows for continuous self-evolution of the preprocessing strategy without manual intervention, avoiding performance degradation over long-term system operation and significantly reducing maintenance costs. Attached Figure Description

[0045] Figure 1 This is a schematic flowchart of a multi-source data processing method for vehicle-road-cloud collaborative perception according to the first embodiment of this application;

[0046] Figure 2 This is a schematic diagram of the framework structure of a multi-source data processing system for vehicle-road-cloud collaborative perception according to the second embodiment of this application. Detailed Implementation

[0047] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0049] The first embodiment of this application provides a multi-source data processing method for vehicle-road-cloud collaborative perception. Driven by real-time scene parameter vectors, it constructs a full-process, fully closed-loop adaptive processing architecture of "multi-source data acquisition → adaptive data source filtering → dynamic preprocessing → precise spatiotemporal alignment → preprocessed data output and quality assessment → closed-loop self-optimization". It addresses the four core defects of existing technologies: high data redundancy, poor scene adaptability, insufficient spatiotemporal alignment accuracy, and lack of self-optimization capability, and realizes scene-based adaptive processing and continuous self-optimization of multi-source data from vehicles, roads, and clouds.

[0050] The overall deployment environment of this embodiment is as follows: the core process of the method is deployed on the roadside edge computing node, and the closed-loop optimization adopts a hierarchical operation mode of "roadside edge local iteration + cloud global optimization". It can be directly compatible with the hardware architecture and communication protocol of the existing vehicle-road-cloud cooperative system without the need to upgrade the hardware of the existing vehicle and roadside equipment.

[0051] For ease of explanation, we will first provide a unified definition for some of the mathematical symbols used:

[0052] Scene parameter vector: These are the core driving parameters for adaptive adjustment throughout the entire process;

[0053] Multiple source raw datasets: It includes raw data collected from all data sources, including vehicle-mounted, roadside, and cloud sources;

[0054] Data efficiency: , is the core quality assessment indicator, representing the proportion of valid data after preprocessing to the original data.

[0055] This embodiment specifically includes six sequentially executed, closed-loop linked steps, the flowchart of which is shown below. Figure 1 As shown, the specific implementation of each step is as follows:

[0056] To facilitate understanding, the data flow path in this embodiment will be explained in advance: (Multi-source raw data from vehicles, roads, and the cloud) → (Valid data source subset) → (Preprocessed valid dataset) → (The spatiotemporally aligned dataset); at the same time The standard preprocessed data package and quality assessment index set are generated after processing in step S5. The quality assessment index set is then updated into a parameter vector through closed-loop self-optimization in step S6, and written back to steps S2, S3, and S4 to form a closed-loop data processing link for the entire process.

[0057] S1 Multi-Source Data Acquisition: The core purpose is to complete the aggregation of multi-source heterogeneous data, the initial unification of formats, and the extraction of scene features, so as to provide standardized inputs and global driving parameters for subsequent full-process adaptive processing, and solve the fundamental problems of inconsistent multi-source data formats and lack of scene feature extraction in existing technologies.

[0058] Specifically, the three types of raw signals collected in this step include:

[0059] Vehicle terminal: 10Hz LiDAR point cloud, 25fps surround view camera image stream, 100Hz IMU six-axis data, 50Hz vehicle speed sensor data;

[0060] Roadside: 20Hz millimeter-wave radar target list, 10Hz pedestrian detection sensor data, 1Hz geomagnetic sensor magnetic field strength data, 1Hz light sensor illuminance data, 25fps high-definition camera video stream, 1Hz traffic signal phase status data;

[0061] Cloud platform: Regional traffic control instructions, historical perception model parameters, and pedestrian early warning push data.

[0062] Furthermore, the data processing in this step specifically involves: parallel access to various data sources via the C-V2XPC5 direct connection interface and the 5G-AUu interface to complete protocol parsing and preliminary format unification, unifying the timestamp format of all data to UTC standard time and the coordinate system to WGS-84 geodetic coordinate system, and aggregating them to form multi-source raw data packets; simultaneously, scene parameters are extracted from light sensor, geomagnetic sensor, GNSS receiver, and traffic control data to construct a real-time scene parameter vector. .

[0063] Optional, scene parameter vector The specific definitions of each parameter are as follows:

[0064] Encode the scene type (1=unobstructed, 2=peak hours, 3=low light, 4=occlusion); Traffic flow density; Light intensity; The occlusion factor is 0 = no occlusion, 1 = complete occlusion. This is the GNSS signal quality factor (0 = complete loss of lock, 1 = full signal).

[0065] This step standardizes the format of multi-source heterogeneous data, extracts the scene parameter vectors that drive the adaptive adjustment throughout the entire process, and outputs the original multi-source dataset. and corresponding scene parameter vectors The data is transmitted to the subsequent processing module of the roadside edge computing node, providing input for step S2.

[0066] S2 Adaptive Data Source Filtering: The core purpose is to eliminate invalid and redundant data through an adaptive filtering mechanism driven by scenario parameters, thereby solving the core problems of serious data redundancy and waste of computing power and bandwidth resources in existing technologies.

[0067] Specifically, the scenario parameter-driven adaptive data source mapping mechanism in this step includes two core components: scenario parameter standardization processing and adaptive filtering rules based on a multi-factor weighted priority scoring function.

[0068] First, scene parameter standardization is performed: the scene parameter vector is... Each component is normalized to unify parameters with different dimensions into the [0,1] interval, eliminating the interference of dimensional differences between different parameters on the screening results. Furthermore, differentiated normalization methods are adopted for the physical characteristics of different parameters. Traffic flow density ρ is normalized to the typical urban density range of [50,300] pcu / h using Min-Max normalization, and light intensity l is logarithmically normalized to [0,100000] lux.

[0069] Next, an adaptive filtering rule based on a multi-factor weighted priority scoring function is executed: for the data source Calculate its priority score in the current scenario. Data sources whose scores exceed the dynamic threshold are selected as a subset of valid data sources.

[0070] Specifically, the overall screening rules are as follows:

[0071] Furthermore, the priority scoring function is defined as follows:

[0072]

[0073] Definitions and values ​​of the parameters in the formula: The scene type weight vector has a value range of [0.6~1.4], with an optimal range of [0.8~1.2], and is determined by optimizing the F1-score based on 1000 Monte Carlo simulations. For data source The prior correlation vectors in four scenarios characterize the degree of sensor adaptability to specific scenarios. They are obtained through experimental calibration of the sensor's physical characteristics and scene matching degree, such as the performance of lidar in low-light scenarios. The camera is ; This is the traffic flow density adjustment coefficient. ), where 115 pcu / h is the typical average density of urban roads in China, and 65 pcu / h is half of the peak-to-low peak density fluctuation range. Used for saturation limiting to prevent coefficient divergence caused by extreme density values; For data source In density The coverage score is determined by actual measurement of the ratio of the sensor's effective detection range to the number of targets. This is the combined light-shading adjustment coefficient. Weights used to enhance the robustness of sensors under low light or high occlusion conditions; For data source In light intensity and occlusion factor The robustness score was obtained through experimental calibration. This is the computing power cost penalty coefficient, ranging from 0.05 to 0.15, with a default value of 0.1. For data source Normalized computing power consumption.

[0074] Furthermore, to adapt to different screening requirements in various scenarios, this step sets a dynamic threshold function, specifically:

[0075]

[0076] Definitions and values ​​of parameters in the formula: 0.65 is the baseline threshold, the optimal operating point determined based on ROC curve analysis; 0.15 is the threshold improvement amount in high-density scenarios, when... Activated at pcu / h to exclude inefficient data sources; 0.2 is the occlusion sensitivity coefficient, the more severe the occlusion, the higher the screening standard; This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise.

[0077] Optionally, this step also includes a screening quality feedback process: recording the redundancy rate and coverage rate of the screening results in this cycle, generating a feedback signal and transmitting it to the closed-loop optimization module in step S6 for subsequent parameter iteration updates.

[0078] This step replaces the existing indiscriminate full-data collection strategy with a scenario-driven, multi-factor weighted scoring mechanism, which can reduce invalid data redundancy by more than half and significantly reduce roadside computing power consumption; the final output is a set of filtered valid data. Each data source Includes its priority score value normalized to the [0,1] interval. (recorded as) (Normalized to the [0,1] interval) to provide input for step S3.

[0079] S3 Dynamic Preprocessing Steps: The core objective is to achieve scenario-driven dynamic sampling and adaptive data cleaning, solving the problems of static and rigid preprocessing strategies and low data efficiency in complex scenarios.

[0080] Specifically, the dynamic preprocessing in this step includes two core components: dynamic adjustment of the data source sampling frequency and adaptive filtering and data cleaning.

[0081] First, dynamically adjust the data source sampling frequency based on traffic flow density. and data source priority A segmented adjustment strategy is employed to dynamically calculate the sampling frequency of each data source, achieving optimized resource allocation by "high-frequency sampling of high-value, high-priority data sources and low-frequency sampling of low-value, low-priority data sources." Specifically, the segmented adjustment formula for the sampling frequency is as follows:

[0082]

[0083] Definitions and values ​​of the parameters in the formula: For data source The reference sampling frequency is determined by the sensor hardware specifications, and the on-board OBU supports dynamic adjustment from 2-8Hz; The upper limit of the sampling frequency is constrained by both the sensor hardware and the communication bandwidth. Prioritize the data source output from step 2 (normalized to) The high-priority data source receives a higher sampling frequency; 0.5 is the priority gain coefficient, which allows the sampling frequency of the high-priority data source to be increased to a maximum of 1.5 times the baseline value; 120 pcu / h is the low-density to medium-density dividing point. Below this value, the scene changes slowly and there is no need to increase the frequency; 300 pcu / h is the extremely high-density saturation point. When this value is exceeded, the upper limit frequency is directly used; 180 is the medium-to-high density range (300-120=180), which is used for linear interpolation normalization.

[0084] Furthermore, the physical logic of this segmented design is as follows: in low-density scenarios, the sampling frequency is reduced to save computing power; in medium- and high-density scenarios, the sampling rate is linearly increased to capture dense targets; and in extremely high-density scenarios, an upper limit frequency is used to ensure full perception coverage.

[0085] Second, perform adaptive filtering and data cleaning: based on light intensity. and occlusion factor The filter threshold is dynamically adjusted to remove noise and false data, thus solving the problems of "too loose a threshold leaves residual noise, while too strict a threshold accidentally deletes valid data".

[0086] Specifically, the formula for dynamically adjusting the filter threshold is:

[0087]

[0088] Definitions and values ​​of the parameters in the formula: The baseline filtering threshold is determined by actual measurements of the noise floor under standard lighting conditions; 0.005 is the light sensitivity coefficient, derived from the actual calibration of the light sensor, with the threshold adjusted by 0.1 for every 20 lux increase. The reference light intensity is 5000 lux by default, corresponding to the typical daytime illuminance in a city; 0.3 is the shading compensation coefficient, which comes from shading simulation experiments. For every 0.1 increase in the shading factor, the threshold needs to be increased by 0.03 to eliminate false signals from multipath reflections.

[0089] Optionally, after adaptive threshold filtering, morphological filtering can be superimposed: combining the building occlusion areas and road structure information marked in the high-precision electronic map, spatial consistency verification is performed on the point cloud and radar data, eliminating false data caused by multipath reflection and ghost targets, and further improving data efficiency.

[0090] This step replaces the fixed-parameter strategy of existing technologies with scene-driven dynamic sampling and adaptive filtering, which can significantly improve data efficiency and accurately remove noise and false targets; the final output is a preprocessed effective dataset. Provide input for step S4.

[0091] S4 Spatiotemporal Alignment: The core objective is to achieve precise unification of spatial coordinates and timestamps from multi-source data, solving the problems of insufficient spatiotemporal alignment accuracy and positioning drift / interruption in occluded scenarios in existing technologies.

[0092] Specifically, the spatiotemporal alignment in this step includes two core components: multi-source fusion spatial alignment and multi-level precise temporal synchronization.

[0093] First, perform multi-source fusion spatial alignment: A GNSS, IMU, and geomagnetic three-source fusion positioning scheme is adopted. Through a weighted fusion algorithm with adaptive signal quality allocation, combined with projection transformation of high-precision electronic maps and intersection-level calibration, the spatial coordinates of multi-source data are unified, solving the problem of single / dual-source positioning failure in occluded scenarios. Specifically, the general formula for spatial alignment is:

[0094]

[0095] Definitions and values ​​of the parameters in the formula: The WGS-84 coordinates output by the GNSS receiver are represented by a three-dimensional vector. ; The calculated position, a three-dimensional vector, is the output of the IMU via inertial navigation. Its error accumulates over time and needs to be complemented by GNSS; For the heading constraint position correction amount assisted by the geomagnetic sensor, a three-dimensional vector It provides auxiliary heading information when GNSS signals attenuate; the fusion weights adopt an adaptive signal quality allocation strategy. (GNSS signal quality factor) (IMU weight increases when GNSS is weakened) (Geomagnetic assist is activated in occluded scenarios, with 0.6 being the upper limit of the calibration gain). These three parameters are normalized to meet the following conditions. ; For high-precision electronic map projection transformation matrix, An orthogonal matrix maps the fused positioning results to the local coordinate system of the intersection; For intersection-level calibration offset, three-dimensional vector The data is estimated online and updated in real time by an extended Kalman filter (EKF) based on the location information of fixed roadside reference points.

[0096] The core advantage of this three-source fusion solution is that when the GNSS signal is good and unobstructed, GNSS is used as the primary positioning source; when the GNSS signal is lost and there is obstruction, it automatically switches to the IMU+Geomagnetic fusion mode to ensure continuous and stable positioning in obstructed scenarios.

[0097] Second, perform multi-level precise time synchronization: adopt a three-level precise time synchronization architecture from cloud to roadside and from roadside to vehicle, and sequentially complete clock synchronization from cloud to roadside and transmission delay correction from roadside to vehicle. At the same time, perform linear interpolation compensation for residual time deviation, realize precise unification of timestamps of multi-source data, and improve the synchronization accuracy from the millisecond level of existing technology to the microsecond level.

[0098] Specifically, the general formula for time synchronization is:

[0099]

[0100] Definitions of the parameters in the formula: A cloud-based UTC reference clock; It provides PTP (IEEE1588) synchronization correction for data from the cloud to the roadside RSU, with an accuracy down to the sub-microsecond level. This is the PC5 interface transmission delay correction amount from the roadside RSU to the vehicle-mounted OBU; The residual error activation threshold is set according to the accuracy requirements of the specific scenario. For linear interpolation compensation, when the residual synchronization error Exceeding the threshold Time-activated, the deviation at the current time is estimated and compensated by linear interpolation of adjacent synchronization points; This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise.

[0101] This step utilizes a three-source fusion spatial alignment and a three-level time synchronization architecture to control spatial positioning deviation within 0.5 meters and improve time synchronization accuracy to the microsecond level, while also resolving the issue of positioning interruption in occluded scenarios; the final output is a spatiotemporally aligned dataset. ,Depend on Each data point is obtained after being calibrated with spatial coordinates and timestamps, providing qualified input data for the standardized packaging and quality assessment in step S5.

[0102] S5 Data Output and Quality Assessment: The core purpose is to complete the standardized encapsulation of data and the quantitative assessment of the entire process quality, providing standard input for downstream integrated sensing, and providing quantitative basis for closed-loop self-optimization.

[0103] Specifically, the data processing in this step consists of two core stages:

[0104] First, standardized encapsulation: The spatiotemporally aligned source data is encapsulated into a standard preprocessed data packet using a unified data format. Furthermore, the format fields of this standard preprocessed data packet include unified coordinates, unified timestamps, sensor identifiers, and confidence labels, ensuring that the encapsulated data can be directly adapted to various downstream fusion sensing algorithms without requiring secondary format conversion.

[0105] Second, quality assessment index calculation: Core quality assessment indicators for this cycle are calculated simultaneously to provide a quantitative basis for closed-loop optimization. Specifically, the calculated indicators include: data effectiveness. The proportion of valid data in this period to the original data; spatial bias. The average deviation between the spatial positioning results and the true values ​​of the roadside reference points in this cycle; the time deviation. Average time synchronization residual error of multi-source data in this period; computing power utilization rate. The computing power occupancy ratio of roadside edge nodes in this cycle.

[0106] This step achieves standardized output of preprocessed data and completes quantitative evaluation of the overall processing quality. The output quality evaluation index is the core input of the closed-loop self-optimization in step S6, which directly determines the direction and magnitude of parameter iteration.

[0107] S6 Closed-Loop Self-Optimization Steps: The core objective is to achieve adaptive iterative optimization of the preprocessing strategy based on the full-process quality assessment results, thereby solving the problems of open-loop processing, long-term performance degradation, and reliance on manual maintenance in existing technologies.

[0108] Specifically, the closed-loop self-optimization process in this step includes five core components:

[0109] First, deviation classification and attribution: compare each quality assessment indicator with the preset target value, classify the indicators that exceed the standard, and match the corresponding parameter optimization path to achieve "precise problem location and targeted parameter optimization".

[0110] Specifically, the attribution rule for bias classification is: if the data is efficient If the value is lower than the target value, the deviation is attributed to the screening parameters of S2 and the preprocessing parameters of S3, forming an optimization feedback for the corresponding parameters; if the spatial deviation... If the deviation exceeds the target range, the bias is attributed to the spatial alignment parameter of S4, forming an optimization feedback for the fusion weights and calibration parameters; if the time deviation... If the deviation exceeds the target range, it is attributed to the time synchronization parameters of S4, forming an optimization feedback for the time compensation parameters; if the computing power utilization rate... If the deviation exceeds the target range, the deviation is attributed to the computing power-related parameters of S2 and S3, forming an optimized feedback on the computing power cost weight and the upper limit of the sampling frequency.

[0111] Second, global loss function construction: Under a unified framework, a global loss function is constructed to weigh all dimensions of quality indicators and quantify the overall quality deviation of the entire preprocessing process, providing a unified optimization target for parameter iteration.

[0112] Specifically, the global loss function formula is as follows:

[0113]

[0114] Definitions of parameters in the formula: weight allocation The importance of the four optimization objectives—data efficiency, spatial precision, temporal precision, and computational cost—is respectively assigned, and a value of [value to be filled in] is recommended. It satisfies the constraints. ; These are the target upper limits for spatial deviation, temporal deviation, and computing power utilization, respectively, used to normalize heterogeneous mass-dimensional indicators to... The range, with specific values ​​set according to the actual engineering requirements of the deployment.

[0115] Third, define the set of parameters to be optimized: clarify the range of parameters that need to be optimized throughout the entire process, and ensure that all core parameters that affect the quality of preprocessing are included in the closed-loop optimization system.

[0116] Specifically, let the set of parameters to be optimized be... Specifically, it includes three types of parameters: filtering parameters S2 includes scene type weight vectors, dynamic threshold function coefficients, traffic flow density adjustment coefficients, etc.; preprocessing parameters. The parameters in S3 include the piecewise sampling frequency function, the filter threshold reference value, the illumination sensitivity coefficient, and the occlusion compensation coefficient; and the spatiotemporal alignment parameters. The parameters in step S4 include the three-source fusion weight gain parameters, EKF calibration parameters, and time synchronization residual error trigger threshold.

[0117] The above three types of parameters together constitute the parameter vector. .

[0118] Gradient-based parameter update rule: Calculate the gradient of the global loss function with respect to the parameter vector, and use a first-order gradient optimization algorithm to complete the iterative parameter update, achieving parameter self-optimization without human intervention.

[0119] Specifically, the parameter update formula is as follows:

[0120]

[0121] The parameters in the formula are defined as follows: The global loss function for the parameter vector The gradient; For the first The effective learning rate for each iteration can be the initial learning rate. Obtained by combining the exponential decay strategy, i.e. This is the attenuation coefficient.

[0122] Preferably, the Adam optimizer is used to automatically adjust the effective step size of each parameter through first-order moment and second-order moment estimation, thereby improving convergence stability.

[0123] Furthermore, to ensure the physical rationality of the parameters, constraints are imposed on the updated parameters: the weight vector must be normalized and each component must be non-negative; the threshold and segment point parameters must be limited to a preset range; and the segment points must satisfy monotonicity constraints.

[0124] Fifth, edge-to-cloud hierarchical update and version management: In order to balance the rapid local adaptation of roadside nodes with the long-term global convergence of the entire network, a hierarchical update strategy is adopted, along with a security rollback mechanism to avoid system performance degradation caused by parameter updates.

[0125] Specifically, the tiered update strategy is as follows:

[0126] Local updates for edge nodes: Roadside edge nodes continuously monitor their own quality indicators. When an indicator deviates from the target range for several consecutive periods, local optimization is triggered. Only the historical data of this node is used to update a subset of local parameters. The updated parameters only apply to the current node, enabling rapid adaptation to local scenarios.

[0127] Global updates are performed in the cloud. The cloud periodically aggregates quality indicators and local parameter updates from multiple intersections, calculates global losses over a longer time scale, and uniformly optimizes shared parameters across the entire network. The new version parameter package is then distributed to each edge node to achieve overall optimization of network performance.

[0128] Optional, a supporting version management and security rollback mechanism: Semantic version numbers are used to manage parameters. New parameter versions are first verified in a local sandbox environment. If core quality indicators are found to be continuously deteriorating, the system will automatically roll back to the previous stable version and record the reasons for update failures to ensure the stability of system operation.

[0129] This step constructs a complete closed loop of "quality indicators → parameter updates → strategy adjustments → quality improvement", replacing the open-loop processing architecture of existing technologies. It can achieve continuous self-optimization of preprocessing strategies without manual intervention, ensuring the stability and accuracy of the system in long-term operation.

[0130] This embodiment also discloses a multi-source data processing system for vehicle-road-cloud collaborative perception, the system structure of which is shown in the figure below. Figure 2 As shown, the system includes a multi-source data acquisition module, an adaptive data source filtering module, a dynamic preprocessing module, a spatiotemporal alignment module, and a data output and quality assessment module, all connected in sequence, as well as a closed-loop self-optimization module that is connected in communication with all the aforementioned modules.

[0131] The multi-source data acquisition module is used to collect multi-source raw data from vehicles, roadsides, and the cloud, generate multi-source raw datasets, and construct real-time scene parameter vectors.

[0132] The adaptive data source filtering module is used to adaptively filter multi-source raw data based on scene parameter vectors and output a subset of valid data sources.

[0133] The dynamic preprocessing module is used to dynamically sample and adjust the effective data source subset based on the scene parameter vector and perform adaptive data cleaning, outputting the preprocessed dataset.

[0134] The spatiotemporal alignment module is used to accurately align the spatial coordinates and timestamps of the preprocessed dataset and output the spatiotemporally aligned dataset.

[0135] The data output and quality assessment module is used to standardize and encapsulate the spatiotemporally aligned dataset for output, and simultaneously calculate and output the quality assessment indicators of data processing.

[0136] The closed-loop self-optimization module is used to iteratively optimize the core operating parameters of the aforementioned modules based on quality assessment indicators, forming a closed-loop self-optimization of the data processing strategy.

[0137] Furthermore, the system adopts a three-layer collaborative deployment architecture consisting of vehicle-mounted terminals, roadside terminals, and the cloud:

[0138] Multi-source data acquisition modules are deployed in a distributed manner at the vehicle-mounted terminal, roadside terminal, and cloud, realizing distributed data acquisition across all scenarios;

[0139] The adaptive data source filtering module, dynamic preprocessing module, spatiotemporal alignment module, and data output and quality assessment module are deployed on the roadside edge computing node to achieve low-latency edge processing of data.

[0140] The closed-loop self-optimization module is deployed in two parts: the local iteration logic is deployed on the roadside edge nodes to achieve rapid local adaptation, and the global optimization and version management logic is deployed on the cloud platform to achieve global optimization across the entire network.

[0141] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any ordinary changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-source data processing method for vehicle-road-cloud collaborative sensing, characterized in that, Includes the following steps: S1, Multi-source data acquisition: Real-time acquisition of multi-source raw data from vehicle-mounted equipment, roadside equipment and cloud, generating multi-source raw datasets and constructing real-time scene parameter vectors; S2, Data source filtering: Based on multi-source original datasets and scene parameter vectors, a scene parameter-driven adaptive mapping mechanism for data sources is constructed. Data is filtered through a multi-factor weighted priority scoring function, and a subset of data sources is output. S3, Dynamic Preprocessing: Based on a subset of the data source and the scene parameter vector, the sampling frequency of the data source is dynamically adjusted, adaptive filtering and data cleaning are performed to obtain the preprocessed dataset; S4, Spatiotemporal Alignment: Based on the preprocessed dataset, spatiotemporal alignment is performed using a multi-source fusion localization and multi-level time synchronization mechanism, and the spatiotemporally aligned dataset is output. S5, Data Output and Quality Assessment: The spatiotemporally aligned dataset is packaged into a standard preprocessed data package in a unified data format and output as a quality assessment index. The data efficiency, spatial bias and temporal bias are calculated simultaneously and output as quality assessment indicators. S6, Closed-loop self-optimization: Based on the quality assessment index, deviation classification and attribution are performed, a global loss function of comprehensive quality assessment index is constructed, and the screening parameters of S2, the preprocessing parameters of S3 and the spatiotemporal alignment parameters of S4 are iteratively optimized and updated to form a closed-loop self-optimization of the data processing strategy.

2. The multi-source data processing method for vehicle-road-cloud collaborative perception according to claim 1, characterized in that, The scene parameter vector in S1 includes scene type encoding, traffic flow density, light intensity, occlusion factor, and GNSS signal quality factor. 3.The method of claim 2, wherein, The adaptive mapping mechanism for data sources in S2 includes scene parameter standardization processing and adaptive filtering rules based on a multi-factor weighted priority scoring function. Scene parameter standardization is used to unify parameters with different mass dimensions in the scene parameter vector to a preset value range; Adaptive filtering rules are used to filter valid data sources based on priority scoring results and dynamic filtering thresholds.

4. The method of claim 3, wherein, The dynamic adjustment of the data source sampling frequency in S3 is performed in segments based on traffic flow density and data source priority scores. The adaptive filtering and data cleaning dynamically adjust the filtering threshold based on light intensity and occlusion factor to remove noise and false data.

5. The method of claim 4, wherein, The multi-source fusion positioning described in S4 is multi-source fusion spatial alignment, which adopts a GNSS, IMU, and geomagnetic three-source fusion positioning scheme. Through a weighted fusion algorithm with adaptive signal quality allocation, combined with the projection transformation of high-precision electronic maps and intersection-level calibration, the spatial coordinates of multi-source data are unified.

6. The vehicle-road-cloud cooperative perception multi-source data processing method according to claim 5, characterized in that, The multi-level time synchronization mechanism in S4 is a three-level precise time synchronization architecture, which sequentially completes clock synchronization from the cloud to the roadside, transmission delay correction from the roadside to the vehicle, and performs linear interpolation compensation for residual time deviations to achieve precise and unified timestamps of multi-source data.

7. The vehicle-road-cloud cooperative perception multi-source data processing method according to claim 6, characterized in that, The standard preprocessed data package in S5 includes uniform coordinates, uniform timestamps, sensor identifiers, and confidence labels. 8.The method of claim 7, wherein, The deviation classification attribution in S6 refers to comparing the quality assessment indicators with the preset target values, classifying the indicators that exceed the standards, and matching the corresponding parameter optimization paths. The global loss function refers to a multi-dimensional quality assessment index that integrates data effectiveness, spatial bias, and temporal bias, used to quantify the overall quality bias of the entire preprocessing process.

9. The vehicle-road-cloud cooperative perception multi-source data processing method according to claim 8, characterized in that, The iterative optimization and update in step S6 includes defining the set of parameters to be optimized, gradient-based parameter update rules, and edge-cloud hierarchical update and version management mechanisms. The set of parameters to be optimized includes the filtering parameters of S2, the preprocessing parameters of S3, and the spatiotemporal alignment parameters of S4. Gradient-based parameter update rules complete parameter iteration by calculating the gradient of the global loss function; The edge-cloud hierarchical update and version management mechanism adopts an update strategy that combines rapid local iteration of roadside edge nodes with long-term global optimization in the cloud, while also providing parameter version management and anomaly rollback mechanisms.

10. A multi-source data processing system for vehicle-road-cloud collaborative perception, characterized in that, It includes a multi-source data acquisition module, an adaptive data source filtering module, a dynamic preprocessing module, a spatiotemporal alignment module, a data output and quality assessment module, and a closed-loop self-optimization module; The multi-source data acquisition module is used to collect multi-source raw data from vehicles, roadsides, and the cloud, generate multi-source raw datasets, and construct real-time scene parameter vectors. The adaptive data source filtering module is used to adaptively filter multi-source raw data based on scene parameter vectors and output a subset of valid data sources. The dynamic preprocessing module is used to dynamically sample and adjust the effective data source subset based on the scene parameter vector and perform adaptive data cleaning, outputting the preprocessed dataset. The spatiotemporal alignment module is used to accurately align the spatial coordinates and timestamps of the preprocessed dataset and output the spatiotemporally aligned dataset. The data output and quality assessment module is used to standardize and encapsulate the spatiotemporally aligned dataset for output, and simultaneously calculate and output the quality assessment indicators of data processing. The closed-loop self-optimization module is used to iteratively optimize the core operating parameters of the aforementioned modules based on quality assessment indicators, forming a closed-loop self-optimization of the data processing strategy.

Images