Automatic driving freight formation cooperative anti-collision early warning method based on multi-sensor data fusion

Abstract

The invention discloses an automatic driving freight formation cooperative anti-collision early warning method based on multi-sensor data fusion, and the method comprises the following steps: 1, obtaining the data collected by a vehicle-mounted Beidou positioning receiver, a vehicle-mounted V2X communication module, a vehicle-mounted visual sensor, and a vehicle-mounted laser radar sensor of a target vehicle, carrying out the preprocessing, and unifying the data space-time reference; step 2, correcting a Beidou positioning error, removing rain and snow noise of the point cloud of the vehicle-mounted laser radar sensor, and improving the target recognition rate of the vehicle-mounted visual sensor in rainy and snowy weather; designing a weight dynamic distribution mechanism, and dynamically adjusting the weights of the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor and the vehicle-mounted V2X communication module; 3, based on the data obtained in the step 2, multi-vehicle collision risk assessment is carried out, and a risk level is obtained through judgment; and 4, according to the risk level obtained in the step 3, triggering early warning and control instructions with different intensities. According to the invention, the problem of insufficient accuracy and reliability of multi-vehicle cooperative anti-collision early warning can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent traffic anti-collision warning methods, especially applicable to freight platoon scenarios, and can realize multi-vehicle cooperative perception, risk assessment and hierarchical collision avoidance in the platoon, solve the problem of collision avoidance warning lag caused by single-vehicle perception limitations and data asynchronization in platoon driving, and specifically relates to a multi-sensor data fusion automatic driving freight platoon cooperative anti-collision warning method. BACKGROUND

[0002] In a rain and snow harsh weather environment, multi-vehicle chain collision accidents are easily caused, which poses a serious threat to road safety. In particular, trucks have a greater risk of chain collision in rain and snow weather due to their large mass and long braking distance.

[0003] Existing vehicle, such as truck, anti-collision systems still have significant deficiencies in actual application. On the one hand, rain and snow weather causes serious degradation of sensor performance, visual sensors are blocked by rain and snow, resulting in low target recognition rate, laser radar point cloud noise ratio is high, and Beidou positioning has significant atmospheric delay error; on the other hand, the multi-sensor fusion strategy is rigid, the fixed weight mechanism cannot dynamically adapt to weather changes, and the existing collision risk assessment model ignores the dynamic changes of road adhesion coefficient and the risk chain propagation effect of multiple vehicles, resulting in low warning accuracy. In addition, the cooperative collision avoidance strategy is single and lacks a hierarchical linkage mechanism, making it difficult to meet the safety needs of vehicle cooperative collision avoidance in rain and snow weather. Moreover, existing anti-collision systems are mostly designed for single vehicles or random multi-vehicle interaction, without considering the characteristics of freight platoons, resulting in problems such as loss of control of following distance, out-of-sync lane changing, and significantly higher risk of chain collision in rain and snow weather for vehicles in platoons than for single vehicles. SUMMARY

[0004] The present application provides a multi-sensor data fusion automatic driving freight platoon cooperative anti-collision warning method to solve the problems existing in the prior art based on vehicle sensor cooperative collision.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0006] The multi-sensor data fusion automatic driving freight platoon cooperative anti-collision warning method has the following process:

[0007] Step 1: The lead vehicle as the core of the platoon needs to first establish a dedicated communication link with all following vehicles through V2X, and clearly define the topology structure of "1 lead vehicle + N following vehicles" in the platoon. The lead vehicle receives sensor data uploaded by the following vehicles in real time, and simultaneously issues driving instructions to the following vehicles, forming a "perception-decision-control" platoon closed loop.

[0008] Then, the original positioning data of the target vehicle collected by the vehicle-mounted Beidou positioning receiver, the communication data collected by the vehicle-mounted V2X communication module, the front road condition image data collected by the vehicle-mounted visual sensor, and the 3D point cloud data of vehicles on the front road collected by the vehicle-mounted laser radar sensor are acquired;

[0009] Then, the data collected by the vehicle-mounted Beidou positioning receiver, the vehicle-mounted V2X communication module, the vehicle-mounted visual sensor, and the vehicle-mounted laser radar sensor are preprocessed, and the data space-time reference is unified.

[0010] Step 2, for the interference of the data collected by the vehicle-mounted Beidou positioning receiver and the vehicle-mounted V2X communication module, the vehicle-mounted visual sensor, and the vehicle-mounted laser radar sensor in rainy and snowy weather, the Beidou positioning error is corrected, the rain and snow noise of the point cloud of the vehicle-mounted laser radar sensor is removed, and the target recognition rate of the vehicle-mounted visual sensor in rainy and snowy weather is improved. A weight dynamic distribution mechanism based on the weather severity factor K is designed to dynamically adjust the weight of the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor, and the vehicle-mounted V2X communication module, and then realize the weighted fusion of the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor, and the vehicle-mounted V2X communication module in the target vehicle. After the head vehicle completes the weighted fusion of the multi-sensor data of itself and all following vehicles, the overall position of the platoon, the relative distance between each vehicle, and the target distribution around the platoon are broadcast to all following vehicles through V2X, helping the following vehicles to compensate for their own perception blind area.

[0011] Step 3, based on the data obtained in step 2, the head vehicle needs to conduct risk assessment from the global perspective of the platoon: on the one hand, it assesses the collision risk of the platoon and external vehicles, and on the other hand, it assesses the collision risk of the vehicles within the platoon; if the local risk level of a following vehicle is too high, the head vehicle needs to separately issue a speed reduction instruction to the two following vehicles, while reminding other vehicles to maintain the original state to avoid affecting the traffic efficiency of the platoon as a whole, and to assess the risk level through multi-vehicle collision risk assessment.

[0012] Step 4, according to the risk level obtained in step 3, trigger different intensity of warning and control instructions.

[0013] If it is the first risk level, the head vehicle sends a voice reminder of "maintain distance" to all following vehicles, and the following vehicle instrument panel displays the current distance of the platoon; if it is the second risk level, the head vehicle driver brakes while issuing a coordinated speed reduction instruction to all following vehicles through V2X, and the following vehicle automatically triggers the braking system to ensure that all vehicles in the platoon are synchronized to reduce speed and avoid rear-end collisions; if it is the third risk level, the head vehicle driver brakes urgently and turns on the double flash, while sending an "emergency avoidance" instruction to all following vehicles, and clearly indicating the avoidance direction of each following vehicle, and broadcasting a "platoon emergency collision avoidance" warning to the vehicles around the platoon through V2X to remind the surrounding vehicles to avoid.

[0014] In further step 1, the original positioning data collected by the vehicle-mounted Beidou positioning receiver is pre-processed by validity inspection, the communication data collected by the vehicle-mounted V2X communication module is pre-processed by filtering, the front road condition image data collected by the vehicle-mounted visual sensor is pre-processed by an image defogging algorithm, and the 3D point cloud data collected by the vehicle-mounted laser radar sensor is pre-processed by noise reduction.

[0015] In further step 1, the time of the vehicle-mounted Beidou positioning receiver is taken as a reference, time stamps are added to the data collected by the vehicle-mounted V2X communication, vehicle-mounted visual sensor and vehicle-mounted laser radar sensor, and clock alignment of the vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted visual sensor and vehicle-mounted laser radar sensor is realized by the vehicle-mounted data processing unit of the target vehicle, so as to complete time synchronization.

[0016] The 3D point cloud data collected by the vehicle-mounted laser radar sensor in its own coordinate system and the front road condition image data collected by the vehicle-mounted visual sensor in its own coordinate system are respectively converted into the coordinate system of the vehicle-mounted Beidou positioning receiver, so as to realize spatial synchronization.

[0017] Finally, through time synchronization and spatial synchronization, the time and space reference of the data collected by the vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted visual sensor and vehicle-mounted laser radar sensor is unified.

[0018] In further step 2, the Beidou positioning error correction process is as follows:

[0019] A1, obtaining the humidity data collected by the vehicle-mounted high-precision meteorological sensor in the target vehicle, the voltage value generated by the raindrop impact of the vehicle-mounted raindrop sensor, and calculating the raindrop density according to the voltage value converted from the raindrop impact signal;

[0020] A2, taking the time of the vehicle-mounted Beidou positioning receiver as a reference, adding time stamps to the humidity data collected by the vehicle-mounted meteorological sensor and the raindrop density data calculated based on the data collected by the vehicle-mounted raindrop sensor, and eliminating time offset by a sliding window algorithm;

[0021] A3, using the 3σ criterion, filtering the abnormal data in the humidity data and raindrop density data after eliminating time offset in step A2, and retaining the corresponding continuous and stable observation data sequence, so as to complete the outlier rejection and realize the time synchronization cleaning of the humidity data of the vehicle-mounted meteorological sensor and the raindrop density data of the vehicle-mounted raindrop sensor;

[0022] A4, calculating the zenith troposphere delay ZTD based on an atmospheric delay model, and then converting the zenith troposphere delay ZTD into troposphere delay TD1 at any satellite elevation angle θ by a mapping function NMF function;

[0023] A5, based on atmospheric humidity data H, raindrop density data D, the atmospheric delay compensation factor δ is dynamically calculated by the least square method, which is used to correct the troposphere delay, that is, the total delay TD after correction is: troposphere delay TD1+ atmospheric delay compensation factor δ;

[0024] A6, the positioning deviation of the vehicle-mounted Beidou positioning receiver is corrected by pseudo-range, and the positioning of the target vehicle is calculated based on the corrected pseudo-range, to obtain the three-dimensional coordinates of the target vehicle; wherein the weighted least square algorithm is used to correct the pseudo-range P corr Assigning different weights to optimize the positioning accuracy, assigning higher weights to pseudo-ranges with smaller errors and lower weights to pseudo-ranges with larger errors, and finally solving the three-dimensional coordinates of the target vehicle with the smallest weighted error square sum.

[0025] Further, in step 2, the process of removing rain and snow noise from the point cloud of the vehicle-mounted laser radar sensor is as follows:

[0026] B1, the invalid points in the original point cloud are preliminarily removed, the points beyond the effective detection range of the laser radar sensor are removed, and the points with reflectivity lower than the reflectivity threshold are filtered;

[0027] B2, time domain filtering is performed on the laser radar sensor point cloud after step B1, to obtain time domain candidate noise points;

[0028] B3, spatial domain filtering is performed on the points after time domain filtering in step B2, to obtain spatial domain candidate noise points;

[0029] B4, for the points marked as candidate noise points in both time domain and spatial domain obtained in steps B2 and B3, they are determined as rain and snow noise points and removed;

[0030] Then, radius filtering is performed on the remaining point cloud to eliminate a small amount of residual isolated noise; and the target cluster is extracted from the point cloud after radius filtering by the Euclidean clustering algorithm, and the reflectivity mean and motion consistency of the points in the cluster are calculated to further purify the point cloud.

[0031] Further, in step 2, the process of improving the target recognition rate of the vehicle-mounted visual sensor in rainy and snowy weather is as follows:

[0032] C1, raindrop / snowflake detection and removal are performed on the image, and the image is dehazed and enhanced, thereby suppressing the rain and snow noise in the image collected by the vehicle-mounted visual sensor, to obtain an image 1 after suppressing rain and snow noise;

[0033] C2, the contrast of the image 1 after suppressing rain and snow noise is enhanced by adaptive histogram equalization, and the details of the image are enhanced by multi-scale Retinex, to obtain an image 2 after enhancing contrast and details;

[0034] C3, color correction is performed on the image 2 after contrast and detail enhancement, and a saliency detection algorithm is combined to highlight the target region in the image.

[0035] In further step 2, a weight dynamic allocation mechanism based on weather severity factor K is designed, and the process of dynamically adjusting the weight and weighted fusion is as follows:

[0036] D1, based on humidity and raindrop density, the weather severity factor K is calculated; according to the weather severity factor K, the weather grade is divided into different grades;

[0037] D2, according to different weather grades, different weights are allocated to the data collected by the vehicle-mounted visual sensor, the data collected by the vehicle-mounted laser radar sensor, and the data collected by the vehicle-mounted V2X communication module;

[0038] D3, based on the allocated weights, the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor and the vehicle-mounted V2X communication module are weighted and averaged;

[0039] And the target confidence output by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor and the vehicle-mounted V2X communication module is multiplied by the corresponding weight and accumulated, and when the accumulated sum is greater than 0.6, it is determined as an effective target. The effective target here refers to a traffic participant or obstacle that is determined to exist and has reliable data after multi-sensor information fusion verification;

[0040] If one or two of the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor and the vehicle-mounted V2X communication module fail, the weight of the failed one is proportionally allocated to the other effective one.

[0041] In further step 3, the process of risk level evaluation and judgment of multi-vehicle collision risk is as follows:

[0042] (3.1), analyze the current state of the target vehicle, calculate the longitudinal safety distance, and evaluate the lateral slip risk;

[0043] (3.2), obtain the front, rear, and side vehicles and environmental information around the target vehicle, calculate the threshold value T through the two-degree-of-freedom collision time threshold model, and calculate the actual collision time TTC between the target vehicle and the front vehicle;

[0044] Compare the actual collision time TTC and the threshold value T, if TTC>T, it is determined that the collision risk is low, and the drivers of the target vehicle and its surrounding vehicles are reminded to pay attention to the road conditions and take measures; if TTC≤T, it is determined that the collision risk is high, and the drivers of the target vehicle and its surrounding vehicles are reminded to brake, change lanes or cooperate with deceleration;

[0045] (3.3), using risk amplification coefficient λ to evaluate the multi-vehicle chain collision risk of the target vehicle and its surrounding vehicles;

[0046] Then the side lane safety gap W is calculated safe , based on the side lane safety gap W safe The feasibility of cooperative collision avoidance is evaluated, if the side lane safety gap W safe >0, the target vehicle initiates a cooperative lane change request. If W safe ≤0, the target vehicle requests the surrounding vehicles to cooperate to slow down, and then changes lanes when the space is released, that is, W safe >0;

[0047] (3.4), based on the relative motion state, distance and snow environment of the two, the risk value R i,i+1 between vehicle i and vehicle i+1 is calculated; then, based on the risk value R i,i+1 , the comprehensive risk R 总 is calculated;

[0048] Based on the comprehensive risk R 总 , the risk level is judged and the corresponding collision warning state is triggered.

[0049] Further, in step 3, when 0.3≤R 总 <0.6, it is the first risk level; when 0.6≤R 总 <0.8, it is the second risk level; when R 总 >0.8, it is the third risk level.

[0050] The application is based on the cooperative anti-collision system of V2X technology and Beidou positioning technology, which realizes multi-vehicle linkage control by real-time acquisition of vehicle position, speed and environment data, and becomes a key means to improve vehicle driving safety.

[0051] The application uses Beidou positioning module, V2X communication module and visual, laser radar multi-sensor module to obtain vehicle information in real time, dynamically corrects Beidou positioning error based on atmospheric humidity and raindrop density, uses space-time domain joint filtering and image enhancement algorithm to process multi-sensor data, and integrates the weight of visual, laser radar and V2X data. The multi-vehicle cooperative risk evaluation model is used to calculate the collision risk, different intensity of warning and control instructions are triggered according to the risk level, and multi-vehicle state interaction and hierarchical cooperative collision avoidance control are realized through V2X communication technology.

[0052] The application can solve the problems of low accuracy and reliability of multi-vehicle cooperative anti-collision warning in rain and snow weather, and finally improve the accuracy and robustness of vehicle anti-collision warning in rain and snow weather. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1This is a flowchart of the method according to an embodiment of the present invention.

[0054] Figure 2 This is a collision warning strategy diagram based on risk level in an embodiment of the present invention.

[0055] Figure 3 This is a flowchart of the real-time trajectory update and optimization process in an embodiment of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0057] like Figure 1 As shown, this embodiment discloses a vehicle collision avoidance warning method, which combines V2X (Vehicle-to-Everything) communication technology, Beidou positioning technology, and vehicle-mounted visual sensors and vehicle-mounted lidar sensors to achieve multi-vehicle collaborative collision avoidance warning, including the following steps:

[0058] Step 1: In this embodiment, the on-board data processing unit of the target vehicle acquires the original positioning data of the target vehicle collected by the on-board Beidou positioning receiver, the communication data collected by the on-board V2X communication module, the road condition image data collected by the on-board vision sensor, and the 3D point cloud data of objects on the road ahead collected by the on-board LiDAR sensor.

[0059] The vehicle-mounted BeiDou positioning receiver is installed in the unobstructed central area of ​​the target vehicle's roof. The antenna phase center of the receiver is calibrated with the vehicle's centroid coordinates to ensure unobstructed satellite signal reception and that the positioning data matches the vehicle's actual location. The vehicle-mounted V2X communication module is fixed to the roof of the target vehicle's cab, ensuring 360° communication coverage. The vehicle-mounted visual sensor is installed in the rearview mirror position on the windshield of the target vehicle. The vehicle-mounted LiDAR sensor is installed at the front of the target vehicle's roof, providing a 120° field of view in front of the vehicle.

[0060] The vehicle-mounted BeiDou positioning receiver simultaneously receives signals from at least six satellites. In addition to communicating with V2X communication modules of surrounding vehicles, the vehicle-mounted V2X communication module also communicates with roadside equipment. The sampling frequency of the vehicle-mounted BeiDou positioning receiver is 10Hz, the sampling frequency of the vehicle-mounted vision sensor is 25fps, the sampling frequency of the vehicle-mounted LiDAR is 10Hz, and the sampling frequency of the vehicle-mounted V2X communication module is 5Hz.

[0061] In this embodiment, the raw positioning data collected by the vehicle-mounted BeiDou positioning receiver includes location information, motion parameters, and auxiliary information. The location information includes the target vehicle's latitude and longitude, and altitude. The motion parameters include the target vehicle's instantaneous speed and direction of travel. The auxiliary information includes the number of positioning satellites, the positioning type, and the timestamp.

[0062] In this embodiment, the communication data collected by the vehicle-mounted V2X communication module includes the basic safety message BSM broadcast by the surrounding vehicles (within a radius of 800 m) through their own V2X communication modules and the road safety information RSI sent by the roadside device. Among them, the basic safety message BSM of the surrounding vehicles contains the Beidou positioning information (including position, speed, heading angle) of the surrounding vehicles, vehicle state information (including brake signal, turn signal state, load level, etc.). The road safety information RSI sent by the roadside device contains the road wetness, snow intensity level, etc.

[0063] In this embodiment, the vehicle-mounted visual sensor collects the front road condition image of the target vehicle (resolution is 1920x1080), and the pixel coordinates of the lane line and obstacle (vehicle, pedestrian) can be obtained from the front road condition image.

[0064] In this embodiment, the vehicle-mounted laser radar sensor collects the 3D point cloud data of the object on the road in front of the target vehicle (point cloud density ≥ 200 points / ㎡), and the distance, azimuth angle, and reflectivity information of the object can be obtained from the 3D point cloud data.

[0065] Then, in this embodiment, the vehicle-mounted data processing unit preprocesses the data collected by the vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted visual sensor, and vehicle-mounted laser radar sensor, and unifies the data space-time reference. Among them:

[0066] In this embodiment, the original positioning data collected by the vehicle-mounted Beidou positioning receiver is preprocessed for validity test, and data with less than 4 satellites or single-point solution type is removed, and RTK fixed solution data is retained.

[0067] In this embodiment, the communication data collected by the vehicle-mounted V2X communication module is preprocessed by filtering, and messages with a delay of >100ms are removed based on the timestamp, and abnormal data is filtered through position legality verification.

[0068] In this embodiment, the front road condition image collected by the vehicle-mounted visual sensor is preprocessed by an image defogging algorithm to eliminate the influence of rain and fog.

[0069] In this embodiment, the 3D point cloud data collected by the vehicle-mounted laser radar sensor is preprocessed by noise reduction to reduce the noise in the point cloud data.

[0070] In this embodiment, when unifying the space-time reference of the data, the UTC time of the vehicle-mounted Beidou positioning receiver is taken as the reference, and time stamps (error ≤10 ms) are added to the data collected by the vehicle-mounted V2X communication, vehicle-mounted visual sensor, and vehicle-mounted laser radar sensor, respectively. The clock of each vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted visual sensor, and vehicle-mounted laser radar sensor is aligned through the PTP (Precision Time Protocol) of the vehicle-mounted data processing unit of the target vehicle, thereby completing the time synchronization. In this embodiment, the 3D point cloud data collected by the vehicle-mounted laser radar sensor in the three-dimensional Cartesian coordinate system and the front road condition image data collected by the vehicle-mounted visual sensor in the pixel coordinate system are converted into the WGS84 geodetic coordinate system of the vehicle-mounted Beidou positioning receiver, respectively, and the conversion error is ≤0.3 m, thereby realizing the space synchronization. Finally, through the time synchronization and the space synchronization, the space-time reference of the data collected by the vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted visual sensor, and vehicle-mounted laser radar sensor is unified.

[0071] In rainy and snowy weather, a single sensor is easily disturbed by the environment, resulting in fluctuations in data reliability. For example, the visual sensor has a decreased recognition rate due to rain and snow, and the laser radar point cloud has increased noise. Therefore, a weighted fusion is selected to solve the problem of uneven data reliability of multiple sensors, to provide a unified and reliable data source for subsequent collision risk assessment, to correct the Beidou positioning error, to remove the rain and snow noise of the vehicle-mounted laser radar sensor point cloud, and to improve the target recognition rate of the vehicle-mounted visual sensor in rainy and snowy weather. A weight dynamic distribution mechanism based on the weather severity factor K is designed to dynamically adjust the weights of the data collected by the vehicle-mounted visual sensor, vehicle-mounted laser radar sensor, and vehicle-mounted V2X communication module, and to realize the weighted fusion of the data collected by the vehicle-mounted visual sensor, vehicle-mounted laser radar sensor, and vehicle-mounted V2X communication module in the target vehicle.

[0072] In this embodiment, the process of correcting the Beidou positioning error is as follows:

[0073] A1. The vehicle-mounted data processing unit of the target vehicle acquires the data collected by the vehicle-mounted high-precision meteorological sensor and vehicle-mounted raindrop sensor in the target vehicle.

[0074] The vehicle-mounted high-precision meteorological sensor is installed on the top of the target vehicle in an unobstructed area to collect atmospheric humidity data (accuracy ±2%).

[0075] The vehicle-mounted raindrop sensor is installed on the outside of the front windshield of the target vehicle and coincides with the field of view of the vehicle-mounted visual sensor to ensure the accuracy of the raindrop impact signal acquisition.

[0076] The vehicle-mounted Beidou positioning receiver is fixed to the center of the roof, and the antenna phase center is calibrated with the vehicle mass center coordinates to ensure that the positioning data is consistent with the actual position of the vehicle.

[0077] And the vehicle-mounted meteorological sensor and the vehicle-mounted raindrop sensor are connected to the vehicle-mounted data processing unit through USB 3.0 interfaces respectively, the transmission rate is greater than or equal to 12 Mbps, and the sampling frequency is set to 1 Hz. The vehicle-mounted Beidou positioning receiver outputs raw observation data including pseudo-range, carrier phase, satellite ephemeris, etc. through an RS232 interface, the baud rate is 9600 bps, the sampling frequency is 10 Hz, and the high-frequency dynamic positioning demand is ensured.

[0078] The vehicle-mounted meteorological sensor collects atmospheric humidity data H in real time and outputs the data to the vehicle-mounted data processing unit. The vehicle-mounted raindrop sensor converts the raindrop impact signal to obtain a voltage value and transmits the voltage value to the vehicle-mounted data processing unit. The vehicle-mounted data processing unit calculates the raindrop density by using a raindrop density calculation formula D = 0.5 * U - 0.1, and the raindrop density is taken as the data output by the vehicle-mounted raindrop sensor, wherein U is the voltage value converted by the vehicle-mounted raindrop sensor, and D is the real-time raindrop density.

[0079] The vehicle-mounted data processing unit obtains the raw positioning data collected by the vehicle-mounted Beidou positioning receiver. The vehicle-mounted Beidou positioning receiver outputs the pseudo P raw (Units: m), carrier phase (Units: cycles), satellite elevation angle θ (Units: °), and signal-to-noise ratio SNR (Units: dB). The vehicle-mounted data processing unit selects satellite data with SNR greater than or equal to 30 dB to participate in subsequent calculation.

[0080] A2, the vehicle-mounted data processing unit takes the UTC time of the vehicle-mounted Beidou positioning receiver as a reference to add a time stamp (error less than or equal to 10 ms) to the humidity data collected by the vehicle-mounted meteorological sensor and the raindrop density data calculated based on the data collected by the vehicle-mounted raindrop sensor, and eliminates time offset by using a sliding window algorithm (window size 100 ms). The process of eliminating time offset by using the sliding window algorithm is as follows:

[0081] (A21) Window initialization.

[0082] The sliding window size is set to 100 ms, that is, the window contains all data points within 100 ms. The UTC time stamp output by the vehicle-mounted Beidou positioning receiver is recorded as T 北斗 , which is taken as a reference time axis.

[0083] Collect the humidity data collected by the vehicle-mounted meteorological sensor and the raindrop density data obtained by the vehicle-mounted raindrop sensor. The local time stamps of the humidity data collected by the vehicle-mounted meteorological sensor and the raindrop density data of the vehicle-mounted raindrop sensor are recorded as T 气象 and T 雨滴 , respectively (there may be clock deviation).

[0084] (A22) Data point mapping and deviation calculation.

[0085] The humidity data and raindrop density data in the window are preliminarily mapped onto the UTC time axis of the vehicle-mounted Beidou positioning receiver according to their respective local time stamps T 气象 , 雨滴 , the preliminary aligned humidity data time stamp T 气象初 = T 气象 + Δt1 and the preliminary aligned raindrop density data time stamp T 雨滴初 = T 雨滴 + Δt2, where Δt1 and Δt2 are initial time offsets to be calibrated.

[0086] The deviation of each data point in the window from the UTC time of the Beidou positioning receiver is calculated as shown in the following formula:

[0087] Humidity data deviation: δt 气象 = |T 气象初 -T 北斗 |,

[0088] Raindrop density data deviation: δt 雨滴 = |T 雨滴初 -T 北斗 |.

[0089] (A23) Offset optimization.

[0090] The sum of the squares of all humidity data deviations and the squares of all raindrop density data deviations in the window is minimized, and Δt1 and Δt2 are iteratively optimized by the least squares method as shown in the following formula:

[0091]

[0092] where n is the number of humidity data in the window, m is the number of raindrop density data in the window, δt 气象,i is the deviation of the i-th humidity data point in the window from the UTC time of the Beidou positioning receiver, and δt 雨滴,j is the deviation of the j-th raindrop density data point in the window from the UTC time of the Beidou positioning receiver.

[0093] When the maximum value of the sum of the squares of the deviations after iteration is ≤10 ms, the final time offset is determined.

[0094] A24) Window sliding and real-time updating.

[0095] The window is slid every 10 ms (sliding step = 10 ms), and steps A22) to A23) are repeated to use the latest optimized time offset for the humidity and raindrop density data newly entering the window.Correct the time stamp, ensure that the deviation of real-time humidity, raindrop density data and UTC time of vehicle-mounted Beidou positioning receiver is always ≤10 ms.

[0096] If a sudden time jump occurs in the window (such as sensor restart causing abnormal time stamp), the window is reset, and the initial offset is recalculated.

[0097] A3, the vehicle-mounted data processing unit adopts 3σ criterion, filters the abnormal data in the humidity data and raindrop density data after eliminating the time offset in step (A2), and retains the continuous and stable corresponding observation data sequence, thereby completing the outlier rejection and realizing the time synchronization cleaning of the humidity data of the vehicle-mounted weather sensor and the raindrop density data of the vehicle-mounted raindrop sensor. In this embodiment, when the humidity data jumps by >10% / s, it is considered as invalid humidity data, and when the raindrop density data jumps by >200 drops / m 2 ·s, it is considered as invalid raindrop density data.

[0098] A4, the vehicle-mounted data processing unit constructs an atmospheric delay model (Saastamoinen model), and calculates the zenith tropospheric delay ZTD based on the atmospheric delay model, as shown in the following formula:

[0099]

[0100] Wherein, P is the air pressure (hPa), φ is the latitude (°), h is the altitude (km), T is the air temperature (℃), e is the water vapor pressure (hPa), and H is the humidity (%).

[0101] Then, the zenith tropospheric delay ZTD is converted into the tropospheric delay TD1 of any satellite elevation angle θ through the mapping function NMF function, that is, Then, the corrected total delay is TD=TD1+δ, wherein θ is the satellite elevation angle, that is, the angle between the satellite and the horizontal plane of the target vehicle (unit: °), δ is the atmospheric delay compensation factor, and k is the empirical coefficient.

[0102] A5, the vehicle-mounted data processing unit dynamically calculates the atmospheric delay compensation factor δ. The atmospheric delay compensation factor δ is related to the humidity data H and the raindrop density data D, and the calculation formula is as follows:

[0103] δ=a×H+b×D+c,

[0104] Wherein, the coefficients a, b and c are obtained by least square fitting.

[0105] The process of obtaining coefficients a, b and c by least square fitting is as follows:

[0106] (A51) The vehicle-mounted data processing unit collects multiple sets of experimental data, including atmospheric humidity data H, raindrop density data D, and corresponding atmospheric delay compensation factor true value δture wherein, δ ture The difference between the data of the Beidou positioning receiver and the data of the CORS reference station is calculated to eliminate the non-atmospheric error factors.

[0107] (A52) Establish a least squares fitting linear model expression.

[0108] Let the least squares fitting linear model be δ = a x H + b x D + c + ε, where ε is an error term, and the goal is to solve the coefficients a, b, and c to minimize the total error sum of squares minQ(a, b, c), that is, where n is the number of samples, i is the i-th group of data, δ ture,i is the true value of the atmospheric delay compensation factor corresponding to the i-th group of data, H i is the atmospheric humidity data in the i-th group of data, D i is the raindrop density data in the i-th group of data.

[0109] (A53) Matrix transformation of the least squares fitting linear model.

[0110] Write the least squares fitting linear model as Y = X·β + ε.

[0111] where the observation vector Y = [δ ture,1 , δ ture,2 , …, δ ture,n ] T , the design matrix The parameter vector β = [a, b, c] T .

[0112] (A54) Solve the coefficients a, b, and c.

[0113] According to the least squares principle, the optimal solution of the parameter vector β is: where X T is the transpose matrix of X, (X T X) -1 is the inverse matrix of X T X.

[0114] Expand the error sum of squares Q, and take the partial derivative of the coefficients a, b, and c respectively and set the partial derivative results to 0 to obtain the following three first-order equations:

[0115]

[0116] Solving the three first-order equations can calculate the coefficients a, b, and c.

[0117] Finally in this embodiment, the vehicle-mounted data processing unit reacquires data every 30s to calculate the coefficients a, b, c (sliding window size 5min), when the weather state mutates (as mentioned in D1 below, when the weather severity factor K≥0.3, i.e. moderate to heavy snow), an instant update is triggered (update time consumption≤500ms), thus realizing dynamic adjustment of the coefficients a, b, c, to ensure that the calculated atmospheric delay compensation factor δ matches the real-time atmospheric state.

[0118] A6, the vehicle-mounted data processing unit corrects the positioning deviation of the vehicle-mounted Beidou positioning receiver, and performs positioning calculation of the target vehicle based on the corrected pseudo-range.

[0119] Specifically, the vehicle-mounted data processing unit corrects the original pseudo-range P raw of each satellite collected by the vehicle-mounted Beidou positioning receiver according to the following formula:

[0120] P corr = P raw -TD,

[0121] wherein θ is the satellite elevation angle; P corr is the corrected pseudo-range.

[0122] Then the vehicle-mounted data processing unit substitutes the corrected pseudo-range P corr into the weighted least squares algorithm to obtain the three-dimensional coordinates (including longitude, latitude, and height) of the target vehicle.

[0123] The weighted least squares algorithm is a method of optimizing positioning calculation accuracy by assigning different weights to different observations (here, the corrected pseudo-range P corr ).

[0124] Beidou positioning needs to calculate the vehicle position through the pseudo-range of at least 4 satellites, but the pseudo-range measurement accuracy of different satellites is different (e.g. low-elevation satellites are more affected by atmospheric interference and have higher errors; high-elevation satellites have shorter signal propagation paths and have smaller errors).

[0125] This algorithm assigns higher weights to pseudo-ranges with smaller errors and lower weights to pseudo-ranges with larger errors, so that the solution result relies more on reliable observations, and finally solves the three-dimensional coordinates of the vehicle that minimize the weighted error sum of squares, which can significantly improve the positioning accuracy compared to the ordinary least squares method (all data are treated equally).

[0126] Let the corrected pseudo-range be P corr,i (i is the satellite number), the corresponding theoretical pseudo-range (based on the geometric distance calculated by the vehicle coordinates), and the weight be ω i (inverse proportional to the pseudo-range error variance, the smaller the error, the larger the ω i ), then the objective function is: Solving the function obtains the optimal three-dimensional coordinates (x, y, z) of the vehicle.

[0127] In this embodiment, the process of removing rain and snow noise from the point cloud of the vehicle-mounted laser radar sensor is as follows:

[0128] B1, point cloud data preprocessing.

[0129] The vehicle-mounted data processing unit loads and formats the original point cloud collected by the vehicle-mounted laser radar sensor. Then, the original point cloud data collected by the vehicle-mounted laser radar sensor is read, and the three-dimensional coordinates, reflectivity, time stamp and laser beam number of each frame of point cloud are extracted.

[0130] Moreover, the vehicle-mounted data processing unit groups the multi-line laser radar data according to the beam, which facilitates subsequent processing according to the height layer (rain and snow particles are mainly distributed in the low-altitude area, and the height difference with the ground and vehicle target is significant).

[0131] Then, the vehicle-mounted data processing unit removes the invalid points in the original point cloud based on the phenomenon that rain and snow particles rarely appear in the extremely close or extremely far area, and the phenomenon that the reflectivity of rain and snow particles is generally lower than that of the road surface, vehicle and other targets. The points beyond the effective detection range of the laser radar sensor are removed, and the points with reflectivity lower than the reflectivity threshold are filtered. The reflectivity threshold can be selected according to the specific model of the laser radar sensor.

[0132] B2, the vehicle-mounted data processing unit performs time domain filtering on the laser radar sensor point cloud removed and filtered in step B1 based on the short-term and motion characteristics of rain and snow particles, to obtain time domain candidate noise points, the process is as follows:

[0133] (B21) The vehicle-mounted data processing unit performs multi-frame point cloud space-time alignment on the laser radar sensor point cloud removed and filtered.

[0134] Selecting three consecutive frames of point cloud, respectively denoted as the previous frame F t-1 , the current frame F t , and the next frame F t+1 , the time interval Δt = 0.1s matches the sampling frequency 10Hz of the vehicle-mounted laser radar sensor.

[0135] Taking the current frame F t as the reference, the points in the previous frame F t-1 and the next frame F t+1 are converted to the coordinate system of the current frame F t by target vehicle motion compensation, eliminating the point cloud offset caused by the motion of the target vehicle itself, thereby completing the multi-frame point cloud space-time alignment.

[0136] The process of multi-frame point cloud space-time alignment by target vehicle motion compensation is as follows:

[0137] B211) Obtain the current frame F from the vehicle-mounted Beidou positioning receiver t , the previous frame F t-1 , the next frame F t+1 Each corresponding target vehicle motion parameter, the motion parameter includes: instantaneous speed v (unit: m / s), reflecting the speed of the target vehicle driving; heading angle θ (unit: °), indicating the included angle between the driving direction of the target vehicle and the north direction; time stamp t, t-1, t+1, the inter-frame time interval Δt = t-(t-1) = (t+1)-t = 0.1s, matching the 10Hz sampling frequency of the vehicle-mounted laser radar sensor.

[0138] B212) Target vehicle motion modeling.

[0139] Assuming that the target vehicle is uniformly linear motion in Δt time, calculate F t-1 to F t , F t to F t+1 The displacement of the target vehicle in the process, including longitudinal displacement Δd (along the heading angle direction) and transverse displacement (perpendicular to the heading angle direction).

[0140] The longitudinal displacement Δd of the target vehicle is calculated as follows:

[0141] Δd = v × Δt

[0142] Since the target vehicle is uniformly linear motion in Δt time, the transverse displacement is 0. If the target vehicle turns, the transverse displacement needs to be corrected in combination with the angular velocity.

[0143] B213) Coordinate transformation matrix construction.

[0144] Taking the coordinate system of the current frame F t as the reference, assuming that the origin is (0, 0, 0), the coordinate transformation matrix of the target vehicle from the previous frame F t-1 , the next frame F t+1 to the current frame F t is constructed, including translation transformation and rotation transformation.

[0145] Among them, the translation transformation includes the translation amount of the target vehicle when the previous frame F t-1 relative to the current frame F t , the translation amount of the target vehicle when the next frame F t+1 relative to the current frame F t . F t-1 relative to F t , because the target vehicle drives forward from t-1 to t, F t-1 is in F tThe translation in the coordinate system is (-Δd×cosθ, -Δd×sinθ, 0), and the negative sign indicates reverse translation. t+1 With respect to F t , the target vehicle drives forward from t to t+1, F t+1 In F t , the translation in the coordinate system is (Δd×cosθ, Δd×sinθ, 0).

[0146] If the heading angle θ of the target vehicle changes between frames, i.e., the target vehicle turns, a rotation transformation matrix is constructed based on the heading angle difference Δθ to correct the direction offset of the point cloud. The rotation transformation matrix O is constructed based on the two-dimensional plane rotation formula, as follows:

[0147]

[0148] B214) Point cloud coordinate conversion.

[0149] For each point P(x, y, z) in the previous frame F t-1 and the next frame F t+1 , the above coordinate transformation matrix is applied to obtain the coordinates (x', y', z') of each point P(x, y, z) in the current frame F t-1 and F t+1 in the coordinate system of the current frame F t .

[0150] Wherein, the coordinates of each point P(x, y, z) in the previous frame F t-1 in the coordinate system of the current frame F t are as follows:

[0151]

[0152] The coordinates of each point P(x, y, z) in the next frame F t+1 in the coordinate system of the current frame F t are as follows:

[0153]

[0154] (B22), the vehicle-mounted data processing unit performs dynamic trajectory consistency detection on the spatio-temporally aligned multiple frames of point clouds, and completes time domain filtering.

[0155] Specifically, for each point P(x, y, z) in the current frame F t , the vehicle-mounted data processing unit searches for matching points from the previous frame F t-1 and the next frame F t+1 by KD tree nearest neighbor search method, thereby completing dynamic trajectory consistency detection, wherein the KD tree nearest neighbor search radius r = 0.3 m.

[0156] The process of finding matching points by KD-tree nearest neighbor search method is as follows:

[0157] B221) KD tree construction.

[0158] For the point cloud data of the previous frame F t-1 and the next frame F t+1 , extract the three-dimensional coordinates of all points in the previous frame F t-1 and the next frame F t+1 converted to the current frame F t coordinate system after target vehicle motion compensation; build a KD tree with three-dimensional coordinates (x, y, z) as the key for efficient nearest neighbor search.

[0159] B222) Target point traversal.

[0160] Traverse each point P (x, y, z) in the current frame F t , record the three-dimensional coordinates and reflectivity, line bundle number and other auxiliary information of each point P for subsequent matching verification.

[0161] B223) Nearest neighbor search and matching judgment.

[0162] With point P (x, y, z) as the center, set the search radius r = 0.3m, find all points falling within the sphere range in the point cloud of the previous frame F t-1 by KD tree, denoted as candidate point set S t-1 ; if S t-1 is empty, it is determined that there is no matching point in the previous frame F t-1 ; if S t-1 is not empty, select the nearest point to point P as the matching point S t-1 in the previous frame F t-1 . Thus, all matching points in the previous frame F t-1 are searched.

[0163] Using the same method, search for the nearest neighbor of point P in the point cloud of the next frame F t+1 , get the candidate point set S t+1 ; if S t+1 is empty, it is determined that there is no matching point in F t+1 ; if S t+1 is not empty, select the nearest point to P as the matching point S t+1 in the next frame F t+1 . Thus, all matching points in F t+1 are searched.

[0164] B224) Mark target points and time domain candidate noise points.

[0165] If a point P has a matched point in the adjacent frame and the motion speed is normal (v≤20m / s), the point P will be preliminarily classified as a "target point" which meets the motion characteristics of a real target.

[0166] If a point P has no matched point in F t-1 and F t+1 (i.e., only appears in a single frame), the point P is marked as a time-domain candidate noise point (rain / snow particles stay in the air for a short time, usually <3 frames); if a point P has a matched point in the adjacent frame, the motion speed v is calculated If v>20m / s (the falling / lateral motion speed of rain / snow particles is usually <10m / s, and a super-high-speed point may be noise), the point P is also marked as a time-domain candidate noise point. Time-domain candidate noise points are listed as potential noise because they exhibit the characteristics of rain / snow particles in the time domain (single-frame appearance or super-high-speed motion), but have not been verified in the spatial domain, so they are retained and await cross-verification with the spatial domain screening results.

[0167] B3. Based on the reflectivity and distribution characteristics of rain / snow particles, the vehicle-mounted data processing unit performs spatial domain filtering on the points after time domain filtering in step B2 to obtain spatial domain candidate noise points, as follows:

[0168] (B31) Analysis of statistical characteristics of reflectivity.

[0169] For the points retained after time domain filtering (i.e., target points and time-domain candidate noise points), the local reflectivity variance of target points and time-domain candidate noise points is calculated.

[0170] Specifically, taking point P (i.e., all points retained after time domain filtering, including target points and time-domain candidate noise points, as point P) as the center, a neighborhood with a radius of r=0.5m is constructed, and the mean μ int and variance σ int of the reflectivity of all points in the neighborhood are calculated.

[0171] If |intensityy P -μ int |>3σ int (indicating that the reflectivity deviates from the overall characteristics of the neighborhood), and intensityy P <0.3 (indicating that it is lower than the typical target reflectivity, and the fixed threshold of 0.3 is designed based on the inherent difference between noise and target reflectivity, and is set in coordination with the reflectivity variance condition to form a more robust spatial domain noise recognition rule), the point P is preliminarily marked as a spatial domain candidate noise point (indicating that the reflectivity of the rain / snow particles of the point P is random and low). Here, intensityy P is the reflectivity value of point P.

[0172] (B32) Spatial distribution discreteness detection.

[0173] Calculate the neighborhood point density p (number of points per unit volume) of point P, as shown in the following formula, where point P refers to all points remaining after the reflectivity characteristic screening of step B31, including both points not marked as spatial domain candidate noise points and points preliminarily marked as spatial domain candidate noise points.

[0174]

[0175] Where N 领域 is the number of all points contained in the spherical neighborhood with point P as the center and a radius r = 0.5 m (unit: pieces), V 领域 is the volume of the spherical neighborhood with point P as the center and a radius r = 0.5 m (unit: m 3 ), and r is the neighborhood radius.

[0176] If p < p0 (p0 is an empirical density threshold), it indicates that there is no continuous target around point P, which meets the discrete distribution characteristic of rain and snow particles, and point P is further confirmed as a spatial domain candidate noise point.

[0177] B4, the vehicle-mounted data processing unit determines the points marked as candidate noise points in both the time domain and the spatial domain obtained through steps B2 and B3 as rain and snow noise points and eliminates them.

[0178] Then, the vehicle-mounted data processing unit performs radius filtering (retaining points with a neighborhood point number ≥ 3) on the remaining point cloud to eliminate a small amount of residual isolated noise; and extracts target clusters from the point cloud after radius filtering through the Euclidean clustering algorithm, calculates the reflectivity mean and motion consistency of the points in the cluster, and further purifies the point cloud.

[0179] Where the process of eliminating a small amount of residual isolated noise through radius filtering is as follows:

[0180] (B401) Construct a KD tree.

[0181] For the remaining point cloud data, a KD tree index structure of a three-dimensional space is constructed. The KD tree is a tree-shaped data structure for storing and quickly querying data points in a k-dimensional space, which can effectively accelerate the subsequent neighborhood point searching process. Traverse the point cloud, take the three-dimensional coordinates (x, y, z) of each point as the node data of the KD tree, and construct the KD tree structure according to a certain partition rule (such as alternately along the x, y, and z axis directions, and selecting the median point for partitioning).

[0182] (B402) Set parameters.

[0183] Determine the neighborhood radius r (set according to the actual point cloud density and scene requirements), and the neighborhood point number threshold N thresh = 3.

[0184] (B403) Traverse the remaining point cloud for filtering.

[0185] Take each point P in the remaining point cloud in turn, and use the constructed KD tree to query all neighborhood points within a radius r range centered on P. Count the number of neighborhood points N, if N≥N thresh , keep the point P; if N<N thresh , determine it as an isolated noise point and remove it from the remaining point cloud.

[0186] (B403) Output the result.

[0187] After the above traversal and judgment, the radius filtering of the remaining point cloud data is completed, and most of the isolated noise points with too few neighborhood points are removed through radius filtering.

[0188] Among them, the target cluster is extracted by the Euclidean clustering algorithm, the reflectivity mean and motion consistency of the points in the cluster are calculated, and the process of further purifying the point cloud is as follows:

[0189] (B411) Initialization.

[0190] Create an empty cluster list to store the extracted target cluster. Create a flag array to mark whether each point in the radius filtered point cloud data has been assigned to a cluster, and initially all points are unassigned (0 represents unassigned, 1 represents assigned).

[0191] (B412) Select seed point.

[0192] Traverse the radius filtered point cloud data to find the first unmarked point P seed , mark the point P seed as a seed point, create a new cluster, and mark the seed point P seed as assigned.

[0193] (B413) Grow cluster.

[0194] Take the seed point P seed as the center, and use the KD tree to query all neighborhood points within a distance threshold d thresh ( determined according to the target size and scene). Add the unmarked points in these neighborhood points to the current cluster and mark them as assigned; then take the newly added point as a new seed point, repeat the above process of querying neighborhood points and growing clusters until there are no new points that can be added to the current cluster.

[0195] (B414) Repeat operation.

[0196] Returning to steps (B412)-(B413), the process of growing the cluster is repeated by continuing to find the next unmarked point as a seed point until all points in the point cloud are marked as assigned, i.e., all points are divided into corresponding clusters.

[0197] (B415) calculating the reflectivity mean value and motion consistency of points in the cluster.

[0198] Specifically, for each extracted cluster, all points in the cluster are traversed to obtain the reflectivity value of each point, and the reflectivity mean value of the cluster is obtained by dividing the sum by the number of points in the cluster. The reflectivity mean value can be used to distinguish different types of targets.

[0199] In combination with the motion information of the vehicle and the position change of the points in the cluster in consecutive frames, whether the motion speed and direction of the points in the cluster are consistent is calculated. If the motion speed and direction of the points in the cluster differ greatly, the cluster may contain noise points or multiple targets with different motion states, which need to be further analyzed and purified.

[0200] (B416) outputting the target cluster.

[0201] After the above steps, a purified target cluster set is obtained, and each cluster corresponds to a potential target, which can be used for subsequent target identification and tracking processing.

[0202] B5, for rain and snow intensity, the vehicle-mounted data processing unit dynamically optimizes the threshold in combination with the weather severity factor K calculated subsequently to realize adaptive adjustment of algorithm parameters.

[0203] The adaptive adjustment logic is as follows: the weather severity factor K reflects the real-time change of rain and snow intensity. When the weather severity factor K is less than 0.3 (light rain and snow), a relatively loose threshold is adopted to reduce the misjudgment of real target points. When K is greater than or equal to 0.3 (medium to heavy snow), real-time parameter adjustment is triggered, and the noise elimination capability is enhanced by using a stricter threshold, while avoiding the loss of real targets caused by excessive filtering.

[0204] According to the weather severity factor K, the threshold parameters of the following steps are adjusted to realize adaptive adjustment of parameters:

[0205] Time domain filtering (step B2): adjust the super high speed point determination threshold, the default threshold is 20 m / s, when K decreases (rain and snow weaken), the threshold can be appropriately increased (such as 25 m / s), when K increases (rain and snow strengthens), the threshold can be appropriately reduced (such as 15 m / s), to more strictly screen abnormal motion points caused by rain and snow interference.

[0206] Spatial domain filtering (step B3): adjust the reflectance threshold, the default reflectance threshold is 0.3, when K decreases (rain and snow weaken), the threshold can be appropriately reduced (such as 0.25), when K increases (rain and snow strengthen), the threshold can be appropriately increased (such as 0.35), and low reflectance rain and snow points are more strictly removed.

[0207] Radius filtering (step B4): adjust the neighborhood point number threshold, by default, points with a point number ≥3 in the neighborhood are retained, when the rain and snow weaken, it can be increased to ≥4, and when the rain and snow is relatively strong, it can be reduced to ≥2, thereby reducing the probability of real target points being mistakenly deleted.

[0208] In this embodiment, the vehicle-mounted data processing unit uses an image enhancement algorithm to improve the target recognition rate of the vehicle-mounted visual sensor in rainy and snowy weather, and the specific process is as follows:

[0209] C1, suppress rain and snow noise in the image collected by the vehicle-mounted visual sensor, comprising the following steps:

[0210] (C11) Raindrop / snowflake detection and removal.

[0211] The morphological characteristics of raindrops (circular / elliptical / high brightness) and snowflakes (irregular blocky / low contrast) are used to locate the rain and snow noise area in the image collected by the vehicle-mounted visual sensor through Top-Hat transformation (extracting bright areas) and Bottom-Hat transformation (extracting dark areas).

[0212] Wherein, the rain and snow noise area positioning process is as follows:

[0213] C111) Preprocessing.

[0214] In rainy and snowy weather, the original image I is first dehazed to enhance the contrast of the image and highlight the morphological characteristics of rain and snow noise, as shown in the following formula:

[0215] I dehaze = Dehaze(I)

[0216] Wherein, I dehaze is the dehazed image, Dehaze() is an image dehazing algorithm, which is used to eliminate image blur caused by water vapor scattering in rainy and snowy weather, enhance image contrast, and provide clearer input for subsequent rain and snow noise detection (such as raindrop / snowflake positioning).

[0217] C112) Structure element design.

[0218] The structure element s is designed according to the morphological characteristics of rain and snow, including raindrops and snowflakes. Wherein, the raindrop is circular or elliptical, and the circular structure element s rain is adopted. The snowflake is irregular blocky, and the square or diamond structure element s snow is adopted.

[0219] C113) Top-Hat transform extracts raindrops (bright regions).

[0220] First, the dehazed image I dehaze is subjected to an opening operation (erosion + dilation) to suppress the background and retain large structures, as shown in the following equation:

[0221] I open = I dehaze · s rain

[0222] where I open is the output image after the opening operation is performed on the dehazed image I dehaze , and s rain is a structural element (typically a circular or square convolution kernel) used for the opening operation.

[0223] Then, the bright regions are calculated by the Top-Hat transform, which subtracts the opening operation result from the original image, to extract raindrop regions that are brighter than the background, as shown in the following equation:

[0224] I tophat = I dehaze - I open .

[0225] where I tophat is the output image obtained by the Top-Hat transform, which reflects regions in the original dehazed image that are brighter than the background.

[0226] C114) Bottom-Hat transform extracts snowflakes (dark regions).

[0227] First, the dehazed image I dehaze is subjected to a closing operation (dilation + erosion) to enhance dark regions and suppress small structures, as shown in the following equation:

[0228] I close = I dehaze · z snow

[0229] where I close is the output image after the closing operation is performed on the dehazed image I dehaze , and s snow is a structural element (typically a circular or square convolution kernel) used for the closing operation.

[0230] Then, the dark regions are calculated by the Bottom-Hat transform, which subtracts the original image from the closing operation result, to extract snowflake regions that are darker than the background, as shown in the following equation:

[0231] I bottomhat = I close - Idehaze .

[0232] where I bottomhat is the output image by Bottom-Hat transform, reflecting the regions in the original haze-removed image that are darker than the background.

[0233] C115) Adaptive inpainting.

[0234] After detecting the noise regions, i.e. the bright regions corresponding to raindrops and the dark regions corresponding to snowflakes, for small-area noise regions, weighted mean filtering is used for image modification, and for large-area noise regions, a texture-based method is used for image inpainting.

[0235] Specifically, for small-area noise regions (such as raindrops with a diameter < 5 pixels), the weighted average value of non-noisy pixels in the neighborhood is used to fill (the closer the distance, the higher the weight). For large-area noise regions (such as snowflake clusters), the Fast Marching Method (FMM) is used to diffuse the surrounding texture information to restore the occluded background details.

[0236] (C12), for scattering blur caused by rain and snow, haze-removal enhancement is performed on image 1, where image 1 refers to the image that has been preliminarily denoised (raindrops / snowflakes removed) but still has scattering blur.

[0237] Specifically, based on the characteristic that in a haze-free image, at least one pixel value of a local region is close to 0, the dark channel (i.e. the minimum value of each pixel in the R, G, B channels in the image) of image 1 and the atmospheric light value are calculated. The application process is as follows:

[0238] C121) Calculate the dark channel image.

[0239] Using the above characteristic, the dark channel of the input image (the image that has been preliminarily denoised but still has scattering blur) is calculated as follows: where I dark (x, y) is the pixel value of the pixel with coordinates (x, y) in the dark channel image; c is the color channel index of the image, taking values {R, G, B}, corresponding to the red, green, and blue channels respectively; (i, j) is the pixel coordinate in the local region of the input image, used to traverse all pixels in the local region centered at (x, y); I c is the R / G / B channel of the image; I c (i, j) is the pixel value of the pixel with coordinates (i, j) in the color channel c of the input image; Ω(x, y) is the local region centered at (x, y), used to calculate the brightness characteristics of the pixels in the region.

[0240] C122) Estimate the atmospheric light value A.

[0241] The highest 0.1% of the luminance of the dark channel image is selected, and the average luminance of the pixels in the original image corresponding to the highest 0.1% of the luminance of the dark channel image is the atmospheric light value A (reflecting the overall brightness of the environment in the fog). This step relies on the characteristic that the dark channel of the non-fog area is close to 0: the high-luminance pixels in the dark channel must come from the fog area, and the luminance of the corresponding original pixels can represent the atmospheric light.

[0242] C123) Calculate the atmospheric transmittance t.

[0243] Based on the dark channel value and the atmospheric light value, the transmittance is estimated by the formula: Where t is the atmospheric transmittance; t(x, y) is the atmospheric transmittance value of the pixel point with coordinates (x, y) in the image, reflecting the local fog / snow concentration at that position; ω is a correction coefficient (usually 0.95), A c is the component value of the atmospheric light in the color channel c (i.e., the red, green, and blue channel intensities of the atmospheric light). Since the dark channel of the non-fog area is close to 0, its transmittance will be close to 1 (no fog); the fog area has a high dark channel value, and the transmittance is reduced (consistent with the characteristics of fog concentration).

[0244] C124) Restore the clear image.

[0245] The transmittance (reflecting the fog concentration) is estimated by the dark channel, and the guided filter is introduced to optimize the smoothness of the transmittance, based on the formula The image 1 is restored to a clear image, thereby completing the dehazing enhancement for scattering blur caused by rain and snow, and improving the scene visibility. Where (x, y) is the two-dimensional coordinates of the pixel in the image; J is the dehazed image, J(x, y) is the pixel value at coordinates (x, y) in the clear image after dehazing, which is the final output enhancement result; I is the original image, I(x, y) is the pixel value at coordinates (x, y) in the input image (i.e., image 1) to be dehazed; A is the atmospheric light value; t is the transmittance, t(x, y) is the transmittance value at coordinates (x, y) after optimization by the guided filter, reflecting the fog / snow concentration at that position (the value closer to 1, the lighter the fog).

[0246] C2, the contrast and details of the image 2 (the image 2 is the clear image J after dehazing output by step C12) are enhanced.

[0247] Where the image is subjected to adaptive histogram equalization (CLAHE) to enhance the contrast. Specifically, to solve the problem of low local contrast of the image in rainy and snowy weather, the image is divided into 8x8 subblocks, and histogram equalization is performed on each subblock. At the same time, by limiting the contrast threshold, the noise amplification is avoided, and the gray difference between the target and the background is enhanced, thereby completing the contrast enhancement.

[0248] And, the image is subjected to multi-scale Retinex enhancement. Specifically, the adaptation mechanism of human eyes to different illumination conditions is simulated, the illumination component (low frequency, reflecting the overall distribution of light) of the image is extracted by using Gaussian filtering, and the strong light area (such as snow reflection) of the image is suppressed by logarithmic transformation; the reflection component (high frequency, reflecting the details of the target) of the image is subjected to dynamic range compression, and the dark details (such as pedestrians in the shadow on a rainy day) are enhanced. The illumination component and the reflection component after fusion processing, thereby improving the overall clarity and color fidelity of the image, and completing the detail enhancement.

[0249] C3, color correction and target region highlighting are performed on the image 3 (the image 3 is the image after the contrast and detail enhancement processing in step C2).

[0250] Among them, since the image is prone to color deviation in rainy and snowy weather (such as blue in rainy days and white in snowy days), the image is subjected to color correction by using the gray world algorithm. Specifically, it is assumed that the average gray values of R, G and B channels in the image are equal, the gain coefficients of each channel are calculated, the color deviation is balanced, and the true color of the target is restored, thereby completing the color correction of the image.

[0251] And, combined with the saliency detection algorithm (spectrum residual method), the possible target region (such as vehicle, pedestrian) in the image is located, the target region is subjected to local sharpening (Laplacian operator), and the background noise is suppressed at the same time, so that the target contour is clearer, thereby completing the target region highlighting of the image, and providing high-quality input for subsequent YOLO, Faster R-CNN and other recognition algorithms.

[0252] C4, the weather severity factor K obtained after the subsequent calculation is introduced into the vehicle-mounted data processing unit for dynamic adjustment, so as to realize adaptive adjustment of algorithm parameters.

[0253] When the weather severity factor K is less than 0.3, it indicates that it is light rain and snow, and the denoising and light contrast enhancement are focused on, so as to avoid excessive processing leading to loss of details. When the weather severity factor K is greater than or equal to 0.3, it indicates that it is medium to heavy snow, and the defogging intensity and local sharpening are strengthened at this time, so as to improve the overall permeability of the image.

[0254] In this embodiment, in rainy and snowy weather, a single sensor is prone to environmental interference, leading to fluctuation of data reliability, such as decreased recognition rate of visual sensor due to rain and snow blur, increased noise of laser radar point cloud, etc. Therefore, weighted fusion is selected to solve the problem of uneven data reliability of multiple sensors. The vehicle-mounted data processing unit designs a weight dynamic distribution mechanism based on the weather severity factor K, dynamically adjusts the weights of the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor and the vehicle-mounted V2X communication module, and then realizes the weighted fusion of the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor and the vehicle-mounted V2X communication module in the target vehicle. The specific process is as follows:

[0255] D1. Calculate the weather severity factor K, including the following steps:

[0256] (D11) The onboard data processing unit collects core parameters. These core parameters include the raindrop density D (unit: drops / m³) obtained from the onboard raindrop sensor. 2 •s), and atmospheric humidity H (unit: %) obtained by the vehicle-mounted meteorological sensor, and echo intensity standard deviation σ obtained by the vehicle-mounted lidar sensor. int echo intensity standard deviation σ int It reflects the level of noise from rain and snow.

[0257] (D12) The weather severity factor K is calculated using a weighted formula, and then normalized to the range [0,1], as shown in the following formula:

[0258]

[0259] In the formula, 200 is the maximum raindrop density; 0.3 is the standard deviation threshold of laser echo intensity; coefficients A, B, and C are assigned according to the sensor reliability weight.

[0260] The specific calculation steps for coefficients A, B, and C need to be combined with the sensor reliability assessment results. The following is a detailed calculation process based on the Analytic Hierarchy Process (AHP) to ensure that the results meet the constraint condition A+B+C=1:

[0261] D121) Define the correspondence between parameters and sensors, as shown in the table below:

[0262]

[0263] D122) Construct a reliability judgment matrix.

[0264] Based on experimental data, the reliability of the three sensors was compared pairwise using a 1-9 scale (1 = equally reliable, 3 = slightly reliable, 5 = significantly reliable, 7 = strongly reliable, 9 = extremely reliable, with the reciprocal indicating a reverse comparison), resulting in a judgment matrix. Where, β HH β HD , β DH β DD , The values ​​are assigned by comparison using the scaling method.

[0265] D123) Calculate the weighting coefficients A, B, and C.

[0266] First, calculate the product M of the elements in each row of the judgment matrix. H M D , As shown in the following formula:

[0267]

[0268]

[0269] After that, the 3rd root of each row product is calculated As shown in the following formula:

[0270]

[0271] Finally, normalization processing is performed to obtain weight coefficients A, B and C:

[0272] Sum

[0273] (D13) According to the weather severity factor K, the weather grade is divided.

[0274] When K < 0.3, it indicates light rain and snow, and the visibility is good, and the sensor is less affected.

[0275] When 0.3≤K<0.7, it indicates moderate rain and snow, and the sensor performance is significantly reduced.

[0276] When K≥0.7, it indicates heavy rain and snow, and the sensor is seriously disturbed.

[0277] D2, dynamically allocate weights.

[0278] When it is light rain and snow, i.e. K < 0.3, the data collected by the vehicle-mounted visual sensor is allocated a weight of 50%, at which time the image is blurred, retaining its advantage in identifying lane lines and traffic signs. The data collected by the vehicle-mounted laser radar sensor is allocated a weight of 30%, at which time the point cloud noise is less, providing accurate distance information. The data collected by the vehicle-mounted V2X communication module is allocated a weight of 20%, which is used to assist verification and supplement long-distance target information.

[0279] When it is moderate rain and snow, i.e. 0.3≤K<0.7, the data collected by the vehicle-mounted visual sensor is allocated a weight of 20%, at which time the image is blurred and only retains part of the texture information. The data collected by the vehicle-mounted laser radar sensor is allocated a weight of 40%, at which time the denoised point cloud is still reliable and becomes the main force of medium-distance perception. The data collected by the vehicle-mounted V2X communication module is allocated a weight of 40%, which is used to break through the line of sight obstruction and provide surrounding vehicle interaction information.

[0280] When heavy rain and snow, K≥0.7, the data collected by the vehicle-mounted visual sensor is assigned a weight of 5%, at which point the image is severely distorted and only serves as a reference for extreme cases. The data collected by the vehicle-mounted laser radar sensor is assigned a weight of 35%, at which point the near-distance point cloud is still effective and assists in verifying the surrounding environment. The data collected by the vehicle-mounted V2X communication module is assigned a weight of 60%, becoming the core perception source, and obtaining unobstructed target data through vehicle-to-vehicle communication.

[0281] D3, based on the assigned weights, the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor, and the vehicle-mounted V2X communication module are fused and optimized. The process is as follows:

[0282] (D31), real-time updating of weights.

[0283] Every 100ms is a period to recalculate the value of the weather severity factor K according to step D1, and the weights of the data collected by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor, and the vehicle-mounted V2X communication module are updated synchronously according to step D2, ensuring matching with the current weather state.

[0284] (D32), weighted fusion of multi-source data.

[0285] The target frame detected by the vehicle-mounted visual sensor, the point cloud center clustered by the vehicle-mounted laser radar sensor, and the coordinates broadcast by the vehicle-mounted V2X communication module are weighted and averaged according to the weights, and Kalman filtering is combined to eliminate jitter, outputting a smooth trajectory.

[0286] And the target confidence output by the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor, and the vehicle-mounted V2X communication module (such as the IOU of the vehicle-mounted visual sensor, the point cloud density of the vehicle-mounted laser radar sensor, and the target confidence parameter of the vehicle-mounted V2X communication module) is multiplied by the corresponding weight and accumulated, and when the cumulative sum is greater than 0.6, it is determined to be an effective target. Here, the effective target refers to a traffic participant or obstacle that is determined to exist and have reliable data after multi-sensor information fusion verification.

[0287] If one or two of the vehicle-mounted visual sensor, the vehicle-mounted laser radar sensor, and the vehicle-mounted V2X communication module fail (such as the frame rate of the vehicle-mounted visual sensor being less than 10fps, the point cloud frame rate of the vehicle-mounted laser radar sensor being less than 5Hz, and the communication of the vehicle-mounted V2X communication module being interrupted), the weight of the failed one is automatically distributed to the other effective one in proportion, ensuring system continuity.

[0288] Step 3, based on the data obtained in step 2, multi-vehicle collision risk assessment is performed to determine the risk level and trigger the corresponding collision warning state. The process is as follows:

[0289] (3.1) Analyze the current state of the target vehicle, calculate the longitudinal safety distance, and assess the lateral slip risk.

[0290] In this embodiment, the vehicle-mounted data processing unit collects the Beidou positioning data after correcting the Beidou positioning error, the point cloud data after removing the rain and snow noise of the vehicle-mounted laser radar sensor point cloud, and the image data after improving the target recognition rate of the vehicle-mounted visual sensor in rainy and snowy weather, calculates the longitudinal safety distance S safe As a preferred, the current speed v0 of the target vehicle is obtained by weighted fusion of the corrected Beidou positioning and V2X broadcast speed, avoiding the deviation of safety distance calculation caused by the speed error of a single sensor.

[0291]

[0292] Wherein: v0 is the current speed of the target vehicle. t0 is the reaction time of the driver of the target vehicle, generally 0.4s-1.0s.

[0293] t1 is the response time of the braking system of the target vehicle. When the target vehicle is in a full load state, an additional distance of 20%-30% of the original basic safety distance needs to be added.

[0294] L is a load correction parameter, which is used to quantify the safety distance increment when the target vehicle is fully loaded.

[0295] K is a weather severity factor, and μ(K) is a road adhesion coefficient related to K. The reason for multiplying the weather severity factor K by the uncertainty coefficient C is that when the weather severity factor K increases, it means that the rainy and snowy weather is more severe, the reaction time of the driver will be relatively prolonged, and the braking performance of the vehicle will also decrease. At this time, after multiplying K by the uncertainty coefficient, it is added to other time parameters, which can reasonably increase the calculated value of the longitudinal safety distance to meet the higher requirements for safety distance in severe weather.

[0296] g is the acceleration of gravity. C is an uncertainty coefficient, which is a compensation coefficient for quantifying the uncertainty of the reaction time of the driver and the braking response of the vehicle in rainy and snowy weather, reflecting the additional impact of weather severity on operation delay.

[0297] In multi-vehicle cooperative collision avoidance, the longitudinal safety distance S safe Based on the speed and acceleration of the ego vehicle, the minimum distance between the ego vehicle and the preceding vehicle can be personalized planned to avoid secondary collision caused by distance misjudgment in the cooperative process.

[0298] In this embodiment, the lateral slip risk calculation formula is:

[0299]

[0300] Wherein: Rturn is the turning radius, calculated in real time from the Beidou positioning trajectory. y is the target vehicle lateral speed. crit is the target vehicle critical speed. sideslip is the lateral slip risk coefficient, dimensionless, the closer the value is to 1, the higher the risk of side slipping.

[0301] (3.2) Based on the front, rear, and side vehicles and environmental information around the target vehicle obtained through the vehicle-mounted V2X communication device, multi-vehicle cooperative analysis is performed to determine the risk.

[0302] Specifically, in this embodiment, a two-degree-of-freedom collision time threshold model is used to calculate the threshold value T, in order to further improve the accuracy of the front vehicle collision risk assessment. And the actual collision time TTC is calculated as follows: wherein, v reld is the weighted fusion speed difference between the target vehicle and the surrounding vehicles, which ensures the accuracy of TTC calculation; d is the relative distance between the target vehicle and the front, rear, or side vehicle after weighted fusion. The actual collision time and the threshold value T are compared, and thus the multi-vehicle cooperative analysis is performed to determine the risk.

[0303] To further realize multi-vehicle cooperative risk determination, the interaction with the rear and side vehicles also needs to be comprehensively evaluated.

[0304] The speed v_rear, braking state, and relative distance d_rear of the rear vehicle are obtained through V2X. The relative speed v_reld_rear between the rear vehicle and the target vehicle is calculated as v_reld_rear = v_rear - target vehicle speed. If v_reld_rear > 0, it means that the rear vehicle is faster than the target vehicle, and there is a risk of rear-end collision. If v_reld_rear > 0, the collision time TTC_rear between the rear vehicle and the target vehicle is calculated as TTC_rear = d_rear / v_reld_rear. If TTC_rear ≤ T, it is determined that the rear vehicle has a high collision risk, and the target vehicle sends a "front vehicle deceleration" warning to the rear vehicle through V2X, and adjusts its own braking strategy, such as avoiding sudden braking, to prevent rear-end collision.

[0305] The speed v_ side of the side vehicle, the turn signal state, the lateral distance d_ lateral to the target vehicle and the longitudinal distance d_ longitudinal are obtained through V2X communication. If the side vehicle turns on the turn signal and changes lanes to the lane of the target vehicle, the lateral relative speed v_reld_ lateral of the side vehicle to the target vehicle is calculated as v_reld_ lateral = lateral speed of the side vehicle - lateral speed of the target vehicle; the lateral time-to-collision TTC_ lateral of the side vehicle to the target vehicle is calculated as TTC_ lateral = d_ lateral / v_reld_ lateral. If v_reld_ lateral > 0, it indicates that the side vehicle is approaching the target vehicle; if TTC_ lateral <= 1.5s, the lateral collision safety threshold is reached, it is determined that the side vehicle has a high collision risk, the target vehicle triggers a lateral pre-warning, the steering wheel vibrates, and a "no lane changing" request is sent to the side vehicle through V2X. In combination with the calculation of the lane safety gap W_safe of the side vehicle, if W_safe <= 0, it indicates that there is no lane changing space in the lane of the side vehicle, the target vehicle avoids changing lanes to the lane to prevent collision with the side vehicle.

[0306] (II) Side vehicle collision risk assessment

[0307] The braking process of the vehicle is divided into four stages, which are the reaction stage of the driver when the pre-warning system sends a pre-warning signal, the stage of eliminating the brake pedal gap, the stage of rising brake master cylinder pressure, and the stage of continuous braking of the brake, and the time used in the four stages is respectively t1, t2, t3 and t4.

[0308] Since the reaction time of different drivers is different, it is difficult to calculate with a simple mathematical model. The reaction time of most drivers is selected, that is, t1 = 1.2s, t2 and t3 are corrected based on the mechanical characteristics of the braking system of the target vehicle, the statistical data of real vehicle experiments and the dynamic environment, mainly including mechanical characteristic correction, real vehicle experiment correction and dynamic environment correction. The heavy truck adopts pneumatic braking, and the pneumatic pressure building speed is slower than that of hydraulic braking, so t2 is corrected to 0.2s and t3 is corrected to 0.5s. In rainy and snowy weather, the braking system components may be jammed due to low temperature and water accumulation, so t2 is increased by 0.05s and t3 is increased by 0.1s, to ensure that the threshold T calculation conforms to the actual braking delay. Under normal hydraulic braking, t2 is 0.15s and t3 is 0.3s.

[0309] In this embodiment, the threshold T is calculated by a two-degree-of-freedom collision time threshold model as shown in the following formula:

[0310] T = t1 + t2 + t3 + t4

[0311] t4 is shown in the following formula:

[0312]

[0313] Wherein, a0 is the rated average deceleration of the braking system of the target vehicle, that is, the average deceleration of the vehicle when braking on a dry road surface under full load at the rated braking pressure. The heavy truck a0 = 4.5m / s2 , bias ± 0.5 m / s 2 ; in rainy and snowy weather, a0 needs to be multiplied by the road adhesion coefficient μ(K).

[0314] Then, the actual collision time TTC and the threshold value T are compared. If TTC>T, the collision risk is low, and the driver of the target vehicle and its surrounding vehicles is reminded to pay attention to the road conditions and take measures; if TTC≤T, the collision risk is high, and the driver of the target vehicle and its surrounding vehicles is reminded to brake, change lanes, or coordinate deceleration.

[0315] (3.3), risk chain propagation calculation.

[0316] The risk amplification coefficient λ is used to evaluate the multi-vehicle chain collision risk of the target vehicle and its surrounding vehicles, and the calculation formula is as follows:

[0317]

[0318] Where: ω i is the collision risk weight of the i-th vehicle around the target vehicle. The attenuation coefficient β=0.1, and the attenuation coefficient γ=0.02.d i is the distance between the target vehicle and the i-th vehicle around it. Δv i is the relative speed between the target vehicle and the i-th vehicle. K is the weather severity factor, K∈[0,1], which is used to quantify the influence of rain and snow intensity on risk propagation.

[0319] Then, the feasibility of cooperative collision avoidance is evaluated, the lane changing feasibility is judged, and the side lane safety gap W safe is calculated, which is expressed as follows:

[0320] W safe =d 前车 -L 本车 -L 旁车 -Δd 缓冲

[0321] Δd 缓冲 =v×1.5

[0322] In the formula:

[0323] d 前车 is the longitudinal distance between the front vehicle on the side lane and the target vehicle.

[0324] L 本车 is the length of the target vehicle.

[0325] L 旁车 is the length of the vehicle on the side lane of the target vehicle, which is obtained by the vehicle-mounted V2X communication module or estimated by the point cloud of the vehicle-mounted laser radar sensor.

[0326] Δd 缓冲For safety redundancy distance in the process of lane changing, used to offset the speed fluctuation and control error when changing lanes.

[0327] As preferred, d 前车 , L 旁车 are all from the weighted fusion data, d 前车 , L 旁车 are weighted by the laser radar denoising and V2X broadcast data, which provides a reliable basis for lane changing feasibility.

[0328] If the derived side lane safety gap W safe > 0, it means that the side lane has enough physical space to accommodate the target vehicle, at which time the target vehicle initiates a cooperative lane changing request. If W safe ≤ 0, the target vehicle requests the surrounding vehicles to cooperate with deceleration through the vehicle-mounted V2X communication module, and then changes lanes after the space is released, i.e. W safe > 0.

[0329] (3.4), comprehensive risk determination and classification

[0330] In multi-vehicle cooperative analysis, based on the relative motion state, distance and snow environment of the target vehicle and the surrounding front and rear vehicles, the risk value R i,i+1 between the associated vehicles i and i+1 is calculated. In order to make the risk value more referential, R i,i+1 is limited between 0 and 1. When the calculation result is greater than 1, it is taken as 1, indicating that there is a high collision risk between the two vehicles; when the calculation result is less than 0, it is taken as 0, indicating that there is no collision risk between the two vehicles. The risk value R i,i+1 is calculated as follows:

[0331] R i,i+1 = min{T / TTC(i,i+1),1}

[0332] Where: T is a two-degree-of-freedom threshold value. TTC(i,i+1) is the actual collision time of vehicles i and i+1.

[0333] By fusing the multi-vehicle interaction risk, lateral slip risk and braking ability risk, the comprehensive risk of collision risk in rainy and snowy weather is realized, and the multi-vehicle cooperative collision risk R 总 is calculated, and the calculation expression is as follows:

[0334]

[0335] a max = μ(K) × g

[0336] Where: a 实际The current actual braking deceleration of the target vehicle is collected in real time by an on-board inertial measurement unit (IMU) and a brake pedal stroke sensor. The IMU directly outputs the instantaneous deceleration by measuring the longitudinal acceleration change of the vehicle. The brake pedal stroke sensor detects the pedal displacement and combines the brake system characteristic curve to assist in calibrating the IMU data, ensuring that a 实际 The brake pedal can be accurately collected even if it is slipping in rainy and snowy weather.

[0337] a max The maximum possible braking deceleration of the road surface is associated with the weather severity factor K.

[0338] μ(K) is the road adhesion coefficient in rainy and snowy weather.

[0339] α is the lateral slip risk weight coefficient, which is based on a large number of real vehicle experiments and simulation data.

[0340] β is the braking margin risk weight coefficient, which is based on real vehicle collision experiment statistics and takes a value of 0.3.

[0341] R sideslip is the lateral slip risk coefficient.

[0342] R 制动裕度 is the braking margin risk coefficient, which reflects the difference between the vehicle braking ability and the road extreme braking ability, and takes a value in the range [0, 1].

[0343] g is the acceleration of gravity.

[0344] As a preferred embodiment, the risk value R i,i+1 Based on the weighted fusion of the relative speed and distance between the two vehicles, the comprehensive risk assessment result is ensured to be accurate.

[0345] According to the calculation value of Rtotal, the risk level is determined and the corresponding collision warning state is triggered, as shown in the following table: Figure 2

[0346] When 0.3≤R 总 <0.6, it is the first risk level. At this time, the dashboard of the target vehicle is yellow, reminding the target vehicle driver that the road is slippery in rainy and snowy weather and to maintain the vehicle distance. A single low-frequency beep (0.5s) is emitted, and a voice broadcast is made to inform the target vehicle driver that the distance to the front vehicle is close. The vehicle within 500m behind the target vehicle is broadcasted to have a low risk warning, and it is suggested to reduce the speed to the current 90%.

[0347] When 0.6≤R 总 ​<0.8, the dashboard of the target vehicle flashes red, countdown reminders, such as too close, 10 seconds to slow down, high-frequency beep (1s cycle), voice broadcast please slow down immediately, brake pedal vibration (2 times / s) and seat light push (5N), send cooperative deceleration instructions to the 3 associated vehicles in front and behind the target vehicle, and suggest a deceleration of ≤0.3g.

[0348] When R 总 >0.8, the third risk level, the target vehicle full light double flash, HUD projection future 2 seconds safe into the left lane, red label high decibel alarm (> 80dB), voice broadcast emergency avoidance, steering wheel vibration, prompt avoidance direction, seat belt pre-tightening (15N), send emergency avoidance request to the associated vehicles in front and behind the target vehicle, upload danger area coordinates to RSU, broadcast warning to subsequent vehicles within 1km.

[0349] (3.5), real-time dynamic updating and optimization.

[0350] The Kalman filter is used to predict the trajectory of each vehicle in the next 3 seconds, and the risk assessment result is updated every 100ms, and when an emergency is detected, the prediction period is shortened to 1 second to improve response speed. The process is shown in Figure 3 .

[0351] Set the state variable, define the target vehicle and the surrounding vehicle motion state vector X K =[x, y, v x , v y , a x , a y ] T , where: x, y are the horizontal and vertical positions of the target vehicle under Beidou positioning; v x , v y are the horizontal and vertical speeds; a x , a y are the horizontal and vertical accelerations.

[0352] Configure the initial parameters, based on the initial data of Beidou positioning (x0, y0), speed sensor data (v x0 , v y0 ), and acceleration initial value set to 0, thus obtaining the initial state estimate X0.

[0353] Initial covariance matrix P0, the diagonal elements of covariance matrix P0 are positioning error (0.3 2 , 0.3 2 ), speed error (0.5 2 , 0.5 2 ), acceleration error (0.2 2 , 0.2 2 ), reflecting the uncertainty of the initial state.

[0354] Considering the randomness of vehicle movement under rain and snow conditions, the process noise covariance Q is set as follows:

[0355] Q=diag([0.01,0.01,0.05,0.05,0.1,0.1]),

[0356] Here, "noise" refers to the uncertainty error in the vehicle's motion state. In rainy or snowy weather, slippery road surfaces cause fluctuations in vehicle acceleration, such as slight sideslip and brake pedal feedback delay. Even with an accurate kinematic model, it's impossible to completely predict the vehicle's state at the next moment. Therefore, process noise is introduced to quantify this uncertainty. The diagonal elements of Q in the formula correspond to:

[0357] Lateral position (y) uncertainty: 0.01 (m) 2 )

[0358] Longitudinal position (x) uncertainty: 0.01 (m) 2 )

[0359] Lateral velocity (v) y Uncertainty: 0.05 (m / s) 2 )

[0360] Longitudinal velocity (v) x Uncertainty: 0.05 (m / s) 2 )

[0361] lateral acceleration (a) y Uncertainty: 0.1 (m / s 2 ) 2 )

[0362] Longitudinal acceleration (a) x Uncertainty: 0.1 (m / s 2 ) 2 )

[0363] In this embodiment, the lateral motion noise (y, v) y ,a y The longitudinal noise is slightly greater than the longitudinal noise to accommodate lateral disturbances caused by slippery road surfaces. The lateral motion noise is slightly greater than the longitudinal noise because slippery road surfaces significantly interfere with the vehicle's lateral control.

[0364] The observation noise covariance R is set and dynamically adjusted according to the sensor type. The BeiDou positioning observation noise R is included in the observation noise covariance R. 北斗 =diag([0.3 2 0.3 2 0.5 2 0.5 2]) ; V2X data observation noise in observation noise covariance R V2X = 1.2 x R 北斗 , V2X data delay is slightly high; laser radar position observation noise in observation noise covariance R 雷达 = diag([0.2 2 , 0.2 2 ]).

[0365] The observation noise of the visual sensor is indirectly included through the "weighted fused data observation noise". In step 2, the visual sensor data has been weighted and fused with the laser radar and V2X data through a dynamic allocation mechanism based on the weight of the weather severity factor K, and the fused target position and speed data already contain the error characteristics of the visual sensor. Therefore, in the Kalman filter observation noise R, there is no need to set the visual sensor noise separately, but it is embodied through the "multi-source fused data observation noise", which is specifically:

[0366] R 融合 = ω 视觉 x R 视觉 + ω 激光雷达 x R 雷达 + ω V2X x R V2X

[0367] where R 视觉 is the visual sensor's individual observation noise, diag([0.5 2 , 0.5 2 , 1.0 2 , 1.0 2 ], position error 0.5 m, speed error 1.0 m / s, and ω is the weight of each sensor, consistent with the weight allocation of D2 in step 2. In actual filtering, the fused data is directly used to construct the observation vector, so there is no need to list the visual sensor noise separately.

[0368] Based on the vehicle kinematic model, the next time state is predicted through the Kalman filter, as shown in the following formula:

[0369] X K|K-1 = F K x X K-1|K-1 + B K x U K

[0370] where X K|K-1 is the predicted state vector at time K based on the observation at time K-1, which is the prediction result of the optimal state at the previous time through the state transition matrix F K .

[0371] X K-1|K-1is the optimal state vector at time k-1, which is the output of the update step in Kalman filter and the input of the prediction at next time.

[0372] B K is the control input matrix.

[0373] U K is the control input vector at time k, in this embodiment U K = 0, the impact of active control on prediction is not considered.

[0374] F K is the state transition matrix, F K is as follows:

[0375]

[0376] The prediction time step Δt = 0.1s matches the sensor sampling frequency.

[0377] The prediction is performed continuously for 30 times to generate a vehicle trajectory prediction sequence of 3 seconds in the future

[0378] {X k+1∣k ,X k+2∣k ,...,X k+30∣k} covering the position and speed change trend of the host vehicle and all target vehicles around.

[0379] Multi-source observation data fusion, new sensor data is received every 100ms, including Beidou positioning, visual sensor, V2X broadcast of surrounding vehicle state, laser radar target position, to construct observation vector

[0380] Z K = [x obs ,y obs ,v x,obs ,v y,obs ,x obs-视觉 ,y obs-视觉 ] T , the observation equation is Z K = H K × X K + v K , and the predicted state is fused, as shown in the following formula:

[0381]

[0382] X k∣k = X k∣k-1 + K K × (Z K -H K × X k∣k-1 )

[0383] where: xobs and y obs are the observed longitudinal and lateral positions of the target vehicle in meters (m) respectively, collected by the lidar sensor and filtered by the rain and snow environment noise.

[0384] v x,obs ,v y,obs are the observed longitudinal and lateral speeds of the target vehicle in meters per second (m / s) respectively, collected by the lidar sensor and filtered by the rain and snow environment noise.

[0385] x obs-视觉 ,y obs-视觉 are the observed longitudinal and lateral positions of the target vehicle in meters (m) respectively, collected by the vision sensor and optimized by image enhancement and target recognition in the rain and snow environment.

[0386] P k∣k-1 is the prediction error covariance matrix based on the observation at time k-1, used to measure the error uncertainty between the predicted state and the true state.

[0387] R K is the observation noise covariance matrix at time k, used to describe the noise characteristics of each observation value in the observation vector Z K and the correlation between the noises.

[0388] Z K is the multi-source observation vector at time k, containing the preprocessed observation information of the target vehicle position, speed, etc. obtained from the Beidou positioning, vision sensor, V2X broadcast, lidar, etc. multi-source sensors.

[0389] v K is the observation noise; H K is the observation matrix; K K is the Kalman gain, which realizes real-time correction of the prediction value and reduces the error caused by sensor noise in the rain and snow weather.

[0390] When the data jump of a sensor exceeds the threshold, it is determined as an outlier by the 3σ criterion, and the observation weight of the sensor is temporarily reduced to avoid contaminating the prediction results.

[0391] Step 4, according to the risk level obtained in step 3, trigger different intensity of warning and control instructions; at the same time, use V2X communication technology to judge the feasibility of lane avoidance, realize multi-vehicle cooperative active deceleration and avoidance.

[0392] The preferred embodiments of the present application are described in detail above with reference to the accompanying drawings, and the embodiments described in the present application are merely a description of the preferred embodiments of the present application, and are not intended to limit the concept and scope of the present application. In the above specific embodiments, various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combination should also be considered as disclosed by the present disclosure, as long as it does not deviate from the concept of the present application. In order to avoid unnecessary repetition, the present application will not further describe various possible combinations.

[0393] The present application is not limited to the specific details described in the above embodiments, and various modifications and improvements to the technical solutions of the present application made by those skilled in the art within the scope of the technical concept of the present application and without departing from the design concept of the present application should fall within the protection scope of the present application. The technical content claimed by the present application has been fully recorded in the claims.

Claims

1. A multi-sensor data fusion-based collaborative collision avoidance warning method for autonomous freight platooning, characterized in that, The process is as follows: Step 1: Determine the composition structure of the freight convoy, and clarify the convoy sequence number and communication priority of the lead vehicle and each following vehicle. The lead vehicle, as the convoy data fusion and strategy decision-making center, first collects its own multi-sensor data, and then receives data collected by the vehicle-mounted Beidou positioning receiver, vehicle-mounted vision sensor, and vehicle-mounted LiDAR sensor of all following vehicles through the V2X communication module to construct a convoy-level global perception dataset. Then, the system acquires the target vehicle's original positioning data collected by the vehicle-mounted Beidou positioning receiver, the communication data collected by the vehicle-mounted V2X communication module, the road condition image data collected by the vehicle-mounted vision sensor, and the 3D point cloud data of vehicles on the road ahead collected by the vehicle-mounted lidar sensor. Then, the data collected by the vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted vision sensor, and vehicle-mounted lidar sensor are preprocessed, and the spatiotemporal reference of the data is unified. Step 2: To address the interference of rain and snow on the data collected by the vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted vision sensor, and vehicle-mounted LiDAR sensor, correct the Beidou positioning error, remove rain and snow noise from the point cloud of the vehicle-mounted LiDAR sensor, and improve the target recognition rate of the vehicle-mounted vision sensor in rain and snow. Furthermore, design a dynamic weight allocation mechanism based on the weather severity factor K to dynamically adjust the weights of the data collected by the vehicle-mounted vision sensor, vehicle-mounted LiDAR sensor, and vehicle-mounted V2X communication module, thereby achieving weighted fusion of the data collected by the vehicle-mounted vision sensor, vehicle-mounted LiDAR sensor, and vehicle-mounted V2X communication module in the target vehicle. Step 3: Based on the data obtained in Step 2, conduct a multi-vehicle collision risk assessment and determine the risk level; Step 4: Based on the risk level obtained in Step 3, trigger warning and control commands of different intensities.

2. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 1, characterized in that, In step 1, the raw positioning data collected by the vehicle-mounted Beidou positioning receiver undergoes validity verification preprocessing. The lead vehicle needs to additionally verify the completeness and timeliness of the data uploaded by each following vehicle, such as removing data with excessive latency, to ensure the consistency of data across all vehicles in the convoy and avoid affecting the overall fusion result due to abnormal data from a single vehicle. The communication data collected by the vehicle-mounted V2X communication module undergoes filtering preprocessing, the road condition images collected by the vehicle-mounted vision sensor undergo image dehazing algorithm preprocessing, and the 3D point cloud data collected by the vehicle-mounted LiDAR sensor undergoes noise reduction preprocessing.

3. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 1, characterized in that, In step 1, using the time of the vehicle-mounted Beidou positioning receiver as a reference, timestamps are added to the data collected by the vehicle-mounted V2X communication, vehicle-mounted vision sensor, and vehicle-mounted LiDAR sensor. The clocks of each vehicle-mounted Beidou positioning receiver, vehicle-mounted V2X communication module, vehicle-mounted vision sensor, and vehicle-mounted LiDAR sensor are aligned through the vehicle-mounted data processing unit of the target vehicle, thereby completing time synchronization. The 3D point cloud data in its own coordinate system collected by the vehicle-mounted lidar sensor and the road condition image data ahead in its own coordinate system collected by the vehicle-mounted vision sensor are respectively converted into the coordinate system of the vehicle-mounted Beidou positioning receiver, thereby achieving spatial synchronization. The lead vehicle associates and stores its own data with the data of all following vehicles after unifying the spatiotemporal reference, generating a global spatiotemporal coordinate system for the formation. Each following vehicle corrects its own local coordinate system deviation based on this coordinate system, ensuring that vehicles in the formation have consistent perception of the position and distance of the same target. Ultimately, by synchronizing time and space, a unified spatiotemporal reference is established for the data collected by the vehicle-mounted BeiDou positioning receiver, the vehicle-mounted V2X communication module, the vehicle-mounted vision sensor, and the vehicle-mounted LiDAR sensor.

4. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 1, characterized in that, In step 2, the process of correcting the BeiDou positioning error is as follows: A1. Obtain humidity data collected by the high-precision meteorological sensor on the target vehicle and voltage value generated by the raindrop sensor when hit by raindrops. Calculate the raindrop density based on the voltage value obtained from the raindrop impact signal. A2. Using the time of the vehicle-mounted Beidou positioning receiver as a reference, add timestamps to the humidity data collected by the vehicle-mounted meteorological sensor and the raindrop density data calculated based on the data collected by the vehicle-mounted raindrop sensor, and eliminate time offset through the sliding window algorithm. A3. Using the 3σ criterion, filter step A2 eliminates abnormal data in humidity data and raindrop density data after time shift, retains continuous and stable corresponding observation data sequences, thereby completing outlier removal and realizing time-synchronized cleaning of humidity data from vehicle-mounted meteorological sensors and raindrop density data from vehicle-mounted raindrop sensors. A4. Calculate the zenith tropospheric delay ZTD based on the atmospheric delay model, and then convert the zenith tropospheric delay ZTD into the tropospheric delay TD1 for any satellite elevation angle θ through the mapping function NMF function. A5. Based on atmospheric humidity data H and raindrop density data D, the atmospheric delay compensation factor δ is dynamically calculated using the least squares method to correct the tropospheric delay. That is, the corrected total delay TD is: tropospheric delay TD1 + atmospheric delay compensation factor δ. A6. The positioning deviation of the vehicle-mounted BeiDou positioning receiver is corrected using pseudorange, and the target vehicle's positioning is calculated based on the corrected pseudorange to obtain the target vehicle's three-dimensional coordinates; among which, the weighted least squares algorithm is used to correct the pseudorange P of different satellites. corr Differentiated weights are assigned to optimize the positioning solution accuracy. Pseudoranges with small errors are given higher weights, while pseudoranges with large errors are given lower weights. Finally, the three-dimensional coordinates of the target vehicle that minimize the sum of squared weighted errors are obtained.

5. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 1, characterized in that, In step 2, the process of removing rain and snow noise from the point cloud of the vehicle-mounted LiDAR sensor is as follows: B1. Perform preliminary removal of invalid points in the original point cloud, removing points that are outside the effective detection range of the lidar sensor and filtering out points with reflectivity below the reflectivity threshold; B2. Perform time-domain filtering on the lidar sensor point cloud after elimination and filtering in step B1 to obtain time-domain candidate noise points. B3. Perform spatial domain filtering on the points filtered in the time domain in step B2 to obtain spatial domain candidate noise points; B4. Points marked as candidate noise points in both the time and spatial domains obtained from steps B2 and B3 are identified as rain and snow noise points and removed. Then, radius filtering is performed on the remaining point cloud to eliminate a small amount of residual isolated noise; and target clusters are extracted from the radius-filtered point cloud using Euclidean clustering algorithm, and the mean reflectance and motion consistency of points within the cluster are calculated to further purify the point cloud.

6. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 1, characterized in that, In step 2, the process of improving the target recognition rate of the vehicle-mounted vision sensor in rainy and snowy weather is as follows: C1. Raindrop / snowflake detection and removal are performed on the image, and the image is dehazed and enhanced, thereby suppressing rain and snow noise in the image acquired by the vehicle vision sensor, resulting in image 1 after suppressing rain and snow noise; C2. By using adaptive histogram equalization, the contrast of image 1 after suppressing rain and snow noise is enhanced, and the details of the image are enhanced by multi-scale Retinex, resulting in image 2 with enhanced contrast and details. C3. Perform color correction on image 2 after contrast and detail enhancement, and combine it with a saliency detection algorithm to highlight the target area in the image.

7. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 4, characterized in that, In step 2, a dynamic weight allocation mechanism based on the weather severity factor K is designed. The process of dynamically adjusting the weights and performing weighted fusion is as follows: D1. Calculate the weather severity factor K based on humidity and raindrop density; classify the weather severity level into different levels according to the weather severity factor K. D2. Assign different weights to the data collected by the vehicle-mounted vision sensor, the data collected by the vehicle-mounted lidar sensor, and the data collected by the vehicle-mounted V2X communication module according to different weather levels. D3. Based on the assigned weights, the data collected by the vehicle vision sensor, vehicle lidar sensor and vehicle V2X communication module are weighted and averaged and fused. D4. The lead vehicle needs to adjust the weights based on the positional characteristics of each vehicle in the formation, and at the same time synchronize the weight allocation strategy to all following vehicles via V2X to ensure that the fusion logic within the formation is consistent. Furthermore, the target confidence scores output by the vehicle-mounted vision sensor, vehicle-mounted LiDAR sensor, and vehicle-mounted V2X communication module are multiplied by their corresponding weights and then summed. When the sum is greater than 0.6, it is determined to be a valid target. Here, a valid target refers to a traffic participant or obstacle that is determined to be real and reliable after multi-sensor information fusion verification. If one or both of the vehicle-mounted vision sensor, vehicle-mounted LiDAR sensor, and vehicle-mounted V2X communication module fail, the weight of the failed one will be proportionally distributed to the other valid ones.

8. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 1, characterized in that, In step 3, the process of conducting a multi-vehicle collision risk assessment to determine the risk level is as follows: (3.1) The lead vehicle assesses the overall driving status of the formation, including the total length of the formation, the distance between each vehicle, and the overall speed. It calculates the multi-level collision risks of "lead vehicle - preceding vehicle", "lead vehicle - following vehicle 1", "following vehicle 1 - following vehicle 2", etc., to form a formation risk chain. The following vehicles only need to focus on assessing the risks with the preceding vehicle (lead vehicle or preceding following vehicle) and the vehicles around the vehicle itself, and feed the assessment results back to the lead vehicle in real time. The lead vehicle then determines whether to trigger the formation-level collision avoidance strategy. (3.2) Analyze the current state of the target vehicle, calculate the longitudinal safety distance, and assess the risk of lateral slippage; (3.3) Obtain information on the vehicles in front, behind, and beside the target vehicle and the environment. Calculate the threshold value T using a two-degree-of-freedom collision time threshold model and calculate the actual collision time TTC between the target vehicle and the vehicle in front. The actual collision time (TTC) is compared with the threshold value (T). If TTC > T, the collision risk is judged to be low, and the drivers of the target vehicle and surrounding vehicles are reminded to pay attention to the road conditions and take measures. If TTC ≤ T, the collision risk is judged to be high, and the drivers of the target vehicle and surrounding vehicles are reminded to brake, change lanes, or slow down in coordination. (3.4) The risk amplification factor λ is used to assess the risk of multi-vehicle collisions involving the target vehicle and its surrounding vehicles; Then calculate the adjacent lane safety clearance W. safe Based on the safety clearance W of the adjacent lane safe Conduct a feasibility assessment of collaborative collision avoidance, and if the resulting safe clearance W in the adjacent lane... safe When W > 0, the target vehicle initiates a coordinated lane change request. If it's a platoon lane change, the lead vehicle must first send its lane change intention to all following vehicles in the platoon via V2X, then calculate the safe lane clearance for each following vehicle, ensuring all following vehicles meet the lane change conditions before the lead vehicle changes lanes first, followed by the following vehicles in sequence, avoiding lane change conflicts within the platoon. If W safe If W ≤ 0, the target vehicle requests coordinated deceleration from surrounding vehicles, and waits for space to be released. safe Change lanes after >0; (3.5) Based on the relative motion state, distance, and rain / snow environment of the two vehicles, calculate the risk value R between vehicle i and vehicle i+1. i,i+1 Then, based on the risk value R i,i+1 Calculate the overall risk R 总 ; Root Comprehensive Risk R 总 It determines the risk level and triggers the corresponding collision warning status.

9. The multi-sensor data fusion-based autonomous driving freight platooning cooperative collision avoidance and early warning method according to claim 8, characterized in that, In step 3, when 0.3 ≤ R 总 When the value is less than 0.6, it is classified as the first level of risk. When 0.6≤R 总 When R < 0.8, it is classified as the second risk level; when R 总 When the value is greater than 0.8, it is classified as the third risk level.

Images