Automatic driving risk quantification method based on conflict risk field

Abstract

The invention relates to an automatic driving risk quantification method based on a conflict risk field. Comprising the steps of 1, constructing a basic risk field; step 2, under a basic risk field framework, constructing a conflict risk field by taking an ADV as a center; step 3, regarding the conflict risk field as a repulsive force acting on the ADV, and using the repulsive force to quantify the dynamic influence of the environmental elements and various TP states on the driving safety of the ADV; the method comprises the following steps: firstly, constructing a basic risk field on the basis of traffic vehicle distribution and motion characteristics, further forming a vehicle forward, lateral, backward and traffic regulation constraint multi-dimensional conflict risk field on the basis, and converting an abstract traffic conflict relationship into a computable repulsive force model; dynamic quantitative characterization of driving risks in a state dimension, a space dimension and a time dimension is realized, and continuous optimization and evolution of an end-to-end automatic driving algorithm are supported by means of a repulsive force model constructed based on a conflict risk field.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically a method for quantifying autonomous driving risks based on conflict risk fields. Background Technology

[0002] With the rapid development of autonomous driving technology, end-to-end decision-making and planning algorithms are gradually becoming an important technical route for intelligent driving systems. These algorithms rely on deep neural networks with "black box" characteristics, directly mapping environmental information perceived by sensors into control commands such as steering, acceleration, and deceleration, structurally weakening the traditional rule base and explicit planning modules. However, the "encapsulation" of the control link by the end-to-end network does not mean that the explicit characterization and identification of driving risks can be ignored. On the one hand, the internal representation of the end-to-end network is difficult to interpret, and once the output is abnormal, it is difficult to determine in a timely manner whether it is near a safety boundary. On the other hand, the actual road traffic environment is highly dynamic, and local risks often evolve rapidly in a very short time. Without online assessment of the current driving risk situation, the system will find it difficult to identify in a timely manner in which specific working conditions the existing end-to-end decision-making strategy has approached or exceeded its applicable boundaries, and it will also be unable to selectively filter out key high-value samples for optimizing the strategy from massive amounts of driving data. Whether the end-to-end algorithm can be continuously optimized largely depends on whether the data used for training and updating has sufficient information gain to drive actual improvements in algorithm performance. Currently, major automakers have generally established data processing centers to receive massive amounts of driving data from autonomous vehicles via the cloud. However, among this vast and diverse natural driving data, how to automatically identify high-value data segments that pose potential driving risks to autonomous vehicles (ADVs) and can thus be used to significantly improve the performance of end-to-end decision-making and planning algorithms remains a critical problem that urgently needs to be solved. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a method for quantifying autonomous driving risks based on conflict risk fields. First, a basic risk field based on traffic vehicles is constructed. Then, a multi-dimensional conflict risk field is further formed, encompassing forward, lateral, and backward traffic, as well as regulatory constraints. This systematically quantifies the potential risks during the interaction between autonomous vehicles and the traffic environment, transforming abstract traffic conflicts into a computable "repulsion force" model. This achieves dynamic, multi-dimensional (state, space, time) risk quantification of driving risks, enabling the use of high-value data fragments to significantly improve the performance of end-to-end decision-making and planning algorithms, thus facilitating the continuous development of end-to-end algorithms.

[0004] The technical solution of this invention is described below in conjunction with the accompanying drawings:

[0005] This invention provides a method for quantifying autonomous driving risks based on conflict risk fields, comprising the following steps:

[0006] Step 1: Construct the basic risk field;

[0007] The underlying risk field includes:

[0008] The state risk potential field of traffic participants (TP) around ADV, hereinafter referred to as TP state risk potential field, is used to characterize the spatial distribution of risks that other traffic participants around ADV may form under the current vehicle state (speed, position, heading angle). This potential field is constructed based on the Gaussian kernel function to characterize the spatial distribution of potential risks induced by each traffic participant.

[0009] Traffic prediction risk potential field, based on the current vehicle state (ADV) and TP (Traffic Risk Field) and their predicted future trajectories, in a given future... Within a one-second time window, the predicted positions of ADV and TP are mapped to a continuous risk distribution on the road ahead, which is used to characterize the risk of traffic conflicts that may occur in the future.

[0010] Step 2: Within the framework of the basic risk field, construct a conflict risk field centered on ADV;

[0011] The conflict risk field includes:

[0012] Lateral conflict risk field, forward conflict risk field, backward conflict risk field, and conflict risk field under traffic regulations;

[0013] The lateral conflict risk field is used to characterize the lateral conflict risk that may be caused by sudden behaviors and lane-changing operations in adjacent lanes.

[0014] The forward conflict risk field is used to identify potential forward conflicts during overtaking, following, and parallel driving by analyzing the spatial overlap between the ADV and the predicted TP trajectory.

[0015] The backward conflict risk field is used to reflect the uncontrollable risks that the driving decisions of the following vehicles bring to ADV;

[0016] The conflict risk field under the constraints of traffic regulations is used to quantify the implicit risks caused by vehicle trajectories deviating from traffic regulations;

[0017] When constructing conflict risk fields formed by lateral conflict risk fields, forward conflict risk fields, backward conflict risk fields, and conflict risk fields under traffic regulations, a safety distance is introduced; the safety distance refers to the minimum spatial distance that must be maintained in a dynamic traffic environment to prevent collisions or other potential dangers between vehicles, including lateral safety distance and backward safety distance.

[0018] Step 3: Treat the conflict risk field as a repulsive force acting on the ADV to quantify the dynamic impact of environmental factors and TP status on driving safety, thereby conducting risk analysis and extracting key test scenarios.

[0019] First, at each moment, the repulsion vector of each risk factor on the ADV is calculated and used as a multidimensional feature variable to form a repulsion time series. Then, principal component analysis is performed on the multidimensional repulsion time series to extract the first principal component score, reducing the multidimensional features to a one-dimensional time series to characterize the comprehensive risk intensity perceived by the ADV. When the surrounding traffic environment changes, the repulsion time series will fluctuate. By setting a threshold for the first principal component score, moments exceeding the threshold are identified as critical driving risk situations, thereby quantifying high-value risk time series segments during ADV operation.

[0020] Furthermore, the specific method for step one is as follows:

[0021] 11) The TP state risk potential field is a dynamic risk assessment model constructed based on the dynamic characteristics of TP at the current moment, namely its velocity, position, and orientation, used to quantify the potential collision risk around ADV; the TP state risk potential field is elliptical in distribution on the plane, and its geometry and scale are jointly determined by the relative velocity and spatial position between TP and ADV: the greater the relative velocity, the larger the major and minor axes of the elliptical risk field, and the wider the coverage of the high-risk area; before defining the risk field of TP, the position is first transformed from the global coordinate system to the vehicle's local coordinate system; The position of vehicle TP in the global coordinate system is The vehicle's orientation angle is ;Will Converting from degrees to radians And construct a rotation matrix accordingly. Used to convert the coordinates in the global coordinate system Rotate to the first In the vehicle's local coordinate system with TP as the reference, the local coordinates are as follows: ;

[0022] (1)

[0023] (2)

[0024] (3)

[0025] Introduce a scaling factor in the local coordinate system. and , used to adjust the scale in the tangential and normal directions; where, This is the scaling factor in the tangential direction, consistent with the vehicle's direction of travel; This is a scaling factor in the normal direction, perpendicular to the vehicle's direction of travel; for according to , Normalization process is performed to obtain Based on this, the risk field of TP state is defined by the following formula:

[0026] (4)

[0027] (5)

[0028] In the formula, For the first The lane type where the vehicle TP is located, i.e., the road condition factor; For the first Vehicle TP equivalent volume factor; For each coordinate point on the lane Compared to the first The Euclidean distance of the TP position; This is a scaling factor used to adjust the weights of the effects of velocity and heading angle on the field strength. for The angle between the vehicle's direction of travel and the direction of its movement;

[0029] 12) The traffic prediction risk field is used to describe the spatial range and corresponding risk intensity distribution that ADV and TP may occupy along the predicted driving trajectory within a given future time window; the traffic prediction risk field is based on the state parameters of ADV and TP at the current time, and under the assumption that the current lane and driving state remain unchanged, it extrapolates the vehicle position in the future period and maps the predicted position to a continuous risk distribution on the road ahead; as shown in formula (6), the traffic prediction risk field Used to characterize the Vehicle transportation theme in the future Within a one-second time window, predict the motion trajectory for discrete points on the road. The degree of potential risk arising therefrom; where, subscript The predicted risk field is used to index all traffic vehicles, including ADVs, that is, to construct corresponding predicted risk fields for both ADVs themselves and other traffic participants; by analyzing data from the initial time... To the future Integrate the predicted trajectories between each time step, and then integrate the predicted trajectories at each time step. Upper position The effects are accumulated to obtain the subject's influence throughout the entire event window. Risk contribution;

[0030] (6)

[0031]

[0032] In the formula, For time integration variables; The standard deviation of the Gaussian kernel function is used to control the spatial diffusion range of the risk field; For the first Individual subject at any moment The predicted location, and the road point The Euclidean distance between them determines the degree of risk attenuation; For a moment The attenuation factor is calculated using formula (7):

[0033] (7)

[0034] In the formula, This is the attenuation coefficient, used to control the rate at which the intensity of the risk field decays over time. and Each does not represent the first The vehicle at the predicted future time Speed ​​and corresponding vehicle mass; to avoid ADV or TP being stationary. When the denominator reaches zero, it causes a singularity problem in the calculation. The smallest positive number was introduced later. Smooth the denominator; This represents the cumulative total length of the vehicle's future trajectory within the predicted time window.

[0035] Furthermore, the specific method for step two is as follows:

[0036] 21) Construct a lateral conflict risk field;

[0037] The lateral conflict risk field refers to the potential risk conflict area that may be caused by the sudden behavior of other TPs or the driving operation of the ADV itself within the adjacent lanes on both sides; the lateral conflict risk field reflects the dynamic safety threats that the ADV may face during lateral interaction.

[0038] The safe zone of the ADV is modeled as a fan-shaped area on both sides of the ADV; the radius of the fan-shaped area is determined by the LSD; during the ADV's movement, the risk of lateral conflict is quantified by the degree of encroachment of the adjacent TP's state risk field on the lateral fan-shaped perception area.

[0039] The lateral safety zone of an ADV is modeled as a sector-shaped area in both directions; the radius of the sector-shaped area is determined by the lateral safety distance; during ADV operation, the risk of lateral conflict is reflected by the degree of encroachment of each traffic participant's state risk field on the sector-shaped safety zone;

[0040] (8)

[0041] (9)

[0042] Formula (8) gives the method for calculating the lateral risk field, where, For the first The lateral collision field strength coefficient introduced by the vehicle TP; The velocity difference adjustment coefficient is used to characterize the effect of the lateral velocity difference on the spatial diffusion range of the risk field. As the lateral velocity difference between ADV and TP increases, the diffusion range of the risk field in space expands accordingly. When the lateral velocity difference is small, the diffusion range of the risk field relatively shrinks.

[0043] Formula (9) further gives The judgment criteria, relative distance Less than the lateral safety distance At that time, it was believed that the first The risk field of the TP vehicle has not yet affected the sector-shaped safety zone of the ADV, at which point the lateral collision risk is recorded as zero; when the relative distance... Less than or equal to At that time, by calculating the maximum value of the TP state risk field within the sector area, the degree of intrusion into the sector safety area is quantified, and the corresponding lateral conflict risk field intensity is determined accordingly.

[0044] The Lateral Safety Distance (LSD) characterizes whether an ADV (Advanced Vehicle) in an adjacent lane (TP) has a lateral approach distance that could potentially affect driving safety. LSD reflects the spatial positional constraints exhibited by lateral vehicles when parallel to or passing an ADV.

[0045] (10)

[0046] In the formula, Width of the ADV vehicle; The lateral velocity of the ADV; This is the maximum lateral acceleration of ADV; This is the minimum lateral braking deceleration for ADV; The first one located on both sides of ADV The lateral speed of the TP vehicle; This is the maximum lateral braking deceleration of TP.

[0047] 22) Construct a forward conflict risk field;

[0048] The forward conflict risk field refers to the overlapping area of ​​the predicted risk field formed when the predicted trajectory of ADV interacts with the predicted trajectory of TP during overtaking or lane changing within a future time window; the overlapping area characterizes the potential risk conflict that ADV may face during forward driving.

[0049] Formula (11) defines the dynamic weighting factor. and They are ADV and the first The relative impact of each TP vehicle in the forward conflict field; among which and ADV and the 1st The speed of the TP vehicle, i.e., ADV and the first The combined velocity of the vehicle's TP in the lateral and longitudinal directions; the design of the dynamic weighting factor can reflect the nonlinear influence of vehicle speed on the field strength contribution, so that vehicles with higher speeds have a greater contribution weight to the PCRF in the overlapping area of ​​the prediction field; Formula (12) Based on this, by weighting and summing each dynamic weighting factor and normalizing it, the forward conflict risk field strength distribution in the overlapping area is obtained; the conflict field is represented by ADV and the first The overlapping portion of the predicted risk fields of each vehicle TP constitutes the calculation area, which corresponds to the spatial range where potential conflicts may occur in the future. By quantifying the risk intensity of the area, a basic support is provided for the identification and screening of key driving risk situations in the future.

[0050] (11)

[0051] (12)

[0052] 23) Construct a backward conflict risk field;

[0053] The rearward conflict risk field refers to the risk caused by the driving decisions and behaviors of the rear TP, and the influencing factors are beyond the direct control of the ADV. As the driving subject, the ADV cannot directly intervene in or affect the speed, driving trajectory or driving decisions of the rear vehicles. Therefore, the rearward risk is defined as a type of uncontrollable conflict risk dominated by other TPs.

[0054] The safe zone behind the ADV is modeled as a triangular region that varies over time, with its scale dynamically determined by the rearward safety distance between the ADV and the rear TP. As the longitudinal driving states of the ADV and TP change, the shape and extent of the rear triangular safe zone are adjusted accordingly, thus forming a rearward safety domain that evolves in real time with the vehicle's state. The intensity of the conflict risk field within the rearward safety domain is quantified by the TP's prediction of the risk field and the degree of encroachment on the triangular safe zone.

[0055] (13)

[0056] (14)

[0057] (15)

[0058] Formula (13) uses the reference point of the backward triangle safety zone. A Gaussian kernel centered on the road plane is used to sample discrete points. Modeling the backward conflict risk field intensity at the location; where, The longitudinal velocity of ADV With the rear longitudinal speed of vehicle TP The difference in longitudinal velocity is used to characterize the longitudinal relative velocity between the two vehicles; The basic diffusion regulation coefficient is used to determine... The baseline diffusion width of the Gaussian kernel at that time; Speed ​​sensitivity coefficient, through Effective control of Gaussian kernel The rate of change thus characterizes the extent of the risk field's diffusion at different longitudinal relative velocity levels;

[0059] According to the definition in formula (14), the backward conflict factor The value of can be divided into three cases:

[0060] ①TP and ADV are in a safe following position;

[0061] The first Absolute value of the longitudinal speed difference between TP and ADV Less than or equal to the following vehicle discrimination threshold , The calculation is shown in formula (15); and the longitudinal relative distance between TP and ADV is... Greater than the rearward safety distance If the TP and ADV are in a safe following state, then the risk of rearward collision is considered negligible.

[0062] Here, ζ is a hyperparameter used to balance the sensitivity of the following distance threshold. In highway scenarios, due to higher vehicle speeds and a larger desired safe distance, ζ is set to a larger value to appropriately relax the safe following distance criterion; in scenarios with high traffic density, such as urban roads, ζ is set to a smaller value to enhance the sensitivity to changes in relative speed.

[0063] ②TP and ADV are in a safe following position.

[0064] When the TP and ADV vehicles Less than This indicates that the following vehicle has entered the danger zone behind the main vehicle. At this point, a sudden acceleration by the TP (Traction Driver) or emergency braking by the ADV (Advanced Driver) will significantly increase the probability of a rear-end collision, therefore this situation is classified as the highest risk level.

[0065] ③TP is located However, in order to meet the requirements for safe following distance...

[0066] When neither of the above two conditions is met, TP and ADV have neither established a stable and safe following pattern nor entered the most dangerous rear-end collision zone; this is an intermediate transitional condition. In this case, the predicted risk field of TP is calculated. With ADV's predefined backward triangle safety area The degree of encroachment (overlap) between them is used to assess the potential conflict risk.

[0067] In the above description, Indicates the rear of the main vehicle and the first The rearward safety distance (RSD) of a vehicle's driver (TP) is a safety distance threshold that measures the potential risks beyond the control range of the driver (ADV) caused by the driving decisions and behaviors of the TP behind it.

[0068] (16)

[0069] The meanings of the parameters in the formula are as follows: Indicated as the rear of the autonomous vehicle The length of a traffic participant This indicates the time delay required between when an autonomous vehicle detects a longitudinal emergency and when it actually takes action. Maximum longitudinal acceleration of traffic participants; The minimum longitudinal braking deceleration for traffic participants; The longitudinal speed of an autonomous vehicle; This refers to the maximum longitudinal braking deceleration of an autonomous vehicle.

[0070] 24) Construct a risk field for conflicts arising from traffic regulations;

[0071] Traffic regulation-constrained conflict risk field refers to a situation where traffic signal regulations themselves do not directly generate physical risks, but if an ADV fails to detect or respond to traffic light status due to defects in its autonomous driving algorithm or sensors, it may be at risk of colliding with other TPs due to violations of traffic regulations.

[0072] This invention argues that the risks posed by traffic lights depend not only on the color change of the lights, but also on the driving strategy of the ADV (Advanced Driver Assistance Vehicle), as shown in formula (17):

[0073] (17)

[0074] (18)

[0075] In the formula, This represents the total time spent in yellow and red lights. This represents the elapsed time since the yellow and red lights came on. , These are the decay exponent and the diffusion coefficient, respectively, which together control the degree of diffusion of the Gaussian potential field in space.

[0076] It takes different values ​​in the two signal phases; in the yellow light or red light phase. Defined as the maximum field strength when the ADV-predicted risk field encroaches on the stop-line area; while during the green light phase, the stop-line area is considered risk-free, at which point... Take 0; therefore, at signal-controlled intersections, the overall risk field of the ADV master vehicle is determined by its own predicted risk field and... The forward conflict risk field formed by the overlap is jointly determined.

[0077] 31) The forward conflict risk field, backward conflict risk field, lateral conflict risk field, and traffic signal constraint risk field constructed above are denoted as... To uniformly characterize the effects of these conflict risk fields on ADV, the effects of various risk fields on ADV in the field theory sense are collectively referred to as "repulsive force".

[0078] (19)

[0079] 32) Spatial gradient of the potential energy field within the potential field framework. Interpreted as a conflict risk field at a point The ADV applies a unit "force density" at a given time. Considering that the ADV only responds to discrete grids within a certain sensing range around it, and that the influence of distant grids should decrease with distance, this invention uses a discrete integral form with distance squared as the weight to summarize the contribution of each grid, thereby obtaining the time-varying value. The total repulsive force felt by ADV.

[0080] (20)

[0081]

[0082] In the formula, This is the repulsive force gain factor, used to adjust the numerical range of the repulsive force. Variation within a specific numerical range facilitates subsequent risk analysis; the strength of the repulsive force is controlled by a point... and Measured by Euclidean distance, the farther away from the ADV vehicle, the weaker the repulsive force on the ADV vehicle;

[0083] 33) ADV at time The four types of repulsive forces experienced constitute a repulsive force vector;

[0084] (twenty one)

[0085] The time series of repulsive forces formed by the ADV being subjected to four types of risk fields throughout the entire driving period: forward conflict risk field, backward conflict risk field, lateral conflict risk field, and traffic regulation constraint risk field. This repulsion force matrix maps the multi-source conflict risk field into a time-varying "repulsion force" index acting on the ADV, achieving a continuous quantitative characterization of the risk level of autonomous driving. In this method, the magnitude of the repulsion force at any given moment is considered the quantitative result of the ADV's driving risk at that moment: the greater the repulsion force, the higher the conflict risk of the traffic environment in which the ADV is located at that moment. By examining the time series of the repulsion force along the time axis, the entire driving process can be divided into high-risk, low-risk, and relatively safe time series segments, thereby completing the quantification of autonomous driving risk based on the conflict risk field and providing a basis for the extraction of key risk driving scenarios.

[0086] The beneficial effects of this invention are as follows:

[0087] 1) The basic risk field in this invention is a comprehensive descriptive model covering ADV, TP, and their interactive risks. It aims to quantitatively describe and dynamically analyze the risks that ADV and surrounding TP may face at different times and locations. In the state dimension, it comprehensively considers the dynamic characteristics of ADV and TP, such as speed and direction, to quantify the risk level of ADV and TP. In the spatial dimension, the basic risk field reflects the relative positional relationship and risk changes of ADV and TP in physical space. In the time dimension, through dynamic time series analysis, it can analyze the evolution trend of risk over time, satisfying the dynamic change characteristics of the risk field in the time dimension.

[0088] 2) In this invention, forward conflict risk refers to the overlapping area of ​​the predicted risk field formed when the predicted trajectory of an autonomous vehicle interacts with the predicted trajectory of traffic participants during overtaking or lane changing within a future time window; this overlapping area reflects the potential risk conflict that the autonomous vehicle may face in forward driving.

[0089] 3) In this invention, lateral risk refers to the potential risk conflict area that may be caused by the sudden behavior of other traffic participants or the autonomous vehicle's own driving operation within the adjacent lanes on both sides. This conflict risk field reflects the dynamic safety threats that autonomous vehicles may face in lateral interactions;

[0090] 4) In this invention, backward risk refers to the risk caused by the driving decisions and behaviors of other traffic participants, and its influencing factors are usually beyond the controllable range of autonomous vehicles. As the driving entity, autonomous vehicles cannot directly intervene in or influence the speed, trajectory, or driving decisions of vehicles behind them. Therefore, backward risk is defined as an uncontrollable conflict risk determined by other traffic participants.

[0091] 5) In this invention, the traffic regulation constraint risk field refers to the potential risks arising from autonomous vehicles violating traffic regulations (such as running red lights). These risks do not originate directly from physical interaction, but rather from hidden safety hazards that may arise due to trajectory planning deviating from regulatory constraints.

[0092] 6) The potential conflict risk field generated by the interaction between ADV and other traffic environment elements (TPs) or traffic regulations (excluding ADV) is collectively referred to as the conflict risk field in this invention. The conflict risk field is characterized by "invasiveness," encompassing both passive and active invasiveness.

[0093] ① Passive encroachment refers to the risk field generated by the predicted trajectory of traffic participants encroaching on the predicted risk field of autonomous vehicles, forcing autonomous vehicles to adjust their driving strategies to avoid potential risks.

[0094] ②Active encroachment refers to the situation where, when an autonomous vehicle switches driving states, its predicted risk field encroaches on the predicted field of traffic participants or the risk area defined by the stop line before traffic lights, thus constituting a potential conflict risk area that the autonomous vehicle may face.

[0095] By analyzing and modeling these encroachment characteristics, the conflict risk field reveals the risk interaction relationship between ADV, traffic participants, and traffic regulations, providing theoretical support for the identification method of driving risk situation of autonomous vehicles.

[0096] 7) This invention comprehensively incorporates traffic regulations and physical rules when constructing the conflict risk during ADV driving. ADV driving risk is defined as the "repulsive force" exerted on the ADV by the conflict risk field constructed in this invention. This "repulsive force" originates from the interaction between the ADV and other road traffic infrastructure (TP). However, not all "repulsive forces" acting on the ADV necessarily constitute actual driving risk. Therefore, based on the construction of the conflict risk field during ADV driving, the conflict risk field is further mapped to the "repulsive force" exerted on the ADV. The magnitude of the "repulsive force" at any given moment is considered the quantification result of the ADV's driving risk at that moment: the greater the repulsive force, the higher the conflict risk of the traffic environment in which the ADV is located at that moment. By examining the time series of the repulsive force along the time axis, the entire driving process can be divided into high-risk, low-risk, and relatively safe time segments, thereby completing the quantification of autonomous driving driving risk based on the conflict risk field. Attached Figure Description

[0097] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0098] Figure 1 This is an architecture diagram of an autonomous driving risk quantification method based on conflict risk field as described in this invention;

[0099] Figure 2a This is a schematic diagram of the TP state risk field of the main vehicle under the straight-road driving condition described in this invention;

[0100] Figure 2b This is a three-dimensional potential field diagram of the passenger vehicle TP1 in the left adjacent rear lane of the ADV described in this invention;

[0101] Figure 2c This is a three-dimensional potential field diagram of motorcycle TP2 in the adjacent auxiliary lane to the right of ADV as described in this invention;

[0102] Figure 2d This is a three-dimensional potential field diagram of the traffic cone TP3, a static obstacle in front of the ADV lane as described in this invention.

[0103] Figure 3 This is a schematic diagram of the traffic prediction risk field described in this invention;

[0104] Figure 4 This is a schematic diagram of the basic risk field based on a T-shaped traffic intersection as described in this invention;

[0105] Figure 5 This is a schematic diagram of the lateral conflict risk field described in this invention;

[0106] Figure 6 This is a schematic diagram of the lateral forward conflict risk field described in this invention;

[0107] Figure 7 This is a schematic diagram of the backward conflict risk field described in this invention;

[0108] Figure 8a This is a schematic diagram illustrating the traffic signal regulation constraint conflict risk field when ADV fails to slow down at a red-light T-junction as described in this invention.

[0109] Figure 8b This is a schematic diagram of the traffic signal regulation constraint conflict risk field under the ADV early deceleration condition at a red-light T-junction as described in this invention.

[0110] Figure 8c This is a schematic diagram illustrating the traffic signal regulation constraint conflict risk field under the normal passage condition of ADV at a green light intersection as described in this invention.

[0111] Figure 9 This is a visual schematic diagram of the repulsive force distribution during the driving process of the ADV described in this invention. Detailed Implementation

[0112] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0113] See Figures 1-9 This invention provides a method for quantifying autonomous driving risks based on conflict risk fields, comprising the following steps:

[0114] Step 1: Construct the basic risk field;

[0115] The underlying risk field includes:

[0116] TP state risk potential field is used to generate the risk distribution that may now form;

[0117] Traffic prediction risk potential field is used to generate potential dividing lines that may form in the future time range; the specific method is as follows:

[0118] See Figures 2a-2dThe TP (Traffic Driver) state risk field is a dynamic risk assessment model built upon the TP's speed, position, and direction of travel, used to quantify the potential collision risk around the ADV (Advanced Vehicle). This risk field is elliptical in shape on a plane, and its geometry and scale are determined by the relative speed and spatial position between the TP and ADV: the greater the relative speed, the larger the major and minor axes of the elliptical risk field, and the wider the coverage area of ​​the high-risk region. Before defining the TP's risk field, its position needs to be transformed from the global coordinate system to the vehicle's local coordinate system. The position of vehicle TP in the global coordinate system is Its vehicle orientation angle is .Will Converting from degrees to radians And construct a rotation matrix accordingly. Used to convert the coordinates in the global coordinate system Rotate to the first In the vehicle's local coordinate system with TP as the reference, its local coordinates are as follows: .

[0119] (1)

[0120] (2)

[0121] (3)

[0122] Introduce a scaling factor in the local coordinate system. and This is used to adjust the scale in the tangential and normal directions. 'b' is the scaling factor in the tangential direction, consistent with the vehicle's direction of travel; 'b' is the scaling factor in the normal direction, perpendicular to the vehicle's direction of travel. Tsgo according to , Normalization process is performed to obtain Based on this, the risk field of TP state can be defined by the following formula:

[0123] (4)

[0124] (5)

[0125] in, Indicates the first The lane type where the vehicle TP is located, i.e., the road condition factor. For the first The equivalent volume factor of a vehicle's TP (Total Volume) Represents each coordinate point on the lane Compared to the first The Euclidean distance of the TP position. This is a scaling factor used to adjust the weights of the influence of velocity and heading angle on the field strength. yes and the direction of vehicle travel (i.e., unit vector) The angle between them.

[0126] Figure 2(a) illustrates the state risk distribution formed by the surrounding traffic cones (TPs) during ADV operation. As shown in Figure 2(b), TP1 travels at a speed of 39.6 km / h, and the elliptical risk field it forms along the direction of travel extends continuously forward in the longitudinal direction. In contrast, the motorcycle TP2 shown in Figure 2(c) travels at a lower speed of 10.8 km / h, and the range of its surrounding elliptical risk field is significantly reduced, indicating that its potential risk impact radius is smaller. Figure 2(d) shows the risk distribution when the traffic cone TP3 is a stationary object. Its risk field is approximately circular, reflecting that the risk range around the stationary target is relatively uniform and has a limited spatial scale.

[0127] See Figure 3 The traffic prediction risk field describes the spatial range and corresponding risk intensity distribution that ADV and TP may occupy along their predicted driving trajectories within a given future time window. Based on the current state parameters (such as speed and position) of ADV and TP, and assuming they maintain their current lanes and driving states, this risk field extrapolates the vehicle positions over a future period and maps these predicted positions to a continuous risk distribution on the road ahead. As shown in formula (6), the traffic prediction risk field... Used to characterize the Vehicle transportation theme in the future Within a one-second time window, its predicted trajectory corresponds to discrete points on the road. The degree of potential risk arising from this. (Subscript) The predicted risk field is used to index all traffic vehicles, including ADVs; that is, a predicted risk field is constructed for both ADVs themselves and other traffic participants. This is achieved by analyzing data from the initial time... To the future Integrate the predicted trajectories between each time step, and then integrate the predicted trajectories at each time step. Upper position The effects are accumulated to obtain the subject's influence throughout the entire event window. Risk contribution.

[0128] (6)

[0129]

[0130] In the formula, For time integration variables, The standard deviation of the Gaussian kernel function is used to control the spatial diffusion range of the risk field. Indicates the first Individual subject at any moment The predicted location, which is related to the road point The Euclidean distance between points determines the degree of risk attenuation at that point. For a moment The attenuation factor is calculated using formula (7):

[0131] (7)

[0132] In the formula, This represents the attenuation coefficient, used to control the rate at which the intensity of the risk field decays over time. and Each does not represent the first The vehicle at the predicted future time The speed and corresponding vehicle mass. To avoid the ADV or TP being stationary ( When the denominator reaches zero, a singularity problem arises in the calculation. The smallest positive number was introduced later. Smooth the denominator; This represents the cumulative total length of the vehicle's future trajectory within the predicted time window.

[0133] This integral formula couples trajectory prediction information in the time dimension with spatial distance relationships to estimate the intensity distribution of the risk field within a future time window. The model models potential future driving trajectories based on the current vehicle state and its historical trajectory. As the prediction time window extends, the spatiotemporal distance between future trajectory points and the current moment gradually increases, and the influence of the current state on the future trajectory points weakens accordingly, thus reducing their contribution to the risk field intensity. This reflects the attenuation effect of ADV and other TP states in the time dimension, and is reflected in the estimation results of the risk field intensity distribution.

[0134] Figure 4This diagram illustrates the basic risk field in a T-junction scenario, comprised of a TP state risk field and a traffic prediction risk field. The scenario includes one ADV (Advanced Driver Vehicle), three surrounding vehicles (TP1, TP2, and TP3), and one traffic cone, depicting the traffic flow at a given moment. At the current moment, the ADV is traveling at 46.8 km / h in its current lane, and its predicted trajectory indicates it will soon overtake the vehicle in front and change lanes to the left of TP1. TP1 is traveling straight in its lane at 25.2 km / h. TP2, located at the T-junction, is traveling at 54 km / h, and its predicted trajectory indicates it will soon cross the intersection and merge into the main road. TP3, in the opposite lane, is traveling at 28.8 km / h, and a traffic cone is positioned in its right lane to indicate a temporary traffic restriction. Based on the above elements, the basic risk field, combined with the current state and future trajectory trends of each vehicle, depicts the spatial distribution of potential risks and their evolution over time, thus revealing the fundamental attributes and spatiotemporal dynamics of the risk field.

[0135] Step 2: Within the framework of the basic risk field, construct a conflict risk field centered on ADV;

[0136] The conflict risk field includes:

[0137] Lateral conflict risk field, forward conflict risk field, backward conflict risk field, and regulatory constraint risk field;

[0138] The lateral conflict risk field is used to characterize the lateral conflict risk that may be caused by sudden behaviors and lane-changing operations in adjacent lanes.

[0139] The forward conflict risk field is used to identify potential forward conflicts during overtaking, following, and parallel driving by analyzing the spatial overlap between the ADV and the predicted TP trajectory.

[0140] The backward conflict risk field is used to reflect the uncontrollable risks that the driving decisions of the following vehicles bring to ADV;

[0141] The conflict risk field under traffic regulations is used to quantify the implicit risks caused by vehicle trajectories deviating from traffic regulations;

[0142] A safety distance is introduced when constructing the conflict risk field formed by the lateral conflict risk field, forward conflict risk field, backward conflict risk field, and traffic regulation constraint risk field. The safety distance refers to the minimum spatial distance that must be maintained in a dynamic traffic environment to prevent collisions or other potential hazards between vehicles, including lateral safety distance and backward safety distance. The specific method is as follows:

[0143] 21) Construct a lateral conflict risk field;

[0144] The Lateral Conflict Risk Field (LCRF) is used to quantify and assess situations where a TP (Driver in an adjacent lane) approaches an ADV (Adventure Vehicle) and suddenly changes lanes before the ADV can react in time and effectively, potentially significantly increasing the ADV's driving risk. This type of risk typically stems from sudden, unpredictable driving maneuvers by other TP drivers, making it difficult for the ADV to take effective evasive action within limited time and space. The LCRF is constructed to quantify the potential conflict risk in this specific driving scenario.

[0145] Figure 5 The lateral safety zone of the ADV shown is modeled as a sector-shaped area (Sec) in both directions; the radius of the sector-shaped area is determined by the lateral safety distance LSD; during the ADV's operation, the risk of lateral conflict is reflected by the degree of encroachment of each traffic participant's state risk field on the sector-shaped safety zone;

[0146] (8)

[0147] (9)

[0148] Formula (8) gives the method for calculating the lateral risk field, where, For the first The lateral collision field strength coefficient introduced by the vehicle TP; The velocity difference adjustment coefficient is used to characterize the effect of the lateral velocity difference on the spatial diffusion range of the risk field. As the lateral velocity difference between ADV and TP increases, the diffusion range of the risk field in space expands accordingly. When the lateral velocity difference is small, the diffusion range of the risk field relatively shrinks.

[0149] Formula (9) further gives The criteria for judgment Represented as ADV and the adjacent lane number The lateral safety distance of the TP; therefore, the ADV and the first Lateral relative distance of vehicles TP Less than ADV and the first Lateral safety distance of TP vehicles At that time, it was believed that the first The risk field of the TP vehicle has not yet affected the sector-shaped safety zone of the ADV, at which point the lateral collision risk is recorded as zero; when the relative distance... Less than or equal to At that time, by calculating the maximum value of the risk field of the TP state within the sector area, the degree of its intrusion into the sector safety area is quantified, and the corresponding lateral conflict risk field intensity is determined accordingly.

[0150] The Lateral Safety Distance (LSD) characterizes whether an ADV (Advanced Vehicle) in an adjacent lane (TP) has a lateral approach distance that could potentially affect driving safety. LSD reflects the spatial positional constraints exhibited by lateral vehicles when parallel to or passing an ADV.

[0151] (10)

[0152] In the formula, Width of the ADV vehicle; The lateral velocity of the ADV; This is the maximum lateral acceleration of ADV; This is the minimum lateral braking deceleration for ADV; The first one located on both sides of ADV The lateral speed of the TP vehicle; This is the maximum lateral braking deceleration of TP.

[0153] Figure 5 In the scenario shown, TP travels at 26 km / h along the lane adjacent to ADV's left and gradually approaches ADV's lane, while ADV maintains a straight-line speed of 24.5 km / h. Because the relative distance between TP and ADV is less than the lateral safety distance, TP's state risk field encroaches on the fan-shaped safety area to ADV's left, thus creating a lateral conflict risk field within that area. The field strength within the fan-shaped area is determined by the degree of encroachment of TP's state risk field within that area.

[0154] 22) Construct a forward conflict risk field;

[0155] The modeling of the Prospective Conflict Risk Field (PCRF) is mainly used to quantify the relationship between ADV and the first... The potential conflict risk of future interactions between the ADV and TP vehicles. The "forward" aspect emphasizes predicting the future movement trajectories of both the ADV and TP vehicles, identifying potential danger zones in advance, and providing a basis for subsequent analysis of key driving risks.

[0156] Formula (11) defines the dynamic weighting factor. and They are ADV and the first The relative impact of each TP vehicle in the forward conflict field; among which and ADV and the 1st The speed of the TP vehicle, i.e., ADV and the first The combined velocity of the vehicle's TP in the lateral and longitudinal directions; the design of the dynamic weighting factor can reflect the nonlinear influence of vehicle speed on the field strength contribution, so that vehicles with higher speeds have a greater contribution weight to the PCRF in the overlapping area of ​​the prediction field; Formula (12) Based on this, by weighting and summing each dynamic weighting factor and normalizing it, the forward conflict risk field strength distribution in the overlapping area is obtained; the conflict field is represented by ADV and the first The overlapping part of the predicted risk fields of each vehicle TP is the calculation area, which corresponds to the spatial range where potential conflicts may occur in the future. By quantifying the risk intensity of the area, it provides basic support for the identification and screening of subsequent key driving risk situations (or key driving time segments).

[0157] (11)

[0158] (12)

[0159] See Figure 6 When the TP travels ahead of the ADV and the predicted trajectories of both overlap in the forward region, the field strength in the overlapping region gradually moves and decays in the longitudinal direction over time, intuitively demonstrating the characteristics of the evolution of the forward conflict wind field with future moments.

[0160] 23) Construct a backward conflict risk field;

[0161] The rearward conflict risk field is a key external factor determining whether a rear-end collision risk will occur, as the driving behavior of the trailing vehicle (TP) is a critical factor. To quantify the potential uncontrollable risks encountered by an adventurous vehicle (ADV) from rear vehicles during its operation, this patent introduces the concept of a rearward conflict risk field (RCRF). This invention models the safe area behind the ADV as a time-varying triangular region, the scale of which is dynamically determined by the rearward safe distance between the ADV and the trailing TP. As the longitudinal driving states of the ADV and TP change, the shape and extent of the rear triangular safe area adjust accordingly, thus forming a rearward safe domain that evolves in real time with the vehicle's driving state. The intensity of the conflict risk field within this region is quantified by the degree to which the TP predicts the encroachment of its risk field on the triangular safe area.

[0162] (13)

[0163] (14)

[0164] Formula (13) uses the reference point of the backward triangle safety zone. A Gaussian kernel centered on the road plane is used to sample discrete points. The intensity of the backward conflict risk field at the location is modeled. In the formula... The basic diffusion adjustment coefficient is used to set the basic diffusion width of the risk field under relatively low velocity conditions, and to avoid numerical problems such as excessive shrinkage or expansion of the diffusion range in extreme cases. and The difference, after being mapped by an exponential function, participates in denominator adjustment, making the effective variance adaptively change with the longitudinal relative speed of the two vehicles. When the relative speed is small, the risk field is mainly concentrated near the triangular safety zone; when the rearmost... When vehicle TP approaches ADV at a higher speed, the effective diffusion width of the Gaussian kernel increases, and the risk intensity within the triangular safe zone and its neighborhood increases significantly. Therefore, formula (13) adjusts for the speed difference factor... The longitudinal speed difference between the following vehicle and the ADV is explicitly written into the effective variance term of the Gaussian kernel. When vehicle TP approaches vehicle ADV at a higher speed, this adjustment factor increases the variance of the Gaussian kernel, thereby expanding the spatial risk distribution centered on the triangular safety zone. When the longitudinal speeds of the two vehicles are close or ADV is faster, the variance remains within a smaller range, and the risk field is mainly concentrated near the safety zone. In this way, the longitudinal relative speed of the following vehicle is directly mapped to the degree of diffusion of the spatial risk field, which is used to characterize the potential rear-end collision risk under high-speed approach conditions.

[0165] According to the definition in formula (14), the backward conflict factor The value of can be divided into three cases:

[0166] ④TP and ADV are in a safe following position;

[0167] The first The absolute value of the longitudinal speed difference between vehicle TP and ADV Less than or equal to the following vehicle discrimination threshold And the longitudinal relative distance between TP and ADV Greater than the rearward safety distance When this condition is met, TP and ADV are determined to be in a safe following state. At this point, the risk of rearward collision is considered negligible, and the following threshold is calculated using the following formula:

[0168] (15)

[0169] Here, ζ is a hyperparameter used to balance the sensitivity of the following distance threshold. In highway scenarios, due to higher vehicle speeds and a larger desired safe distance, ζ is set to a larger value to appropriately relax the safe following distance criterion; in scenarios with high traffic density, such as urban roads, ζ is set to a smaller value to enhance sensitivity to changes in relative speed. The meanings of the other parameters have been given above and will not be repeated here.

[0170] ⑤TP and ADV are in a safe following position.

[0171] When the TP and ADV vehicles Less than This indicates that the following vehicle has entered the danger zone behind the main vehicle. At this point, a sudden acceleration by the TP (Traction Driver) or emergency braking by the ADV (Advanced Driver) will significantly increase the probability of a rear-end collision, therefore this situation is classified as the highest risk level.

[0172] ⑥TP is located However, in order to meet the requirements for safe following distance...

[0173] When neither of the above two conditions is met, TP and ADV have neither established a stable and safe following pattern nor entered the most dangerous rear-end collision zone; this is an intermediate transitional condition. In this case, the predicted risk field of TP is calculated. With ADV's predefined backward triangle safety area The degree of encroachment (overlap) between them is used to assess the potential conflict risk.

[0174] Among them, the rearward safety distance (RSD) is a safety distance threshold that measures the potential risks caused by the driving decisions and behaviors of the rear TP that are beyond the control range of the ADV.

[0175] (16)

[0176] The meanings of the parameters in the formula are as follows: Indicated as the rear of the autonomous vehicle The length of a traffic participant This indicates the time delay required between when an autonomous vehicle detects a longitudinal emergency and when it actually takes action. Maximum longitudinal acceleration of traffic participants; The minimum longitudinal braking deceleration for traffic participants; The longitudinal speed of an autonomous vehicle; This refers to the maximum longitudinal braking deceleration of an autonomous vehicle.

[0177] like Figure 7 As shown, the ADV's rearward safety zone is modeled as a triangular region facing the opposite direction of the ADV's travel. The maximum extension distance of this triangular area in the longitudinal rearward direction of the vehicle is determined by the rearward safety distance between TP and ADV. Decision. In the current scenario, the result is calculated according to equation (16). The distance is approximately 6.5m, which determines the longitudinal upward extension of the triangular safety zone behind the ADV. In this condition, the TP travels at 39.6 km / h behind the ADV, while the ADV travels at a constant speed of 32.4 km / h in a straight line. At this point, the longitudinal relative distance between the two vehicles is 2.0m, significantly greater than the safe distance. However, the longitudinal speed difference between the two vehicles was approximately 2 m / s, exceeding the following vehicle detection threshold. The velocity is approximately 0.975 m / s, therefore this state does not meet the conditions for safe following. In this "unsafe following" state, the risk of rearward collision cannot be simply considered zero, but rather determined by calculating the TP (Total Threat) to predict the risk field and the triangular safety region. The degree of overlap is quantified, that is, when TP is at a safe distance. Assess the potential risk of rear-end collisions.

[0178] 24) Construct a conflict risk field under the constraints of traffic regulations;

[0179] Traffic signal regulations themselves do not directly create physical risks, but if an ADV fails to detect or respond to traffic light status due to defects in its autonomous driving algorithm or sensors, it may be at risk of colliding with other TPs due to violations of traffic regulations.

[0180] This invention argues that the risks posed by traffic lights depend not only on the color change of the lights, but also on the driving strategy of the ADV (Advanced Driver Assistance Vehicle), as shown in formula (17):

[0181] (17)

[0182] (18)

[0183] In the formula, This represents the total time spent in yellow and red lights. This represents the elapsed time since the yellow and red lights came on. , These are the decay exponent and the diffusion coefficient, respectively, which together control the degree of diffusion of the Gaussian potential field in space.

[0184] It takes different values ​​in the two signal phases; in the yellow light or red light phase. Defined as the maximum field strength when the ADV-predicted risk field encroaches on the stop-line area; while during the green light phase, the stop-line area is considered risk-free, at which point... Take 0; therefore, at signal-controlled intersections, the overall risk field of the ADV master vehicle is determined by its own predicted risk field and... The forward conflict risk field formed by the overlap is jointly determined.

[0185] like Figure 8aThe ADV shown is traveling at 50 km / h, approaching a red-light T-junction without slowing down in advance. Its predicted risk field has crossed the stop line, forming a trapezoidal high-intensity distribution at the stop line with maximum field strength, indicating that continued driving would violate traffic regulations and create a risk without physical significance; (See attached image) Figure 8b The ADV shown begins to decelerate at 15 km / h a certain distance from the stop line. Although the traffic light is still red, its predicted risk field has not been crossed, indicating that traffic regulations have not yet posed a risk to the ADV at this point. (Attached) Figure 8c When an ADV (Advanced Driver) travels through a green light intersection at a speed of 50 km / h, its predicted risk field does not conflict with the stop line, indicating that traffic signal regulations do not pose a risk to ADV.

[0186] Step 3: Treat the conflict risk field as a repulsive force acting on the ADV (Advanced Driver Vehicle) to quantify the dynamic impact of environmental factors and TP (Traffic Terminal) status on driving safety, thereby conducting risk analysis and extracting key test scenarios, as detailed below:

[0187] First, at each moment, the repulsion vector of each risk factor on the ADV is calculated and used as a multidimensional feature variable to form a repulsion time series. Then, principal component analysis is performed on the multidimensional repulsion time series to extract the first principal component score, reducing the multidimensional features to a one-dimensional time series to characterize the comprehensive risk intensity perceived by the ADV. When the surrounding traffic environment changes, the repulsion time series will show significant fluctuations. By setting a threshold for the first principal component score, moments exceeding the threshold are identified as critical driving risk situations, thereby achieving automatic screening of high-value risk time series segments during ADV operation.

[0188] 31) In the preceding text, forward conflict risk fields, backward conflict risk fields, lateral conflict risk fields, and conflict risk fields under traffic signal constraints were constructed, denoted as follows: To uniformly characterize the effects of these conflict risk fields on ADV, this invention collectively refers to the effects of various risk fields on ADV in the field theory sense as "repulsive force".

[0189] (19)

[0190] 32) Spatial gradient of the potential energy field within the potential field framework. Interpreted as a conflict risk field at a point The ADV applies a unit "force density" at a given time. Considering that the ADV only responds to discrete grids within a certain sensing range around it, and that the influence of distant grids should decrease with distance, this invention uses a discrete integral form with distance squared as the weight to summarize the contribution of each grid, thereby obtaining the time-varying value. The total repulsive force felt by ADV.

[0191] (20)

[0192]

[0193] In the formula, This is the repulsive force gain factor, used to adjust the numerical range of the repulsive force. Variation within a specific numerical range facilitates subsequent risk analysis; the strength of the repulsive force is controlled by a point... and Measured by Euclidean distance, the farther away from the ADV vehicle, the weaker the repulsive force on the ADV vehicle;

[0194] 33) ADV at time The four types of repulsive forces experienced constitute a repulsive force vector;

[0195] (twenty one)

[0196] The time series of repulsive forces formed by the ADV being subjected to four types of risk fields throughout the entire driving period: forward conflict risk field, backward conflict risk field, lateral conflict risk field, and traffic regulation constraint risk field. This repulsion force matrix maps the multi-source conflict risk field into a time-varying "repulsion force" index acting on the ADV, achieving a continuous quantitative characterization of the risk level of autonomous driving. In this method, the magnitude of the repulsion force at any given moment is considered the quantitative result of the ADV's driving risk at that moment: the greater the repulsion force, the higher the conflict risk of the traffic environment in which the ADV is located at that moment. By examining the time series of the repulsion force along the time axis, the entire driving process can be divided into high-risk, low-risk, and relatively safe time series segments, thereby completing the quantification of autonomous driving risk based on the conflict risk field and providing a basis for the extraction of key risk driving scenarios.

[0197] like Figure 9 As shown, the simulation scenario is built on the autonomous driving virtual simulation software VTD (Virtual Test Drive), and an end-to-end autonomous driving decision-making strategy is deployed on it to obtain physically meaningful perception and control outputs. This invention constructs a driving segment containing typical dangerous situations under straight-line conditions to test whether the proposed conflict risk field-repulsion force calculation method can provide a significant repulsion force response at dangerous moments. The figure shows the distribution of the repulsion force exerted on the ADV by the forward, lateral, and backward conflict risk fields over time. Since this scenario does not include intersections and signal control, this section focuses on analyzing the contribution of the above three types of risk fields to the repulsion force.

[0198] Three representative hazardous scenarios were set up in this straight track operation:

[0199] 1) At approximately t = 4 s, ADV attempts to change lanes to the right, while TP2 in the adjacent lane on the right simultaneously accelerates forward, creating a potential risk of lateral and forward conflict.

[0200] 2) At approximately t = 12s, vehicle TP1 suddenly changes lanes to the right and merges into the lane where ADV is located, while ADV maintains a constant speed and continues straight, which could easily lead to a forward rear-end collision.

[0201] 3) At approximately t = 14s, TP1 accelerates and its speed exceeds that of ADV. As the longitudinal distance between the two vehicles decreases rapidly, TP1 decelerates. This process corresponds to a significant evolution of the risk of a rear-end collision.

[0202] As can be observed from Figure 8, during the three time periods that pose a potential threat to ADV driving safety, the repulsion force curves corresponding to various conflict risk fields all show sudden increases in amplitude or abrupt changes in shape, clearly reflecting the generation and mitigation process of dangerous situations. This indicates that the repulsion force calculation method based on traffic conflict risk fields proposed in this invention can effectively quantify the comprehensive risk effects experienced by the vehicle at different times, providing a reliable basis for subsequent identification of key risk periods and automatic extraction of high-value driving segments.

[0203] In summary, this invention first constructs a basic risk field based on the distribution and motion characteristics of traffic vehicles. Building upon this, it further forms a multi-dimensional conflict risk field encompassing vehicle forward, lateral, and backward traffic, as well as traffic regulation constraints. The abstract traffic conflict relationships are transformed into a computable "repulsion force" model, enabling dynamic quantitative representation of driving risks across state, space, and time dimensions. By leveraging the repulsion force model built upon the conflict risk field, high-value data fragments that can enhance the performance of end-to-end decision-making and planning algorithms can be extracted from massive amounts of autonomous driving operational data, thus supporting the continuous optimization and evolution of end-to-end autonomous driving algorithms.

[0204] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for quantifying the driving risk of autonomous vehicles based on conflict risk fields, characterized in that, Includes the following steps: Step 1: Construct the basic risk field; The underlying risk field includes: The TP state risk potential field is used to characterize the spatial distribution of risks that other traffic participants around the ADV may form under the current vehicle state. This potential field is constructed based on the Gaussian kernel function to characterize the spatial distribution of potential risks induced by each traffic participant. Traffic prediction risk potential field, based on the current vehicle state (ADV) and TP (Traffic Risk Field) and their predicted future trajectories, in a given future... Within a one-second time window, the predicted positions of ADV and TP are mapped to a continuous risk distribution on the road ahead, which is used to characterize the risk of traffic conflicts that may occur in the future. Step 2: Within the framework of the basic risk field, construct a conflict risk field centered on ADV; The conflict risk field includes: Lateral conflict risk field, forward conflict risk field, backward conflict risk field, and conflict risk field under traffic regulations; The lateral conflict risk field is used to characterize the lateral conflict risk that may be caused by sudden behaviors and lane-changing operations in adjacent lanes. The forward conflict risk field is used to identify potential forward conflicts during overtaking, following, and parallel driving by analyzing the spatial overlap between the ADV and the predicted TP trajectory. The backward conflict risk field is used to reflect the uncontrollable risks that the driving decisions of the following vehicles bring to ADV; The conflict risk field under the constraints of traffic regulations is used to quantify the implicit risks caused by vehicle trajectories deviating from traffic regulations; When constructing conflict risk fields formed by lateral conflict risk fields, forward conflict risk fields, backward conflict risk fields, and conflict risk fields under traffic regulations, a safety distance is introduced; the safety distance refers to the minimum spatial distance that must be maintained in a dynamic traffic environment to prevent collisions or other potential dangers between vehicles, including lateral safety distance and backward safety distance. Step 3: Treat the conflict risk field as a repulsive force acting on the ADV to quantify the dynamic impact of environmental factors and TP status on driving safety, thereby conducting risk analysis and extracting key test scenarios. First, at each moment, the repulsion vector of each risk factor on the ADV is calculated and used as a multidimensional feature variable to form a repulsion time series. Then, principal component analysis is performed on the multidimensional repulsion time series to extract the first principal component score, reducing the multidimensional features to a one-dimensional time series to characterize the comprehensive risk intensity perceived by the ADV. When the surrounding traffic environment changes, the repulsion time series will fluctuate. By setting a threshold for the first principal component score, moments exceeding the threshold are identified as critical driving risk situations, thereby quantifying high-value risk time series segments during ADV operation.

2. The method for quantifying autonomous driving risks based on conflict risk fields according to claim 1, characterized in that, The specific method for step one is as follows: 11) The TP state risk potential field is a dynamic risk assessment model built on the dynamic characteristics of TP at the current moment, namely velocity, position and orientation, and is used to quantify the potential collision risk around ADV. The risk potential field of the TP state is elliptical in shape on the plane, and its geometry and scale are determined by the relative velocity and spatial position between the TP and ADV: the greater the relative velocity, the larger the major and minor axes of the elliptical risk field, and the wider the coverage of the high-risk area; before defining the risk field of the TP, the position is first transformed from the global coordinate system to the vehicle's local coordinate system; The position of vehicle TP in the global coordinate system is The vehicle's orientation angle is ;Will Converting from degrees to radians And construct a rotation matrix accordingly. Used to convert the coordinates in the global coordinate system Rotate to the first In the vehicle's local coordinate system with TP as the reference, the local coordinates are as follows: ; (1) (2) (3) Introduce a scaling factor in the local coordinate system. and , used to adjust the scale in the tangential and normal directions; where, 'b' is the scaling factor in the tangential direction, consistent with the vehicle's direction of travel; 'b' is the scaling factor in the normal direction, perpendicular to the vehicle's direction of travel; according to , Normalization process is performed to obtain Based on this, the risk field of TP state is defined by the following formula: (4) (5) In the formula, For the first The lane type where the vehicle TP is located, i.e., the road condition factor; For the first Vehicle TP equivalent volume factor; For each coordinate point on the lane Compared to the first The Euclidean distance of the TP position; This is a scaling factor used to adjust the weights of the effects of velocity and heading angle on the field strength. for The angle between the vehicle's direction of travel and the direction of its movement; 12) The traffic prediction risk field is used to describe the spatial range and corresponding risk intensity distribution that ADV and TP may occupy along the predicted driving trajectory within a given future time window; the traffic prediction risk field is based on the state parameters of ADV and TP at the current time, and under the assumption that the current lane and driving state remain unchanged, it extrapolates the vehicle position in the future period and maps the predicted position to a continuous risk distribution on the road ahead; as shown in formula (6), the traffic prediction risk field Used to characterize the Vehicle transportation theme in the future Within a one-second time window, predict the motion trajectory for discrete points on the road. The degree of potential risk arising therefrom; where, subscript The predicted risk field is used to index all traffic vehicles, including ADVs, that is, to construct corresponding predicted risk fields for both ADVs themselves and other traffic participants; by analyzing data from the initial time... To the future Integrate the predicted trajectories between each time step, and then integrate the predicted trajectories at each time step. Upper position The effects are accumulated to obtain the subject's influence throughout the entire event window. Risk contribution; (6) ； In the formula, For time integration variables; The standard deviation of the Gaussian kernel function is used to control the spatial diffusion range of the risk field; For the first Individual subject at any moment The predicted location, and the road point The Euclidean distance between them determines the degree of risk attenuation; For a moment The attenuation factor is calculated using formula (7): (7) In the formula, This is the attenuation coefficient, used to control the rate at which the intensity of the risk field decays over time. and Each does not represent the first The vehicle at the predicted future time Speed ​​and corresponding vehicle mass; to avoid ADV or TP being stationary. When the denominator reaches zero, it causes a singularity problem in the calculation. The smallest positive number was introduced later. Smooth the denominator; This represents the cumulative total length of the vehicle's future trajectory within the predicted time window.

3. The method for quantifying autonomous driving risks based on conflict risk fields according to claim 1, characterized in that, The specific method for step two is as follows: 21) Construct a lateral conflict risk field; The lateral conflict risk field refers to the potential risk conflict area that may be caused by the sudden behavior of other TPs or the driving operation of the ADV itself within the adjacent lanes on both sides; the lateral conflict risk field reflects the dynamic safety threats that the ADV may face during lateral interaction. The safe zone of the ADV is modeled as a fan-shaped area on both sides of the ADV; the radius of the fan-shaped area is determined by the LSD; during the ADV's movement, the risk of lateral conflict is quantified by the degree of encroachment of the adjacent TP's state risk field on the lateral fan-shaped perception area. The lateral safety zone of an ADV is modeled as a sector-shaped area in both directions; the radius of the sector-shaped area is determined by the lateral safety distance; during ADV operation, the risk of lateral conflict is reflected by the degree of encroachment of each traffic participant's state risk field on the sector-shaped safety zone; (8) (9) Formula (8) gives the method for calculating the lateral risk field, where, For the first The lateral collision field strength coefficient introduced by the vehicle TP; The velocity difference adjustment coefficient is used to characterize the effect of the lateral velocity difference on the spatial diffusion range of the risk field. As the lateral velocity difference between ADV and TP increases, the diffusion range of the risk field in space expands accordingly. When the lateral velocity difference is small, the diffusion range of the risk field relatively shrinks. Formula (9) further gives The judgment criteria, relative distance Less than the lateral safety distance At that time, it was believed that the first The risk field of the TP vehicle has not yet affected the sector-shaped safety zone of the ADV, at which point the lateral collision risk is recorded as zero; when the relative distance... Less than or equal to At that time, by calculating the maximum value of the TP state risk field within the sector area, the degree of intrusion into the sector safety area is quantified, and the corresponding lateral conflict risk field intensity is determined accordingly. The lateral safety distance is used to characterize whether there is a lateral approach distance that may affect driving safety when the adjacent lane TP approaches the ADV; LSD reflects the spatial position constraint characteristics of lateral vehicles when they are parallel to or pass the ADV. (10) In the formula, Width of the ADV vehicle; The lateral velocity of the ADV; This is the maximum lateral acceleration of ADV; This is the minimum lateral braking deceleration for ADV; The first one located on both sides of ADV The lateral speed of the TP vehicle; This is the maximum lateral braking deceleration of TP; 22) Construct a forward conflict risk field; The forward conflict risk field refers to the overlapping area of ​​the predicted risk field formed when the predicted trajectory of ADV interacts with the predicted trajectory of TP during overtaking or lane changing within a future time window; the overlapping area characterizes the potential risk conflict that ADV may face during forward driving. Formula (11) defines the dynamic weighting factor. and They are ADV and the first The relative impact of each TP vehicle in the forward conflict field; among which and ADV and the 1st The speed of the TP vehicle, i.e., ADV and the first The combined velocity of the vehicle's TP in the lateral and longitudinal directions; the design of the dynamic weighting factor can reflect the nonlinear influence of vehicle speed on the field strength contribution, so that vehicles with higher speeds have a greater contribution weight to the PCRF in the overlapping area of ​​the prediction field; Formula (12) Based on this, by weighting and summing each dynamic weighting factor and normalizing it, the forward conflict risk field strength distribution in the overlapping area is obtained; the conflict field is represented by ADV and the first The overlapping portion of the predicted risk fields of each vehicle TP constitutes the calculation area, which corresponds to the spatial range where potential conflicts may occur in the future. By quantifying the risk intensity of the area, a basic support is provided for the identification and screening of key driving risk situations in the future. (11) (12) 23) Construct a backward conflict risk field; The rearward conflict risk field refers to the risk caused by the driving decisions and behaviors of the rear TP, and the influencing factors are beyond the direct control of the ADV. As the driving subject, the ADV cannot directly intervene in or affect the speed, driving trajectory or driving decisions of the rear vehicles. Therefore, the rearward risk is defined as a type of uncontrollable conflict risk dominated by other TPs. The safe zone behind the ADV is modeled as a triangular region that varies over time, with its scale dynamically determined by the rearward safe distance between the ADV and the rear TP. As the longitudinal driving states of the ADV and TP change, the shape and extent of the rear triangular safe zone are adjusted accordingly, thus forming a rearward safe domain that evolves in real time with the vehicle's state. The intensity of the conflict risk field within the rearward safe domain is quantified by the TP's prediction of the risk field and the degree of encroachment on the triangular safe zone. (13) (14) (15) Formula (13) uses the reference point of the backward triangle safety zone. A Gaussian kernel centered on the road plane is used to sample discrete points. Modeling the backward conflict risk field intensity at the location; where, The longitudinal velocity of ADV With the rear longitudinal speed of vehicle TP The difference in longitudinal velocity is used to characterize the longitudinal relative velocity between the two vehicles; The basic diffusion regulation coefficient is used to determine... The baseline diffusion width of the Gaussian kernel at that time; Speed ​​sensitivity coefficient, through Effective control of Gaussian kernel The rate of change thus characterizes the extent of the risk field's diffusion at different longitudinal relative velocity levels; According to the definition in formula (14), the backward conflict factor The value of can be divided into three cases: ①TP and ADV are in a safe following position; The first Absolute value of the longitudinal speed difference between TP and ADV Less than or equal to the following vehicle discrimination threshold , The calculation is shown in formula (15); and the longitudinal relative distance between TP and ADV is... Greater than the rearward safety distance If the TP and ADV are in a safe following state, then the risk of rearward collision is considered negligible. Here, ζ is a hyperparameter used to balance the sensitivity of the following distance threshold. In highway scenarios, due to the higher vehicle speed and the larger expected safe distance, ζ is taken as a larger value to appropriately relax the safe following distance criterion. In scenarios with high traffic density, such as urban roads, ζ is taken as a smaller value to enhance the sensitivity to changes in relative speed. ②TP and ADV are in a safe following position; When the TP and ADV vehicles Less than When the following vehicle has entered the danger zone behind the main vehicle, it indicates that the following vehicle has entered the danger zone behind the main vehicle. At this time, if the TP suddenly accelerates or the ADV brakes suddenly, the probability of a rear-end collision will increase significantly. Therefore, this situation is classified as the highest risk level. ③TP is located However, in order to meet the requirements for safe following; When neither of the above two conditions is met, TP and ADV have neither formed a stable and safe following pattern nor entered the most dangerous rear-end collision zone, which is an intermediate transitional condition; at this time, the predicted risk field of TP is calculated. With ADV's predefined backward triangle safety area The degree of encroachment between them is used to assess the potential risk of conflict; in, The rear of the main vehicle and the first The rearward safety distance of a TP is a safety distance threshold that measures the potential risks beyond the control range of the ADV caused by the driving decisions and behaviors of the TP behind it. (16) The meanings of the parameters in the formula are as follows: Indicated as the rear of the autonomous vehicle The length of a traffic participant This indicates the time delay required between when an autonomous vehicle detects a longitudinal emergency and when it actually takes action to address the situation; Maximum longitudinal acceleration of traffic participants; The minimum longitudinal braking deceleration for traffic participants; The longitudinal speed of an autonomous vehicle; The maximum longitudinal braking deceleration of an autonomous vehicle; 24) Construct a risk field for conflicts arising from traffic regulations; Traffic regulations constrain conflict risk fields, which refer to situations where traffic signal regulations themselves do not directly generate physical risks, but if an ADV fails to detect or respond to traffic light status due to defects in its autonomous driving algorithm or sensors, it may be at risk of colliding with other TPs due to violations of traffic regulations. The risks posed by traffic lights depend not only on the color change of the lights, but also on the driving strategy of the ADV, as shown in formula (17): (17) (18) In the formula, This represents the total time spent in yellow and red lights. This represents the elapsed time since the yellow and red lights came on. , These are the decay exponent and the diffusion coefficient, respectively, which together control the degree of diffusion of the Gaussian potential field in space. It takes different values ​​in the two signal phases; in the yellow light or red light phase. Defined as the maximum field strength when the ADV-predicted risk field encroaches on the stop-line area; while during the green light phase, the stop-line area is considered risk-free, at which point... Take 0; therefore, at signal-controlled intersections, the overall risk field of the ADV master vehicle is determined by its own predicted risk field and... The forward conflict risk field formed by the overlap is jointly determined.

4. The method for quantifying autonomous driving risks based on conflict risk fields according to claim 1, characterized in that, The specific method for step three is as follows: 31) The constructed forward conflict risk field, backward conflict risk field, lateral conflict risk field, and risk field under traffic signal constraints are denoted as... The effects of various risk fields on ADV in the field theory sense are collectively referred to as "repulsive forces". (19) 32) Spatial gradient of the potential energy field within the potential field framework. Interpreted as a conflict risk field at a point The unit "force density" applied at ADV; Considering that ADV only responds to discrete grids within a certain sensing range, and that the influence of distant grids should decrease with distance, a discrete integral form with distance squared as the weight is used to summarize the contributions of each grid, thereby obtaining the time step. The total repulsive force felt by ADV at that time; (20) ； In the formula, This is the repulsive force gain factor, used to adjust the numerical range of the repulsive force. Variation within a specific numerical range facilitates subsequent risk analysis; the strength of the repulsive force is controlled by a point... and Measured by Euclidean distance, the farther away from the ADV vehicle, the weaker the repulsive force on the ADV vehicle; 33) ADV at time The four types of repulsive forces experienced constitute a repulsive force vector; (21) The time series of repulsive forces formed by the ADV being subjected to four types of risk fields during the entire driving period: forward conflict risk field, backward conflict risk field, lateral conflict risk field, and conflict risk field under traffic regulations. The repulsion force matrix maps the multi-source conflict risk field into a time-varying "repulsion force" index acting on the ADV, realizing a continuous quantitative representation of the risk level of autonomous driving. The magnitude of the repulsion force at any moment is regarded as the quantitative result of the ADV's driving risk at that moment: the greater the repulsion force, the higher the conflict risk of the traffic environment in which the ADV is located at that moment. Examining the time series of the repulsion force along the time axis, that is, dividing the entire driving process into high-risk, low-risk and relatively safe time series segments, thereby completing the quantification of autonomous driving risk based on the conflict risk field and providing a basis for the extraction of key risk driving scenarios.

Images