Vehicle intelligent auxiliary driving system correction device and system

Abstract

The invention discloses a vehicle intelligent auxiliary driving system correction system and correction method. A vehicle measurement module, a multi-mode combination identifier, a field end truth value measurement module, an evaluation and correction module and a report generation module are arranged. The vehicle measurement module is used for completing sensor self-inspection and ensuring attitude initialization when a vehicle enters a parking lot; the multi-modal combination identifier can be identified by multiple types of equipment at the same time, so that a cross-modal unified target is realized; the field end truth value measurement module obtains the real positions and postures of the obstacle and the vehicle through various means; and the evaluation and correction module establishes geometric, semantic and cross-modal consistency indexes, and online correction is realized by adopting a progressive small-step correction and rollback mechanism when deviation is found. By means of the design, automatic checking and continuous optimization of the intelligent driving system can be efficiently achieved in a real site environment, the sensing precision and recognition stability of the sensor are improved, and the problems that in the prior art, unified identification is lacked, an independent verification channel is lacked, and correction is not stable are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle intelligent driving and environment perception system, in particular to a vehicle intelligent auxiliary driving system correction system and a correction method. BACKGROUND

[0002] With the rapid development of intelligent driving and advanced driving assistance system (ADAS), the vehicle's ability to perceive the environment has become a key link to ensure driving safety and improve driving experience. Current mainstream vehicles generally carry cameras, laser radars, millimeter wave radars, and inertial measurement units, and realize target detection, obstacle recognition and path planning through data fusion. However, sensors are prone to problems such as external parameter drift, time synchronization deviation and recognition accuracy decline during long-term operation, which leads to a gradual deviation between the vehicle's perception results and the actual environment, and further affects the reliability of the auxiliary driving system.

[0003] Existing sensor calibration and correction methods mostly rely on single channel or offline methods. For example, the common practice is to arrange specific targets in the laboratory conditions to calibrate single sensors or part of the sensors independently. This kind of method often cannot cover the complexity of real road environment, and it is difficult to correct in real time during vehicle operation. Some research attempts to use on-board data for online calibration, but due to the lack of external reference and unified multi-modal target, the correction results are often unstable and have poor repeatability.

[0004] Therefore, the existing technology generally has the following deficiencies: first, there is a lack of a unified identification that can be recognized by multiple modal sensors, resulting in low correction efficiency; second, the vehicle's own recognition results lack independent verification by external monitoring channels, which is prone to deviation accumulation; third, the existing correction process is mostly one-time large-scale modification, lacking gradual optimization and safety rollback mechanism, which is prone to introduce new errors.

[0005] In view of the above problems, it is urgent to propose a vehicle intelligent auxiliary driving system correction system that can be used in real site environment. SUMMARY

[0006] The purpose of the present application is to provide a vehicle intelligent auxiliary driving system correction system and a correction method to solve the problems raised in the background.

[0007] To achieve the above purpose, the present application provides the following technical solution: a vehicle intelligent auxiliary driving system correction system, comprising a vehicle measurement module, a multi-modal combined identification, a field end true value measurement module, an evaluation and correction module, and a report generation module:

[0008] A vehicle measurement module is arranged on the vehicle itself, including a camera, a laser radar, a millimeter wave radar, an inertial measurement unit and a vehicle positioning unit, for collecting vehicle running state and environmental perception data, and serving as a basis for correction input;

[0009] A multi-modal combined identifier is of an integrated structure, provided with a visual coding pattern for camera recognition, a high reflectivity element or intensity coding structure for laser radar recognition, an angle reflector for millimeter wave radar recognition, and an ultra-wideband ranging module or radio frequency identification module for positioning or identity confirmation, each sub-identifier maintains a fixed relative position and can be attached or installed on the surface of the vehicle or obstacle;

[0010] A field end true value measurement module is used to obtain the spatial position and attitude information of the vehicle and the obstacle with the multi-modal combined identifier in real time through a GNSS-RTK base station, a UWB anchor point, an optical motion capture array or a total station;

[0011] An evaluation and correction module is used to compare the recognition results of the vehicle measurement module with the field end true value, calculate consistency indicators including geometric indicators, semantic indicators and cross-modal physical quantities, and when the consistency indicators exceed a threshold, perform gradual small-step fine tuning on the sensor external parameters or acquisition time delay parameters within the scope of a safety list, and retain a rollback point;

[0012] A report generation module is used to output a report containing scene arrangement, consistency indicator statistical results, correction parameter change range and traceability information.

[0013] A correction method based on a vehicle intelligent auxiliary driving system correction system, comprising the following steps:

[0014] S1, vehicle measurement: collecting data of the camera, laser radar, millimeter wave radar, inertial measurement unit and vehicle positioning unit by the vehicle measurement module, to complete the initial acquisition of the vehicle's own running state and environmental perception;

[0015] S2, identifier arrangement: attaching a multi-modal combined identifier on the vehicle and the surface of the predetermined obstacle, the combined identifier containing a visual coding pattern, a laser radar high reflectivity element or intensity coding structure, a millimeter wave radar angle reflector and an ultra-wideband ranging module or radio frequency identification module;

[0016] S3, true value measurement: obtaining the spatial position and attitude information of the vehicle and the obstacle with the combined identifier in real time by the field end true value measurement module using a GNSS-RTK base station, a UWB anchor point, an optical motion capture array or a total station;

[0017] S4, online comparison: comparing the perception recognition results of the vehicle measurement module with the field end true value, and calculating consistency indicators including geometric indicators, semantic indicators and cross-modal physical quantities;

[0018] S5, progressive correction: when the consistency index exceeds the threshold, gradually fine-tune the vehicle sensor extrinsic parameters or acquisition time delay parameters within the safe list range, and keep the rollback point to restore;

[0019] S6, report output: generate and output a report containing scene arrangement, consistency index statistics, correction parameter change range and traceability information.

[0020] Preferably, the multi-modal combined identifier is an integrated sheet structure, comprising:

[0021] A visual coding pattern for camera recognition;

[0022] A high-reflectivity sphere or intensity coding structure for lidar recognition;

[0023] A pluggable trihedral reflector for millimeter wave radar recognition;

[0024] An ultra-wideband ranging module or radio frequency identification module for positioning or identity confirmation,

[0025] Wherein each sub-identifier maintains a fixed relative position and can be installed on the surface of the vehicle or obstacle through a sticky layer or magnetic element.

[0026] Preferably, the obstacle includes a soft column, a mannequin model, a box or a reflector, and the multi-modal combined identifier is attached to the surface of the obstacle; the arrangement of the obstacle can be a regular grid, a ring, a fan-shaped arrangement, or a non-regular random placement to simulate the distribution of obstacles in various driving environments.

[0027] Preferably, the scene end true value measurement module includes one or more of the following: GNSS-RTK base station, UWB anchor point, optical motion capture array, laser total station or laser tracker, for real-time acquisition of spatial position and attitude information of the vehicle and the obstacle with multi-modal combined identifier.

[0028] Preferably, the consistency index includes:

[0029] Geometric index: lateral error, longitudinal error and heading angle error of the vehicle and the obstacle;

[0030] Semantic index: detection recall rate, detection accuracy rate and intersection over union of the identified target box and the true value box;

[0031] Cross-modal physical quantity index: difference distribution between camera recognition box center, lidar clustering center and millimeter wave radar ranging result, and matching rate of radar echo intensity and reflection cross-sectional area.

[0032] Preferably, the report generated by the report generation module includes:

[0033] The scene arrangement mode and the obstacle type;

[0034] Statistical results of the consistency index and error distribution curves;

[0035] Triggering number of times, parameter change range and rollback record of progressive correction;

[0036] Traceability information for regulatory compliance verification.

[0037] Preferably, the progressive correction comprises:

[0038] Based on the deviation result of the consistency index, the sensor extrinsic matrix or the time synchronization parameter is iteratively corrected according to a preset small step;

[0039] The maximum change is limited in a single correction, and the true value range is gradually approached;

[0040] When the cumulative correction exceeds the safe range, the parameter rollback is triggered to restore to the last effective parameter set.

[0041] Preferably, the multi-modal combined identifier is an integrated thin plate structure, and a reusable adhesive layer or a magnetic fixing member is arranged on the back surface, which is used for being attached to the surface of the vehicle or the surface of the obstacle, the relative positions between the sub-identifiers are fixed, and the thin plate has a surface treatment layer for waterproofing, anti-reflection and anti-interference, so as to ensure stable identification in different environments.

[0042] Preferably, the report generation module is further used for uploading the correction result to a cloud server, the cloud server establishes a parameter file and a threshold template according to different vehicle models, and the correction data of each vehicle is uniformly stored, called and updated, so as to facilitate batch consistency verification and compliance traceability.

[0043] Compared with the prior art, the vehicle intelligent auxiliary driving system correction system provided by the application realizes double-channel comparison and correction of vehicle self-recognition and external monitoring recognition through the cooperation of the vehicle measurement module, the field true value measurement module and the multi-modal combined identifier. The system not only performs tire correction and direction alignment when the vehicle enters the field, but also continuously collects the recognition data of the vehicle and the outside during the test process, and uses a consistency index system for comprehensive comparison, and when deviation is found, a progressive correction method is used to optimize the sensor parameters in small steps.

[0044] Compared with the prior art, the advantages of the present application are as follows: firstly, integrated multi-modal combined identification is adopted, which can be recognized by cameras, laser radars, millimeter wave radars and UWB modules at the same time, reducing the cumbersome process of separate arrangement of different targets and ensuring the unity of multi-modal observation; secondly, a double-channel comparison mechanism is provided, and the external monitoring device not only records the true value, but also independently identifies the obstacles and calculates the avoidance effect, thereby ensuring the objectivity and reliability of the correction result; thirdly, the gradual correction and rollback mechanism is used to effectively suppress the drift problem of the sensor in long-term operation under the premise of ensuring the safety boundary.

[0045] Therefore, the present application can efficiently complete the automatic checking and correction of the vehicle intelligent auxiliary driving system in the actual site environment, improve the vehicle perception accuracy and recognition stability, and solve the problems of correction process relying on a single channel, low correction efficiency and large repeated workload in the prior art. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] The present application provides a technical solution: a vehicle intelligent auxiliary driving system correction system, comprising a vehicle measurement module, a multi-modal combined identification, a field end true value measurement module, an evaluation and correction module and a report generation module:

[0048] The vehicle measurement module is arranged in the vehicle itself and comprises a camera, a laser radar, a millimeter wave radar, an inertial measurement unit and a vehicle positioning unit, which is used to collect vehicle running state and environmental perception data and serve as the basic input for correction;

[0049] The vehicle measurement module not only includes basic perception components such as cameras, laser radars, millimeter wave radars, inertial measurement units and vehicle positioning units, but also further combines the mechanical state of the vehicle itself as a correction input. When the vehicle enters the correction site, the system first detects whether the tire angle of the vehicle is in the straightening state, confirms whether the front wheel of the vehicle is deflected through the angle sensor or the steering wheel angle encoder; if deflection is found, the driver is automatically prompted to straighten, or in the case of steer-by-wire, the control unit directly corrects. Secondly, the vehicle measurement module also uses the vehicle positioning unit to compare with the standard alignment point preset in the site, judges whether the longitudinal direction of the vehicle is consistent with the reference line, so as to ensure that the overall posture of the vehicle is aligned with the standard reference direction.

[0050] Through the above steps, the vehicle has completed a posture initialization before entering the subsequent identification arrangement and the true value measurement link, which can reduce the need for repeated correction caused by vehicle posture deviation. At the same time, the vehicle measurement module can also combine the inertial measurement unit and wheel speed sensor data to judge the static state and slight displacement of the vehicle, avoiding measurement errors caused by the vehicle not being completely stationary. This module is executed as the first step in the entire correction process, ensuring that the comparison of subsequent perception data and external true values has a consistent reference benchmark.

[0051] The multi-modal combined identification is an integrated structure, which is provided with a visual coding pattern for camera recognition, a high reflectivity element or intensity coding structure for laser radar recognition, an angle reflector for millimeter wave radar recognition, and an ultra-wideband ranging module or radio frequency identification module for positioning or identity confirmation. Each sub-identification maintains a fixed relative position and can be attached or installed on the surface of a vehicle or an obstacle;

[0052] The multi-modal combined identification is an integrated thin plate structure, the main body of which is made of lightweight composite materials or engineering plastics to ensure strength and weather resistance. The overall size of the thin plate can be adjusted according to the installation position on the surface of the vehicle and the obstacle, usually in the A4-A3 size range, with a thickness controlled within 5-10 mm, which can provide sufficient rigidity without affecting the attachment and handling.

[0053] The combined identification is composed of multiple sub-identifications, each of which maintains a fixed relative position to ensure simultaneous recognition by different sensors in the same coordinate system:

[0054] Visual identification: high-contrast AprilTag or ArUco patterns are printed on the plate surface, and anti-fouling and anti-fading coatings are provided on the edges to ensure that the camera can accurately recognize in strong light, weak light, and polluted environments.

[0055] Laser radar identification: high-reflectivity spheres or laser intensity coding patterns are embedded on the plate surface, and their diameters and distribution positions are precisely designed to form obvious high-reflective points or recognizable coding structures in the radar point cloud, facilitating coordinate extraction.

[0056] Millimeter wave radar identification: three-sided angle reflectors are inserted or fixed in the reserved mounting hole positions on the plate surface. The structure is made of aluminum alloy or high-conductive materials, which can produce stable large reflection cross-section (RCS) in the millimeter wave frequency band, ensuring the recognition reliability of the vehicle radar.

[0057] Positioning / identity identification: UWB active tags or RFID / NFC modules are embedded in the plate body to provide centimeter-level ranging information under shielding conditions and complete obstacle identity confirmation through a unique ID.

[0058] In terms of installation, the combined identification back is provided with a reusable adhesive layer or a magnetic attraction structure, which can be quickly attached to the surface of a vehicle or an obstacle, and supports multiple disassembly and movement, facilitating flexible arrangement in different test scenarios. To improve long-term stability, a waterproof layer, an anti-reflection coating, and an electromagnetic interference shielding layer are provided on the surface of the thin plate to adapt to rainy days, strong light, dust, or multi-source electromagnetic environment, ensuring consistent identification of the identification under various environments.

[0059] Through the design of the combined identification, a unified common view target can be provided for the camera, laser radar, millimeter wave radar, and UWB module at the same time, spatial alignment of cross-modal perception data is realized, the problem of repeated arrangement of single-modal calibration tools is reduced, and the correction efficiency is greatly improved.

[0060] The field true value measurement module is used to obtain the spatial position and attitude information of the vehicle and the obstacle with the multi-modal combined identification in real time through the GNSS-RTK base station, the UWB anchor point, the optical motion capture array, or the total station;

[0061] The field true value measurement module is used to provide high-precision spatial position and attitude reference for the vehicle and the obstacle, and different measurement devices can be flexibly configured according to the environmental conditions of the test site to ensure that reliable "true value" coordinates can be obtained in various scenarios.

[0062] GNSS-RTK base station: suitable for open outdoor environment, uses differential positioning technology to provide centimeter-level position accuracy. By arranging fixed base stations at the edges of the site, the receivers on the vehicle and the obstacle can receive correction signals in real time, thereby obtaining high-precision three-dimensional position for establishing the reference coordinate system of the overall scene.

[0063] UWB anchor point system: suitable for semi-indoor sites with more obstructions or urban canyon environments. Several UWB anchor points are arranged around the site, and the combined identification on the obstacle and the vehicle is provided with a UWB tag, which forms a ranging link with the anchor points, thereby realizing centimeter-level relative positioning under the condition of no satellite signal coverage.

[0064] Optical motion capture array: suitable for small-range tests that require high-precision attitude calculation. By installing reflective balls or identification points on the vehicle and the obstacle, infrared cameras arranged around the site can capture three-dimensional coordinates and rotational attitude in real time, with an accuracy of millimeters, which is commonly used for high-precision verification of vehicle sensor external parameter correction.

[0065] Laser total station or laser tracker: suitable for engineering-level precision verification scenarios. The vehicle and the obstacle are positioned through reflective prisms or corner reflectors, which can maintain a measurement accuracy of millimeters within a range of tens of meters or even hundreds of meters, and can be used to verify the accuracy of other positioning systems.

[0066] In practical applications, the field-end ground truth measurement module can use a single device or multiple sources in a fusion manner. For example, in an open environment, GNSS-RTK is preferred, while in a partially blocked area, UWB positioning is automatically switched to. In a precise experimental condition, an optical motion capture array and a total station can be used together for cross-validation and data redundancy. Through such flexible combination, continuous, stable and reliable ground truth data can be provided in different scenarios.

[0067] The ground truth measurement module not only obtains the three-dimensional position, but also outputs the obstacle and vehicle attitude angle information, so that the evaluation and correction module can perform more comprehensive comparison, especially in the vehicle steering, braking and acceleration state, the attitude change can be captured in real time and reflected in the calculation of the consistency index.

[0068] The evaluation and correction module is used to compare the recognition results of the vehicle measurement module with the field-end ground truth, calculate the consistency index including geometric index, semantic index and cross-modal physical quantity, and when the consistency index exceeds the threshold, gradually adjust the sensor extrinsic parameters or acquisition time delay parameters within the safety list, and reserve the rollback point;

[0069] The evaluation and correction module is used to compare the recognition results of the vehicle measurement module with the results of the field-end ground truth measurement module in real time, so as to generate a consistency index, and correct the vehicle sensor parameters when necessary.

[0070] Specifically, the evaluation and correction module first establishes a unified coordinate system, and converts the vehicle perception results (camera detection frame, laser radar point cloud clustering, millimeter wave radar ranging and angle data, etc.) to the reference system consistent with the field-end ground truth. Then, based on the following dimensions, the consistency index is calculated:

[0071] Geometric consistency: including the deviation of the vehicle and the obstacle in the lateral, longitudinal and heading angle, using root mean square error (RMSE) and maximum error value as evaluation parameters;

[0072] Semantic consistency: comparing the target class matching relationship between the vehicle recognition result and the ground truth label, calculating the detection recall rate, accuracy rate, and the intersection over union (IoU) of the recognition frame and the ground truth frame;

[0073] Cross-modal consistency: comparing the differences between the camera recognition frame center, laser radar clustering center and millimeter wave radar ranging result respectively, calculating the distribution consistency of multi-source perception results, and combining radar echo intensity and the standard radar cross section (RCS) of the corner reflector for matching degree test.

[0074] When the index of any dimension exceeds the preset threshold, the evaluation and correction module will trigger the correction process. The correction process adopts a gradual small step adjustment method, which specifically includes:

[0075] External parameter correction: Based on the geometric deviation results, the external parameter matrix between camera, lidar and millimeter wave radar is iteratively optimized by small step, ensuring that the observations of different sensors in the unified coordinate system are aligned.

[0076] Time synchronization correction: When there is a significant time delay error in the cross-modal recognition results, adjust the data acquisition timestamps of each sensor to ensure that multi-source data is fused under the same time reference.

[0077] Safety limit: The maximum change threshold is set for each adjustment to avoid excessive deviation of sensor parameters. If necessary, automatically roll back to the last valid parameter set to ensure that the system operates within a safe range.

[0078] In addition, the evaluation and correction module also has adaptive ability, which can call corresponding parameter templates according to different vehicle models and sensor configurations, so that different vehicles can quickly converge when correcting in the same place, reducing the number of repeated calibration. Through the above design, this module can effectively reduce the error accumulation caused by sensor drift in long-term operation, and improve the stability and reliability of intelligent auxiliary driving system in complex environments.

[0079] The report generation module is used to output a report containing scene arrangement, consistency index statistical results, correction parameter change range and traceability information.

[0080] The report generation module is used to automatically generate a data-based review report after the correction process is completed, and supports uploading to the cloud for unified storage and calling.

[0081] Specifically, the report generation module includes a local processing unit and a cloud service interface:

[0082] Local processing unit: After the vehicle completes a correction process, the system integrates the collected sensor recognition results, scene true value data and processing records of the evaluation and correction module to form a standardized test record file. This file is indexed by timestamp and includes parameter comparison before and after correction, consistency index statistical results and error distribution curve.

[0083] Report content:

[0084] Scene information: type, number and arrangement of obstacles (regular grid, ring or random placement);

[0085] Consistency index: including lateral error, longitudinal error, heading angle deviation, detection recall rate, accuracy, IoU value, cross-modal difference distribution and radar RCS matching rate;

[0086] Correction record: number of correction triggers, parameter change range, single iteration adjustment amount and rollback situation;

[0087] Environmental conditions: background information such as light, weather, temperature, etc. during testing, for subsequent reproduction of experimental environment;

[0088] Regulatory compliance information: Provide traceability fields that meet the regulations related to autonomous driving (such as UNECE R157, etc.), including test procedure number, version number, and issuer.

[0089] Cloud service interface: After report generation, it can be automatically uploaded to the cloud server. The cloud establishes independent archives for different vehicle models and generates corresponding threshold templates for different sensor configurations. Through unified storage, it can support consistency verification of batch vehicles in factory detection, road testing, and long-term operation and maintenance, facilitating comparison of performance fluctuations of different batches of vehicles.

[0090] Statistics and analysis: The cloud server has the ability to statistically analyze historical data and trends, automatically generate error variation curves, parameter adjustment frequency charts, and annual or phased evaluation reports, providing decision-making basis for R&D, production, and after-sales.

[0091] Through the above design, the report generation module not only realizes the full-process traceability of the test process, but also serves as an important basis for compliance verification, ensuring that the vehicle is fully verified before being put into use, and improving the safety and reliability of the intelligent auxiliary driving system.

[0092] A correction method based on a vehicle intelligent auxiliary driving system correction system, comprising the following steps:

[0093] S1, vehicle measurement: Collect data from cameras, lidar, millimeter wave radar, inertial measurement unit, and vehicle positioning unit by the vehicle measurement module to complete the initial acquisition of the vehicle's own running state and environmental perception;

[0094] S2, label arrangement: Paste multi-modal combined labels on the vehicle and the surface of the pre-set obstacles, which contain visual coding patterns, laser radar high reflectivity elements or intensity coding structures, millimeter wave radar corner reflectors, and ultra-wideband ranging modules or radio frequency identification modules;

[0095] S3, true value measurement: Use GNSS-RTK base stations, UWB anchor points, optical motion capture arrays, or total stations to obtain the spatial position and attitude information of the vehicle and the obstacles with combined labels in real time through the field true value measurement module;

[0096] S4, online comparison: Compare the sensing recognition results of the vehicle measurement module with the field true value, and calculate the consistency indicators including geometric indicators, semantic indicators, and cross-modal physical quantities;

[0097] S5, progressive correction: when the consistency index exceeds the threshold value, the vehicle sensor extrinsic parameter or acquisition time delay parameter is gradually fine-tuned in the safe list range, and the rollback point is reserved for recovery;

[0098] S6, report output: generate and output a report containing scene arrangement, consistency index statistics, correction parameter change range and traceability information.

[0099] The specific completion process of the vehicle intelligent auxiliary driving system correction method provided by the application comprises the following steps:

[0100] S1 vehicle measurement and initialization

[0101] After the vehicle enters the correction site, the vehicle measurement and initialization step is first performed, which aims to unify the vehicle's own attitude, direction and time reference before any external layout and comparison begins, so as to reduce the subsequent multiple repeated corrections.

[0102] (1) Sensor self-checking

[0103] The vehicle measurement module includes:

[0104] Camera

[0105] Laser radar (LiDAR)

[0106] Millimeter wave radar (Radar)

[0107] Inertial measurement unit (IMU, Inertial Measurement Unit)

[0108] Vehicle positioning unit (GNSS-RTK receiver or GNSS+odometer)

[0109] The system detects whether the data link of each sensor is normal at time t0 (initial time point), and confirms whether the time stamp of the data output is consistent with the unified time base.

[0110] (2) Tire alignment

[0111] The front wheel angle δ is read through the steering wheel angle sensor or the front wheel angle sensor:

[0112] When |δ|≤δ0, it is determined that the front wheel is in the alignment state.

[0113] When |δ|>δ0, the driver is prompted to align or the alignment is automatically performed by the steer-by-wire system.

[0114] Wherein:

[0115] δ represents the current front wheel angle value, unit: degree (°);

[0116] δ0 denotes the allowed threshold of turning angle, preferably set to 0.5°.

[0117] (3) Direction alignment reference point

[0118] The vehicle longitudinal direction is compared with the standard reference line L0 of the site:

[0119] Define the current heading angle of the vehicle as ψ (psi, unit: degree);

[0120] Define the reference line direction as ψ0;

[0121] When |ψ-ψ0|≤Δψ0, the vehicle direction is considered to be aligned with the reference point;

[0122] When |ψ-ψ0|>Δψ0, the system prompts adjustment, and Δψ0 is preferably 0.5°.

[0123] (4) Static determination

[0124] Read angular velocity and acceleration from IMU;

[0125] Read wheel speed from wheel speed sensor;

[0126] When angular velocity <ω0 (such as 0.01 rad / s), acceleration <a0 (such as 0.01 m / s 2 ), wheel speed

[0127] v0 (such as 0.05 m / s), the vehicle is determined to be in a static state.

[0128] Where:

[0129] ω0 is the angular velocity threshold;

[0130] a0 is the acceleration threshold;

[0131] v0 is the speed threshold.

[0132] (5) Time unification and parameter set loading

[0133] The data timestamps of each sensor need to be aligned to a unified reference clock:

[0134] If an external timing system (NTP or PTP) is used, all devices are aligned to the global clock;

[0135] If hardware trigger synchronization is used, t0 is used as the reference.

[0136] The initial time offset is defined as Δt0:

[0137] Δt0=t sensor -t0

[0138] Where t sensorSensor data timestamp, t0 represents the reference clock time.

[0139] The system also loads the last stored initial parameter set P(0), which includes the extrinsic matrix of each sensor and the time synchronization parameter, as the starting point of correction.

[0140] End condition of S1:

[0141] All sensors are online and the timestamps are unified;

[0142] Front wheel alignment (|δ|≤δ0);

[0143] Heading alignment (|ψ-ψ0|≤Δψ0);

[0144] Vehicle is stationary;

[0145] Parameter set P(0) is successfully loaded.

[0146] So far, the vehicle has completed the initialization of the posture before correction.

[0147] S2 Combination identification and site layout

[0148] After the vehicle completes the initialization, obstacles with multi-modal combination identification need to be arranged in the site so that the vehicle and the external monitoring equipment can recognize at the same time, and the subsequent comparison and correction can be realized.

[0149] (1) Structure of combination identification and arrangement of sub-identifications

[0150] The multi-modal combination identification is an integrated sheet structure with a thickness of h_b = 5–10 mm and a size range of A4–A3 (about 210×297 mm to 297×420 mm). The main material is lightweight composite material or engineering plastic, and the surface is covered with a waterproof, anti-reflection and anti-electromagnetic interference layer.

[0151] Sub-identifications include:

[0152] Visual identification V_tag: printed high-contrast AprilTag or ArUco pattern;

[0153] LiDAR identification L ref : high-reflective sphere (diameter d s = 50–100 mm) or intensity coding area;

[0154] Millimeter wave radar identification R corner : pluggable trihedral corner reflector, providing stable radar cross section (RCS);

[0155] UWB identification U_id: active tag or RFID / NFC module, outputting a unique ID and supporting centimeter-level ranging.

[0156] The relative positions of each sub-identifier are fixed, forming a unified rigid structure to ensure that the same coordinate system is used when identified in any mode.

[0157] (2) Obstacle types

[0158] The obstacle O can be:

[0159] a soft pole (Soft Pole);

[0160] a dummy model (Dummy);

[0161] a box (Box);

[0162] a reflective board (Reflective Board).

[0163] Each obstacle surface is fixedly installed with at least one combined identifier.

[0164] (3) Layout method

[0165] The layout method is divided into two types:

[0166] 1. Regular mode (Regular Mode)

[0167] Arranged in a grid on both sides of the main driving line, with a grid spacing of g = 2-5 m;

[0168] or arranged in a ring / fan shape, with a ring radius r = 5-25 m and a fan angle θ_s = 30°-90°.

[0169] 2. Random mode (Random Mode)

[0170] Randomly scattered outside the safety boundary, with a minimum static gap s min ≥ 1.0-1.5 m;

[0171] The random distribution is generated by using "test number + time stamp" as a random seed to ensure reproducibility.

[0172] (4) Coverage index

[0173] To ensure that the vehicle sensor always has enough reference targets during driving, the field of view (FOV) of the vehicle sensor is divided into zones:

[0174] Azimuth angle binning: Δθ = 10°;

[0175] Distance binning: Δr = 5 m.

[0176] The coverage rate C = N eff / N tot , where:

[0177] N eff = the number of bins with presence of the label;

[0178] N tot = the total number of bins.

[0179] When C≥80%, determine that the layout coverage is up to standard.

[0180] (5) Safety constraints

[0181] All obstacles need to satisfy:

[0182] The nearest static distance d to the main line of vehicle travel safe ≥s min ;

[0183] Does not affect the emergency avoidance path of the vehicle;

[0184] Each obstacle is equipped with a remote stop or soft material to ensure safety in the event of vehicle deviation.

[0185] (6) Record and storage

[0186] After the layout is completed, a layout file needs to be established for each obstacle, including:

[0187] Obstacle ID (given by UWB or visual code);

[0188] Coordinate position (x w , y w , z w );

[0189] Installation height h o ;

[0190] Normal vector of the board surface n o ;

[0191] Allowed displacement range Δpos (for movable obstacles).

[0192] All data is uploaded to the site database as input for subsequent comparison.

[0193] End condition of S2:

[0194] Obstacle layout is complete, coverage C≥80%;

[0195] Each obstacle information is recorded in the database;

[0196] Safety distance and emergency stop measures meet the requirements.

[0197] S3 field end true value and external re-identification

[0198] After the vehicle and obstacles are deployed, the scene end true value measurement channel and the external re-identification channel need to be established. The purpose is to obtain high-precision vehicle and obstacle position and attitude information in the world coordinate system, and to form a second observation channel independent of the vehicle's own identification result for subsequent comparison and correction.

[0199] (1) Scene end true value measurement module

[0200] The scene end true value module can flexibly combine different measurement methods according to the site conditions:

[0201] 1. GNSS-RTK base station

[0202] Provide centimeter-level three-dimensional position.

[0203] Install GNSS antennas on vehicles and obstacles to receive base station differential correction signals.

[0204] Output coordinates p g =(x g ,y g ,z g ).

[0205] 2. UWB anchor point system

[0206] Deploy N anchor points (recommended N≥4) in the scene.

[0207] Combine the identification built-in UWB tag to measure the distance by Time of Flight (TOF) or Time Difference of Arrival (TDOA).

[0208] Calculate the relative position p u =(x u ,y u ,z u ).

[0209] 3. Optical motion capture array

[0210] Paste reflective balls on obstacles and vehicles, and deploy M infrared cameras (recommended M≥8).

[0211] Obtain accurate coordinates and attitude through multi-view triangulation.

[0212] Output pose T m =[R m |t m ], where R m is the rotation matrix and t m m is the translation vector.

[0213] 4. Laser total station or laser tracker

[0214] Install a prism or corner reflector on the obstacle.

[0215] The total station angle and distance measurement can achieve millimeter level accuracy.

[0216] Output coordinates p l = (x l , y l , z l ).

[0217] Through a weighted fusion algorithm (such as extended Kalman filter EKF), the final true value coordinates and attitude are obtained:

[0218] T truth = f(p g , p u ·Tm, p l ), ∑ truth = Cov(T truth )

[0219] Where T truth is the 6-DOF pose (position + attitude) of the vehicle or obstacle, and ∑ truth is its covariance matrix.

[0220] (2) External re-identification module

[0221] In addition to the true value measurement, the external monitoring device also needs to independently identify the combined identifier on the obstacle, forming an external channel O2 corresponding to the vehicle identification channel O1:

[0222] 1. External camera array

[0223] o Decode the visual code (AprilTag / ArUco) on the combined identifier to obtain the pose T v ;

[0224] o The pose solution uses the PnP algorithm combined with the intrinsic matrix K and distortion parameters.

[0225] 2. External LiDAR

[0226] o Capture high-reflectivity point cloud, use least squares sphere fitting to obtain the sphere center coordinates p s ;

[0227] o Or identify the intensity-encoding plane to obtain the feature plane parameters.

[0228] 3. External millimeter wave radar

[0229] o Scan the trihedral corner reflector and extract the maximum radar cross section (RCSmax);

[0230] o Output distance d r , azimuth angle θ r .

[0231] 4. UWB anchor re-identification

[0232] o Confirm the obstacle identity directly by tag ID, measure the distance d u as an independent ranging channel.

[0233] All external re-identification data is unified to the world coordinate system through the coordinate transformation matrix T w (world coordinate system→device coordinate system) to form the obstacle external observation pose T ext .

[0234] (3) Object association

[0235] To ensure one-to-one correspondence between the vehicle-end identification channel O1 and the external channel O2, object matching is needed:

[0236] Time alignment: all data are aligned to a unified reference time t0, allowing a time deviation

[0237] Δt≤10ms;

[0238] ID priority: if the obstacle contains a unique UWB / RFID ID, it is directly matched;

[0239] Geometric correlation: if there is no ID, the Mahalanobis distance criterion is used:

[0240]

[0241] When D M ≤D thr , it is considered that O1 and O2 belong to the same obstacle;

[0242] . Where x1 is the vehicle-end identification coordinate, x2 is the external observation coordinate, Σ is the joint covariance matrix, and D thr is the threshold value.

[0243] (4) End condition of S3

[0244] · The vehicle and all obstacles obtain true value poses T truth ;

[0245] · The external re-identification channel O2 outputs obstacle poses T ext ;

[0246] . Each obstacle successfully establishes a unique association between O1 and O2.

[0247] S4 online comparison and evaluation

[0248] When the vehicle is traveling at low speed along the predetermined route, the system needs to compare the results of the vehicle-side identification channel O1 with the external monitoring channel O2 frame by frame, calculate a series of indicators, form a deviation spectrum, and determine whether to trigger correction.

[0249] (1) Geometric and distance indicators

[0250] • Closest Point of Approach (CPA)

[0251] Perform time-series analysis on the vehicle trajectory and obstacle trajectory to determine the closest point between the vehicle and the obstacle.

[0252] The minimum distance is defined as:

[0253]

[0254] Where p veh (t) is the position vector of the vehicle at time t, P obs (t) is the position vector of the obstacle at time t.

[0255] · Lateral clearance distance

[0256] At time CPA, using the vehicle velocity direction as a reference, calculate the interval perpendicular to the velocity direction:

[0257]

[0258] in It is a unit velocity vector.

[0259] Time-To-Collision (TTC)

[0260] When the relative velocity Δv points towards the obstacle:

[0261]

[0262] (2) Trajectory and Avoidance Effect

[0263] • Avoidance angle θ avoid

[0264] At time CPA, calculate the vehicle velocity vector. Vehicle-obstacle line vector

[0265] The included angle:

[0266]

[0267] When θ avoid If the angle is greater than a preset threshold (e.g., 8°), it is considered a valid avoidance.

[0268] • Trajectory deviation Δ traj

[0269] The vehicle planned trajectory (estimated by car end recognition) is compared with the external observed trajectory, and the maximum deviation is recorded as:

[0270]

[0271] (3) Semantic consistency index

[0272] • Category consistency If the car end recognition category C O1 is consistent with the external recognition category C O2 , it is recorded as correct, otherwise as error.

[0273] • Intersection over Union (IoU)

[0274] For camera detection box and external ground truth box:

[0275]

[0276] Where A det is the detection box area, and A truth is the ground truth box area.

[0277] . Detection recall rate / accuracy is calculated according to the standard definition of TP (True Positive), FP (False Positive), and FN (False Negative).

[0278] (4) Cross-modal consistency index

[0279] . Camera-LiDAR deviation

[0280] Δ cum-lidar = ||c cam -c lidaγ ||

[0281] Where c cam is the center point of the camera detection box, and c lidaγ is the LiDAR cluster centroid.

[0282] • Camera-Radar deviation

[0283]

[0284] • Radar echo intensity matching degree The deviation of the actual measured echo intensity RCS meas and the theoretical value RCS ref is:

[0285]

[0286] where s rcs ∈[0,1], the closer to 1 indicates the higher the matching degree.

[0287] (5) Comparison and decision-making logic

[0288] · Hard threshold judgment:

[0289] o If |d min , O1-d min , O2|>ε d , it is recorded as out-of-bound;

[0290] o If |θ avoid , O1-θ avoid , O2|>ε θ , it is recorded as out-of-bound;

[0291] o If IoU<τ (such as 0.5), and the categories are inconsistent, it is judged as a semantic error.

[0292] . Soft score fusion:

[0293] Define the comprehensive score S:

[0294] S=w1·f(d min )+w2·f(θ avoid )+w3·f(IoU)+w4·f(Δ cam-lidar , Δ cam-radar )+w5·S rcs

[0295] where w i is the weight, When S<S thr , trigger correction.

[0296] (6) End condition of S4

[0297] · All indicators (geometry, avoidance, semantics, cross-modal) complete comparison between O1 / O2 channels;

[0298] . Samples exceeding the threshold are accurately labeled as out-of-bound;

[0299] · Output "bias spectrum" (mean, root mean square error RMSE, 95 percentile, maximum).

[0300] S5 gradual correction execution

[0301] In the comparison results of S4, if the difference between the vehicle recognition channel O1 and the external monitoring channel O2 exceeds the preset threshold, the gradual correction is triggered, and the parameters of the vehicle sensor are iteratively corrected in small steps until the error converges or reaches the safety boundary.

[0302] (1) Correction trigger condition

[0303] When any of the following conditions are met, the system enters the correction state:

[0304] · Geometric difference:

[0305] | d min , O1-d min , O2 | > ε d

[0306] · Avoidance angle difference:

[0307] | θ avoid , O1-θ avoid , O2 | > ε θ

[0308] · Semantic inconsistency: category error or IoU < τ;

[0309] ● Insufficient cross-modal score:

[0310] S < S thr

[0311] (2) Correction target

[0312] The core objects of correction include:

[0313] · Time synchronization parameter Δt

[0314] Indicates the difference between the sensor data timestamp and the reference clock, defined as:

[0315] Δt = t sensor -t ref

[0316] · External parameter matrix T extrinsic

[0317] Indicates the rotation and translation relationship between the two sensors:

[0318]

[0319] Where R is a 3 × 3 rotation matrix and t is a 3 × 1 translation vector.

[0320] . Necessary internal parameter fine-tuning (small correction of camera principal point and focal length).

[0321] (3) Optimization objective function

[0322] Define the comprehensive error function J:

[0323] J = w1 · E geo +w2 · E avoid +w3 · E sem +w4 · E cross +w5 · Etime

[0324] where:

[0325] · E geo = |d min , O1-d min , O2 |) ;

[0326] · E avoid = |θ avoid , O1-θ avoid , O2 | (avoidance angle error) ;

[0327] · E sem = 1 - IoU or class error penalty;

[0328] · E cross = Δ cam - lidar + Δ cam - radar + (1 - S rcs ) (cross-modal error) ;

[0329] · E time = |Δ t - Δ tref | (time synchronization error).

[0330] The weight coefficient w i satisfies:

[0331]

[0332] (4) Small step iteration

[0333] A block-wise optimization (e.g. Levenberg–Marquardt or Ceres Solver) is employed, with the adjustment magnitude in each iteration limited to:

[0334] Δ step ≤ Δ max

[0335] ∑Δ step ≤ Δ max

[0336] The single step size Δ step limits minor changes and avoids large jumps;

[0337] The cumulative change ∑ max safeguards the boundary.

[0338] (5) Safety checklist and rollback mechanism

[0339] • All modifiable parameters must be listed in the safety checklist L safe ;

[0340] If parameter adjustment exceeds the allowed range, the system immediately suspends;

[0341] If n consecutive iterations do not make J decrease, or a new out-of-boundary occurs, roll back to the last valid parameter set P(k-1).

[0342] Rollback condition formalization:

[0343]

[0344] Trigger rollback and record.

[0345] (6) End condition of S5

[0346] · The integrated error J converges to the threshold value J thr Below;

[0347] · Or the cumulative change ∑ max To the upper limit;

[0348] · Or trigger the rollback mechanism and freeze the current parameter set.

[0349] At this time, a new valid parameter set P(k) is obtained as the corrected reference.

[0350] S6 Report Generation and Storage

[0351] After the correction process is completed, the system needs to organize all data and parameter change conditions into a standardized report and upload it to the cloud archive. This can not only be used for regulatory traceability, but also provide a reference for the rapid correction of subsequent similar vehicles.

[0352] (1) Report structure

[0353] The report generation module organizes data into five categories of fields:

[0354] 1. Scene field F scene

[0355] o Site type: open / semi-indoor / calibration track;

[0356] o Obstacle arrangement method: regular mode / random mode;

[0357] o Grid spacing g, ring radius r, safety boundary s min

[0358] o Obstacle list: ID, coordinates (x w , y w , z w ) of each obstacle, installation height h o , normal vector

[0359] 2. Indicator field F metric

[0360] o Geometric indicators: minimum distance d min Clearance d lat , TTC

[0361] o Trajectory and evasion: evasion angle θ avoid , trajectory deviation Δ traj

[0362] o Semantic indicators: class consistency, recall, precision, IoU

[0363] o Cross-modal indicators: Δ cam -lidar, Δ cam -radar, RCS matching degree S rcs

[0364] o Output statistics: mean (Mean), root mean square error (RMSE), 95th percentile value (P95), maximum value (Max).

[0365] 3. Correction field F corr

[0366] o Number of triggers N corr

[0367] o Adjusted parameter categories (Δ t , extrinsic matrix T extrinsic , camera intrinsic parameters)

[0368] o Single step size Δ step , cumulative change ∑Δ

[0369] o Rollback record: whether triggered, trigger time t rollback , restored parameter set

[0370] P(k-1)

[0371] o Convergence: final comprehensive error J final

[0372] 4. Compliance field F_reg

[0373] o Test procedure number ID_flow

[0374] o Software / firmware version number V_soft

[0375] o Operator ID_op and auditor ID_audit

[0376] o Environmental conditions: light, weather, temperature

[0377] ​​​​o Essential traceability items to meet regulatory requirements (e.g. UNE CER 157 verification items for ALKS system).

[0378] 5. Cloud field F_cloud

[0379] o Report hash value H_report (for tamper-proof traceability);

[0380] o Archive number ID_arch (generated by combining vehicle model + time + test number);

[0381] o Threshold template version V_temp;

[0382] o Historical trend link (call past data comparison for the same vehicle model).

[0383] (2) Report generation process

[0384] • All fields first generate a JSON / XML format file locally with a timestamp t gen ;

[0385] . Encrypted by the signature module and generate hash value H_report;

[0386] • A copy is retained locally and one is uploaded to the cloud for archiving.

[0387] (3) Cloud storage and model update

[0388] The report uploaded to the cloud is written to the corresponding vehicle model archive library:

[0389] • The archive library is classified by vehicle ID_vehicle + sensor configuration ID_sensor;

[0390] • The system compares the current report with the historical report, draws error trend curve d min (t), θ avoid(t) , IoU(t);

[0391] • If the same vehicle model repeatedly exceeds the threshold in similar scenarios is detected, the cloud will update the threshold template V temp , and push suggestions (such as increasing the IoU threshold, tightening the Δt range).

[0392] (4) End condition of S6

[0393] • The report is successfully generated and the fields are complete;

[0394] • The cloud confirms storage and generates a unique archive number ID_arch;

[0395] • Return the link for subsequent traceability.

[0396] Specifically, the multi-modal combined identification is an integrated thin plate structure, comprising:

[0397] a visual coding pattern for camera recognition;

[0398] a high-reflectivity sphere or intensity coding structure for laser radar recognition;

[0399] a pluggable trihedral reflector for millimeter wave radar recognition;

[0400] a super-wideband ranging module or radio frequency identification module for positioning or identity confirmation,

[0401] wherein each sub-identification maintains a fixed relative position and can be installed on the surface of a vehicle or an obstacle through an adhesive layer or a magnetic attraction member.

[0402] The multi-modal combined identification is designed as an integrated thin plate structure, with a total size generally ranging from A4 to A3 (210mm x 297mm to 297mm x 420mm) and a thickness controlled at 5-10mm to balance rigidity and light weight. The identification main body is made of weather-resistant engineering plastic or composite material, with a surface coated with a waterproof, anti-reflective, and anti-electromagnetic interference layer to ensure stable identification in rainy days, strong light, dust, or multi-source electromagnetic environments.

[0403] The combined identification is composed of multiple sub-identifications, each maintaining a fixed spatial relationship to form a rigid whole, ensuring that vehicles and external monitoring equipment can obtain consistent reference targets when observing different modalities. Among them:

[0404] · Visual identification: high-contrast AprilTag or ArUco patterns are used to achieve unique identification through black and white matrix coding. The pattern is coated to prevent fading or staining caused by ultraviolet light.

[0405] · Laser radar identification: high-reflectivity spheres with diameters between 50-100mm are evenly distributed on the surface of the thin plate, which can form strong return point cloud clusters in the laser radar point cloud. Optionally, laser intensity coding areas can also be designed on the plate surface to form identifiable patterns in the point cloud intensity values under specific arrangements.

[0406] · Millimeter wave radar identification: trihedral reflectors made of conductive materials such as aluminum alloy or copper plates are fixed on the plate body, with sizes meeting the effective reflection conditions of the millimeter wave frequency band to ensure a significant high-scattering cross-section area (RCS) in the vehicle radar return.

[0407] · Positioning / identity identification: UWB active tags or RFID / NFC chips are embedded inside or on the back of the plate body, providing centimeter-level ranging in obstructed environments and enabling obstacle identity confirmation through unique IDs.

[0408] In order to ensure installation convenience, the back of the combined identification is provided with a reusable adhesive layer or a magnetic fixing member, so that it can be quickly attached to the surface of a vehicle or an obstacle, and supports multiple detachments and movements. In actual use, it can be attached to the front, side or top of the obstacle according to the scene needs, to ensure that it can be seen within the field of view of the vehicle and the observation range of the external monitoring equipment.

[0409] In further embodiments, to enhance the adaptability of the identification, a transparent protective layer and an anti-fog coating layer can be additionally designed on the surface of the sheet to avoid visual recognition failure caused by rainwater or dust accumulation. For sites exposed to the outdoors for a long time, the identification material can also be selected from polycarbonates with ultraviolet resistance to prolong the service life.

[0410] Through the above structural design, the multi-modal combined identification can be recognized by cameras, laser radars, millimeter wave radars and UWB modules at the same time, achieving cross-modal unified observation targets. This design greatly reduces the cumbersome steps of arranging different modal targets in traditional calibration methods, ensures that different perception channels complete comparison and calibration under a unified target, and improves the efficiency and accuracy of the correction process.

[0411] Specifically, the obstacles include soft columns, dummy models, boxes or reflective plates, and the multi-modal combined identification is attached to the surface of the obstacles; the arrangement of the obstacles can be a regular grid, a ring, a fan-shaped arrangement, or a non-regular random placement, to simulate the distribution of obstacles in various driving environments.

[0412] In order to simulate the perception and obstacle avoidance effect in various driving environments, the obstacles not only exist as physical bodies, but also have the multi-modal combined identification attached to their surfaces, so that the vehicle measurement module and the external monitoring equipment can recognize at the same time.

[0413] The types of obstacles include but are not limited to:

[0414] • Soft column: made of flexible materials such as foam plastic or rubber, which will not be damaged even if the vehicle has slight contact, commonly used for simulation of road edges or narrow passages.

[0415] • Dummy model: manufactured according to human body size, used to simulate the scene of pedestrians crossing the road, and the surface can be attached with a combined identification for recognition by perception devices.

[0416] • Box: regular shape, commonly used as a low obstacle or a substitute for a stationary vehicle, with a flat surface suitable for installing large size identification.

[0417] . Reflective plate: can be used alone or as a carrier for combined identification, to enhance the recognition effect in long distance or night environment.

[0418] The arrangement of obstacles is divided into two categories:

[0419] 1. Regular arrangement:

[0420] Arrange in a preset grid or annular / fan array in the venue.

[0421] o When using grid arrangement, the grid spacing g is preferably 2-5 m, so that the vehicle always has multiple obstacles entering the sensor field of view during driving.

[0422] o When using annular arrangement, the radius band r of the obstacle is preferably in the range of 5-25 m, and the annular coverage angle is 180°-360°.

[0423] o When using fan arrangement, the sector angle θ s can be set to 30°-90° to simulate the presence of concentrated obstacles in part of the field of view.

[0424] 2. Random arrangement:

[0425] Under the premise of maintaining a safe boundary, use a random number generation algorithm to determine the placement of obstacles.

[0426] o The minimum safe spacing s min 1.0-1.5 m ensures sufficient static clearance between the vehicle's main driving path and the obstacles.

[0427] o The seed of the random arrangement is generated by "test number + timestamp", so that the arrangement is reproducible.

[0428] o Through random arrangement, the irregular distribution of pedestrians or obstacles in actual roads can be simulated, so as to test the performance of vehicle perception and avoidance algorithms under non-ideal conditions.

[0429] In further embodiments, in order to quantify the arrangement effect, the field of view (FOV) of the vehicle sensor is divided into angular and distance intervals, such as azimuth angle binning Δθ = 10° and radial binning Δr = 5 m, and it is counted whether each bin contains at least one obstacle with a combined identifier in the entire arrangement. Define coverage C = N eff / N tot , where N eff is the number of effective bins, N tot is the total number of bins, and when C ≥ 80%, the arrangement is determined to be qualified.

[0430] Through the above arrangement methods, not only can the vehicle observe obstacles in different directions and distances, but also by flexibly selecting regular or random arrangement mode, test conditions that are controllable and close to actual road environment can be simulated, so as to comprehensively verify the perception and correction capabilities of vehicle intelligent auxiliary driving system.

[0431] Specifically, the field-end true value measurement module includes one or more of the following: GNSS-RTK base station, UWB anchor point, optical motion capture array, laser total station, or laser tracker, for real-time acquisition of spatial position and attitude information of the vehicle and the obstacle with multi-modal combined identification.

[0432] The field-end true value measurement module is used to provide high-precision spatial position and attitude reference for the vehicle and the obstacle. According to different test environments, one or more of GNSS-RTK base station, UWB anchor point, optical motion capture array, laser total station, or laser tracker can be flexibly selected for combination.

[0433] In an open outdoor site, a GNSS-RTK base station is preferably used, fixed base stations are arranged at the edges of the site, and GNSS antennas are installed on the vehicle and the obstacle to receive differential signals, thereby obtaining centimeter-level three-dimensional coordinates p g =(x g ,y g ,z g ).

[0434] In an environment with building obstructions or semi-indoor, a UWB anchor point system can be used, N anchor points (N≥4) are arranged around the site, and UWB tags are built into the identification plates of the obstacle and the vehicle. The relative position p u =(x u ,y u ,z u ) is calculated by time of flight (TOF) or time difference of arrival (TDOA), which can still maintain centimeter-level positioning accuracy in severe obstructions.

[0435] In a small range test requiring fine attitude solution, an optical motion capture array can be used, reflective balls are arranged on the surfaces of the vehicle and the obstacle, M infrared cameras (M≥8) are arranged around the site, and high-precision position and rotation matrix T m =[R m |t m ] is obtained by multi-view triangulation and bundle adjustment, where R m is the rotation matrix, and t m is the translation vector, and the solution accuracy can reach millimeter level.

[0436] Under engineering-level verification conditions, a laser total station or a laser tracker can also be used. Reflective prisms or corner reflectors are fixed on the surface of the obstacle, and the measurement system obtains millimeter-level coordinate data p l =(x l ,y l ,z l ) by laser angle measurement, which can be used as the highest accuracy reference.

[0437] To ensure data continuity, the field end true value measurement module supports multi-source data fusion. For example, when GNSS-RTK is used as the main source in an open field, it is automatically switched to UWB positioning when entering a blocked area. If equipped with optical motion capture or a total station, it can be used as a high-precision check. The fusion process uses extended Kalman filtering (EKF) to obtain the final true value pose according to the covariance of each measurement source:

[0438] T truth = f(p g , p u , T m , p l ), ∑ truth = Cov(T truth )

[0439] where T truth represents the six-degree-of-freedom pose (position + attitude) of the vehicle or obstacle, and ∑ truth represents the corresponding covariance matrix, which describes the uncertainty of the true value.

[0440] Through the above design, the field end true value measurement module can be flexibly configured according to the characteristics of the site, ensuring that continuous, stable, and reliable reference information can be obtained under open, semi-indoor, or engineering precision requirements, providing a reliable reference for vehicle correction.

[0441] Specifically, the consistency indicators include:

[0442] Geometric indicators: lateral error, longitudinal error, and heading angle error of the vehicle and obstacles;

[0443] Semantic indicators: detection recall rate, detection accuracy rate, and intersection over union of recognized target boxes and true value boxes;

[0444] Cross-modal physical quantity indicators: difference distribution between camera recognition box center, laser radar clustering center, and millimeter wave radar ranging results, and matching rate of radar echo intensity and reflection cross-sectional area.

[0445] The evaluation and correction module establishes a multi-dimensional consistency index system to comprehensively evaluate the differences between vehicle recognition results and external true values. This index system considers geometric accuracy, semantic accuracy, and cross-modal consistency.

[0446] In terms of geometry, the system focuses on the lateral and longitudinal deviations between vehicles and obstacles, as well as the differences between vehicle heading angles and true value heading angles. These indicators can reflect the accuracy of vehicles in spatial position and attitude, and are commonly used to determine whether vehicles are running on the correct road center line and desired driving direction.

[0447] In terms of semantics, the system evaluates the accuracy and recall of the vehicle's identification of obstacles, confirming whether there are any missed or false detections. At the same time, it also calculates the degree of overlap between the vehicle's identification box and the external true value box, measuring the precision of the identification in the form of intersection over union. When the intersection over union is too low or the class identification is incorrect, it is considered that the vehicle's perception result has a significant deviation.

[0448] In terms of cross-modal, the system compares the identification results of the same obstacle by camera, lidar, and millimeter wave radar, checking whether their spatial positions are consistent. If the center of the camera detection box deviates too much from the centroid of the lidar point cloud, or the orientation and distance of the camera and millimeter wave radar differ significantly, it indicates that there is a problem with the cross-modal data fusion. In addition, the system also evaluates the matching degree between the actual echo strength measured by the millimeter wave radar and the theoretical value of the standard corner reflector. If the difference is large, it is determined that the radar identification stability is insufficient.

[0449] Through the calculation of these consistency indicators, the system can find perception deviations in multiple aspects, not only limited to geometric errors, but also reflecting the reliability of semantic identification and the coordination between multiple sensors. This provides comprehensive data support for subsequent correction, making the correction decision more targeted and complete.

[0450] Specifically, the report generated by the report generation module includes:

[0451] The arrangement of the scene and the type of the obstacle;

[0452] The statistical results of the consistency indicators and the error distribution curve;

[0453] The number of triggers, parameter change range, and rollback records of the gradual correction;

[0454] Traceability information for regulatory compliance verification.

[0455] The report generation module is used to automatically output a complete test and review report after the vehicle completes the correction process. The report records the vehicle operation data, external monitoring results, and correction steps in a unified format, ensuring that the test process is traceable, reproducible, and meets the regulatory compliance requirements.

[0456] In specific implementation, the report will first describe the test scene, including the type of the site, the arrangement, number, and location of the obstacles, as well as environmental conditions such as weather and lighting. Subsequently, the report details the comparison results between the vehicle measurement module and the field true value measurement module, covering geometric errors, avoidance effects, semantic consistency, and cross-modal comparison indicators. All indicators are provided with statistical values such as mean, root mean square error, percentile, and maximum deviation to facilitate a comprehensive analysis of system performance.

[0457] The content related to the correction is also recorded in detail, including the number of times the correction is triggered, the parameter category involved in each adjustment, the amplitude of the single adjustment, the cumulative change, and whether a rollback operation occurs. If a rollback occurs, the report will clearly mark the rollback time and the restored parameter set to ensure that all processes leave clear traces.

[0458] The final report is stored locally for the test personnel to view on site and is also uploaded to the cloud database. The cloud establishes separate archives for different vehicle models and sensor configurations, associates the current report with historical reports, and forms long-term performance curves and trend analyses. When the cloud finds that a certain type of vehicle repeatedly exhibits similar deviations in similar scenarios, it generates optimization suggestions or threshold adjustment prompts to assist subsequent batch corrections.

[0459] In this way, the report generation module not only provides a complete record for a single test but also enables comparison and tracking across test scenarios and vehicle models, making the entire correction system standardized, systematic, and sustainable for optimization.

[0460] Specifically, the gradual correction includes:

[0461] Based on the deviation results of the consistency indicators, the sensor extrinsic matrix or time synchronization parameters are iteratively corrected according to a preset small step size;

[0462] The maximum change is limited in a single correction to gradually approach the true value range;

[0463] When the cumulative correction exceeds the safe range, the parameters are rolled back to the last valid parameter set.

[0464] The gradual correction method adjusts the vehicle's perception parameters through small steps and gradual optimization, thereby avoiding instability caused by large-scale modifications. The trigger condition for correction comes from the comparison results of the aforementioned consistency indicators. When any indicator exceeds the set threshold, the system enters the correction mode.

[0465] During the correction process, the time synchronization deviation between sensors is first corrected to ensure that the data from different modalities are compared and fused under the same time reference. Subsequently, the system fine-tunes the extrinsic relationships between cameras, lidar, and millimeter wave radar, such as small adjustments to rotation angles or translation positions, to make the recognition results of the same target by different sensors more consistent. If necessary, small corrections can also be made to the internal parameters of the camera, such as fine-tuning the principal point position or focal length, to compensate for possible hardware drifts after long-term operation.

[0466] The adjustment range of each correction is strictly limited, and the single correction range does not exceed the preset safety step, and the cumulative change must also be kept within the total threshold. If multiple iterations fail to improve the indicators or introduce new recognition biases during the correction process, the system will automatically roll back to the last valid parameter set to avoid retaining incorrect correction results for a long time.

[0467] Through this gradual correction method, the perception system of the vehicle can gradually optimize without interrupting normal operation, ensuring system stability and effectively suppressing the drift problem caused by long-term sensor operation, thereby improving the reliability and consistency of the assisted driving function.

[0468] Specifically, the multi-modal combined identifier is an integrated thin plate structure with a reusable adhesive layer or magnetic fixing element on the back for mounting on the surface of the vehicle or obstacles. The relative positions between the sub-identifiers are fixed, and the thin plate has a waterproof, anti-reflection, and anti-interference surface treatment layer to ensure stable recognition in different environments.

[0469] The multi-modal combined identifier can be reliably identified in different environments, and the identifier is optimized in installation method and surface treatment.

[0470] In terms of installation, the back of the identifier is pre-integrated with a reusable fixing structure. For temporarily laid sites, an adhesive layer or industrial magic tape can be used to allow test personnel to quickly mount the identifier on the surface of the obstacle, and facilitate disassembly and reuse after the test is completed. For long-term fixed sites or environments that need to withstand wind load, magnetic structure, screw fixing seat or embedded card slot are preferred to allow the identifier to stably maintain its position and avoid loosening during long-term testing or multiple vehicle passes. The installation height is usually selected within the main field of view of the vehicle sensor, such as between 0.5 meters and 1.5 meters, to ensure that it is within the effective observation area during vehicle recognition and external monitoring.

[0471] In terms of surface treatment, the identifier is covered with a transparent protective film and coated with a waterproof and anti-fog coating to prevent rain, dust, or temperature difference condensation from causing recognition failure. At the same time, the visual code area uses anti-ultraviolet ink or film materials to ensure that it does not fade or become blurred during long-term outdoor use. For the reflective elements of laser radar and millimeter wave radar, the surface is treated with a high reflectivity plating or conductive coating to maintain stable echo characteristics. If necessary, the entire identifier plate body can be equipped with a lightweight protective cover that can block external impacts without affecting the identification of sensor features.

[0472] Through the above installation and surface treatment methods, the multi-modal combined identification not only realizes convenient layout and stable fixation, but also maintains the consistency and reliability of identification in complex environments such as rain, snow, strong wind, strong light or electromagnetic interference, thereby ensuring the continuity and accuracy of the correction process.

[0473] Specifically, the report generation module is also used to upload the correction results to a cloud server, and the cloud server establishes parameter archives and threshold templates according to different vehicle models, and uniformly stores, calls and updates the correction data of each vehicle, so as to facilitate batch consistency verification and compliance traceability.

[0474] The system will store all the data during the vehicle test locally after the correction is completed, and synchronize it to the cloud archive library to realize long-term preservation and subsequent analysis.

[0475] In terms of local storage, the system will index the vehicle identification data, external monitoring data, consistency index results, correction trigger records and final parameter set with test task numbers, and save them in complete packages. All files have time stamps and digital signatures to ensure that the test process is traceable and tamper-proof. Test personnel can directly call the data through the vehicle interface or external terminal for on-site analysis and verification.

[0476] In terms of cloud management, the server establishes an independent archive for each vehicle according to the vehicle model number, sensor configuration and test scene type. Each test report and raw data uploaded will be automatically archived and associated with historical data. The cloud system will automatically generate error trend curves, parameter adjustment frequency statistics tables and avoidance effect comparison charts to help the R&D team and quality inspection department understand the stability of the vehicle in long-term operation. If similar identification deviations or correction patterns repeatedly occur in different batches of tests for a certain type of vehicle, the cloud system will trigger a reminder and generate optimization suggestions, such as updating the default threshold template or adjusting the parameter initialization strategy.

[0477] In addition, the cloud archive library supports permission management, and users with different roles can access it as needed. For example, test engineers can view detailed correction data, manufacturers can call overall trend statistics, and regulatory agencies can directly verify whether the system meets the verification requirements of relevant regulations. Through this hierarchical management method, the security and privacy of the data are guaranteed, and the practical value of the archive library in R&D, production and supervision is improved.

[0478] Through the above design, the data storage and cloud archive management module not only realizes the saving of the results of a single test, but also establishes a continuous database across time and across vehicle models, so that the correction results of the vehicle intelligent auxiliary driving system can be tracked and continuously optimized for a long time.

[0479] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A vehicle intelligent auxiliary driving system correction system, comprising a vehicle measurement module, a multi-modal combined identifier, a field end true value measurement module, an evaluation and correction module, and a report generation module, characterized in that: the vehicle measurement module is arranged on the vehicle itself, comprising a camera, a laser radar, a millimeter wave radar, an inertial measurement unit, and a vehicle positioning unit, for collecting vehicle running state and environmental perception data, and serving as the basis input for correction; the multi-modal combined identifier is an integrated structure, provided with a visual coding pattern for camera recognition, a high reflectivity element or intensity coding structure for laser radar recognition, an angle reflector for millimeter wave radar recognition, and an ultra-wideband ranging module or radio frequency identification module for positioning or identity confirmation, each sub-identifier maintains a fixed relative position, and can be attached or installed on the vehicle or the surface of the obstacle; the field end true value measurement module is used to obtain the spatial position and attitude information of the vehicle and the obstacle with the multi-modal combined identifier in real time through a GNSS-RTK base station, a UWB anchor point, an optical motion capture array, or a total station; the evaluation and correction module is used to compare the recognition results of the vehicle measurement module with the field end true value, calculate the consistency index including geometric index, semantic index, and cross-modal physical quantity, and when the consistency index exceeds the threshold, perform gradual small-step fine tuning on the sensor external parameter or acquisition time delay parameter within the safe list range, and reserve the rollback point; the report generation module is used to output a report containing the scene arrangement mode, the consistency index statistical result, the correction parameter change range, and the traceability information. The method comprises the following steps: S1, vehicle measurement: collecting data of the camera, laser radar, millimeter wave radar, inertial measurement unit, and vehicle positioning unit by the vehicle measurement module, to complete the initial acquisition of the vehicle's own running state and environmental perception; S2, identifier arrangement: attaching the multi-modal combined identifier on the vehicle and the surface of the preset obstacle, which contains visual coding pattern, laser radar high reflectivity element or intensity coding structure, millimeter wave radar angle reflector, and ultra-wideband ranging module or radio frequency identification module; S3, true value measurement: obtaining the spatial position and attitude information of the vehicle and the obstacle with the combined identifier in real time through the field end true value measurement module using GNSS-RTK base station, UWB anchor point, optical motion capture array, or total station; S4, online comparison: comparing the perception recognition results of the vehicle measurement module with the field end true value, and calculating the consistency index including geometric index, semantic index, and cross-modal physical quantity; 2. The correction method of a correction system based on a vehicle intelligent auxiliary driving system according to claim 1, characterized in that, S5, gradual correction: when the consistency index exceeds the threshold, performing gradual small-step fine tuning on the vehicle sensor external parameter or acquisition time delay parameter within the safe list range, and reserving the rollback point for recovery; S6, report output: generating and outputting a report containing the scene arrangement mode, the consistency index statistical result, the correction parameter change range, and the traceability information. The multi-modal combined identifier is an integrated thin plate structure, comprising: a visual coding pattern for camera recognition; a high reflectivity sphere or intensity coding structure for laser radar recognition; a pluggable three-sided angle reflector for millimeter wave radar recognition; ​ 3. The correction system of claim 1, wherein ​ ​ ​ ​ An ultra-wideband ranging module or a radio frequency identification module for positioning or identity confirmation, Wherein the fixed relative positions between the sub-identifiers are maintained, and the sub-identifiers can be installed on the surface of a vehicle or an obstacle through an adhesive layer or a magnetic attraction member.

4. The correction system of claim 1, wherein: The obstacle includes a soft column, a dummy model, a box, or a reflector, and the multi-modal combined identifier is attached to the surface of the obstacle; the obstacle can be arranged in a regular grid, a ring, a fan shape, or irregularly and randomly, to simulate the distribution of obstacles in various driving environments.

5. The correction system of claim 1, wherein, The field end true value measurement module includes one or more of the following: a GNSS-RTK base station, a UWB anchor point, an optical motion capture array, a laser total station, or a laser tracker, for real-time acquisition of spatial position and attitude information of the vehicle and the obstacle with the multi-modal combined identifier.

6. The correction method of a correction system based on a vehicle intelligent auxiliary driving system according to claim 2, characterized in that, The consistency indicators include: Geometric indicators: lateral error, longitudinal error, and heading angle error of the vehicle and the obstacle; Semantic indicators: detection recall rate, detection accuracy rate, and intersection over union of the recognized target box and the true value box; Cross-modal physical quantity indicators: difference distribution between the camera recognition box center, the laser radar clustering center, and the millimeter wave radar ranging result, and matching rate of radar echo intensity and reflection cross-sectional area.

7. The correction method of a correction system based on a vehicle intelligent auxiliary driving system according to claim 2, characterized in that, The report output by the report generation module includes: Scene arrangement and obstacle type; Statistical results of consistency indicators and error distribution curves; Triggering number of progressive corrections, parameter change range, and rollback records; Traceability information for regulatory compliance verification.

8. The correction method of a correction system based on a vehicle intelligent auxiliary driving system according to claim 2, characterized in that, The progressive correction includes: Based on the deviation results of the consistency indicators, the sensor extrinsic matrix or the time synchronization parameter is iteratively corrected in a small step; Limit the maximum change in a single correction to gradually approach the true value range; When the cumulative correction exceeds the safe range, trigger parameter rollback to restore to the last effective parameter set.

9. The correction system of claim 1, wherein: The multi-modal combined identifier is an integrated thin plate structure with a reusable adhesive layer or magnetic attraction fixing member on the back for attachment to the surface of a vehicle or an obstacle. The relative positions between the sub-identifiers are fixed, and the thin plate has a waterproof, anti-reflective, and anti-interference surface treatment layer to ensure stable recognition in different environments.

10. The correction method of a correction system based on a vehicle intelligent auxiliary driving system according to claim 2, characterized in that, The report generation module is also used to upload the correction results to a cloud server, and the cloud server establishes parameter profiles and threshold templates for different vehicle models to uniformly store, call, and update the correction data of each vehicle, facilitating batch consistency verification and compliance traceability.

Images