A highway tunnel-oriented dynamic visual range compensation method for an autonomous vehicle

Abstract

The application discloses a kind of highway tunnel-oriented automatic driving vehicle dynamic range compensation method, belong to highway tunnel vehicle-road cooperation technical field, this method includes the following steps: S1, data acquisition and preparation before entering tunnel;S2, tunnel data fusion and collaborative perception field construction;S3, over-the-horizon perception and risk quantification;S4, decision calculation and control instruction generation.The application adopts the above-mentioned highway tunnel-oriented automatic driving vehicle dynamic range compensation method, constructs time and space continuous collaborative perception field, realizes vehicle-road data efficient fusion, accurately completes trajectory prediction and risk quantification, balances tunnel traffic safety and efficiency, effectively extends the over-the-horizon range of automatic driving vehicle.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of highway tunnel vehicle-road cooperation, and particularly relates to a dynamic visual range compensation method for an automatic driving vehicle in a highway tunnel. BACKGROUND

[0002] When an automatic driving vehicle runs in a typical closed scene of a highway tunnel, its environmental perception and driving decision-making capability face severe challenges. The internal structure of the tunnel causes satellite navigation signals to fail, and although short-term dead reckoning can be performed by relying on an inertial measurement unit, the positioning error will accumulate over time. At the same time, factors such as internal light changes and air turbidity can significantly reduce the effective detection distance and reliability of vehicle-mounted visual sensors, laser radars and other perception elements, forming a perception blind area.

[0003] However, the prior art has the following disadvantages: the over-the-horizon information provided by the roadside unit and the local perception data on the vehicle differ in time and space reference, data format and confidence level, lack an efficient fusion mechanism, and cannot construct a continuous and unified environmental situation field; the accuracy of traditional trajectory prediction and behavior recognition models decreases in the scene where tunnel information is incomplete, making it difficult to provide reliable basis for decision control; the safety distance and vehicle speed calculation of the adaptive cruise control model are based on fixed parameters, and the real-time road adhesion coefficient, dynamic visibility and over-the-horizon potential risks in the tunnel are not fully considered, resulting in conservative (affecting efficiency) or aggressive (laying hidden dangers) control instructions; the single vehicle sensor is affected by the tunnel environment, and the effective detection distance is limited, making it difficult to realize over-the-horizon risk prediction.

[0004] Therefore, a new method is urgently needed. SUMMARY

[0005] The purpose of the present application is to provide a dynamic visual range compensation method for an automatic driving vehicle in a highway tunnel, which constructs a time and space continuous cooperative perception field, realizes efficient fusion of vehicle and road data, accurately completes trajectory prediction and risk quantification, balances tunnel passing safety and efficiency, and effectively extends the over-the-horizon range of the automatic driving vehicle.

[0006] To achieve the above purpose, the present application provides a dynamic visual range compensation method for an automatic driving vehicle in a highway tunnel, comprising the following steps:

[0007] S1, obtaining real-time driving state, tunnel design parameters, real-time traffic flow information, environmental parameters and maps to generate continuous positioning data; the real-time driving state, positioning data, tunnel design parameters, traffic flow information, environmental parameters and high-precision maps are all transmitted to S2;

[0008] S2. Collect local environmental data through vehicle-mounted sensors and simultaneously receive beyond-line-of-sight information transmitted by roadside units. Preprocess and verify the effectiveness of the local environmental data and beyond-line-of-sight information respectively. Based on the linear and nonlinear characteristics of the tunnel scene and the data uncertainty, dynamically select extended Kalman filter, unscented Kalman filter, Bayesian estimation, and deep learning model to perform vehicle-road data fusion. Construct a spatiotemporally continuous collaborative perception field based on the fused data. The collaborative perception field is then transmitted to S3.

[0009] S3. Based on the collaborative perception field, extract trajectory prediction features and behavioral intention features, obtain the trajectory prediction results and behavioral intention recognition conclusions of the vehicle in front through the trajectory prediction model and behavioral intention recognition model respectively, fuse the trajectory prediction results and behavioral intention recognition conclusions, and dynamically adjust the beyond-line-of-sight range in combination with the lighting brightness in the environmental parameters to realize the quantitative perception of potential risks within the beyond-line-of-sight range. The risk quantification results and the adjusted beyond-line-of-sight range are transmitted to S4.

[0010] S4. Integrate the risk quantification results, real-time tunnel environmental parameters, and vehicle power performance parameters to calculate the safe following distance and correct the safe speed. Based on the comparison results of the safe following distance, safe speed, and the vehicle's current driving state, generate corresponding control commands.

[0011] Preferably, in S1, a preset distance before entering the tunnel needs to be calculated, and the formula for calculating the preset distance is:

[0012] ;

[0013] In the formula, Real-time driving speed; Preparation time for perception; Minimum safe warning distance;

[0014] The real-time driving speed The speed is obtained by fusing speed data from vehicle-mounted wheel speed sensors and vehicle-mounted radar through extended Kalman filtering. The speed measurement accuracy is ≤ ±0.1 m / s, and the fused real-time driving speed is output every 100 ms. The initial value of the noise covariance matrix of the extended Kalman filtering process is set to... The measurement noise covariance matrix is ​​set as follows: .

[0015] Preferably, in S1, the sensing preparation time The value ranges from 3 to 10 seconds, and is dynamically adjusted based on the tunnel design speed and real-time traffic density.

[0016] When the tunnel design speed is ≥80km / h or the real-time traffic density is ≥20pcu / (km·lane), Take 6~10s, according to "design speed every increase 10km / h, Increase 1s; vehicle density every increase 5pcu / (km·lane), Increase 0.5s" rule dynamic value;

[0017] When the tunnel design speed <80km / h and real-time traffic density <20pcu / (km·lane), Take 3~5s, according to "design speed every decrease 10km / h, Decrease 0.5s; vehicle density every decrease 5pcu / (km·lane), Decrease 0.3s" rule value;

[0018] Minimum safety warning distance The default value is 50m, if there is a curve with a radius of curvature <500m or a construction area at the tunnel entrance, according to "the radius of curvature every decrease 100m, Increase 5m; when there is a construction area, Increase 10m" rule correction, ensure ≥50m;

[0019] The calculation frequency of the preset distance is every 200ms, when the distance between the vehicle and the tunnel entrance is equal to the preset distance, the data acquisition instruction is triggered;

[0020] The continuous positioning data is generated by combining vehicle-mounted global navigation satellite system and inertial measurement unit positioning, and the positioning accuracy in open environment is ≤0.5m;

[0021] The environmental parameters include visibility, road temperature, road adhesion coefficient and lighting brightness, which are collected by corresponding sensors according to the frequency of "visibility every 30s, road temperature every 1min, road adhesion coefficient every 2min, lighting brightness every 10s".

[0022] Preferably, in S2, the vehicle-mounted sensor includes laser radar, forward-looking high-definition camera, side-view camera, millimeter wave radar and ultrasonic radar;

[0023] The laser radar collects obstacle point cloud data within 0-50m range, and the update frequency is 10Hz;

[0024] The forward-looking camera collects the front 0-50m environmental illumination intensity and obstacle category, and the side-view camera collects the lateral 0-30m obstacle information;

[0025] The millimeter wave radar collects 0-50m obstacle speed and distance information, and the update frequency is 20Hz;

[0026] The ultrasonic radar is installed on the front and rear bumpers of the vehicle, collects 0-5m close-range obstacle information, and has an update frequency of 10Hz, and is used for supplementing perception in a low-speed scene with a vehicle speed of less than 20km / h.

[0027] Preferably, in S2, all vehicle-mounted sensors package and transmit the collected local environment data every 50ms;

[0028] The preprocessing of the local environment data includes denoising, time synchronization and space calibration, the denoising mode is laser radar statistical filtering denoising and camera Gaussian filtering denoising; after space calibration, it is unified into the WGS84 geodetic coordinate system, and the calibration accuracy is less than or equal to 5cm, and recalibration is performed once every 3 months;

[0029] The over-the-horizon information is received through a 5G-V2X or LTE-V2X vehicle-road cooperative communication module, and the validity verification rule of the over-the-horizon information is:

[0030] After format conversion, it is judged according to the confidence, the confidence is calculated by weighting the accuracy of the roadside sensor and the data consistency, and the weights are 0.6 and 0.4 respectively;

[0031] When the confidence is greater than or equal to 0.8, it is directly determined to be valid, when the confidence is between 0.5 and 0.8, the standard deviation of 3 consecutive frames of data needs to meet “position standard deviation ≤2m or speed standard deviation ≤1km / h” to be valid, and when the confidence is less than 0.5, it is directly excluded;

[0032] The rule for dynamically selecting the fusion algorithm is:

[0033] In a scene with high linear approximation degree, an extended Kalman filter is used, and the process noise covariance matrix is set to , and the measurement noise covariance matrix is adjusted according to the data confidence;

[0034] In a strong nonlinear scene, an unscented Kalman filter is used, wherein the number of sigma points is set to 2n+1, wherein n is the state dimension, and the scale factors are ;

[0035] In a multi-sensor data uncertainty known scene, a Bayesian estimation is used, the prior probability is set based on the sensor accuracy, the posterior probability is calculated through the Bayesian formula, and the fusion result with the maximum posterior probability is selected;

[0036] In a complex dynamic scene, a deep learning model is used; the training data set contains multi-sensor data in different scenes in a tunnel, and is verified by 5-fold cross-validation; the model input is a feature vector, local features are extracted through ResNet50, global features are fused through a Transformer encoder, the loss function is cross-entropy loss, and the fusion accuracy is greater than or equal to 92%;

[0037] The collaborative perception field covers 0-200m, wherein 0-50m is a local perception field, and 50-200m is an over-the-horizon perception field.

[0038] In S3, the trajectory prediction result is fused with the behavior intention recognition conclusion, specifically:

[0039] The trajectory prediction features include the position coordinate sequence, vehicle speed sequence, acceleration sequence, and current lane state of the preceding vehicle in the past 3 seconds. The trajectory prediction model is an improved LSTM model, which outputs the motion trajectory in the future 1-5 seconds, and the position error is ≤0.5 meters.

[0040] The behavior intention features include the headlamp state, steering wheel angle data, and vehicle speed change rate of the preceding vehicle. The behavior intention recognition model is a convolutional neural network combined with a Transformer model, which outputs 6 standardized behavior intentions, including normal driving, preparing to decelerate, preparing to change lanes, changing lanes, emergency deceleration, and temporary parking.

[0041] The calculation formula of risk quantification is:

[0042] Risk value = obstacle distance weight × 0.4 + relative speed weight × 0.3 + behavior intention weight × 0.3.

[0043] Obstacle distance weight = 1 - (obstacle distance / maximum range of super vision);

[0044] Relative speed weight = relative speed / vehicle speed of the host vehicle.

[0045] Among them, the behavior intention weight is set as “emergency deceleration, temporary parking = 1.0; preparing to decelerate, preparing to change lanes = 0.6; normal driving, changing lanes = 0.2”, and the risk value range is 0-1.

[0046] The super vision range adjustment rule is:

[0047] When the lighting brightness < 200 lux, the super vision range = road side detection distance × 0.8;

[0048] When the lighting brightness is 200-500 lux, the super vision range = road side detection distance × 0.9;

[0049] When the lighting brightness > 500 lux, the super vision range = road side detection distance × 1.0.

[0050] In S4, the dynamic performance parameters of the host vehicle include the maximum braking deceleration, maximum acceleration, and steering system response delay.

[0051] The formula for the maximum braking deceleration is:

[0052] ;

[0053] In the formula, is the acceleration of gravity, ; For the road adhesion coefficient;

[0054] The maximum acceleration is adjusted according to "when the electric quantity is less than 50% or the rotating speed is greater than 2000 rpm, it is set to , otherwise it is set to ".

[0055] The steering system response delay is adjusted according to "when the speed is less than 30 km / h, it is set to 0.2 s, and when the speed is greater than or equal to 30 km / h, it is set to 0.15 s".

[0056] Preferably, in S4, the safety following distance calculation formula is:

[0057] Safety following distance = reaction distance + shortest braking distance under current environment + safety redundancy distance;

[0058] Reaction distance = current vehicle speed (m / s) x system reaction time;

[0059] Wherein, the system reaction time is dynamically adjusted: 0.2 seconds for low risk, 0.3 seconds for medium risk, and 0.5 seconds for high risk;

[0060] The shortest braking distance under the current environment is represented as:

[0061] ;

[0062] In the formula, S is the shortest braking distance, is the initial speed of the vehicle;

[0063] Safety redundancy distance = shortest braking distance x redundancy percentage k;

[0064] Wherein, k is set according to the risk value: risk value <0.5, take 10%-20%, default 15%; 0.5≤risk value<0.8, take 30%-50%, risk value increases by 0.1, k increases by 5%; risk value≥0.8, take 60%-80%, default 70%;

[0065] When the calculation result is less than 10m, take 10m, and recalculate every 100ms;

[0066] The safe speed is based on the tunnel speed limit and corrected according to the risk level:

[0067] Risk value <0.5, take 100% of the speed limit;

[0068] 0.5≤risk value<0.8, take 80%-90% of the speed limit;

[0069] Risk value≥0.8, take 60%-70% of the speed limit;

[0070] Linear transition adjustment when the difference between the safe speed and the current speed is greater than 10 km / h, updated every 50 ms;

[0071] The control instructions include acceleration, deceleration, constant speed, lane keeping and emergency instructions. After the emergency instruction is triggered, the "priority avoidance lane changing, emergency braking if there is no space" is executed.

[0072] Therefore, the application adopts the above-mentioned automatic driving vehicle dynamic visual range compensation method for highway tunnels. Compared with the prior art, the technical scheme of the application has the following beneficial effects:

[0073] (1) The technical means of "data preprocessing (denoising, calibration, format conversion) + dynamic fusion algorithm selection" (adapted to linear, strong non-linear, complex dynamic and other different scenes) is adopted, which overcomes the technical problem of unreliable perception fusion, and further achieves the technical effects of constructing a time and space continuous cooperative perception field covering 0-200 meters and realizing efficient fusion of vehicle-mounted and roadside data;

[0074] (2) The technical means of improved LSTM model to predict the trajectory of the preceding vehicle 1-5 seconds and CNN+Transformer model to identify 6 types of standardized behavior intentions (combined with the results of long-range perception) is adopted, which overcomes the technical problem of low prediction and recognition accuracy, and further achieves the technical effects of trajectory prediction position error ≤0.5 meters and precise risk quantification;

[0075] (3) The technical means of dynamically adjusting system reaction time, safety redundancy distance and vehicle speed based on risk level, and integrating real-time road adhesion coefficient, visibility and other environmental parameters is adopted, which overcomes the technical problem of rigid decision control strategy, and further achieves the technical effect of balancing the safety and efficiency of automatic driving vehicles in tunnel passing;

[0076] (4) The technical means of "vehicle-road-tunnel" cooperative perception is adopted, which overcomes the technical problem of blind area in the tunnel (limited by the visual range of single vehicle sensor), and further achieves the technical effect of extending the available visual range of automatic driving vehicles from 0-50 meters to 50-200 meters of long-range visual range.

[0077] The technical scheme of the application will be further described in detail below through the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 The flowchart of the embodiment of the application, a dynamic visual range compensation method for automatic driving vehicles for highway tunnels;

[0079] Figure 2 The flowchart of the embodiment of the application, a dynamic visual range compensation method for automatic driving vehicles for highway tunnels;

[0080] Figure 3 A flowchart of the super-visual range risk perception of an embodiment of the dynamic visual range compensation method for an automatic driving vehicle in a highway tunnel of the present application is shown in the figure;

[0081] Figure 4 A decision and control flowchart of tunnel driving of an embodiment of the dynamic visual range compensation method for an automatic driving vehicle in a highway tunnel of the present application is shown in the figure. DETAILED DESCRIPTION

[0082] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. All other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application. Unless otherwise defined, the technical terms or scientific terms used in the present application should be understood as the usual meanings understood by a person having ordinary skills in the art to which the present application belongs.

[0083] Embodiment One

[0084] As shown in the figure, the present embodiment provides a dynamic visual range compensation method for an automatic driving vehicle in a highway tunnel. It should be understood that the specific parameters, models and protocols mentioned in the present embodiment are only examples for helping a person skilled in the art to understand the present application, and are not a limitation on the present application. Figures 1-4

[0085] The dynamic visual range compensation method for an automatic driving vehicle in a highway tunnel of the present application comprises the following steps:

[0086] S1, acquiring real-time driving state of the vehicle, tunnel design parameters, real-time traffic flow information, environmental parameters and L4 high-precision map, and generating continuous positioning data, comprising the following steps:

[0087] S101, calculating a preset distance before the tunnel entrance every 200 ms , the formula is:

[0088] ;

[0089] In the formula, is the real-time driving speed; is the perception preparation time; is the minimum safety warning distance;

[0090] The real-time driving speed is obtained by fusing the speed measurement data of the vehicle-mounted wheel speed sensor and the vehicle-mounted radar through extended Kalman filtering, and the speed measurement accuracy is ≤±0.1 m / s;

[0091] ​Wherein, the process noise covariance matrix initial value is set as , and the measurement noise covariance matrix is set as according to the sensor manual; the fused real-time driving speed is output once every 100 ms ;

[0092] The perception preparation time is 3-10 s, and the tunnel design speed and real-time traffic density are first read by the vehicle-road cooperation unit:

[0093] When the tunnel design speed is greater than or equal to 80 km / h or the real-time traffic density is greater than or equal to 20 pcu / (km·lane), 6-10 s are taken, and according to the rule that “the design speed is increased by 10 km / h, 1 s is added; the traffic density is increased by 5 pcu / (km·lane), 0.5 s is added”;

[0094] When the tunnel design speed is less than 80 km / h and the real-time traffic density is less than 20 pcu / (km·lane), 3-5 s are taken, and according to the rule that “the design speed is reduced by 10 km / h, 0.5 s is reduced; the traffic density is reduced by 5 pcu / (km·lane), 0.3 s is reduced”;

[0095] The minimum safe warning distance is 50 m by default, and if there is a curve with a radius of curvature less than 500 m or a construction area at the tunnel entrance, according to the rule that “the radius of curvature is reduced by 100 m, 5 m is added; when there is a construction area, 10 m is added”, to ensure that ≥ 50 m, which is used to offset the errors caused by sensor data transmission delay and vehicle-road cooperation unit wake-up delay;

[0096] When the distance between the vehicle and the tunnel entrance is equal to , a data acquisition instruction is triggered, and the trigger signal is directly transmitted to S103 to start the acquisition of dynamic traffic environment information and high-precision maps;

[0097] S102, combined positioning of vehicle-mounted global navigation satellite system and inertial measurement unit, open environment positioning accuracy ≤0.5 meters, in the tunnel entrance shielding area, the inertial navigation of the inertial measurement unit compensates for the loss of vehicle-mounted global navigation satellite system signals, to ensure the continuity of positioning; generate continuous positioning data (including position coordinates, positioning accuracy identifier);

[0098] S103, acquire L4 level high-precision map and multi-dimensional information in the tunnel through the vehicle-road cooperation unit, the multi-dimensional information including traffic flow and obstacle information, environmental parameters.

[0099] The traffic flow and obstacle information includes: vehicle flow density of each lane, lane occupancy, obstacle position (geodetic coordinate system), speed, and size, which are collected by the laser radar and high-definition camera of the roadside unit;

[0100] The environmental parameters include: visibility (collected by a roadside visibility instrument every 30 seconds), road temperature (collected by a buried road temperature sensor every 1 minute), road adhesion coefficient (collected by a roadside friction coefficient sensor every 2 minutes), and lighting brightness (collected by a light sensor every 10 seconds);

[0101] All environmental parameters are transmitted to the vehicle in real time through the vehicle-road cooperation unit;

[0102] The high-precision map is delivered to S206 for sensing the boundary constraints of the field; the traffic flow and obstacle information is delivered to S202 for supplementing the beyond-visual-range information; the visibility and road adhesion coefficient in the environmental parameters are synchronously delivered to S401 for safety distance and vehicle speed calculation, and the lighting brightness is delivered to S304 for beyond-visual-range range adjustment;

[0103] S2, collect local environmental data through the vehicle-mounted sensor, synchronously receive the beyond-visual-range information transmitted by the roadside unit, pre-process and validity verify the local environmental data and the beyond-visual-range information, dynamically select a fusion algorithm for data fusion according to the current scene, and construct a spatiotemporally continuous collaborative sensing field based on the fused data, including the following steps;

[0104] S201, collect 0-50m local data through a combination of vehicle-mounted multiple sensors:

[0105] The laser radar collects obstacle point cloud data within 0-50m around the vehicle, with a data update frequency of 10Hz;

[0106] Two front-view high-definition cameras and two side-view cameras are selected for the camera, the front-view cameras collect environmental illumination intensity and obstacle categories within 0-50m in front, and the side-view cameras collect obstacle information within 0-30m laterally;

[0107] The millimeter wave radar collects obstacle speed and distance information within 0-50m around the vehicle, with a data update frequency of 20Hz;

[0108] Eight ultrasonic radars are selected for the ultrasonic radar, which are installed on the front and rear bumpers of the vehicle, and collect near-range (0-5m) obstacle information, which is used for supplementing perception in low-speed scenes (<20km / h), with a data update frequency of 10Hz;

[0109] All vehicle-mounted sensors are time-synchronized through the Precision Time Protocol, with a synchronization accuracy of ≤1ms. Every 50ms, the collected local environment perception data is packaged in the format of "sensor ID + timestamp + perception data" and transmitted to the vehicle-mounted domain controller, and directly delivered to S203 for preprocessing.

[0110] S202, based on the 5G-V2X or LTE-V2X vehicle-road cooperative communication module, synchronously receives the tunnel over-the-horizon information transmitted by the roadside unit, including the positions, speeds, and trajectories of multiple vehicles ahead, the distribution of mid-to-long-range obstacles, and global visibility parameters, with a communication delay of ≤100ms and a packet loss rate of ≤5%;

[0111] The simultaneously received over-the-horizon information is directly delivered to S204 for format conversion and validity verification.

[0112] S203, the local environment perception data includes the vehicle's position, speed, near-range obstacle features (outline, category), and ambient light intensity. The collected local environment perception data is denoised, time-synchronized, and spatially calibrated, with abnormal data removed and unified into the geodetic coordinate system:

[0113] The denoising method is statistical filtering denoising for laser radar and Gaussian filtering denoising for cameras.

[0114] The time synchronization method is to align different sensor data to the same time axis based on the timestamp of the Precision Time Protocol, with data delayed by more than 5ms supplemented by linear interpolation to ensure time consistency.

[0115] The spatial calibration method is to calibrate the laser radar and camera through a calibration board to obtain the rotation matrix and translation vector, and convert the camera pixel coordinates to laser radar three-dimensional coordinates. The millimeter wave radar and laser radar are calibrated through real vehicle field testing to ensure that the spatial coordinates are unified into the geodetic coordinate system WGS84. All calibration parameters are stored in the vehicle-mounted controller, and recalibration is performed every 3 months with a calibration accuracy of ≤5cm.

[0116] Set reasonable ranges for each sensor data (laser radar detection distance 0-50m, camera light intensity 0-255, millimeter wave radar speed 0-120km / h), and mark and remove data outside the range as abnormal. At the same time, through the difference value judgment of adjacent two frames of data, the sudden abnormal data is removed.

[0117] The effective local data after preprocessing is delivered to S205 for fusion with roadside data.

[0118] S204, format conversion and validity verification of received over-the-horizon information, specifically:

[0119] The over-the-horizon information format is converted and verified according to the confidence, and the accuracy and data consistency of the roadside sensor are weighted for calculation, and the weights are respectively set to 0.6 and 0.4;

[0120] The confidence is greater than or equal to 0.8, and the direct determination is effective data; the confidence is between 0.5 and 0.8, and the same type of information of three consecutive time frames needs to be extracted, the standard deviation is calculated, and if the position standard deviation is less than or equal to 2m or the speed standard deviation is less than or equal to 1km / h, the verification is passed and the effective data is included; the confidence is less than 0.5, and the direct elimination is performed;

[0121] The effective over-the-horizon data after verification is transmitted to S205 for fusion with local data;

[0122] S205, according to the current scene, dynamically selects a fusion algorithm, and performs time and space alignment and feature fusion on the preprocessed vehicle local data and the roadside effective data, specifically:

[0123] The scene with high linear approximation degree (the linear correlation coefficient of local perception data and over-the-horizon information is greater than or equal to 0.8) uses extended Kalman filtering, and the process noise covariance matrix is set to , and the measurement noise covariance matrix is adjusted according to the data confidence;

[0124] The strong nonlinear scene (linear correlation coefficient is less than 0.6) uses unscented Kalman filtering, wherein the number of sigma points is set to 2n+1 (n is the state dimension), and the scale factors are , to ensure the nonlinear data fusion accuracy;

[0125] The multi-sensor data uncertainty known scene uses Bayesian estimation, the prior probability is set based on the sensor accuracy, the posterior probability is calculated through the Bayesian formula, and the fusion result with the maximum posterior probability is selected;

[0126] The complex dynamic scene (there are multiple lane changes, sudden obstacles, and light mutations in the tunnel) uses a deep learning model; the training data set contains multi-sensor data in different scenes in the tunnel (sample size is greater than or equal to 100,000 frames), and 5-fold cross-validation is performed; the model input is a feature vector, the local feature is extracted through ResNet50, the global feature is fused through the Transformer encoder, the training batch size is 32, the learning rate is 0.001, the optimizer is Adam, and the loss function is cross-entropy loss, and the fusion accuracy is greater than or equal to 92%;

[0127] The unified data after fusion is transmitted to S206 for constructing a collaborative perception field;

[0128] S206, based on the fusion data transmitted by S205, constructs a time and space continuous perception field covering 0-200m, and dynamically maps the vehicle, obstacle and environmental parameter changes; wherein, 0-50m is a local time and space continuous perception field, and 50-200m is an over-the-horizon time and space continuous perception field;

[0129] The constructed tunnel environment collaborative perception field is delivered to S301 as a whole to provide a data basis for feature extraction and risk perception;

[0130] S3, based on the collaborative perception field, extract trajectory prediction features and behavior intention features, obtain the front vehicle motion trajectory prediction result and the behavior intention recognition conclusion through the trajectory prediction model and the behavior intention recognition model respectively, fuse the trajectory prediction result and the behavior intention recognition conclusion, dynamically adjust the over-the-horizon range combined with the lighting brightness in the environmental parameters, realize the quantitative perception of potential risks in the over-the-horizon range, including the following steps;

[0131] S301, based on the collaborative perception field delivered by S206, synchronously extract trajectory prediction features and behavior intention features; wherein the trajectory prediction features include the position coordinate sequence, the vehicle speed sequence, the acceleration sequence and the current lane state of the front vehicle in the past 3 seconds; the behavior intention features include the front vehicle turn signal state, the steering wheel angle data and the vehicle speed change rate;

[0132] The extracted trajectory prediction features are delivered to S302, and the behavior intention features are delivered to S303;

[0133] S302, an improved LSTM model is used to input the trajectory prediction features delivered by S301, and the future 1-5 second motion trajectory with a position error ≤0.5 meters is output; the trajectory prediction result is delivered to S304 for risk fusion judgment;

[0134] S303, a model combining convolutional neural network (CNN) and Transformer is used to perform feature fusion and semantic analysis on the behavior intention features delivered by S301, and 6 standardized behavior intention results (normal driving, preparation for deceleration, preparation for lane change, lane change, emergency deceleration, temporary parking) are output;

[0135] The model structure is specifically as follows:

[0136] CNN feature extraction layer: 2 convolutional layers are set, the first layer has a convolution kernel size of 3×3, a number of 32, a step of 1, a padding method of "same", and a ReLU activation function; the second layer has a convolution kernel size of 3×3, a number of 64, a step of 1, a padding method of "same", and a ReLU activation function; a maximum pooling layer is connected after the convolutional layer to extract local spatial features;

[0137] Transformer encoder layer: 2 encoder layers are set, each layer contains a multi-head attention mechanism and a feedforward neural network to fuse local features and global time sequence features;

[0138] Output layer: a fully connected layer is used, the output dimension is set to 6, the activation function is Softmax, and the class with the maximum probability is the final result.

[0139] The behavior intention recognition conclusion is delivered to S304 for risk fusion judgment;

[0140] S304, fuses the trajectory prediction result delivered by S302 and the behavior intention recognition conclusion delivered by S303, and realizes three-dimensional position positioning, dynamic moving trend deduction and potential risk quantitative perception of the obstacle in the beyond-visual-range of the front of the vehicle based on the tunnel environment perception boundary constraint;

[0141] The risk quantitative calculation formula is:

[0142] Risk value = obstacle distance weight x 0.4 + relative speed weight x 0.3 + behavior intention weight x 0.3;

[0143] Obstacle distance weight = 1 - (obstacle distance / maximum beyond-visual-range);

[0144] Relative speed weight = relative speed / vehicle speed;

[0145] Among them, the behavior intention weight is set according to the category (emergency deceleration, temporary parking = 1.0; preparation for deceleration, preparation for lane change = 0.6; normal driving, lane changing = 0.2), and the risk value range is 0-1 (the greater the value, the higher the risk);

[0146] The upper limit of the beyond-visual-range is dynamically adjusted in combination with the lighting brightness delivered by S103:

[0147] When the lighting brightness < 200 lux, the beyond-visual-range = road side detection distance x 0.8;

[0148] When the lighting brightness is 200-500 lux, the beyond-visual-range = road side detection distance x 0.9;

[0149] When the lighting brightness > 500 lux, the beyond-visual-range = road side detection distance x 1.0;

[0150] The beyond-visual-range perception result (including obstacle information, dynamic trend and risk value) is delivered to S401 as the core input of decision control;

[0151] S4, integrates the risk quantitative result, the tunnel real-time environment parameter and the vehicle dynamic performance parameter, calculates the safe following distance and corrects the safe speed, generates the corresponding control instruction based on the comparison result of the safe following distance, the safe speed and the current driving state of the vehicle, including the following steps:

[0152] S401, integrates multiple source core data, including the beyond-visual-range perception result delivered by S304, the tunnel real-time environment parameter (road adhesion coefficient, visibility) delivered by S103 and the vehicle dynamic performance parameter;

[0153] The power performance parameters include maximum braking deceleration, maximum acceleration, and steering system response delay, and the determination rules are as follows:

[0154] The maximum braking deceleration is dynamically adjusted according to the road adhesion coefficient to ensure that there is no lock during braking, and the formula is:

[0155] ;

[0156] In the formula, is the acceleration of gravity, ; is the road adhesion coefficient;

[0157] The maximum acceleration is adjusted according to the current power of the vehicle or the engine speed. When the power is <50% or the speed is >2000rpm, ; otherwise ;

[0158] The steering system response delay is pre-calibrated by bench test. When the speed is <30km / h, , when the speed is ≥30km / h, , and is stored in the parameter library for calling;

[0159] The integrated complete data is transmitted to S402 and S403, respectively, for safety following distance calculation and safety speed correction;

[0160] S402, based on the improved adaptive cruise control model, calculates the safety following distance combined with the data integrated by S401, and recalculates every 100ms, and the formula is:

[0161] Safety following distance = reaction distance + shortest braking distance under current environment + safety redundancy distance;

[0162] Reaction distance = current vehicle speed (m / s) × system reaction time;

[0163] Among them, the system reaction time is dynamically adjusted according to the risk level transmitted by S304: 0.2s for low risk, 0.3s for medium risk, and 0.5s for high risk;

[0164] The shortest braking distance under the current environment is represented as:

[0165] ;

[0166] In the formula, S is the shortest braking distance (unit: meter), is the initial speed of the vehicle (unit: m / s), is the road adhesion coefficient;

[0167] The current speed of the vehicle Read from the vehicle CAN bus, updated every 50ms; road adhesion coefficient From the integrated environmental parameters in S401, if , the calculation result needs to be multiplied by a correction factor of 1.2;

[0168] Safety redundancy distance = shortest braking distance x redundancy percentage ;

[0169] Where, According to the risk value setting: low risk (<0.5) takes 10%-20% (default 15%), medium risk (0.5≤risk value<0.8) takes 30%-50% (risk value increases by 0.1, k increases by 5%), high risk (≥0.8) takes 60%-80% (default 70%);

[0170] If the calculation result is <10m, take 10m to avoid too close following, and the final safe following distance is transmitted to S404 for control command generation;

[0171] S403, based on the tunnel speed limit standard, combined with the risk level transmitted by S304 to correct the vehicle speed: low risk (risk value <0.5) executes at 100% of the speed limit, medium risk (0.5≤risk value<0.8) is reduced to 80%-90% of the speed limit (and ≥20km / h), high risk (risk value ≥0.8) is reduced to 60%-70% of the speed limit;

[0172] If the difference between the safe vehicle speed and the current vehicle speed is >10km / h, use linear transition adjustment, update the target vehicle speed every 50ms, to avoid the influence of sudden acceleration or sudden deceleration on comfort;

[0173] The corrected safe vehicle speed is transmitted to S404 for control command generation;

[0174] S404, compare the safe following distance transmitted by S402 with the current following distance, and the safe vehicle speed transmitted by S403 with the current vehicle speed, to generate corresponding control commands:

[0175] Following distance trigger: current following distance < safe following distance + 5m, generate "acceleration command" (acceleration acceleration = ); current following distance > safe following distance - 5m, generate "deceleration command" (deceleration acceleration = ); within the range of safe following distance ±5m, generate "constant speed command";

[0176] Vehicle speed trigger: current vehicle speed < safe vehicle speed - 2km / h, combined with the following distance condition to generate "acceleration command"; current vehicle speed > safe vehicle speed + 2km / h, combined with the following distance condition to generate "deceleration command"; within the range of safe vehicle speed ±2km / h, generate "constant speed command";

[0177] Lane keeping instruction: if the front vehicle trajectory prediction has no lane change risk, control the steering system to maintain the current lane center driving, and the steering angle is dynamically adjusted according to the lane curvature (the steering angle is increased by 1° for every 100 m reduction in the curvature radius), and the steering response delay is ≤ ;

[0178] Based on the over-the-horizon perception result transmitted in S304, if an unyielding sudden obstacle is detected, an emergency process is triggered:

[0179] First, the adjacent lane space is judged by cooperative perception field, and if the safety condition (lateral distance m, longitudinal distance from the front vehicle in the adjacent lane , and lateral distance from the rear vehicle in the adjacent lane) is met, a “yielding lane change instruction” is generated, and the lane change trajectory is planned by a quintic polynomial (lateral acceleration ≤ 0.3 m / s 2 , and lane change time ≥ 3 s);

[0180] If there is not enough space in the adjacent lane, an “emergency braking instruction” (braking acceleration = ) is generated, and at the same time, a danger warning flash light is triggered, and the wheel rotation speed is monitored to avoid locking;

[0181] After the execution of the instruction, the execution result is collected by the vehicle-mounted sensor every 50 ms and fed back to the closed-loop control module, and if the deviation is > 10%, the instruction parameters are corrected to ensure the control accuracy.

[0182] Therefore, the automatic driving vehicle dynamic horizon compensation method for highway tunnels adopts the above method, which constructs a time and space continuous cooperative perception field, realizes efficient fusion of vehicle and road data, accurately completes trajectory prediction and risk quantification, balances tunnel passing safety and efficiency, and effectively extends the over-the-horizon range of the automatic driving vehicle.

[0183] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0184] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A highway tunnel-oriented dynamic range compensation method for an autonomous vehicle, characterized in that, The method comprises the following steps: S1, acquiring real-time driving state of the vehicle, tunnel design parameters, real-time traffic flow information, environmental parameters and a map to generate continuous positioning data; the real-time driving state, positioning data, tunnel design parameters, traffic flow information, environmental parameters and high-precision map are all transmitted to S2; S2, collecting local environmental data through a vehicle-mounted sensor, synchronously receiving over-the-horizon information transmitted by a roadside unit, pre-processing and validity verifying the local environmental data and the over-the-horizon information respectively, dynamically selecting extended Kalman filtering, unscented Kalman filtering, Bayesian estimation and a deep learning model for vehicle-road data fusion according to linear characteristics, nonlinear characteristics and data uncertainty of a tunnel scene, and constructing a time-space continuous collaborative perception field based on the fused data; the collaborative perception field is transmitted to S3; S3, extracting trajectory prediction features and behavior intention features based on the collaborative perception field, acquiring a front vehicle motion trajectory prediction result and a behavior intention recognition conclusion through a trajectory prediction model and a behavior intention recognition model respectively, fusing the trajectory prediction result and the behavior intention recognition conclusion, dynamically adjusting an over-the-horizon range combined with lighting brightness in the environmental parameters, realizing quantitative perception of potential risks in the over-the-horizon range, and transmitting the risk quantization result and the adjusted over-the-horizon range to S4; S4, integrating the risk quantization result, tunnel real-time environmental parameters and vehicle dynamic performance parameters, calculating a safe following distance and correcting a safe speed, and generating corresponding control instructions based on a comparison result of the safe following distance, the safe speed and a current driving state of the vehicle. 2.The highway tunnel-oriented dynamic range compensation method for an autonomous vehicle according to claim 1, wherein, In S1, a preset distance before a tunnel entrance needs to be calculated, and the calculation formula of the preset distance is: ； In the formula, is the real-time driving speed; is the perception preparation time; is the minimum safety warning distance; The real-time driving speed The real-time driving speed is obtained by fusing the data of the vehicle-mounted wheel speed sensor and the vehicle-mounted radar speed measurement through an extended Kalman filter, the speed measurement accuracy is ≤±0.1 m / s, and the fused real-time driving speed is output once every 100 ms; the initial value of the process noise covariance matrix of the extended Kalman filter is set as ; and the measurement noise covariance matrix is set as . 3.The highway tunnel-oriented dynamic range compensation method for an autonomous vehicle according to claim 2, wherein, S1, perception preparation time 3~10s, according to the tunnel design speed and real-time traffic density dynamic adjustment: When the tunnel design speed ≥ 80 km / h or the real-time traffic density ≥ 20 pcu / (km·lane), Take 6~10 s, according to "design speed every increase of 10 km / h, Increase 1 s; The traffic density increases by 5 pcu / (km·lane), The rule of increasing 0.5 s" is dynamically valued; When the tunnel design speed < 80 km / h and the real-time traffic density < 20 pcu / (km·lane), Take 3~5s, according to "design speed every 10km / h, Reduce 0.5s; every 5pcu / (km·lane), Reduce 0.3s" rule value; Minimum safety warning distance The default value is 50 m. If there is a curve with a radius of curvature < 500 m or a construction area at the tunnel entrance, the rule is modified as "5 m is added for each 100 m reduction in radius of curvature, up to a maximum of 10 m when there is a construction area." ≥ 50 m;​​ The calculation frequency of the preset distance is once every 200 ms, and a data acquisition instruction is triggered when the distance between the vehicle and the tunnel entrance is equal to the preset distance; Continuous positioning data is generated by combining vehicle-mounted global navigation satellite system and inertial measurement unit positioning, and the positioning accuracy in an open environment is ≤0.5 meters; The environmental parameters include visibility, road temperature, road adhesion coefficient and lighting brightness, which are collected by corresponding sensors at a frequency of "collecting visibility once every 30 seconds, collecting road temperature once every 1 minute, collecting road adhesion coefficient once every 2 minutes, and collecting lighting brightness once every 10 seconds". 4.The highway tunnel-oriented dynamic range compensation method for an autonomous vehicle according to claim 1, wherein, In S2, the vehicle-mounted sensor includes a laser radar, a front-view high-definition camera, a side-view camera, a millimeter wave radar and an ultrasonic radar; The laser radar collects obstacle point cloud data within a range of 0-50 m, and the update frequency is 10 Hz; The front-view camera collects front 0-50 m environmental light intensity and obstacle categories, and the side-view camera collects lateral 0-30 m obstacle information; The millimeter wave radar collects 0-50 m obstacle speed and distance information, and the update frequency is 20 Hz; The ultrasonic radar is installed on the front and rear bumpers of the vehicle, collects 0-5 m close-range obstacle information, and the update frequency is 10 Hz, which is used for supplementary perception in a low-speed scene with a vehicle speed <20 km / h.

5. The method of claim 1, wherein, In S2, all vehicle-mounted sensors transmit the collected local environmental data every 50 ms. The preprocessing of local environment data includes denoising, time synchronization and spatial calibration. The denoising method is laser radar statistical filter denoising and camera Gaussian filter denoising. After spatial calibration, it is unified to the WGS84 geodetic coordinate system, and the calibration accuracy is ≤5 cm. It is recalibrated once every 3 months; The over-the-horizon information is received through the 5G-V2X or LTE-V2X vehicle-road cooperative communication module. The validity verification rule of the over-the-horizon information is: After format conversion, it is judged according to the confidence. The confidence is calculated by weighting the accuracy of the roadside sensor and the data consistency. The weights are 0.6 and 0.4 respectively. When the confidence is ≥0.8, it is directly determined to be valid. When the confidence is between 0.5 and 0.8, it is valid only when the standard deviation of 3 consecutive frames of data meets the "position standard deviation ≤2m or speed standard deviation ≤1km / h". When the confidence is <0.5, it is directly excluded. The rule for dynamically selecting the fusion algorithm is: The linear approximation degree high scene uses the extended Kalman filter, the process noise covariance matrix is set as , and the measurement noise covariance matrix is adjusted according to data confidence. For strong nonlinear scenarios, unscented Kalman filter is used, where the number of sigma points is set to 2n+1, where n is the state dimension, and the scaling factors are ; In the known uncertainty scenario of multi-sensor data, Bayesian estimation is used. The prior probability is set based on the sensor accuracy, and the posterior probability is calculated through the Bayesian formula. The fusion result with the maximum posterior probability is selected. In complex dynamic scenarios, a deep learning model is used. The training data set contains multi-sensor data in different scenarios in the tunnel, which is verified by 5-fold cross-validation. The model input is a feature vector. Local features are extracted through ResNet50, and global features are fused through a Transformer encoder. The loss function is cross-entropy loss, and the fusion accuracy is ≥92%. The collaborative perception field covers 0-200m, of which 0-50m is the local perception field and 50-200m is the over-the-horizon perception field.

6. The method of claim 1, wherein, In S3, the trajectory prediction result and the behavior intention recognition conclusion are fused, specifically: The trajectory prediction features include the position coordinate sequence, speed sequence, acceleration sequence, and current lane state of the preceding vehicle in the past 3 seconds. The trajectory prediction model is an improved LSTM model, which outputs the motion trajectory in the future 1-5 seconds with a position error ≤0.5 meters. The behavior intention features include the turn signal state, steering wheel angle data, and speed change rate of the preceding vehicle. The behavior intention recognition model is a convolutional neural network combined with a Transformer model, which outputs 6 standardized behavior intentions, including normal driving, preparing to decelerate, preparing to change lanes, changing lanes, emergency deceleration, and temporary parking. The calculation formula for risk quantification is: Risk value = obstacle distance weight × 0.4 + relative speed weight × 0.3 + behavior intention weight × 0.3; Obstacle distance weight = 1 - (obstacle distance / maximum over-the-horizon range); Relative speed weight = relative speed / vehicle speed; Among them, the behavior intention weight is set as "emergency deceleration, temporary parking = 1.0; preparing to decelerate, preparing to change lanes = 0.6; normal driving, changing lanes = 0.2", and the risk value range is 0-1. The over-the-horizon range adjustment rule is: When the lighting brightness is <200 lux, the over-the-horizon range = roadside detection distance × 0.8; When the lighting brightness is 200-500 lux, the over-the-horizon range = roadside detection distance × 0.9; When the lighting brightness is >500 lux, the over-the-horizon range = roadside detection distance × 1.

0. ​ 7. The method of claim 1, wherein, In S4, the vehicle dynamic performance parameters include maximum braking deceleration, maximum acceleration, and steering system response delay. The formula for maximum braking deceleration is: ； wherein is the gravitational acceleration, ; is the road adhesion coefficient; Maximum acceleration is set to "when battery level is <50% or engine speed is >2000 rpm". Otherwise set to "Adjustment; The steering system response delay is adjusted according to "when the speed is < 30 km / h, set to 0.2 s, when the speed is ≥ 30 km / h, set to 0.15 s". 8.The highway tunnel-oriented dynamic range compensation method for an autonomous vehicle according to claim 7, wherein, In S4, the safety following distance calculation formula is: Safety following distance = reaction distance + shortest braking distance under current environment + safety redundancy distance; Reaction distance = current vehicle speed (m / s) × system reaction time; Wherein, the system reaction time is dynamically adjusted: low risk 0.2 seconds, medium risk 0.3 seconds, high risk 0.5 seconds; The shortest braking distance under the current environment is represented as: ； In the formula, S is the shortest braking distance, is the initial speed of the vehicle; Safety redundancy distance = shortest braking distance × redundancy percentage k; Wherein, k is set according to the risk value: risk value < 0.5, take 10%-20%, default 15%; 0.5 ≤ risk value < 0.8, take 30%-50%, risk value increases by 0.1, k increases by 5%; risk value ≥ 0.8, take 60%-80%, default 70%; When the calculation result is < 10m, take 10m, recalculate every 100ms; The safe speed is based on the tunnel speed limit and corrected according to the risk level: Risk value < 0.5, take 100% of the speed limit; 0.5 ≤ risk value < 0.8, take 80%-90% of the speed limit; Risk value ≥ 0.8, take 60%-70% of the speed limit; When the difference between safe speed and current speed is > 10km / h, linear transition adjustment, update every 50ms; control instructions include acceleration, deceleration, constant speed, lane keeping and emergency instruction, after the emergency instruction is triggered, "preferential avoidance lane change, no space then emergency braking" is executed.

9. A computer device, comprising: Including: A processor is coupled with a memory to read and execute instructions and / or program codes in the memory to perform the method of any one of claims 1-8.

10. A computer readable medium characterized by The computer readable medium stores computer program codes which, when executed on a computer, cause the computer to perform the method of any one of claims 1-8.

Images