Rsm quality monitoring method and system based on socialized probes and video arbitration

Abstract

The application discloses a kind of RSM quality monitoring method and system based on socialized probe and video arbitration, which comprises the following steps: obtaining the roadside safety message broadcast by the target roadside unit and the perception data packet reported by the socialized probe vehicle, and performing space-time alignment;Based on the dynamic confidence model, the probe perception data is screened by group voting, and suspected false negative or false positive events are identified;For high-confidence suspected events, automatically retrieve roadside video stream for arbitration and root cause analysis;Based on the arbitration result, generate operation and maintenance work order, inject algorithm training set or adjust the weight of probe, and perform closed-loop optimization operation.The system correspondingly includes a socialized mobile probe group, roadside facilities, a video arbitration subsystem and a cloud-based quality monitoring platform.The present application uses a wide range of social vehicles as probes, combined with video true value arbitration, to achieve low-cost, wide-coverage, and normalized quality monitoring and accurate problem diagnosis of roadside safety messages, and continuously optimizes itself through a multi-closed-loop driving system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicles and vehicle-road cooperation technology, specifically involving an RSM quality monitoring method and system based on socialized probes and video arbitration. Background Technology

[0002] With the large-scale deployment of vehicle-to-everything (C-V2X) communication, roadside perception systems and their broadcast Resource Signals (RSMs) have become core data sources for ensuring traffic safety. The accuracy of RSMs directly determines the reliability of upper-layer applications such as collision warning and blind spot warning. Currently, monitoring RSM quality mainly faces the following bottlenecks: Coverage versus cost contradiction: relying on fixed high-precision truth systems (such as lidar arrays) or a small number of professional test vehicles makes it difficult to achieve wide-area, continuous coverage, and the construction and maintenance costs are extremely high.

[0003] Unable to perform routine maintenance: Existing methods are mostly project-based spot checks, which cannot continuously monitor the health of massive RSU nodes 24 / 7, and performance degradation or sudden failures cannot be detected in a timely manner.

[0004] Problem localization is difficult: After an anomaly is detected, it is difficult to quickly distinguish whether it is a sensor hardware failure, a perception algorithm defect, a communication link interruption, or environmental interference, resulting in low operation and maintenance efficiency and the inability to improve system reliability in a closed loop.

[0005] Therefore, the industry urgently needs a low-cost, scalable, and accurate RSM quality routine monitoring solution that can achieve accurate problem diagnosis. Summary of the Invention

[0006] The technical problem to be solved by this invention is that existing roadside safety message (RSM) quality monitoring methods are costly, have limited coverage, cannot be routinely maintained, and are difficult to locate problems.

[0007] This invention aims to overcome the aforementioned shortcomings and provides a method and system for monitoring RSM quality based on socialized probes and video arbitration. It utilizes widely available connected vehicles as mobile probes and combines roadside video for authoritative arbitration, thereby achieving a low-cost, wide-coverage, routine monitoring and closed-loop optimization method and system for RSM data quality. See below for details: This invention provides an RSM quality monitoring method based on social probes and video arbitration, comprising the following steps: Acquire roadside safety messages broadcast by the target roadside unit and data packets reported by probe vehicles in the socialized mobile probe group. The data packets include at least vehicle identification, timestamp and location information, vehicle-side sensing target list and received original roadside safety messages. Based on a unified spatiotemporal reference, the data packets reported by the probe vehicle and the roadside safety messages broadcast by the target roadside unit are aligned in time and space. Based on a dynamic confidence model, a group voting screening is performed on the probe vehicle perception data that has been spatiotemporally aligned to identify missed or false alarm events of the roadside safety messages relative to the probe perception results, and to generate high-confidence suspected abnormal events. For the high-confidence suspected abnormal events, the corresponding roadside video stream is automatically retrieved for arbitration to confirm the authenticity of the event and analyze the root cause of the anomaly. Based on the video arbitration confirmation result and the root cause of the anomaly, perform at least one of the following operations: generate an operation and maintenance work order, inject the arbitration confirmed case data into the algorithm training set, or dynamically adjust the dynamic confidence weight of the probe vehicle.

[0008] Optionally, the socialized mobile probe group includes connected vehicles equipped with dual-mode smart rearview mirrors, ADAS functions, or advanced intelligent driving functions.

[0009] Optionally, the group voting screening based on the dynamic confidence model includes: Assign dynamic confidence weights to each probe vehicle based on its sensor configuration, vehicle automation level, and historical data performance. For the same spatiotemporal region, if the following conditions are met: the sum of the products of the perception confidence of each probe vehicle for the same target A and its current dynamic confidence weight is greater than or equal to a preset threshold, and the roadside safety message does not contain target A, then it is determined to be a missed event for target A.

[0010] Optionally, the group voting screening based on the dynamic confidence model further includes: For the same spatiotemporal region, if the roadside safety message contains target B, and the sum of the products of the perception confidence of each probe vehicle for target B and its current dynamic confidence weight is less than a preset threshold, it is determined to be a false alarm event for target B.

[0011] Optionally, the dynamic adjustment of the dynamic confidence weight of the probe vehicle includes: Based on the final result of the video arbitration, the weights of the probe vehicles involved in the event determination will be retrospectively adjusted. If the perceived target reported by the probe vehicle is confirmed by video arbitration, its dynamic confidence weight is increased; if the perceived target reported by the probe vehicle is not confirmed by video arbitration, its dynamic confidence weight is decreased.

[0012] Optionally, the analysis of the root cause of the anomaly includes at least one of the following: determining whether the sensor is obstructed or damaged, determining whether the perception algorithm has ghosting or erroneous tracking, and determining whether the traffic event recognition algorithm has failed.

[0013] Optional, also includes: Based on the missed and false alarm events confirmed by video arbitration, the detection rate and false alarm rate of roadside safety messages for the target roadside unit are calculated.

[0014] Optionally, the preset threshold is dynamically adjusted based on at least one of the following factors: traffic flow at the intersection or road segment, time and ambient lighting conditions, and target type.

[0015] Optionally, the automatic retrieval of the corresponding roadside video stream for arbitration includes: Computer vision algorithms are used to preprocess video clips to generate an analysis report containing a list of targets and comparison results with probe data and roadside safety messages; The analysis report, along with the original video, probe data, and roadside safety messages, are pushed to the adjudicator's interface so that the adjudicator can make a ruling within a preset time.

[0016] This invention also provides an RSM quality monitoring system based on socialized probes and video arbitration, comprising: Socialized mobile probe groups consist of various intelligent connected vehicles connected to the vehicle-road-cloud platform, used to report their own perception data and received roadside safety messages via cellular networks; Roadside sensing and communication facilities, including the roadside units to be evaluated, associated sensors, and edge computing units; The roadside video arbitration subsystem consists of high-definition cameras at intersections, video storage servers, and video analytics services. Cloud-based quality monitoring platform; The cloud-based quality monitoring platform is used to receive data from the socialized mobile probe group and the target roadside unit in the roadside sensing and communication facility, and to perform spatiotemporal alignment on the data; The cloud-based quality monitoring platform is used to perform group voting screening on the aligned probe vehicle perception data based on a dynamic confidence model in order to identify suspected abnormal events in the roadside safety messages broadcast by the target roadside unit. The cloud-based quality monitoring platform is used to trigger the roadside video arbitration subsystem to retrieve video streams for arbitration and root cause analysis in response to identified high-confidence suspected abnormal events. The cloud-based quality monitoring platform is used to perform closed-loop optimization operations based on the results of video arbitration confirmation. The cloud-based quality monitoring platform or the roadside video arbitration subsystem is pre-configured with a roadside device mapping database. The roadside device mapping database records the geographical location, field of view, and correlation between cameras and roadside units, and is used to query and lock the corresponding camera based on the spatiotemporal coordinates of suspected abnormal events in order to retrieve the video stream.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Achieved low-cost, wide-coverage, and scalable routine monitoring: This invention constructs a widely distributed and dynamically supplementable monitoring network by using a massive number of socialized intelligent connected vehicles as mobile sensing probes and defining a unified data interface. This fully utilizes existing vehicle resources, avoids the huge costs associated with deploying and maintaining fixed high-precision truth-checking systems, and makes it possible to conduct continuous 24 / 7 quality monitoring of all roadside units within the area. Furthermore, as the penetration rate of connected vehicles increases, the density and accuracy of the monitoring network can automatically improve.

[0018] 2. Improved automation and accuracy of anomaly screening: This invention assigns and dynamically optimizes confidence weights to probe vehicles with different capabilities, and uses group voting decisions based on the product of weights and perceived confidence. This allows the system to efficiently and automatically identify high-confidence suspected missed or false alarm events from massive amounts of data. This mechanism not only achieves automated preliminary screening of roadside safety messages but also enables the probe network to have a "self-cleaning" capability, automatically amplifying the impact of high-reliability data sources and suppressing noisy data, thereby ensuring the continuous improvement of the reliability of the screening foundation and the efficiency of assessment.

[0019] 3. Provides authoritative arbitration and precise problem localization capabilities: For high-confidence suspected events identified through group voting, the system automatically triggers and retrieves the corresponding roadside high-definition video stream for final arbitration, upgrading quality assessment from probabilistic statistical inference to factual determination based on visual truth. Combined with video footage, it enables rapid and accurate preliminary diagnosis of the cause of the anomaly, precisely pinpointing whether the problem stems from hardware failure, algorithmic defects, or environmental interference, greatly improving the targetedness and efficiency of maintenance responses.

[0020] 4. A multi-loop system driving continuous optimization and self-evolution has been established: Based on the results of video arbitration confirmation and root cause analysis, the system of this invention can automatically generate actionable maintenance work orders, forming an "operation and maintenance loop." Simultaneously, typical cases confirmed by arbitration (including video, sensor data, and error messages) are automatically labeled and injected into the roadside perception algorithm training set, driving targeted optimization of the algorithm model, forming an "algorithm loop." Furthermore, a mechanism for dynamically adjusting probe vehicle weights based on arbitration results continuously optimizes the evaluation quality of the probe network itself, forming a "probe network optimization loop." These three loops work together to enable the entire monitoring system and its evaluation object (roadside perception system) to continuously learn and improve using defects discovered in real-world operation, achieving continuous self-evolution and enhancing the system's intelligence level.

[0021] In summary, this invention organically combines social resource utilization, intelligent collaborative perception, authoritative video verification, and data-driven closed-loop, which not only significantly lowers the implementation threshold for routine monitoring of high-quality roadside safety information, but also makes the monitoring conclusions authoritative and actionable, and ultimately promotes the development of the entire vehicle-road cooperative system towards becoming more reliable and intelligent with use. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a complete system framework diagram provided by the present invention.

[0024] Figure 2 This is a complete method flowchart provided by the present invention.

[0025] Figure 3 This is a timing diagram of video arbitration triggered by group voting provided by the present invention.

[0026] Figure 4 This is a schematic diagram of scenario example one provided by the present invention. Detailed Implementation

[0027] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 This embodiment provides an RSM quality monitoring method based on socialized probes and video arbitration. (See also...) Figure 2 The method flowchart shown below specifically includes the following steps: 1. Standardized access to multi-source data The cloud-based quality monitoring platform continuously receives Roadside Safety Message (RSM) streams broadcast from specific target Roadside Units (RSUs) to be evaluated. Simultaneously, the platform receives data packets reported from a broad network of socialized mobile probes. Data reporting is triggered by specific events, primarily through the following mechanisms: when a probe vehicle receives an RSM message broadcast by a target RSU, it triggers a data report containing the associated context; and when the probe vehicle's onboard perception system identifies a pre-defined traffic event (such as an obstacle, pedestrian, or abnormal parking), it also triggers a data report.

[0029] This probe group consists of various intelligent connected vehicles connected to the existing vehicle-road-cloud platform. These vehicles, acting as mobile perception probes, include vehicles equipped with specific dual-mode intelligent rearview mirrors, vehicles with Advanced Driver Assistance Systems (ADAS) functions, and vehicles with advanced intelligent driving functions. The probe vehicles receive messages broadcast by roadside units via the PC5 direct communication interface or the Uu cellular network interface and report data via the Uu cellular network.

[0030] The platform defines a unified interface specification for vehicle-side perception data reporting, based on the industry standard T / CSAE 53-2020. The data packets reported by probe vehicles must include at least the following fields: a message header containing a message ID, timestamp, unique vehicle identifier, and geographic location; vehicle status information including speed, heading angle, acceleration, and braking status; a list of perceived targets generated by the vehicle-side perception system; and the original roadside safety messages and their source identifiers received synchronously by the probe vehicle. Each target record in the perceived target list includes a target ID, target type (vehicle, person, cyclist, etc.), coordinates relative to the vehicle (X, Y, Z offset), relative speed, absolute speed, heading angle, target dimensions (length, width, height), and the presence confidence level given by the vehicle-side perception algorithm.

[0031] 2. Spatiotemporal alignment of accessed multi-source data To ensure accurate comparison of data from different sources, all data accessed to the platform, including message streams from target roadside units, data packets reported by each probe, and time indexes from the roadside video subsystem, are synchronized based on satellite timing signals. The cloud platform performs millisecond-level time and space alignment on all data under a unified spatiotemporal reference. This step ensures that all subsequent analyses are conducted on traffic scenarios within the same geographic area at the same time, laying a solid foundation for accurate assessment.

[0032] 3. Preliminary screening of group voting based on dynamic confidence model This step utilizes a large amount of aligned probe sensing data to perform automated and intelligent preliminary anomaly detection on messages broadcast by the target roadside unit. Its core is to build a dynamic confidence model with self-learning capabilities for group decision-making.

[0033] First, the system assigns and maintains a dynamic confidence weight for each probe vehicle or each type of probe vehicle. The initial weight is set based on several factors: the vehicle's sensor hardware configuration (e.g., whether it is equipped with LiDAR), the level of automation as defined in the national standard "Classification of Driving Automation for Automobiles" (GB / T 40429-2021) that its autonomous driving function conforms to, and the comprehensive score of the vehicle's historical data performance. In a preferred embodiment, the mapping relationship between the initial weight and the level of automation is as follows: The initial weight of L1 (basic reference layer) level vehicles (including aftermarket smart rearview mirrors, etc.) is set to 0.4; The initial weight of L2 (baseline) level vehicles is set to 0.6; The initial weight for L3 (Enhanced Trust Layer) vehicles is set to 0.8; The initial weight of L4 (high confidence level) vehicles is set to 0.9.

[0034] Sensor configuration differences can be fine-tuned based on the weighting of this level, but the overall configuration follows the benchmarks mentioned above.

[0035] These weights are not fixed but are continuously and dynamically optimized based on actual performance. The system establishes a continuous learning loop, retrospectively adjusting the weights of the probe vehicles participating in each event determination based on the final factual results of video arbitration in subsequent steps. The adjustment rule adopts a linear update method with boundary control, as follows: When a perceived target reported by a vehicle is confirmed by video arbitration, its weight is increased by the first reward value. ; When a perceived target reported by a vehicle is disproven by video arbitration, its weight is reduced by the second penalty value. ,and To strengthen the suppression of erroneous data; Even if a vehicle does not report a target, but an unreported event confirmed by arbitration occurs within the spatiotemporal area it passes through, and the vehicle has the ability to perceive the target, a small reward may be given for the contribution of the unreported event. .

[0036] In a preferred embodiment, set , , And set a weight cap. Lower limit The updated weights are restricted to this range.

[0037] For example, if an L2 vehicle with an initial weight of 0.6 reports a target that is confirmed once, the weight is updated to min(0.6+0.02, 1.5)=0.62; if it is disproven once, the weight is updated to max(0.6-0.05, 0.2)=0.55.

[0038] To scientifically evaluate the historical data performance of vehicles, the system constructs a comprehensive evaluation score calculated by weighting arbitration consensus rate and effective contribution rate, which serves as one of the bases for initial weight allocation and a reference for dynamic adjustment.

[0039]

[0040] Where W1 = 0.7 and W2 = 0.3.

[0041] The evaluation window N adopts a dynamic window strategy based on the amount of data. However, when the amount of data is insufficient, the maximum backtracking time limit is enabled, and data within a maximum of 90 days can be backtracked.

[0042] Based on the aforementioned dynamic weights, the system performs a group voting screening. For a specific spatiotemporal slice, such as the state of an intersection within a few seconds, the platform aggregates all probe data passing through that area. The screening logic is precisely expressed through a mathematical model: (1) Screening for suspected underreporting events: If for a real target A, the following formula is satisfied: If target A is not included in the roadside safety message broadcast by the target roadside unit, a high-confidence suspected missed event is generated.

[0043] (2) Screening for suspected false alarms: If the roadside safety message broadcast by the target roadside unit reports target B, but meets the following conditions... If so, a high-confidence suspected false alarm event is generated.

[0044] It should be noted that the preset threshold Instead of a globally fixed value, a multi-dimensional dynamic adjustment strategy is adopted to adapt to the perception difficulty and probe density in different scenarios: a. Adjust based on intersection / segment characteristics: Set basic thresholds And adjusted by the flow correction factor, i.e. If the base threshold T_base is set to 3 and the busy intersection correction factor is set to 1.5, then the actual threshold is 4.5, which is rounded up to 5.

[0045] b. Adjust according to time and ambient light: Differentiate between daytime, nighttime, and inclement weather, and set... , , .

[0046] c. Adjust by target type: Set different thresholds for different types of traffic participants. , , For objectives that are more difficult to perceive, a higher degree of consensus in group voting is required.

[0047] The threshold setting adopts a phased strategy of experience-based startup and data-driven optimization: In the initial stage of system deployment, a conservative initial threshold is determined based on domain knowledge and traffic simulation models to ensure safe system startup; as the system operates, real cases confirmed by video arbitration are continuously accumulated, and statistical learning or lightweight machine learning models (such as decision trees) are used, with intersection type, time period, weather, target type, probe density, etc. as input features, to automatically calculate and dynamically adjust the optimal threshold for each scenario with the goal of optimizing monitoring effectiveness and false alarm rate. The logical framework, safety boundaries, and initial values ​​for threshold adjustment are defined manually, while the iterative optimization of specific values ​​is automatically completed by data-driven processes.

[0048] This mechanism enables the screening network to "self-cleanse," automatically reinforcing highly reliable data sources, suppressing noise, and continuously improving the efficiency and accuracy of the entire assessment network over time.

[0049] 4. Video arbitration linkage and precise root cause localization See Figure 3 The diagram shows the sequence of events triggered by the group vote. For the aforementioned high-confidence suspected anomalies, the cloud platform automatically triggers the video arbitration process. The platform automatically generates precise video retrieval instructions, which include information such as the time of the incident, geographical location, and the angle to be observed, and requests the original video stream or video clips from the designated roadside camera during a specific time period before and after the incident from the roadside video arbitration subsystem.

[0050] The continuous high-definition video stream provided by the roadside video arbitration subsystem serves as the undisputed source of ultimate truth. Using the objective facts presented in the video footage as the basis for arbitration, the authenticity of suspected events is confirmed, thereby elevating quality assessment from a "speculation" level based on statistical probability to a "fact-finding" level based on visual evidence. This is the key to the authoritative conclusion reached in this invention.

[0051] Based on the arbitration confirming the authenticity of the incident, further preliminary root cause analysis can be conducted using video footage to quickly pinpoint the source of the RSM quality problem. In this embodiment, the analysis of abnormal root causes includes at least one of the following: (1) When the video clearly shows the presence of a traffic participant (such as a pedestrian or non-motorized vehicle), but the RSM does not report it, the field of view of the corresponding roadside sensor in the video can be further checked for physical obstruction, lens damage, or abnormal lighting conditions such as overexposure or underexposure.

[0052] (2) When the video confirms that there is no object at a certain location, but the RSM continues to report the presence of a target (such as a stationary vehicle), it can be determined that the roadside perception algorithm is generating a persistent "ghost" or erroneous tracking under specific environmental conditions (such as nighttime, changes in light and shadow).

[0053] (3) When the video footage and the reports from most high-weight probe vehicles show that a certain road segment has serious congestion or accidents, but the RSM does not report any traffic incidents, it can be determined that the traffic incident recognition algorithm of the roadside system has failed.

[0054] This step extends simple anomaly detection to problem diagnosis.

[0055] 5. Quality Quantitative Assessment and Visual Diagnosis Based on a series of verified cases of missed and false alarms confirmed by video arbitration, the system can calculate a set of authoritative and quantifiable quality assessment indicators to objectively measure the performance of the target roadside unit. These indicators include, but are not limited to: (1) Detection rate (Recall), calculated using the following formula: TP represents actual alarms, and FN represents missed alarms.

[0056] (2) False Alarm Rate, calculated using the following formula: FP represents a false alarm, and TN represents a real no-alarm situation.

[0057] In addition, this embodiment also statistically analyzes the distribution of average positioning error and data update delay.

[0058] The calculation results are displayed through a cloud-based visualization dashboard, allowing maintenance personnel to perform drill-down analysis and global overviews based on multiple dimensions such as intersections, time periods, and event types, thereby accurately grasping the health status and performance trends of the entire roadside perception system.

[0059] 6. Three-loop optimized execution Based on the video arbitration results and preliminary root cause analysis, the system initiates three optimization loops in parallel, driving the entire monitoring system and the monitored roadside systems to continuously evolve towards higher performance.

[0060] (1) Operation and maintenance closed loop: The platform automatically generates diagnostic reports and specific operation and maintenance work orders. The work orders adopt a structured data format and contain information in at least the following dimensions: Problem location information: unique work order identifier, associated intersection / road segment identifier, target roadside unit identifier, and problem occurrence timestamp; Problem diagnostic information: Fault type (including missed or false alarms), affected target type, suspected root cause label; Handling guidance information: Key evidence chains, including associated video clip Uniform Resource Locators, problematic data snapshots, or original message fragments; Work order management information: work order generation time, dispatch recipient, current status, and required completion deadline.

[0061] Work orders can be automatically pushed to the existing roadside facility management platform via application programming interface, or exported as structured data files (such as JSON / XML) for manual distribution.

[0062] Maintenance personnel perform precise repairs such as sensor calibration, equipment cleaning, and algorithm troubleshooting based on the work order, thereby quickly restoring the system to health.

[0063] (2) Algorithm Closed Loop: The platform uses an automated human-machine collaborative process to process typical cases confirmed by arbitration, forming a continuously enriched dataset of challenging scenarios for roadside perception algorithms. Specifically: Automatic data extraction and packaging: Based on the arbitration event identifier, the system automatically extracts the associated original roadside safety messages, original outputs of roadside sensors (such as image detection boxes and radar point cloud lists), arbitration video clips, and data reported by probe vehicles to form a case data package; Preliminary machine annotation: Using a pre-trained visual perception model, target detection and tracking are performed on video frames to generate preliminary annotation files containing target bounding boxes, categories, labels, and motion attributes; Manual review and confirmation: The annotation personnel review, correct, or supplement the preliminary annotation results to form a high-precision standard truth value annotation file; One-click import: Import the complete case data package and its standard annotation files, after review and confirmation, into the algorithm training platform through a unified interface.

[0064] A typical data packet contains at least the following fields: a. Metadata: Unique case identifier, intersection identifier, roadside unit identifier, sensor identifier, precise timestamp, weather conditions, lighting conditions; b. Problem input: The full text of the original erroneous roadside safety message that triggered the arbitration, and the original perception results output by the roadside perception system (including the original detection list of each sensor without fusion). c. Arbitration truth: The associated video clips or video frame sequences, and the standardized annotation file generated based on the video content, which records the pixel-level bounding boxes, categories, target identifiers, motion attributes and other information of all traffic targets in the spatiotemporal area; d. Related context: The perception data packet reported by the probe vehicle that triggered this arbitration must contain at least its list of perceived targets; e. Diagnostic labels: Problem types and suspected root cause labels automatically marked by the system based on the root cause analysis results, in a format such as "Missed Reports - Pedestrians - East Side Camera - Low Light".

[0065] This dataset is continuously used to drive the directional optimization, iterative training, and release of new versions of the roadside perception model, thereby directly improving the perception accuracy and robustness of the roadside system at the algorithm level.

[0066] (3) Probe Network Optimization Closed Loop: As mentioned above, the system continuously and dynamically adjusts the confidence weights of the probe vehicles participating in the judgment based on the final result of each video arbitration. This process is a continuous calibration and optimization of the "evaluation tool itself"—that is, the socialized probe network. It ensures that the evaluation basis of the entire monitoring system can automatically identify and strengthen high-reliability data sources, suppress the interference of noise sources, and enable the reliability and accuracy of the monitoring system itself to continuously improve with the increase of operating time, realizing the self-evolution of the evaluation tool.

[0067] To illustrate the above method and process in more detail, this embodiment will be described below with two typical scenario examples.

[0068] (1) Detection and handling of roadside radar missed electric bicycle incidents (Scenario Example 1).

[0069] See Figure 4 At a city intersection, the radar deployed on its south side experienced a decrease in sensitivity to detect smaller electric bicycles due to mechanical displacement, resulting in missed detections.

[0070] During a one-hour monitoring period, eight probe vehicles with high initial weights (such as those equipped with LiDAR) passed through the area. The vehicle-mounted perception systems of these vehicles independently reported the presence of electric bicycles in the west-side non-motorized vehicle lane. The cloud platform spatiotemporally aligned these reports and performed a weighted summation based on the current dynamic weights of each vehicle, resulting in a value far exceeding the preset threshold T. However, the RSM messages broadcast by the target roadside units at the intersection during the same period did not contain any electric bicycle targets.

[0071] Therefore, the system triggered a high-confidence suspected missed event. The platform then automatically retrieved the video stream from the camera on the west side of the intersection during that time period. After arbitration, it was confirmed that there was indeed a continuous cyclist in the video, thus confirming the RSM missed event.

[0072] The closed-loop optimization mechanism responded immediately: a specific calibration work order was generated for the radar on the west side of the intersection and pushed to the operation and maintenance system (operation and maintenance closed loop). Simultaneously, the video clips from this confirmed case, the original radar point cloud data, and the RSM messages for the missing targets were injected into the "non-motorized vehicle detection" specific training set (algorithm closed loop). Furthermore, the confidence weights of the eight probe vehicles that accurately reported the issue received a slight increase due to the reliability of their data; while system logs showed that another vehicle with license plate number VIN_101, although initially having a high weight, had its weight automatically reduced after this incident due to multiple false alarms recently reported that were not verified by video (probe network optimization closed loop). Through this complete process, not only was the specific problem located and output, but the probe network was also optimized simultaneously, making future screening of this area more accurate.

[0073] (2) Identification and suppression of false alarms of "ghosting" at night by roadside algorithm (Scenario Example 2) Under conditions of insufficient lighting at night, a roadside unit continuously broadcasts a non-existent "stationary vehicle" target to passing vehicles, whose location is fixed at a certain point on the roadside.

[0074] During several days of monitoring, although the RSU continuously reported this target, no probe vehicle passing through the location reported its presence in its reported data packets. Based on the false alarm judgment logic in the group voting screening, the system triggered a high-confidence suspected false alarm event.

[0075] The platform automatically retrieved the video stream from the corresponding time point and location at night for arbitration. The video clearly showed that the location was empty, confirming it as a false alarm. Based on scene characteristics, root cause analysis determined that this was a misjudgment of a persistent "ghosting" effect caused by the roadside perception algorithm in low-light conditions, resulting in a fixed background or shadow.

[0076] Closed-loop optimization was then initiated: generating algorithm problem investigation work orders (operational closed loop). More importantly, the nighttime radar point cloud data, video images, and erroneous RSM messages corresponding to this "ghosting" feature were added as high-quality negative samples to the false alarm suppression library of the perception algorithm training set (algorithm closed loop). This will be used in subsequent algorithm model iterations to specifically enhance the ability to distinguish such low-light "ghosting" scenes, fundamentally suppressing the occurrence of similar false alarms in the future.

[0077] Example 2 This embodiment provides an RSM quality monitoring system based on socialized probes and video arbitration for implementing the method described in Embodiment 1. See also... Figure 1 The system architecture diagram shown indicates that the system mainly includes the following components: Socialized Mobile Probe Cluster: This probe cluster consists of various intelligent connected vehicles widely connected to the existing vehicle-road-cloud platform, forming the system's wide-area perception layer. These vehicles, acting as distributed mobile sensor nodes, are specifically used to periodically report their perception data and received roadside safety messages to the cloud via cellular networks. The probe vehicles cover connected vehicles with different technical configurations, such as vehicles equipped with smart rearview mirrors featuring PC5 and Uu dual-mode communication capabilities (whose perception capabilities are substantially no different from ordinary ADAS vehicles, but can receive messages broadcast by roadside units through the PC5 interface), vehicles with ADAS functions, and vehicles with advanced intelligent driving functions, thus forming a hierarchical and scalable mobile probe network with enhanced perception capabilities.

[0078] Roadside sensing and communication facilities: This part of the facilities is one of the data sources. It includes the target roadside unit whose broadcast message quality is to be evaluated, various sensors that work in conjunction with the target roadside unit, and edge computing units deployed on the roadside. Sensors include cameras, radar, lidar, etc., which are responsible for raw environmental perception; edge computing units are used to perform preliminary processing of the perceived data and generate roadside safety messages.

[0079] Roadside Video Arbitration Subsystem: This subsystem consists of high-definition cameras deployed at key intersections or road sections, cloud or edge video storage servers, and hybrid collaborative video analysis services. To achieve accurate video evidence association, the system pre-builds and continuously maintains a digital twin mapping database for roadside equipment. This database records the precise geographical location, orientation angle, field of view model, and inter-device relationships of each camera and roadside unit.

[0080] Its main function is to provide searchable historical video stream data, and when it receives instructions from the cloud quality monitoring platform to locate the corresponding camera through spatial query based on the spatiotemporal coordinates of a suspected event, it retrieves historical video clips from storage within a specific time window before and after the event and executes an arbitration process that combines machine pre-analysis with online human adjudication.

[0081] Specifically, the workflow of the hybrid collaborative video analytics service is as follows: When an arbitration request is received, the system first automatically calls computer vision algorithms to preprocess the associated video segments and generate an "analysis brief" that includes a list of targets in the image, their motion trajectories and preliminary comparison results with probe data and roadside safety messages. Subsequently, the briefing, along with the original video, probe data, and roadside safety messages, was pushed to the dedicated adjudicator's interface. The adjudicator comprehensively reviews all information and makes a final ruling within a preset time (usually set to 30 to 60 seconds), confirming the authenticity of missed or false reports and classifying the root causes.

[0082] This hybrid approach not only significantly improves the efficiency of processing massive amounts of video by leveraging machines, but also ensures the authority and reliability of the arbitration results by relying on human intelligence, thus avoiding the circular reasoning problem of confidence levels caused by pure algorithm comparison.

[0083] Cloud-based quality monitoring platform: This platform communicates with the socialized mobile probe network, roadside sensing and communication facilities, and the roadside video arbitration subsystem. The cloud-based quality monitoring platform is specifically configured to perform a series of operations to coordinate the work of the above components and implement complete monitoring logic: (1) The platform is used to receive data packets reported from the socialized mobile probe group and raw roadside safety message streams broadcast from the target roadside unit in the roadside sensing and communication facility.

[0084] (2) The platform is used to perform millisecond-level time and space alignment operations on the received multi-source heterogeneous data based on a unified spatiotemporal reference for satellite time synchronization.

[0085] (3) The platform is used to run a dynamic confidence model, assign and dynamically adjust weights to each probe vehicle, and perform group voting calculation on the aligned probe perception data based on this model, so as to intelligently identify suspected missed or false alarm events of target roadside unit messages.

[0086] (4) The platform is used to automatically generate instructions and trigger the roadside video arbitration subsystem to retrieve the corresponding video stream for high-confidence suspected abnormal events, and then conduct event authenticity arbitration and preliminary root cause analysis based on the video.

[0087] (5) The platform is used to perform three core closed-loop optimization operations based on the final result of video arbitration confirmation: automatically generating and dispatching maintenance work orders, injecting confirmed cases into the algorithm training dataset, and dynamically adjusting the confidence weight of the probe network.

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for RSM quality monitoring based on socialized probes and video arbitration, characterized in that, Includes the following steps: Acquire roadside safety messages broadcast by the target roadside unit and data packets reported by probe vehicles in the socialized mobile probe group. The data packets include at least vehicle identification, timestamp and location information, vehicle-side sensing target list and received original roadside safety messages. Based on a unified spatiotemporal reference, the data packets reported by the probe vehicle and the roadside safety messages broadcast by the target roadside unit are aligned in time and space. Based on a dynamic confidence model, a group voting screening is performed on the probe vehicle perception data that has been spatiotemporally aligned to identify missed or false alarm events of the roadside safety messages relative to the probe perception results, and to generate high-confidence suspected abnormal events. For the high-confidence suspected abnormal events, the corresponding roadside video stream is automatically retrieved for arbitration to confirm the authenticity of the event and analyze the root cause of the anomaly. Based on the video arbitration confirmation result and the root cause of the anomaly, perform at least one of the following operations: generate an operation and maintenance work order, inject the arbitration confirmed case data into the algorithm training set, or dynamically adjust the dynamic confidence weight of the probe vehicle.

2. The method according to claim 1, characterized in that, The socialized mobile probe group includes connected vehicles equipped with dual-mode smart rearview mirrors, ADAS functions, or advanced intelligent driving functions.

3. The method according to claim 1, characterized in that, The group voting screening based on the dynamic confidence model includes: Assign dynamic confidence weights to each probe vehicle based on its sensor configuration, vehicle automation level, and historical data performance. For the same spatiotemporal region, if the following conditions are met: the sum of the products of the perception confidence of each probe vehicle for the same target A and its current dynamic confidence weight is greater than or equal to a preset threshold, and the roadside safety message does not contain target A, then it is determined to be a missed event for target A.

4. The method according to claim 3, characterized in that, The group voting screening based on the dynamic confidence model also includes: For the same spatiotemporal region, if the roadside safety message contains target B, and the sum of the products of the perception confidence of each probe vehicle for target B and its current dynamic confidence weight is less than a preset threshold, it is determined to be a false alarm event for target B.

5. The method according to claim 1, characterized in that, The dynamic adjustment of the dynamic confidence weight of the probe vehicle includes: Based on the final result of the video arbitration, the weights of the probe vehicles involved in the event determination will be retrospectively adjusted. If the perceived target reported by the probe vehicle is confirmed by video arbitration, its dynamic confidence weight is increased; if the perceived target reported by the probe vehicle is not confirmed by video arbitration, its dynamic confidence weight is decreased.

6. The method according to claim 1, characterized in that, The analysis of the root causes of anomalies includes at least one of the following: determining whether the sensor is obstructed or damaged, determining whether the perception algorithm has ghosting or erroneous tracking, and determining whether the traffic event recognition algorithm has failed.

7. The method according to claim 1, characterized in that, Also includes: Based on the missed and false alarm events confirmed by video arbitration, the detection rate and false alarm rate of roadside safety messages for the target roadside unit are calculated.

8. The method according to claim 3, characterized in that, The preset threshold is dynamically adjusted based on at least one of the following factors: traffic flow at the intersection or road segment, time and ambient lighting conditions, and target type.

9. The method according to claim 1, characterized in that, The automatic retrieval of the corresponding roadside video stream for arbitration includes: Computer vision algorithms are used to preprocess video clips to generate an analysis report containing a list of targets and comparison results with probe data and roadside safety messages; The analysis report, along with the original video, probe data, and roadside safety messages, are pushed to the adjudicator's interface so that the adjudicator can make a ruling within a preset time.

10. An RSM quality monitoring system based on socialized probes and video arbitration, characterized in that, include: Socialized mobile probe groups consist of various intelligent connected vehicles connected to the vehicle-road-cloud platform, used to report their own perception data and received roadside safety messages via cellular networks; Roadside sensing and communication facilities, including the roadside units to be evaluated, associated sensors, and edge computing units; The roadside video arbitration subsystem consists of high-definition cameras at intersections, video storage servers, and video analytics services. Cloud-based quality monitoring platform; The cloud-based quality monitoring platform is used to receive data from the socialized mobile probe group and the target roadside unit in the roadside sensing and communication facility, and to perform spatiotemporal alignment on the data; The cloud-based quality monitoring platform is used to perform group voting screening on the aligned probe vehicle perception data based on a dynamic confidence model in order to identify suspected abnormal events in the roadside safety messages broadcast by the target roadside unit. The cloud-based quality monitoring platform is used to trigger the roadside video arbitration subsystem to retrieve video streams for arbitration and root cause analysis in response to identified high-confidence suspected abnormal events. The cloud-based quality monitoring platform is used to perform closed-loop optimization operations based on the results of video arbitration confirmation. The cloud-based quality monitoring platform or the roadside video arbitration subsystem is pre-configured with a roadside device mapping database. The roadside device mapping database records the geographical location, field of view, and correlation between cameras and roadside units, and is used to query and lock the corresponding camera based on the spatiotemporal coordinates of suspected abnormal events in order to retrieve the video stream.

Images