Vehicle collision risk prediction device and method
Through the vehicle collision risk prediction device, the damage risk of vehicles and occupants is calculated using perception and prediction components, which solves the problems of slow response in traditional systems and indetailed data records, real-time risk assessment and early warning of vehicles and occupants, and reduces accidents damage and subsequent processing complexity.
Patent Information
- Application Number
- CN202310000753.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-03
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-01-03
AI Technical Summary
Traditional collision safety systems react slowly after the accident, and lack real-time records of detailed data on the accident site, which cannot effectively predict the risk of vehicle collision caused by driver operation uncertainty, resulting in the occupant protection system absorbing a large amount in a very short time. Energy, causing damage.
Design a vehicle collision risk prediction device, including perception components, basic components and prediction components, collect personnel, environment and vehicle information through perception components, use digital human body modules and vehicle prediction modules to predict collision risks, and combine personnel and vehicle damage prediction modules, Calculate reliability indicators for collision risk and damage risk for early warning and airbag control.
Real-time and accurate prediction of vehicle and occupants' damage risks is achieved, pre-collision system is started in advance, occupants' protection time is extended, energy absorption per unit time is reduced, and detailed damage information is provided for subsequent maintenance and insurance processes.
Smart Images

Figure CN116013110B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent automobile design, and in particular to a vehicle collision risk prediction device and method based on the fusion of model design and data drive. Background Art
[0002] Collision safety technology absorbs the energy from the collision between the vehicle and the target object through the vehicle structure, restraint system, pedestrian protection device, etc. within a limited time and space, so as to avoid or reduce the injuries to people inside and outside the vehicle in the collision.
[0003] The control of the traditional collision safety occupant protection system relies only on several external acceleration or pressure sensors to identify whether a collision accident has occurred. The limitation of perception ability causes the airbag to be detonated too late or unable to detonate in some actual road traffic accidents, or to detonate in working conditions where it should not be detonated (misfire). The traditional collision safety development standards are limited to limited test conditions, dummies of standard body shape and sitting posture. At the same time, in an environment where human driving is the main and intelligent driving is the auxiliary, the "human-vehicle-human" interactive response, the driver shows different stress reactions and operations (operation uncertainty) due to the different "people, time and place" based on the operating status of the other vehicle and the road environment during driving. Body shape, sitting posture, stress reaction, etc. will have a significant impact on human injuries. Traditional collision safety only works within about 150ms after the accident occurs, which means that the occupant protection system needs to absorb the collision energy in a very short time. The energy transferred to the driver and passengers per unit time is high, and the damage caused is high. Traditional collision safety uses the Vehicle Event Data Recording System (DER) to record relevant vehicle data at the time of the collision after an accident occurs. It lacks the detailed vehicle damage and personal injury information required for accident rescue (E-Call), after-sales maintenance and insurance claims. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology, consider the operational uncertainty caused by the driver's stress response during vehicle driving, and propose a method for quickly predicting the risk of vehicle collision and the risk of vehicle and personnel injury in the event of a hypothetical collision: predict the risk of vehicle and occupant injury based on the parameters captured by the perception component at the current moment, and consider the uncertainty of the vehicle state caused by the driver's operational uncertainty at the current moment, and calculate the reliability index of the collision risk, vehicle and personnel injury risk. Based on the above indicators, a safety assessment signal or instruction with clear guidance can be sent for application scenarios such as vehicle warning in the pre-collision stage, activation of the pre-collision system, active intervention of the auxiliary / automatic driving system, ignition of the airbag controller in the collision stage, and driving risk assessment.
[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0006] A vehicle collision risk prediction device comprises: a perception component, a basic component, a prediction component and a decision component; the perception component is used to collect information about personnel, environment and vehicles; the basic component comprises a digital human body module and a vehicle prediction module, the digital human body module is used to predict personnel position parameters from the current moment to a certain moment in the future according to the scene parameters where the personnel are located, the vehicle prediction module comprises a state parameter preprocessing unit, a trajectory prediction unit and a collision detection unit, the state parameter preprocessing unit numbers and classifies the target object according to the target object locked by the perception component, the trajectory prediction unit performs motion state correction on the locked target object or uses a kinematic or dynamic model to predict the driving trajectory of the self-vehicle and / or the target object within a preset time range, the collision detection unit is used to detect whether the self-vehicle and the target object collide at a specific moment; the prediction component comprises a personnel injury prediction module, a vehicle injury prediction module and a reliability prediction module, the personnel injury prediction module is used to predict the personnel position parameters from the current moment to a certain moment in the future according to the collision waveform, personnel state parameters and restraint system state parameters when the self-vehicle collides The damage curve of the occupant after a hypothetical collision is predicted in real time, and the occupant injury risk value is calculated based on the damage curve; the vehicle damage prediction module includes a deformation prediction unit and a collision waveform prediction unit, the deformation prediction unit predicts the vehicle damage risk value after a hypothetical collision in real time according to the collision scene parameters at the collision moment, the collision waveform prediction unit predicts the collision waveform after a hypothetical collision according to the collision scene parameters, and the predicted collision waveform is used as one of the inputs of the personnel injury prediction module; the reliability prediction module predicts the reliability indicators ∑CR, ∑HIR, ∑VIR of the expected parameters CR, HIR, VIR according to the scene parameters and the expected parameters when a hypothetical collision occurs, the expected parameters including the collision risk CR, the occupant injury risk value HIR, and the vehicle injury risk value VIR; the decision component obtains the expected parameters CR, HIR, VIR and their corresponding reliability indicators ∑CR, ∑HIR, ∑VIR, while taking into account the extreme values and change trend indicators of the predicted signal in the time domain, performs a comprehensive collision risk assessment, and outputs a control instruction.
[0007] Furthermore, the perception component includes a personnel perception module, an environment perception module and a vehicle perception module; the personnel perception module is used to collect personnel status parameters in real time, and the personnel status parameters are obtained through a camera and / or a TOF camera, including basic occupant parameters and occupant position parameters, the basic occupant parameters include the height, weight and age of the occupant, the occupant position parameters include the occupant's sitting posture and the occupant's sitting position, the occupant's sitting posture includes the occupant's forward tilt angle and roll angle, and the occupant's sitting position refers to the occupant's absolute position relative to the vehicle body; the environment perception module is used to collect target object parameters in real time, the target object parameters include the mass, size, type and kinematic parameters of the target object, and the kinematic parameters of the target object include the speed, acceleration, yaw angle and yaw angular velocity of the target object; the vehicle perception module is used to collect self-vehicle parameters and restraint system status parameters in real time, the self-vehicle parameters include the mass, size and kinematic parameters of the self-vehicle, and the kinematic parameters of the self-vehicle include the speed, acceleration, yaw angle and yaw angular velocity of the self-vehicle.
[0008] Furthermore, the state parameter preprocessing unit numbers and classifies the target objects according to the target objects locked by the sensing component. Specifically, the target objects are locked and numbered according to the sensing component, the collision time TTC of each target object is calculated, and the target objects are arranged in ascending order according to the TTC value of each target object; a first collision time threshold TTC is preset. thres1 and the second collision time threshold TTC thres2 , if the collision time TTC of a target object satisfies TTC thres1 ≤TTC≤TTC thres2 , define it as a potential dangerous target, classify it into the "potential dangerous target set", and continuously monitor its motion state; if the collision time TTC of a target satisfies TTC<TTC thres1 , it is defined as a dangerous target and classified into the "dangerous target set". The trajectory prediction unit corrects the motion state of the target in the "potentially dangerous target set". The trajectory prediction unit uses a kinematic or dynamic model to predict the driving trajectory of the vehicle and the target in the "dangerous target set" within a preset time range. The collision detection unit is used to detect whether the vehicle and the target collide at a specific time. If the vehicle and the target are in contact in space, it is defined as a "collision", and this time is defined as the collision time T c , collision risk CR=1, and record T c The collision scenario parameters at the moment include the vehicle speed, the target speed, the collision overlap rate, the collision angle, the target mass and the target type; if no contact is detected, the collision risk CR=0.
[0009] The detection of the collision detection unit includes a rough detection and a fine detection stage. The rough detection is to calculate the distance L between the geometric center of the vehicle and the geometric center of the target object.i , and calculate the minimum enveloping circles of the ego vehicle and the target object respectively. The centers of the enveloping circles are at their respective geometric centers. When the enveloping circles of the ego vehicle and the target object are tangent, the distance between the geometric centers of the ego vehicle and the target object is recorded as L crit , when L i <αL crit When , fine detection is performed, and a is the gain coefficient. Fine detection is used to detect whether the vehicle and the target object overlap in space. If there is overlap, it is defined as contact. The specific method of fine detection is the separation axis STA method or the GJK method.
[0010] Furthermore, the motion state correction performs trajectory correction through extended Kalman filtering, and the dynamic model is a constant angular velocity and constant acceleration motion model, a constant acceleration model or a constant speed model.
[0011] Further, the occupant injury prediction module is an offline occupant injury prediction model obtained by training based on a occupant injury data set, wherein the occupant injury data set refers to a database whose input is a collision waveform, occupant state parameters, and restraint system state parameters, and whose output is an occupant injury curve, which can be obtained through restraint system numerical calculation or collision test;
[0012] The deformation prediction unit is an offline deformation prediction model obtained by training based on the vehicle damage data set. The collision waveform prediction unit is an offline collision waveform prediction model obtained by training based on the vehicle damage data set. The vehicle damage data set refers to the input of the collision time T c The collision scene parameters are output as a database of vehicle deformation and waveform characteristic values for the collision waveform prediction unit, which are obtained through whole vehicle finite element numerical calculation, whole vehicle collision test or collision accident collection;
[0013] The reliability prediction module is an offline reliability prediction model obtained by training a reliability quantified index database, wherein the reliability quantified index database refers to a database whose input is scenario parameters and expected parameters and whose output is reliability indexes in the current scenario.
[0014] Furthermore, the deformation prediction model, collision waveform prediction model and reliability prediction model adopt the machine learning integration algorithm of random forest and XGBOOST, or the neural network algorithm of multi-layer perceptron, CNN and RNN.
[0015] Furthermore, the calculation method of the reliability prediction module is as follows: according to the scene parameters and expected parameters at the current moment, after considering the driver's operation uncertainty, the driver's operation uncertainty is set to have U discrete values, corresponding to u different driving trajectories, and according to the vehicle prediction module, the personnel injury prediction module, and the vehicle damage prediction module, U different vehicle collision risk prediction results, U personnel injury risk prediction results, and U vehicle damage risk prediction results can be obtained respectively, which are recorded as: [(CR k_1 , HIR k_1 , VIR k_1 ), (CR k_2 , HIR k_2 , VIR k_2 ),...,(CR k_u , HIR k_U , VIR k_U )],
[0016] The reliability quantitative index of the kth target is expressed as:
[0017]
[0018] When ∑CR is smaller, CR is more accurate and more likely to occur. Conversely, it means that the driver's operation is more unstable and less likely to occur when the driver intervenes.
[0019] When ∑HIR is smaller, the HIR is more accurate and more likely to occur. On the contrary, it means that the driver's operation is more unstable and less likely to occur when the driver intervenes.
[0020] The smaller ∑VIR is, the more accurate VIR is and the more likely it is to occur. On the contrary, the smaller ∑VIR is, the more unstable the driver's operation is and the less likely it is to occur.
[0021] Furthermore, the driver operation uncertainty is quantitatively characterized as follows: the acceleration a and the yaw angular velocity ω in the driver operation behavior database are numerically discretized under the premise that physical conditions permit, and the acceleration a is set to have ζ discrete values and the yaw angular velocity ω is set to have η discrete values, that is, ω=[ω1,ω1,...,ω η ], establish a posterior model for predicting driver operation behavior and generate different trajectories.
[0022] Furthermore, the decision-making component outputs control instructions through fuzzy control, empirical calibration or logical judgment.
[0023] The decision-making component refers to obtaining six indicators, namely, risk value (CR, HIR and VIR) and its corresponding reliability (∑CR, ∑HIR and ∑VIR), and taking into account the extreme value and change trend of the predicted signal in the time domain, and directly using these indicators or derived new indicators to conduct a comprehensive collision risk assessment as input for other specific application scenario instructions. Its control method can output control instructions through fuzzy control, empirical calibration, or logical judgment.
[0024] As a specific example, the action instruction can be: 0 means no action, 1 means level 1 reminder, such as visual reminder, 2 means level 2 reminder, such as sound reminder, 3 means tactile reminder, such as vibration for safety, 4 means level 1 adjustment for pre-collision, such as tightening for safety with a lower force value, 5 means level 2 adjustment for pre-collision, such as tightening for safety with a medium force value, and adjusting the seat position to a standard position (such as zero gravity seat return) at the same time, 6 means level 3 adjustment for pre-collision, such as tightening for safety with a higher force value, activating the pre-collision safety airbag, and adjusting the algorithm of the airbag controller and the threshold limit for airbag detonation to detonate the airbag in advance.
[0025] The above indicators can also be used for path selection, driving risk assessment and accident rescue in autonomous driving scenarios.
[0026] A vehicle collision risk prediction method, using the vehicle collision risk prediction device of the present invention to predict vehicle collision risk, comprises the following steps:
[0027] Step 1: Input the information of people, environment and vehicles collected by the perception component into the basic component;
[0028] Step 2: The vehicle prediction module of the basic component outputs the collision risk CR value and collision scenario parameters based on the environment and vehicle information, and the digital human body module of the basic component obtains the personnel status parameters at the time of collision based on the personnel information and vehicle information;
[0029] Step 3: The vehicle damage prediction module of the prediction component outputs the vehicle damage risk value VIR and the collision waveform according to the collision scene parameters, and inputs the collision waveform, the occupant state parameters at the time of collision, and the restraint system state parameters input by the perception component into the occupant damage prediction module to obtain the occupant injury risk value HIR;
[0030] Step 4: Input the collision scene parameters and expected parameters CR, HIR, and VIR into the reliability prediction module to obtain corresponding reliability indicators ∑CR, ∑HIR, and ∑VIR;
[0031] Step 5: The decision component performs a comprehensive collision risk assessment based on the expected parameters CR, HIR, VIR and their corresponding reliability indicators ∑CR, ∑HIR, ∑VIR, and takes into account the extreme values and change trend indicators of the predicted signal in the time domain, and outputs a control instruction;
[0032] Step six, according to the collision waveform recorded by the airbag controller, the detailed damage of the vehicle and personnel is obtained through the vehicle damage prediction module and the personnel damage prediction module, and sent to the cloud server through the wireless network, and then notified to the relevant parties, including maintenance, insurance, and rescue.
[0033] Beneficial effects of the present invention:
[0034] 1. The present invention takes into account the uncertainty of driver operation, comprehensively considers the information of the driver, the environment and the vehicle, and predicts the collision risk and the possible damage risk caused thereby in real time and accurately, including vehicle damage and personal damage, as well as the reliability index of the collision risk and the damage risk. According to the expected parameters and the reliability index corresponding to the expected parameters, control instructions can be provided for each component of the occupant protection system in the pre-collision stage, and the pre-collision system can be triggered in a graded manner. For low-level risks, no action is taken; for medium risks, warnings are given to the driver and the occupants, such as tactile, auditory, and visual reminders; for high-level risks, the pre-collision safety system is started in advance, such as zero-gravity seats, pre-collision airbags, etc., to extend the occupant protection time and reduce the energy absorbed by the restraint system per unit time. At the same time, the restraint system detonation strategy is adjusted to adapt to the current scene and minimize the damage risk. When the automatic driving enters a dangerous scene, the device can calculate the collision risk, vehicle damage risk and occupant damage risk under different planned paths to select the optimal path to reduce losses. And when a collision occurs, the damage information of the people and vehicles after the collision can be provided in time and sent to the cloud server through the wireless network, so as to notify the relevant parties, including maintenance, insurance, and rescue, implement precise rescue, and carry out subsequent maintenance and insurance processes in time.
[0035] 2. The present invention takes into account the uncertainty of driver operation, and the reliable quantitative indicators of the collision risk and damage risk prediction results, and comprehensively considers the robustness and change trend of the prediction results. It can be applied to the graded triggering of the pre-collision system, advances the time of occupant protection, and increases the user's perceptible interactive experience "before collision". At the same time, by adjusting the detonation strategy of the restraint system, the ignition control accuracy of the traditional restraint system can be improved.
[0036] 3. The present invention establishes a method that combines model design and data-driven: on the one hand, based on the information observed by the perception system, a physical model (vehicle kinematics / dynamics model) is established to predict the future vehicle trajectory; on the other hand, a pre-trained, offline machine / deep learning model is used to predict the risk of injury to personnel and vehicles, and to predict the collision risk caused by the uncertainty of the driver's operation and the reliability index of the risk of injury to personnel and vehicles corresponding to the hypothetical collision, thereby ensuring the prediction accuracy and greatly reducing the real-time calculation time.
[0037] 4. Both the personnel injury prediction module and the vehicle injury prediction module of the present invention can use a high-precision finite element simulation database as training / verification samples. The database has extremely high confidence damage labels, which reduces the noise of the data and the variance of the model. The collision waveform prediction model adopts a method of predicting waveform eigenvalues and then reconstructing the waveform from the waveform eigenvalues. A small amount of samples can be used to train the model to a better accuracy and realize the prediction of the collision waveform. The encoder-decoder architecture adopted by the personnel injury prediction model adopts a method of combining manual extraction and machine automatic extraction to extract features, so that the model integrates human prior knowledge of injuries, improves the model prediction accuracy, and reduces the sample size required for model training. The vehicle damage prediction model and the personnel injury prediction model can timely predict the accurate information of vehicle body damage and personnel injury, so that 4S stores and insurance companies can proactively provide customers with accurate after-sales and claims services, ensuring that injured people get the best rescue and medical services in the first place. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 This is a schematic diagram of the vehicle collision risk prediction device of the present invention.
[0039] Figure 2 Schematic diagram of trajectory prediction module.
[0040] Figure 3 This is the definition diagram of the rebound moment.
[0041] Figure 4 Schematic diagram of OLC calculation.
[0042] Figure 5 Schematic diagram of vehicle damage prediction module.
[0043] Figure 6 Schematic diagram of waveform eigenvalue calculation. DETAILED DESCRIPTION
[0044] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.
[0045] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0046] See also Figure 1 , the vehicle collision risk prediction device shown in the figure includes: a perception component S100, a basic component S200, a prediction component S300 and a decision component S400; the perception component is used to collect information about people, environment and vehicles as input for subsequent components. The perception component S100 includes a person perception module S110, an environment perception module S120 and a vehicle perception module S130.
[0047] The personnel sensing module S110 is used to collect personnel status parameters in real time. The personnel status parameters are obtained through a camera and / or a TOF camera, including basic passenger parameters and passenger position parameters. The basic passenger parameters include the height, weight, and age of the passenger. The passenger position parameters include the passenger's sitting posture and the passenger's sitting position. The passenger's sitting posture includes the passenger's recline angle and leaning angle. The passenger's sitting position refers to the passenger's absolute position relative to the vehicle body. The personnel status parameters are realized through sensors such as cameras and TOF cameras.
[0048] The environment perception module S120 is used to collect target object parameters in real time, and the target object parameters include the mass (m o )、Size(l o 、w o 、h o ), type and kinematic parameters, wherein the kinematic parameters of the target object include the velocity v of the target object o , acceleration a o , yaw angle θ o , yaw angular velocity ω o , where the mass of the target object is m oIt cannot be obtained directly, but can be estimated by the size and type of the target object. The remaining parameters can be collected by ADAS sensors such as lidar, camera, millimeter-wave radar, etc.
[0049] The vehicle perception module S130 is used to collect vehicle parameters and restraint system state parameters in real time. The vehicle parameters include the mass of the vehicle (m e )、Size(l e 、w e 、h e ) and kinematic parameters, where the mass and size do not need to be collected, and the pre-stored values are directly read. The kinematic parameters of the ego vehicle include the velocity v of the ego vehicle e , acceleration a e , yaw angle θ e , yaw angular velocity ω e The state parameters of the restraint system include whether the seat belt is fastened, whether the airbag is deployed, etc., which can be realized through switch sensors, collision sensors, etc.
[0050] The basic component S200 includes a digital human body module S210 and a vehicle prediction module S220. The digital human body module S210 is used to predict the position parameters of the personnel from the current moment to a certain moment in the future according to the scene parameters in which the personnel are located. For example, in a dangerous scene, when the driver takes braking or steering operations to avoid an accident, the driver and passengers will be out of position due to the longitudinal and / or lateral acceleration field, and the posture and sitting position of the driver and passengers will be out of position relative to the start time of braking and / or steering. The vehicle prediction module S220 can predict the posture and sitting position after the out-of-position occurs. The scene parameters refer to the vehicle parameters, target parameters and dynamic parameters. The vehicle parameters mainly include the vehicle kinematic parameters, the vehicle mass (m e )、Size(l e 、w e 、h e ), target object parameters refer to the kinematic parameters, target object type, mass and size of the target object, and dynamic parameters refer to the parameters calculated based on the kinematic parameters of the ego vehicle and the target object, such as collision time TTC (Time To Collision), headway THW (Time Headway), relative distance (dis_rel), relative speed (v_rel), etc. The TTC refers to the time when the ego vehicle collides with the target object when the ego vehicle is traveling at the current speed and acceleration and the target object's motion state remains unchanged; the headway THW refers to the ratio of the distance between the ego vehicle and the target object to the ego vehicle's speed.
[0051] See also Figure 2 The vehicle prediction module S220 includes a state parameter preprocessing unit S221, a trajectory prediction unit S222 and a collision detection unit S223.
[0052] The state parameter preprocessing unit S221 numbers and classifies the targets locked by the sensing component, calculates the collision time TTC of each target according to the target locked by the sensing component, and arranges the targets in ascending order according to the TTC value of each target; presets the first collision time threshold TTC thres1 and the second collision time threshold TTC thres2 , if the collision time TTC of a target object satisfies TTC thres1 ≤TTC≤TTC thres2 , define it as a potential dangerous target, classify it into the "potential dangerous target set", and continuously monitor its motion state; if the collision time TTC of a target satisfies TTC<TTC thres1 , it is defined as a dangerous target and classified into the “dangerous target set”;
[0053] The trajectory prediction unit S222 corrects the motion state of the target in the "potentially dangerous target set". The trajectory prediction unit uses a kinematic or dynamic model to predict the movement of the vehicle and the target in the "dangerous target set" within a preset time range (TTC). thres1 ). The predicted trajectory can predict the future vehicle's driving trajectory based on the scene parameters at the current moment through the kinematic / dynamic model. The motion state correction can be performed through the extended Kalman filter (EKF: Extended Kalman Filter) or other vehicle trajectory correction methods, with the aim of improving the accuracy of the perceived information within a certain range. The kinematic model is: constant turn rate and acceleration motion model (CTRA: Constant Turn Rate and Acceleration), constant acceleration model (CA: Constant Acceleration, if the yaw rate measurement is zero or there is no valid record) or constant velocity model (CV: Constant Velocity, if both the acceleration and yaw rate measurements are zero or there is no valid record). The kinematic model in this embodiment adopts CTRA. The kinematic model is adopted because of its high calculation rate, extreme time of dangerous scenes, and limited operating space for the driver. The kinematic model can obtain accuracy comparable to that of the dynamic model. It is worth noting that the trajectory prediction unit S222 can also be replaced by a dynamic model.
[0054] The collision detection unit S223 is used to detect whether the vehicle collides with the target object at a specific time. If the vehicle and the target object are in contact in space, it is defined as a "collision", and this time is defined as the collision time T c , collision risk CR=1, and record T c The collision scene parameters at the moment, the collision scene parameters include the vehicle speed v e_c、Target speed v o_c , collision overlap rate δ, collision angle γ, target mass m o and the type of the target object type; if no contact is detected, CR=0, CR∈{0,1}.
[0055] The detection of the collision detection unit includes a rough detection and a fine detection stage. The rough detection is to calculate the distance L between the geometric center of the vehicle and the geometric center of the target object. i , and calculate the minimum enveloping circles of the ego vehicle and the target object respectively. The centers of the enveloping circles are at their respective geometric centers. When the enveloping circles of the ego vehicle and the target object are tangent, the distance between the geometric centers of the ego vehicle and the target object is recorded as L crit .
[0056] When L i <αL crit When , fine detection is performed, and a is the gain coefficient. Fine detection is used to detect whether the vehicle and the target object overlap in space. If there is overlap, it is defined as contact. The specific method of fine detection is the Separating Axis Theorem (STA) or GJK (Gibert-Johnson-Keerthi) method.
[0057] The prediction component S300 includes a personnel damage prediction module S310, a vehicle damage prediction module S320 and a reliability prediction module S330.
[0058] The human injury risk model (HIRM) S310 is used to predict the injury curve of the occupant after a hypothetical collision in real time based on the collision waveform, occupant state parameters and restraint system state parameters when the vehicle collides, and calculate the occupant injury risk value HIR based on the injury curve. This module is an offline occupant injury prediction model trained based on the occupant injury dataset.
[0059] The construction of the personnel injury prediction module is completed online, including the following steps:
[0060] SH1 builds a personnel injury database. First, randomly sample the same space composed of collision waveform, driver and passenger information, and restraint system status information to generate a passenger injury matrix table. The collision waveform refers to the acceleration curves in the X, Y, and Z directions of the vehicle and the rotational displacement curves rotating around the X, Y, and Z axes in the vehicle coordinate system during the collision process. They are all time history curves and are obtained from the vehicle damage database. Then, a high-precision restraint system model is constructed. The high-precision restraint system model is a finite element model or a multi-rigid body model. Numerical simulation calculations are performed according to the passenger injury matrix table, and human body injury curves are extracted from the calculation results to form a human body injury database whose input (features) are collision waveforms, personnel state parameters, and restraint system state parameters, and output (labels) are injury curves of various parts of the human body. The injury curves of various parts of the human body refer to the acceleration curve of the head, the axial force curve of the neck, the bending moment curve of the neck, the injury index curve of the neck, the compression curve of the chest, the acceleration curve of the chest, the axial force curve of the femur, the axial force curve of the tibia, the shear force curve of the tibia, and the bending moment curve of the tibia around the x-axis and y-axis.
[0061] SH2 artificially extracted features. The artificially extracted features are based on human experience in studying injuries and calculate the parameters that affect human injuries based on the collision waveform, including but not limited to: vehicle collision speed v e_c , rebound time t rebound , collision energy The maximum acceleration a in the X, Y, and Z directions x_max 、a y_max 、a z_max , the average acceleration in the X, Y, and Z directions a x_mean 、a y_mean 、a z_mean , the maximum rotation displacement around the X, Y, and Z axes d x_max d y_max d z_max , sliding window average acceleration SM w , occupant load index OLC and acceleration enhancement index ASI. In specific implementation, if the acceleration curves in the Y and Z directions and the rotation displacement curves around the Y and Z axes cannot be obtained, the characteristic value may not include a y_max , a z_max 、a y_mean 、a z_mean d y_max d z_max The collision waveform refers to the acceleration curves in the X, Y, and Z directions of the vehicle and the rotational displacement curves around the X, Y, and Z axes in the vehicle coordinate system during the collision, both of which are time history curves.
[0062] The vehicle collision speed v e_c The collision time T cThe speed of the vehicle at time t can be obtained through the vehicle damage prediction module, and the collision speed curve v(t) is obtained by integrating the X-direction collision acceleration curve obtained by the vehicle damage prediction module.
[0063] The speed change Δv refers to v e_c The collision end time T end The vehicle speed v e_end Difference,
[0064] The rebound time t rebound Refers to the moment when the vehicle's speed reverses. Figure 3 As shown in , if the direction of the vehicle speed does not reverse after the collision, that is, the speed does not return to zero, the variance of the vehicle speed at 10 consecutive sampling points can be calculated. The speed curve samples one point every 1ms to obtain the speed variance curve. If the variance values of 10 consecutive points on the speed variance curve (the time interval corresponding to these 10 points is defined as [t, t+10]ms) are all less than the threshold value θ, then the rebound time t rebound =t+5ms, θ can be taken as 0.005.
[0065] The collision energy
[0066] The maximum value of the acceleration in the X direction, a x_max =max(abs(a x (t)), a y_max 、a z_max Can be solved similarly.
[0067] The average value of the acceleration in the X direction, a y_mean 、a z_mean Can be solved similarly.
[0068] The maximum rotation displacement around the X axis d x_max =max(abs(d x (t)), d y_max d z_max Can be solved similarly.
[0069] The sliding window average acceleration W represents the window width of the sliding window, which is generally 3ms.
[0070] The occupant load index OLC, the OLC index was proposed by Dr. Kübler, such as Figure 4As shown, in the initial stage of the collision, the restraint system has no restraint effect on the human body (free flight stage), and the free movement reaches time t1, and the movement distance is s. After that, the restraint system starts to work, moving at a constant deceleration ds. This stage is the restraint system working, and the value of this constant deceleration is the OLC value. Usually, d is the X-direction distance between the driver's chest and the steering wheel (dashboard), and s is 60mm. OLC is obtained by iterative solution according to the following formula.
[0071]
[0072]
[0073] v e_c -OLC·(t2-t1)=v(t2),
[0074] Among them, v e_c is the initial collision velocity, and v(t) is the collision velocity curve.
[0075] The acceleration intensification index ASI is calculated as follows:
[0076] Among them, SM 50ms,x , SM 50ms,y , SM 50ms,z Respectively represent the average acceleration values of the 50ms sliding window in the X, Y, and Z directions of the vehicle.
[0077] SH3 builds a personnel injury prediction model and completes training. The personnel injury prediction model is based on an encoder-decoder architecture. The encoder fuses the features extracted manually and the features automatically extracted by the machine. Specifically, the encoder can be implemented in two ways: the encoder consists of a scalar processing module, a vector processing module, and a scalar / vector fusion module. The vector processing module maps the original collision waveform to the vector hidden state, which can use 1D convolution + pooling or RNN units. The scalar processing module maps the scalar value to the scalar hidden state, which can be a structural unit such as a multi-layer perceptron (MLP). The fusion module splices the vector hidden state and the scalar hidden state, and completes the fusion through transposed convolution; the scalar input and the vector input can also be represented by an embedding layer, and the embedding layer outputs of all scalars and vectors are added, and then the features are extracted using a time series model such as TCN and RNN; the decoder is responsible for mapping the fusion features of the encoder to the injury value of the occupant, using an MLP structure. After the model is built, the model is initialized and trained until the accuracy meets the requirements.
[0078] SH4 real-time prediction, according to the real-time collision waveform, occupant state parameters and restraint system state parameters, calculates artificial feature parameters according to SH2, and inputs these artificial feature parameters into the pre-trained occupant injury prediction model to obtain the injury curve of each part of the occupant, and calculates the occupant injury risk value HIR according to the occupant injury risk function defined by NHTSA, HIR∈[0,1]:
[0079] HIR C =1-(1-P head )*(1-P neck )*(1-P chest )*(1-P femur )*(1-P tibia );
[0080] Among them, p head 、p neck , P chest , P femur , P tibia , representing the injury risk values of the head, neck, chest, thigh and calf respectively.
[0081] Taking the head as an example (the calculation of injury risk values for other parts is similar), calculate the injury risk value p head , the head injury index HIC is calculated according to the predicted head acceleration curve:
[0082]
[0083] Among them, a g (t) is the acceleration of the occupant's head center of mass at the time of collision, expressed in gravitational acceleration g; t0 and t T are respectively the start time and the end time of the collision; the collision time t1 and the collision time t2 are respectively the start time and the end time for calculating HIC, and the difference between t1 and t2 is less than a preset threshold, for example, the preset threshold may be 15ms, that is, t2-t1 is less than 15ms.
[0084] According to the AIS3 head injury risk curve: Here, Φ represents the cumulative normal distribution.
[0085] The damage index expresses the quantitative value of the damage of a certain part. A part can be represented by multiple damage indicators. When calculating its loss risk value, the maximum value of the damage risk value corresponding to each damage indicator is taken.
[0086] See also Figure 5The vehicle damage prediction module S320 includes a deformation prediction unit S321 and a collision waveform prediction unit S322. The deformation prediction unit S321 predicts the vehicle damage risk value after a hypothetical collision in real time according to the collision scene parameters at the collision moment. The collision waveform prediction unit S322 predicts the collision waveform after a hypothetical collision according to the collision scene parameters. The predicted collision waveform serves as one of the inputs of the personnel injury prediction module.
[0087] The deformation amount prediction unit S321 calculates the deformation amount according to T c The collision scene parameters at the moment of the simulation are used to predict the damage risk value VIR of the vehicle after a hypothetical collision in real time. This is an offline deformation prediction model trained based on the vehicle damage dataset.
[0088] The construction of the deformation prediction model is completed online, which mainly consists of three steps: SD1 to build a vehicle damage database, SD2 to build a deformation prediction model and complete training, and SD3 to make real-time predictions.
[0089] SD1 builds a vehicle damage database. First, determine the scene parameters that affect the collision waveform and the boundaries of these scene parameters. The boundaries of the scene parameters can be determined by analyzing the traffic accident database, and generate a vehicle damage matrix table based on random sampling of the same space composed of the scene parameters; then, build a high-precision finite element vehicle model, update the vehicle model according to the vehicle damage matrix table and perform numerical simulation calculations; finally, extract the vehicle deformation and collision waveform from the simulation results, and calculate the eigenvalue F1 based on the collision waveform true 、F2 true , the input (features) are scene parameters, and the output (labels) are a database of vehicle deformation and waveform feature values.
[0090] The waveform eigenvalue refers to simplifying the waveform into an approximate waveform. The eigenvalue is the only parameter that determines the simplified waveform. The method of simplifying the waveform can use a half-sine wave, a second-order square wave, a trapezoidal wave, a Fourier waveform, etc. A specific example is given below. Figure 6 :
[0091] It is mainly divided into two parts: one is to extract the eigenvalues of the real acceleration curve in the X, Y, and Z directions, and the other is to extract the eigenvalues of the real rotation displacement curve around the X, Y, and Z axes.
[0092] 1. Extraction of eigenvalues of acceleration curve: The acceleration curve has many small peaks and valleys. According to Fourier theory, any waveform can be obtained by superposition of multiple sine waves, and these multiple sine waves are unique. In order to better fit the trend of the acceleration curve, the Fourier equivalent method is used. Taking the X direction as an example (the extraction of eigenvalues in the Y and Z directions is similar to that in the X direction), the real acceleration curve A(t) is obtained from the finite element calculation results, and the adjacent points (tj-1 ,A(t) t=j-1 ) and (t j ,A(t) t=j )’s slope m j ,
[0093]
[0094] Calculate the effective duration of the collision t k , t k =t A(t)=0 , that is, the time when the acceleration curve returns to zero after time 0.
[0095] By m j and t k Calculate the i-th eigenvalue a i ,
[0096]
[0097] Where k is the total number of discrete points contained in the original waveform curve, and the natural angular frequency is calculated
[0098] Finally, reconstruct the acceleration curve and convert a i Substituting ω into:
[0099] The characteristic values of the acceleration waveforms in the X, Y, and Z directions are recorded as:
[0100] F1 true =[a x1 , a x2 , ..., a xn , a y1 , a y2 , ..., a yn , a z1 , a z2 , ..., a zn ].
[0101] 2. Extraction of characteristic values of the rotational displacement curve: During the collision process, the rotational displacement curve is either monotonically increasing or has only a single peak. There is no such thing as an acceleration waveform with many peaks. This is because the effective collision time is usually very short (about 150ms), and in such a short time, the vehicle usually moves in one direction around a certain axis. Therefore, for the rotational displacement curve, directly taking the points on the original curve can well characterize the original curve. Taking the X direction as an example (the extraction of characteristic values in the Y and Z directions is similar to that in the X direction), D(t) is the true rotational displacement curve in the X direction, and its characteristic values are directly taken from the points on the original curve. Take n points to characterize the original curve, and the first and last points are taken from t0 and t1 of the original curve, respectively. kThe rotation displacement value corresponding to the moment, t0 = 0, and the remaining n-2 points are the corresponding points on the original curve at equal time intervals. The eigenvalue is r x1 , r x2 , ..., r xn Similarly, the eigenvalues r in the Y and Z directions can be obtained respectively y1 , r y2 , ..., r yn and r z1 , r z2 , ..., r zn , the vector formed by these eigenvalues [r x1 , r x2 , ..., r xn , r y1 , r y2 , ..., r yn , r z1 , r z2 , ..., r zn ] is recorded as F2 true .
[0102] The characteristic value of the rotation displacement curve in the X, Y, and Z directions is recorded as: F2 true =[r x1 , r x2 , ..., r xn , r y1 , r y2 , ..., r yn , r z1 , r z2 , ..., r zn ].
[0103] SD2 builds a deformation prediction model and completes the training. The model is trained by taking the scene parameters as input (features) and the vehicle deformation as output (labels). The model can use machine learning integrated algorithms such as random forest and XGBOOST, or neural network algorithms such as multi-layer perceptron, CNN, and RNN.
[0104] SD3 real-time prediction. After obtaining the scene parameters in real time, they are input into the offline pre-trained deformation prediction model to obtain the predicted deformation and calculate the relative deformation, that is, the ratio of the body deformation to the vehicle size. The vehicle deformation is decomposed into longitudinal, lateral and vertical. The relative deformation takes the maximum value of the lateral, longitudinal and vertical relative deformations, VIR∈[0,1].
[0105] The collision waveform prediction unit S322 calculates the collision waveform according to the collision time T c The collision scene parameters are used to predict the collision waveform of the vehicle after a hypothetical collision in real time. This is an offline waveform prediction model trained based on the vehicle damage dataset.
[0106] The construction of the waveform prediction model is completed online, and mainly consists of three steps: SP1 building a vehicle damage database, SP2 building a waveform prediction model and completing training, and SP3 real-time prediction.
[0107] SP1 builds the vehicle damage database, which has been completed in SD1.
[0108] SP2 builds a waveform prediction model and completes training. The model is trained by training samples whose input (feature value) is the scene parameter and whose output (label) is the waveform feature value. The model can be the method described in SD2.
[0109] SP3 real-time prediction. The predicted eigenvalue F1 is obtained based on the collision scene parameter prediction. pre and F2 pre , the predicted collision waveform is obtained based on the predicted feature value. Taking the acceleration curve in the X direction as an example (similar to the Y and Z directions), take F1 pre The first n eigenvalues, according to Get the acceleration curve in the X direction, according to F2 pre The first n eigenvalues constitute the ordinate of the rotation displacement curve around the X-axis, and the corresponding abscissa is Constitute the rotation displacement curve in the X direction.
[0110] The reliability prediction module S330 predicts the reliability indicators ∑CR, ∑HIR, ∑VIR of the expected parameters CR, HIR, and VIR based on the scenario parameters and the expected parameters when a hypothetical collision occurs, wherein the expected parameters include the collision risk CR, the occupant injury risk value HIR, and the vehicle damage risk value VIR; the decision component obtains the expected parameters CR, HIR, and VIR and their corresponding reliability indicators ∑CR, ∑HIR, and ∑VIR, while taking into account the extreme values and change trend indicators of the predicted signals in the time domain, performs a comprehensive collision risk assessment, and outputs control instructions.
[0111] The reliability prediction module S330 is an offline reliability prediction model obtained by training a reliability quantified index database, wherein the reliability quantified index database refers to a database whose input is scenario parameters and expected parameters and whose output is reliability indexes in the current scenario.
[0112] The reliability index is the credibility of CR, HIR and VIR after considering the driver's reaction to the current vehicle operation state, that is, considering the uncertainty of the driver's operation at the current moment. The trajectory prediction module does not take into account the uncertainty of the driver's operation. According to the statistical analysis of traffic accidents, in dangerous scenes, most drivers are not able to make effective evasive operations in a very short time. Usually, there is no operation before the collision. The vehicle's driving trajectory only evolves from the initial state in time. This trajectory is defined as the presumptive true-path, that is, the vehicle's driving trajectory predicted by the trajectory prediction module is the most likely driving trajectory, so T c The collision risk value CR, personal injury risk value HIR and vehicle damage risk value VIR predicted at each moment are the expected risk values, so CR, HIR and VIR are described as expected parameters. ∑CR, ∑HIR and ∑VIR respectively characterize the degree of deviation of different operating behaviors from the expected values CR, HIR and VIR, that is, reliability indicators.
[0113] The construction of the reliability prediction model is completed online, mainly through three steps: SR1 builds a reliability quantification indicator database, SR2 builds a reliability prediction model and completes training, and SR3 performs real-time prediction.
[0114] SR1 builds a reliability quantification index database. The reliability quantification index database is a database whose input (features) are scene parameters and expected parameters (CR, HIR and VIR), and whose output (labels) are reliability indicators (∑CR, ∑HIR and ∑VIR) in the current scene.
[0115] Construction method: 1. Establish a driver operation behavior database. For possible collision scenarios, simplify the driver operation behavior (braking, acceleration and steering) into the longitudinal and lateral movements of the vehicle, which are represented by longitudinal acceleration a and yaw angular velocity ω respectively. Numerically discretize a and ω under the premise of physical conditions, and establish a driving behavior database with input as scenario parameters and output as a and ω; 2. Establish a quantitative representation of driver operation uncertainty based on the driver operation behavior database; 3. For each group of specific scenario parameters, on the basis of the assumed development trajectory, for each discrete moment (the prediction time range T is discretized into N values, T = [T0, T1, ..., T N-1 ], T N-1 =TTC thres1 ), all consider the random state caused by the driver's operation behavior, and generate a large number of possible vehicle trajectories. According to the vehicle prediction model, the personnel injury prediction model and the vehicle damage prediction model, the collision risk value CR of each driving trajectory and the k-th dangerous target is obtained respectively. k_r , Personnel Injury Risk Value HIR k_r and vehicle damage risk value VIRk_r ; 4. Calculate the original collision risk value (CR k_r , HIR k_r , VIR k_r ) corresponds to the reliability quantification index (∑CR k ,∑HIR k ,∑VIR k ), similarly, the reliability quantitative indexes under all scenario parameters can be calculated, thus completing the construction of the driver operation uncertainty database.
[0116] The reliability quantification index is calculated as follows: according to the scene parameters and expected parameters at the current moment, after considering the driver's operation uncertainty (assuming that the driver's operation uncertainty has U discrete values corresponding to U different driving trajectories), according to the vehicle prediction module, the personnel injury prediction module and the vehicle damage prediction module, U different vehicle collision risk prediction results, U personnel injury risk prediction results and U vehicle damage risk prediction results can be obtained respectively, which are recorded as:
[0117] [(CR k_1 , HIR k_1 , VIR k_1 ), (CR k_2 , HIR k_2 , VIR k_2 ),...,(CR k_U , HIR k_U , VIR k_U )], then in statistical terms, the reliability quantification index for the kth target object can be expressed as:
[0118]
[0119] When ∑CR is smaller, CR is more accurate and more likely to appear. Otherwise, it means that when the driver intervenes, it is more unstable and less likely to appear. ∑HIR and ∑VIR have the same interpretability, that is, when ∑HIR is smaller, HIR is more accurate and more likely to appear. Otherwise, it means that when the driver intervenes, it is more unstable and less likely to appear. When ∑VIR is smaller, VIR is more accurate and more likely to appear. Otherwise, it means that when the driver intervenes, it is more unstable and less likely to appear.
[0120] The quantitative representation of the driver's operation uncertainty (taking the vehicle as an example, the representation of the target object is similar): First, the acceleration a and yaw rate ω in the driver's operation behavior database are numerically discretized under the premise that physical conditions allow. Assume that a has discrete values, ω has η discrete values, namely ω=[ω1,ω1,...,ω η]; Then, the driver operation behavior prediction posterior model can be established and different trajectories can be generated in the following three ways:
[0121] A: Assume that the acceleration a and the yaw rate ω obey uniform distribution That is, the probability of the driver taking a certain action at any time is equal, regardless of the scene parameters. In the process of generating trajectories, at each discrete moment, according to and Probabilistic models are available Different operation behaviors can eventually obtain Different trajectory curves.
[0122] B: Assuming that acceleration a and yaw rate ω are related to scene parameters (S), establish a driver operation posterior model of acceleration a, yaw rate ω and scene parameters: a, ω = f(S). The f can be a time series model such as a hidden Markov model (HMM), RNN, etc.
[0123] C: Establish a normal distribution probability model for a and ω:
[0124]
[0125]
[0126] Assuming that the acceleration a and the yaw angular velocity ω are independent of each other, a fused normal distribution probability model can be established:
[0127]
[0128] Will Substituting into the kinematic model, we can obtain [T i , T i+1 ] The average driving trajectory of the driver’s operation during the time period is obtained, and T i+1 The probability distribution of the vehicle state at time T (characterizing the changes caused by different driver operations) i+1 The probability distribution of the vehicle state at the time is then randomly scattered within the vehicle state distribution range based on the “6σ principle”. Each point represents T i+1 Similarly, at each discrete moment, the probability distribution of the vehicle state at the current moment is calculated and randomly scattered until T N If ξ points are scattered at each discrete moment, the maximum number of points that can be obtained is ξ N Tracks.
[0129] Method A does not consider the impact of different driving conditions on the driver's operation, ignores the change of time, and simplifies the driver's operation probability model; Method B, based on Method A, considers the impact of driving conditions on the driver's operation, uses environmental parameters as hidden state variables, and establishes a posterior model of the driver's operation; and Method C, based on Method B, establishes a probability model for the driver's operation and uses the parameterized form of the probability model to perform trajectory recursive calculation. Methods A and B must recursively calculate the state of each trajectory sample at each discrete moment, and their computational complexity can be simply expressed as The computational complexity of method C can be expressed as O(ξ N ).
[0130] SR2 builds a reliability prediction model and completes the training. The model is trained by training samples whose input is scene parameters and expected parameters (CR, HIR and VIR) and whose output (label) is reliability indicators (∑CR, ∑HIR and ∑VIR). The model can be the method described in SD2.
[0131] SR3 real-time prediction. The scene parameters obtained by the perception component and the expected parameters (CR, HIR and VIR) predicted by the vehicle prediction module, the personnel damage prediction module and the vehicle damage prediction module are input into the pre-trained offline reliability prediction model to obtain the predicted reliability indicators (∑CR, ∑HIR and ∑VIR).
[0132] The decision-making component S400 outputs control instructions through fuzzy control, empirical calibration or logical judgment, that is, on the basis of the above-mentioned safety indicators, it takes into account the extreme values of the predicted signal in the time domain, the change trend and other indicators, and comprehensively performs collision risk assessment. Therefore, according to empirical rules and test calibration, the fuzzy form of the evaluation signal that can be output includes but is not limited to a series of grade evaluations such as "safety", "medium probability collision, medium damage", "high probability collision, serious damage", etc., which constitute an evaluation fuzzy set. Thus, a fuzzy logic control process is established, and finally a descriptive signal about the collision risk assessment is output. At the same time, the output should also include a description of the corresponding locked target position. Take the prediction result (CR k , HIR k , VIR k ), (∑CR k ,∑HIR k ,∑VIR k ), the control method is explained by an example of logical judgment:
[0133] First, HIR k and VIR k and its reliability index ∑HIR k and∑VIR kWeighted to comprehensive loss risk index IR k :
[0134] IR k =w h *(HIR k *∑HIR k )+w v *(VIR k *∑VIR k ), w h 、w v are the weights of occupant damage and vehicle damage, respectively, h 、w v ∈[0, 1], w h +w v =1. h 、w v The specific value of can be determined based on experience. If more attention is paid to personnel damage, w h A larger weight value; if you pay more attention to vehicle damage, give w v The former has a greater weight than the latter, and is used when there are drivers and passengers in the car, while the latter can be used for autonomous driving vehicles without drivers and passengers in the car to reduce vehicle damage.
[0135] Then, considering the robustness in the time domain, monitoring CR k M consecutive prediction signal values, (CR k,1 , CR k,2 , ..., CR k,M ), and calculate its variance
[0136] Pre-set thresholds Set Threshold IR thres1 IR thres2 , when IR>IR thres2 , define IR level as 0, when IR thres1 ≤IR≤IR thres2 , define IR level as 1, when IR<IR thres1 , define IR level as 2; set threshold ∑CR thres1 ,∑CR thres2 , when ∑CR<∑CR thres1 , define ∑CR level as 0, when ∑CR thres1 ≤∑CR≤∑CR thres2 , define ∑CR level as 1, when ∑CR>IR thres2 , define the ∑CR level as 2.
[0137] when Enter the following logical judgment (CR, IR and ∑CR refer to the current moment, that is, moment M):
[0138] When CR=1: IR level is 0, no matter what the value of ∑CR is, the output action instruction is 0; when the IR level is 1 and the ∑CR level is 0, the output action instruction is 1; when the IR level is 1 and the ∑CR level is 1, the output action instruction is 2; when the IR level is 1 and the ∑CR level is 2, the output action instruction is 3; when the IR level is 2 and the ∑CR level is 0, the output action instruction is 4; when the IR level is 2 and the ∑CR level is 1, the output action instruction is 5; when the IR level is 2 and the ∑CR level is 2, the output action instruction is 6.
[0139] When CR=0: IR level is 0, no matter what the value of ∑CR is, the output action instruction is 0; in other cases, the input action instruction is reduced by 1 level compared with the case of CR=1. That is, when the IR level is 1 and the ∑CR level is 0, the output action instruction is 0; when the IR level is 1 and the ∑CR level is 1, the output action instruction is 1; when the IR level is 1 and the ∑CR level is 2, the output action instruction is 2; when the IR level is 2 and the ∑CR level is 0, the output action instruction is 3; when the IR level is 2 and the ∑CR level is 1, the output action instruction is 4; when the IR level is 2 and the ∑CR level is 2, the output action instruction is 5.
[0140] The corresponding action instructions are: 0 means no action, 1 means level 1 reminder, such as visual reminder, 2 means level 2 reminder, such as sound reminder, 3 means tactile reminder, such as vibrating seat belts, 4 means level 1 adjustment for pre-collision, such as tightening the safety belt with a lower force, 5 means level 2 adjustment for pre-collision, such as tightening the safety belt with a medium force, and adjusting the seat position to the standard position (such as 0 gravity seat return), 6 means level 3 adjustment for pre-collision, such as tightening the safety belt with a higher force, activating the pre-collision airbag, adjusting the algorithm of the airbag controller, and adjusting the threshold limit of the airbag detonation to detonate the airbag in advance.
[0141] The damage risk value and its reliability index can also be used for path selection in an autonomous driving scenario, that is, calculating the damage risk index IR under each path and selecting the path with the lowest damage risk.
[0142] The damage risk value and its reliability index can also be used to calculate the driver's driving style and judge the driving risk:
[0143]
[0144] Among them, t1 and t2 are the start time and end time of driving risk assessment, which can be defined according to actual conditions. risk The higher it is, the more aggressive the driving is, and vice versa, the more conservative it is.
[0145] The occupant injury prediction module can also be used for accident rescue. Specifically, after a collision occurs, the collision waveform recorded by the airbag controller, and the state parameters of the occupant and the state parameters of the restraint system at time T0 are input into the occupant injury prediction model to obtain a predicted occupant injury curve, and the occupant injury risk value or occupant injury level is calculated based on the injury curve.
[0146] The injury level is represented by AIS, and its calculation method refers to NEHTSA's definition of injury level. Taking the head injury level as an example: first calculate the HIC value, and then convert the HIC value into the risk probability corresponding to different injury levels according to the following formula. When a certain injury level is greater than a threshold (such as 25%) for the first time, an injury of that level has occurred.
[0147]
[0148]
[0149]
[0150]
[0151]
[0152]
[0153] Assuming HIC = 1000, substituting HIC into the above formula, we can obtain the risk values of head injuries of levels 1, 2, 3, 4, 5, and 6, which are 99.15%, 89.19%, 52.75%, 16.80%, 2.37%, and 0.11%, respectively, which means that the head suffers a level 3 injury.
[0154] A vehicle collision risk prediction method, using the vehicle collision risk prediction device of the present invention to predict vehicle collision risk, comprises the following steps:
[0155] Step 1: input the information of personnel, environment and vehicle collected by the perception component S100 into the basic component S200;
[0156] Step 2: The vehicle prediction module S220 of the basic component S200 outputs the collision risk CR value and the collision scene parameters based on the environment and vehicle information, and the digital human body module S210 of the basic component S200 obtains the personnel status parameters at the collision moment based on the personnel information and the vehicle information;
[0157] Step 3: The vehicle damage prediction module S320 of the prediction component S300 outputs the vehicle damage risk value VIR and the collision waveform according to the collision scene parameters, and inputs the collision waveform, the occupant state parameters at the time of collision, and the restraint system state parameters input by the perception component into the occupant damage prediction module to obtain the occupant injury risk value HIR;
[0158] Step 4: Input the collision scene parameters and expected parameters CR, HIR, VIR to the reliability prediction module S330 to obtain corresponding reliability indicators ∑CR, ∑HIR, ∑VIR;
[0159] Step 5: The decision component S400 performs a comprehensive collision risk assessment based on the acquired expected parameters CR, HIR, VIR and their corresponding reliability indicators ∑CR, ∑HIR, ∑VIR, taking into account the extreme values and change trend indicators of the predicted signal in the time domain, and outputs a control instruction;
[0160] Step six, based on the collision waveform recorded by the airbag controller, the detailed damage to the vehicle and personnel is obtained through the vehicle damage prediction module S320 and the personnel injury prediction module S310, and sent to the cloud server through the wireless network, and then notified to the relevant parties, including maintenance, insurance, and rescue.
[0161] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or change made by a person skilled in the art based on the present invention is within the protection scope of the present invention.
Claims
1. A vehicle collision risk prediction device, characterized in that: include: Perception component, basic component, prediction component and decision component; The perception component is used to collect information about people, environment and vehicles; The basic components include a digital human body module and a vehicle prediction module. The digital human body module is used to predict the position parameters of the personnel from the current moment to a certain moment in the future according to the scene parameters in which the personnel are located. The vehicle prediction module includes a state parameter preprocessing unit, a trajectory prediction unit and a collision detection unit. The state parameter preprocessing unit numbers and classifies the target objects locked by the perception component. The trajectory prediction unit corrects the motion state of the locked target objects or uses a kinematic or dynamic model to predict the driving trajectory of the vehicle and / or the target object within a preset time range. The collision detection unit is used to detect whether the vehicle and the target object collide at a specific moment. The state parameter preprocessing unit numbers and classifies the target objects according to the target objects locked by the sensing component. Specifically, the sensing component locks the target objects and numbers the target objects, calculates the collision time TTC of each target object, and arranges the targets in ascending order according to the TTC value of each target object; presets a first collision time threshold TTC thres1 and the second collision time threshold TTC thres2 , if the collision time TTC of a target object satisfies TTC thres1 ≤TTC≤TTC thres2 , define it as a potential dangerous target, classify it into the "potential dangerous target set", and continuously monitor its movement status; If the collision time TTC of a target object satisfies TTC<TTC thres1 , it is defined as a dangerous target and classified into the "dangerous target set"; The trajectory prediction unit corrects the motion state of the target in the "potentially dangerous target set", and the trajectory prediction unit uses a kinematic or dynamic model to predict the driving trajectory of the vehicle and the target in the "dangerous target set" within a preset time range; The collision detection unit is used to detect whether the vehicle collides with the target object at a specific time. If the vehicle and the target object come into contact in space, it is defined as a "collision", and this time is defined as the collision time T c , collision risk CR=1, and record T c The collision scenario parameters at the time, wherein the collision scenario parameters include the vehicle speed, the target speed, the collision overlap rate, the collision angle, the target mass and the target type; if no contact is detected, the collision risk CR=0; The detection of the collision detection unit includes a rough detection and a fine detection stage. The rough detection is to calculate the distance L between the geometric center of the vehicle and the geometric center of the target object. i , and calculate the minimum enveloping circles of the ego vehicle and the target object respectively. The centers of the enveloping circles are at their respective geometric centers. When the enveloping circles of the ego vehicle and the target object are tangent, the distance between the geometric centers of the ego vehicle and the target object is recorded as L crit , when L i <αL crit When , fine detection is performed, and α is the gain coefficient. Fine detection is used to detect whether the vehicle and the target object overlap in space. If there is overlap, it is defined as contact. The specific method of fine detection is the separation axis STA method or the GJK method. The prediction component includes a personnel injury prediction module, a vehicle injury prediction module and a reliability prediction module. The personnel injury prediction module is used to predict the injury curve of the occupant after a hypothetical collision in real time according to the collision waveform, personnel state parameters and restraint system state parameters when the vehicle collides, and calculate the occupant injury risk value according to the injury curve; the vehicle damage prediction module includes a deformation prediction unit and a collision waveform prediction unit. The deformation prediction unit predicts the vehicle damage risk value after a hypothetical collision in real time according to the collision scene parameters at the collision moment. The collision waveform prediction unit predicts the collision waveform after a hypothetical collision according to the collision scene parameters, and the predicted collision waveform is used as one of the inputs of the personnel injury prediction module; the reliability prediction module predicts the reliability indicators ∑CR, ∑HIR, ∑VIR of the expected parameters CR, HIR, and VIR according to the scene parameters and the expected parameters when a hypothetical collision occurs, the expected parameters including the collision risk CR, the occupant injury risk value HIR, and the vehicle injury risk value VIR; The calculation method of the reliability prediction module is as follows: according to the scene parameters and expected parameters at the current moment, after considering the driver's operation uncertainty, the driver's operation uncertainty is set to have U discrete values, corresponding to U different driving trajectories, and according to the vehicle prediction module, the personnel injury prediction module, and the vehicle damage prediction module, U different vehicle collision risk prediction results, U types of personnel injury risk prediction results, and U types of vehicle damage risk prediction results can be obtained respectively, which are recorded as: [(CR k_1 ,HIR k_1 ,VIR k_1 ),(CR k_2 ,HIR k_2 ,VIR k_2 ),…,(CR k_U ,HIR k_U ,VIR k_U )], The reliability quantitative index of the kth target is expressed as: When ∑CR is smaller, CR is more accurate and more likely to occur. Conversely, it means that the driver's operation is more unstable and less likely to occur when the driver intervenes. When ∑HIR is smaller, the HIR is more accurate and more likely to occur. Conversely, it means that the driver's operation intervention is more unstable and less likely to occur. When ∑VIR is smaller, VIR is more accurate and more likely to occur. Conversely, it means that the driver's operation intervention is more unstable and less likely to occur. The driver operation uncertainty is quantitatively characterized as follows: establishing a driver operation behavior database, simplifying the driver operation behavior (braking, acceleration and steering) into the longitudinal and lateral movements of the vehicle for possible collision scenarios, and representing them with longitudinal acceleration a and yaw angular velocity ω respectively, and numerically discretizing a and ω under the premise that physical conditions permit, and establishing a driving behavior database with scene parameters as input and a and ω as output; numerically discretizing the acceleration a and yaw angular velocity ω in the driver operation behavior database under the premise that physical conditions permit, setting the acceleration a to have ζ discrete values and the yaw angular velocity ω to have η discrete values, that is, a=[a1, a2, …, a ζ ],ω=[ω1,ω1,…,ω η ], establish a posterior model for predicting driver operation behavior and generate different trajectories; The decision component obtains the expected parameters CR, HIR, VIR and their corresponding reliability indicators ∑CR, ∑HIR, ∑VIR, while taking into account the extreme values and change trend indicators of the predicted signal in the time domain, performs a comprehensive collision risk assessment, and outputs control instructions.
2. The vehicle collision risk prediction device according to claim 1, characterized in that: The perception component includes a personnel perception module, an environment perception module and a vehicle perception module; The personnel sensing module is used to collect personnel status parameters in real time. The personnel status parameters are obtained through a camera and / or a TOF camera, including basic parameters of the occupant and occupant position parameters. The basic parameters of the occupant include the height, weight and age of the occupant. The occupant position parameters include the sitting posture of the occupant and the sitting position of the occupant. The sitting posture of the occupant includes the forward tilt angle and the side tilt angle of the occupant. The sitting position of the occupant refers to the absolute position of the occupant relative to the vehicle body. The environment perception module is used to collect target object parameters in real time, wherein the target object parameters include the mass, size, type and kinematic parameters of the target object, and the kinematic parameters of the target object include the speed, acceleration, yaw angle and yaw angular velocity of the target object; The vehicle perception module is used to collect vehicle parameters and constraint system state parameters in real time. The vehicle parameters include the mass, size and kinematic parameters of the vehicle. The kinematic parameters of the vehicle include the speed, acceleration, yaw angle and yaw angular velocity of the vehicle.
3. The vehicle collision risk prediction device according to claim 1, characterized in that: The motion state correction performs trajectory correction by extended Kalman filtering, and the dynamic model is a constant angular velocity and constant acceleration motion model, a constant acceleration model or a constant speed model.
4. The vehicle collision risk prediction device according to claim 1 or 2, characterized in that: The occupant injury prediction module is an offline occupant injury prediction model obtained by training based on a occupant injury dataset, wherein the occupant injury dataset refers to a database whose input is a collision waveform, occupant state parameters, and restraint system state parameters, and whose output is an occupant injury curve, which can be obtained through restraint system numerical calculation or collision test; The deformation prediction unit is an offline deformation prediction model obtained by training based on the vehicle damage data set. The collision waveform prediction unit is an offline collision waveform prediction model obtained by training based on the vehicle damage data set. The vehicle damage data set refers to the input of the collision time T c The collision scene parameters are output as a database of vehicle deformation and waveform characteristic values for the collision waveform prediction unit, which are obtained through whole vehicle finite element numerical calculation, whole vehicle collision test or collision accident collection; The reliability prediction module is an offline reliability prediction model obtained by training a reliability quantified index database, wherein the reliability quantified index database refers to a database whose input is scenario parameters and expected parameters and whose output is reliability indexes in the current scenario.
5. The vehicle collision risk prediction device according to claim 4, characterized in that: The deformation prediction model, collision waveform prediction model and reliability prediction model adopt the machine learning integration algorithm of random forest and XGBOOST, or adopt the neural network algorithm of multi-layer perceptron, CNN and RNN.
6. The vehicle collision risk prediction device according to claim 1 or 2, characterized in that: The decision-making component outputs control instructions through fuzzy control, empirical calibration or logical judgment.
7. A vehicle collision risk prediction method, characterized in that: The vehicle collision risk prediction device according to any one of claims 1 to 6 is used to predict the vehicle collision risk, comprising the following steps: Step 1: Input the information of people, environment and vehicles collected by the perception component into the basic component; Step 2: The vehicle prediction module of the basic component outputs the collision risk CR value and collision scenario parameters based on the environment and vehicle information, and the digital human body module of the basic component obtains the personnel status parameters at the time of collision based on the personnel information and vehicle information; Step 3: The vehicle damage prediction module of the prediction component outputs the vehicle damage risk value VIR and the collision waveform according to the collision scene parameters, and inputs the collision waveform, the occupant state parameters at the time of collision, and the restraint system state parameters input by the perception component into the occupant damage prediction module to obtain the occupant injury risk value HIR; Step 4: Input the collision scene parameters and expected parameters CR, HIR, and VIR into the reliability prediction module to obtain corresponding reliability indicators ∑CR, ∑HIR, and ∑VIR; Step 5: The decision component performs a comprehensive collision risk assessment based on the expected parameters CR, HIR, VIR and their corresponding reliability indicators ∑CR, ∑HIR, ∑VIR, and takes into account the extreme values and change trend indicators of the predicted signal in the time domain, and outputs a control instruction; Step six, according to the collision waveform recorded by the airbag controller, the detailed damage of the vehicle and personnel is obtained through the vehicle damage prediction module and the personnel damage prediction module, and sent to the cloud server through the wireless network, and then notified to the relevant parties, including maintenance, insurance, and rescue.
