An aerial photography data set-oriented intelligent automobile key test scene extraction method

Abstract

The application relates to an intelligent automobile key test scene extraction method for aerial photography data sets, and steps include key analysis on vehicle position distribution around the automobile, determination of a boundary, i.e. a range domain [theta 1, theta n ], based on a field strength theory, scene key behavior analysis, determination of a time domain [t1, t N ] of a scene key quantization model, and key scene extraction. The method of the application uses a part that may collide with the automobile as a CDCE collection area, and proposes an evaluation method of key behaviors of vehicles in a scene to screen vehicle trajectories, and further extracts meaningful and valuable test scenes. The application is more sensitive to lateral motion objects, and indicates key scenes in a time dimension ahead of TTC, can search scenes that are challenging for intelligent automobiles, and the method is oriented to real vehicle trajectories to ensure the authenticity of the extracted scenes.

Description

TECHNICAL FIELD

[0001] The application relates to an intelligent automobile key test scene extraction method, in particular to an intelligent automobile key test scene extraction method for aerial photography data sets. BACKGROUND

[0002] Intelligent automobile technology provides a new solution for solving the problems of environmental pollution, driving safety and traffic congestion, but how to ensure the safety of various elements in the human-vehicle-road environment under any condition has become a hot research topic, so a sound intelligent automobile test evaluation system is needed to ensure that the intelligent automobile meets the driving requirements in the complex environment of the actual road. Traditional automobiles often use mileage-based test methods to complete the above tasks, however, relevant authoritative research shows that hundreds of billions of kilometers of driving data are needed to improve the intelligent automobile algorithm to make the intelligent automobile reach the same driving ability as a human driver, which is time-consuming and laborious and cannot meet the needs of the product cycle in actual industrial production. At present, the scene-based intelligent automobile test method is a hot research topic, however, the actual driving scene has a long tail problem, that is, there are a large number of safety scenes, which cannot help the researchers identify the functional defects of the intelligent automobile, so it is necessary to design scenes that can efficiently test the intelligent automobile. The existing test scene extraction method mainly targets the intelligent automobile perception function module, cannot search for scenes that are challenging to the intelligent automobile planning control module, or is based on non-real vehicle trajectories, and cannot guarantee the authenticity of the scene. SUMMARY

[0003] In order to solve the above technical problems, the application provides an intelligent automobile key test scene extraction method for aerial photography data sets. The aerial photography data refers to the feature data of the road vehicle collected and extracted by the unmanned aerial vehicle device at the overlooking angle. Compared with the data collected indoors, the aerial photography data can better reflect the dynamic interaction between various traffic participants in the complex traffic environment. The aerial photography data set extracts the motion characteristics of each frame of the vehicle through an advanced image recognition algorithm to obtain the basic vehicle motion characteristics, including the position, shape characteristics, lateral and longitudinal speed, yaw angular velocity, offset of the center point from the current lane center line, distance from the front vehicle and position vehicle serial number information in eight directions, and simultaneously calculates the abstract representation parameters, including the headway (DHW) and time to collision (TTC) parameters.

[0004] The specific steps of the application are as follows:

[0005] 1. Key analysis of the position distribution of the vehicles around the vehicle to determine the boundary, that is, the range [theta1, theta2] of the test scene concerned. n

[0006] ​Eight different position information relative to the ego vehicle are extracted in the aerial data set, i.e. left front, front, right front, left side, right side, left rear, rear and right rear; different position information of vehicles poses different degrees of challenge to the ego vehicle, a polar coordinate system is introduced to describe the scene key, the bird's-eye two-dimensional perspective in the aerial data set is converted to a one-dimensional feature space, and the effective contour distance curve of the ego vehicle (CDCE) is used to describe the position distribution of the possible collision parts of the surrounding vehicles relative to the ego vehicle, the centroid of the ego vehicle is taken as the origin of the curve, the road direction of the vehicle moving forward is taken as the x axis, the clockwise direction is taken as positive, and the contour information of the surrounding vehicles within 360° of the vehicle is mapped to the polar coordinate system;

[0007] The aerial data set records rich vehicle motion information, including the position feature information of vehicle i (x i , y i ), width j , height j , yaw ; for position information, coordinate transformation is needed, taking the left front end of the circumscribed rectangle frame of vehicle j as an example:

[0008]

[0009]

[0010]

[0011] In the formula, s j is the horizontal coordinate of vehicle j after rotation transformation during lane changing; t j is the vertical coordinate of vehicle j after rotation transformation during lane changing; θ1 is the angle between the line connecting the mass center of the ego vehicle i and the left front end of the target vehicle j and the forward direction of the ego vehicle; d1 is the distance between the mass center of the ego vehicle i and the left front end of the target vehicle j.

[0012] The coordinate transformation of other position information of the circumscribed rectangle frame of vehicle j is the same; the angle and distance output by all the possible collision parts relative to the ego vehicle i jointly constitute θ ij and r ij , and are input to the field intensity formula in the second section.

[0013] The part that may collide with the vehicle is selected as the collection area of the CDCE, and the area that may collide is scanned from the centroid of the vehicle i, and the intersection of the scanning ray and the circumscribed rectangular frame of the other vehicle is recorded as a collection point in the polar coordinate system;

[0014] 2. Scene criticality quantification based on field strength theory:

[0015] Based on the field theory of driving risk under complex traffic conditions, a criticality quantification model of vehicle behavior in the scene is established, and it is proposed that the field strength n formed by the vehicle i and other vehicles j at time t is defined as:

[0016]

[0017] In the formula, M i and M j are equivalent masses, the values of which are related to vehicle type and mass, and are set as specific constants; R is a road condition parameter, which is set as a specific constant; k1, k2 and G are all constants greater than 0; θ ij is the angle between the connecting line of the centroid of the host vehicle i and each point on the possible collision profile of the target vehicle j and the forward direction of the host vehicle, which is consistent with the calculation method of θ defined in the previous step; v ij is the relative speed of the two vehicles, and the specific expression is The relative speed of each point on the same vehicle outer frame line is equal; it can be seen that the greater the field strength, the higher the scene criticality, and the scene criticality at time t n is represented by the integral of θ1 and θ n in the range:

[0018]

[0019] Since R i , M j and M i are constants relative to θ at the same time, the above formula is simplified as:

[0020]

[0021] Then the scene criticality from t1 to t N is determined by summing up each timestamp, and the specific expression is:

[0022]

[0023] 3. Scene critical behavior analysis to determine the time domain [t1, t N ] of the scene criticality quantification model:

[0024] The trajectory of the vehicle is screened by the evaluation method of the key behavior of the vehicle in the scene, and then the meaningful and valuable test scene is extracted. Specifically, the follow-up scene and the front vehicle intrusion into the lane line are analyzed, and the time domain [t N ] of the scene key quantization model is determined, and the range domain [θ n ] needs to be determined by the position distribution in the first step, and the corresponding scene in the aerial data set is quantified, and the scenes with test value are screened and applied.

[0025] 3.1 Follow-up scene behavior analysis:

[0026] The follow-up scene grading key quantization model is designed, and the follow-up scene warning threshold D W and the follow-up scene emergency braking warning threshold D B are set to screen the follow-up scene into three states and quantify the key of the follow-up scene;

[0027] When the distance L between the vehicle and the front vehicle is greater than D W , this kind of scene belongs to a safe scene; when L is greater than D B and less than or equal to D W , a warning signal will be triggered to remind the driver that there is a vehicle in front, and this kind of scene has a certain potential risk; when L is less than D B , the vehicle should be braked with the maximum braking deceleration as the control target, and this kind of scene has a high risk;

[0028] The method of specifically determining D W and D B needs to describe the braking process of the vehicle. The common braking process deceleration and braking time relationship of intelligent vehicles includes: the first stage is the sampling stage of the intelligent vehicle perception system, corresponding to τ1; the second stage is the brake braking process, corresponding to τ2, with the movement of the brake pedal, the braking acceleration of the vehicle to the maximum braking deceleration a Bmax needs to pass a period of time; the third stage is the stage of constant braking acceleration, corresponding to τ3; the fourth stage is the brake recovery process, corresponding to τ4, which refers to the process of gradually recovering to zero after the brake pedal is released.

[0029] Based on the above process, the follow-up scene warning threshold D W and the follow-up scene emergency braking warning threshold D B are calculated:

[0030] D W =D B +v ego ·t w

[0031]

[0032] where v ego is the speed of the ego vehicle; a Bmax is the maximum deceleration of the ego vehicle, which is filtered in the dataset; f0is the minimum detection frequency of the target detection module and the ranging module of the intelligent vehicle; t w is the warning time set by the system;

[0033] Therefore, when quantifying the key scene of the following vehicle scene, the time t W corresponding to L≤D fol-beg is taken as the initial time of the following vehicle scene quantification model. At this time, the distance between the ego vehicle and the front vehicle will gradually decrease, and the relative speed will also decrease until the relative speed is 0. Therefore, the time t fol-end corresponding to the relative speed of 0 between the front and rear vehicles is taken as the end time of the scene key quantification model.

[0034] 3.2 Analysis of the front vehicle lane invasion scene behavior

[0035] The decoupled front vehicle invasion lane change scene is analyzed to obtain the upper and lower bounds of the lane change scene key quantification model. For the lateral decoupling of the front vehicle invasion lane change scene, the lane line pressed by the invasion lane change to the ego vehicle lane is taken as the invasion lane line, and the lateral invasion distance q of the front vehicle is taken as an important parameter to determine the start of the lane change. The specific expression is:

[0036]

[0037]

[0038] In the formula, is the lateral offset distance between the vehicle i center line and the ego vehicle lane center line, which is directly obtained in the dataset; Width lane is the lane line width; d is the lateral distance between the ego vehicle center line and the invasion lane line; where y i and y j are the horizontal coordinate values of ID i and j in the dataset; height j is the width of vehicle j; is the heading angle of vehicle j, which is calculated by the angle parameter and the orientation parameter provided in the dataset, w b is the wheelbase of the vehicle; when q is 0, it is defined as the start time of the front vehicle invasion scene, and when the offset amount of the front vehicle from the lane line center line of the ego vehicle is less than or equal to 0.1 m, it is defined as the end time of the front vehicle invasion scene;

[0039] 4. Key scene extraction:

[0040] The open source data sets commonly used for intelligent vehicle testing and verification are the HighD data set and the Aerial Dataset for China Congested Highway and Expressway (AD4CHE) proposed by DJI. The method proposed in the patent can process these two data sets to obtain key test scenarios. The specific extraction process is as follows: two important trajectory feature quantities, the number of lane changes (numLaneChanges) and the minimum time to collision (minTTC), are determined through the overall trajectory statistical information in the aerial data set, which are used for the preliminary screening of the front vehicle intrusion scene and the following vehicle scene, respectively; for the front vehicle intrusion scene, when the number of lane changes is not 0 and there is a rear vehicle after the lane line (LaneId) occupied by the intrusion vehicle changes, it can be defined as a front vehicle intrusion scene, after determining the intrusion scene, the vehicle trajectory before the LaneId changes is analyzed, the time corresponding to the front vehicle lateral intrusion distance q is calculated as the starting time of the key quantitative model, the time corresponding to the lane line offset of the intrusion vehicle less than or equal to 0.1m is found as the ending time of the key quantitative model, and then the key of the front vehicle intrusion scene is quantified; for the following vehicle scene, TTC is a commonly used index to express the key of the following vehicle scene, first, scenes where the front vehicle does not exist are filtered out, and then the upper and lower limits in the key quantitative model are determined using the method in section 3.1 to calculate the key of the following vehicle scene.

[0041] The beneficial effects of the present application are:

[0042] The method can search for scenes that are challenging for intelligent vehicles, and the method is based on real vehicle trajectories to ensure the authenticity of the scenes. The application uses the part that is easy to collide with the vehicle as the collection area of the CDCE, and the method has the advantages of describing the vehicle position distribution, including: 1. Scale invariance: The size of the surrounding vehicles does not affect the shape of the CDCE curve, so it can be used for any bird's eye view dataset, and it has strong universality; 2. Rotational invariance: When the UAV rotates by a certain angle, the polar coordinate records the distance relationship at different angles, which simplifies the expression of the scene and is not affected by the motion of the collection device itself; 3. Morphological differentiation: The polar coordinate also records the morphology of different vehicles, such as different morphologies of cars and trucks, which have different curve expressions in the polar coordinate; 4. Test of matching sensor function: For camera sensors, for example, different field of view (FOV) should be selected when selecting test scenes, and the FOV range of the tested camera should be selected for function test of other vehicle trajectory information. The motion information of objects beyond the FOV range of the tested camera is useless, i.e. no test value, which can improve the test pertinence. The application proposes an evaluation method for the key behaviors of vehicles in the scene to filter the vehicle trajectories, and then extracts meaningful and valuable test scenes. The application is more sensitive to lateral moving objects and indicates the key scene in the time dimension before TTC. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The application is a schematic diagram of aerial data collection;

[0044] Figure 2 The application is a schematic diagram of aerial extraction results;

[0045] Figure 3 The application is a schematic diagram of the effective contour distance curve of the vehicle;

[0046] Figure 4 The application is a schematic diagram of CDCE;

[0047] Figure 5 The application is a schematic diagram of the intrusion scene quantification method;

[0048] Figure 6 The application is a schematic diagram of the follow-up scene classification key quantification model;

[0049] Figure 7 The application is a schematic diagram of the relationship between vehicle braking deceleration and braking time;

[0050] Figure 8 The application is a schematic diagram of lateral decoupling;

[0051] Figure 9 The application is a schematic diagram of the main extraction steps of the key scene;

[0052] Figure 10 The quantitative results of the method of the application and TTC for the same scene are compared. DETAILED DESCRIPTION

[0053] The application provides a smart car key test scene extraction method for aerial photography data sets. The aerial photography data refers to the collection and extraction of feature data of road vehicles at a bird's eye view angle through a drone device. Compared with indoor collected data, aerial photography data can better reflect the dynamic interaction between various traffic participants in a complex traffic environment. Figure 1 And 2 As shown in the drawings, the aerial photography data set extracts the motion characteristics of each frame of the vehicle through an advanced image recognition algorithm to obtain basic vehicle motion characteristics, such as the position, shape characteristics, lateral and longitudinal speed, yaw angular velocity, offset of the center point from the current lane center line, distance from the front vehicle, and position vehicle serial number in eight directions. At the same time, abstract representation parameters such as the headway (DHW) and the time to collision (TTC) can also be indirectly calculated.

[0054] The specific steps of the application are as follows:

[0055] 1. Key analysis of the position distribution of vehicles around the ego vehicle to determine the boundary of the test scene, i.e., the range [θ1, θ2] of the scene. n ]:

[0056] Eight different position information relative to the ego vehicle (Egovehicle) is extracted in the aerial photography data set, namely left front, front, right front, left side, right side, left rear, rear, and right rear. Vehicles at different positions pose different degrees of challenge to the ego vehicle, but traditional Euclidean distance description methods cannot meet the description of scene criticality. For example, a left lane line vehicle with a longitudinal speed similar to that of the ego vehicle should not be defined as a scene with high criticality even if it is close to the ego vehicle. The application introduces a polar coordinate system to describe the scene criticality. The advantage is that it converts the bird's eye view two-dimensional perspective in the aerial photography data set to a one-dimensional feature space. The application proposes an effective contour distance curve of the ego vehicle (CDCE) to describe the position distribution of the possible collision parts of the surrounding vehicles relative to the ego vehicle. The centroid of the ego vehicle is taken as the origin of the curve, the road direction of the vehicle moving forward is taken as the x-axis, and the clockwise direction is taken as positive. The contour information of the surrounding vehicles within 360° of the vehicle is mapped to the polar coordinate system, as shown in the drawings. Figure 3 .

[0057] The aerial photography data set records rich vehicle motion information. The position characteristic information of the vehicle with ID i includes (xi , y i ), width j Corresponding to the length of the vehicle j, height j Corresponding to the width of the vehicle j, Corresponding to the yaw angle of the vehicle j; for position information, coordinate transformation is needed, taking the left front end of the circumscribed rectangle frame of the vehicle j as an example:

[0058]

[0059]

[0060]

[0061] In the formula, in the formula, s j The transverse coordinate of the vehicle j after rotation transformation during the lane changing process; t j The longitudinal coordinate of the vehicle j after rotation transformation during the lane changing process; θ1 is the included angle between the line connecting the mass center of the host vehicle i and the left front end of the target vehicle j and the forward direction of the host vehicle; d1 is the distance between the mass center of the host vehicle i and the left front end of the target vehicle j;

[0062] The coordinate transformation of other position information of the circumscribed rectangle frame of the vehicle j is the same; all the angles and distances output relative to the host vehicle i that may collide constitute θ ij And r ij , and are input as the field strength formula in the second section.

[0063] The present application selects the part that may collide with the vehicle as the collection area of the CDCE, and the intersection of the scanning ray and the circumscribed rectangle frame of the other vehicle is recorded in the polar coordinate system as a collection point, to obtain the CDCE schematic diagram as shown in Figure 4 .

[0064] 2. Scene key quantification based on field strength theory:

[0065] The driving risk field under complex traffic conditions is used to represent the influence degree of human-vehicle-road interaction and various factors on driving safety. The present application establishes a key quantification model of vehicle behavior in the scene based on the theory, and proposes that the field strength n formed by the host vehicle i and other vehicles j at t time is defined as:

[0066]

[0067] In the formula, M i And M jis the equivalent mass, the value is related to the type of vehicle, mass, in the present application is set to a specific constant; R is the road condition parameter, in the present application is set to a specific constant; k1, k2 and G are all constants greater than 0, in the research of the present application G is set to 1, and k1 and k2 are both set to 2; theta ij is the angle between the line connecting the mass center of the host vehicle i and each point on the possible collision profile of the target vehicle j and the forward direction of the host vehicle, which is calculated in the same way as theta defined in the previous step; v ij is the relative speed of the two vehicles, and the specific expression is The relative speeds of the points on the same vehicle outer frame line are equal; therefore, the greater the field strength, the higher the scene criticality at t n ; the scene criticality at t n is represented by the integral of theta

[0068]

[0069] Since R i , M j and M i are constants relative to theta at the same time, the above formula can be simplified as:

[0070]

[0071] The scene criticality from t1 to t N is determined by summing up each timestamp, and the specific expression is:

[0072]

[0073] A front vehicle cutting into a scene is taken as an example to quantify it, two vehicles in the aerial data set are analyzed, in this scene, the host vehicle drives forward, and in a certain frame of image in the data set, the invading vehicle changes lanes from the other lane to the host vehicle lane and continues to drive in front of the host vehicle, the scene is quantitatively calculated by the above method, and a schematic diagram is shown in Figure 5 .

[0074] 3. Scene critical behavior analysis, determine the time domain [t1, t N ] of the scene criticality quantification model:

[0075] The previous step explores how to quantify the scene criticality, but traversing all scenes in the data set using the above method will consume a large amount of computing resources, the present application proposes a method of evaluating the key behavior of vehicles in the scene to filter the trajectories of vehicles, and then extracts meaningful and valuable test scenes. Specifically, two kinds of scenes, following vehicle scene and front vehicle invading host lane scene, are analyzed to determine the time domain [t1, tN ] and the range domain [θ1, θ n ] needs to be determined by the position distribution of the first step, and the corresponding scene in the aerial data set is quantified, and the scene with test value is screened out for application.

[0076] 3.1 Following scene behavior analysis

[0077] Following driving is a common scene, which requires intelligent vehicles to perform appropriate longitudinal control under the premise of ensuring a safe distance. Therefore, to ensure driving safety, the following scene poses a great challenge to the collision avoidance system of intelligent vehicles. The current common intelligent vehicle collision avoidance system is composed of a warning reminder system and an automatic emergency braking system triggered by hearing, touch or vision. The present application designs a following scene classification and key quantification model, sets a following scene warning threshold D W and a following scene emergency braking warning threshold D B , and classifies the following scene into three states to quantify the key of the following scene.

[0078] The specific scene classification is shown in Figure 6 When the distance L between the vehicle and the preceding vehicle is greater than D W , this type of scene belongs to a safe scene; when L is greater than D B and less than or equal to D W , a warning signal will be triggered to remind the driver that a vehicle is present in front, and this type of scene has a certain potential risk; when L is less than D B , the vehicle should brake with the maximum braking deceleration as the control target, and this type of scene has a high risk.

[0079] The method for specifically determining D W and D B needs to describe the braking process of the vehicle, as shown in Figure 7 the common braking process deceleration and braking time relationship curve of an intelligent vehicle. The first stage is the sampling stage of the intelligent vehicle perception system, corresponding to τ1; the second stage is the brake braking process, corresponding to τ2, and as the brake pedal moves, the braking acceleration of the vehicle to the maximum braking deceleration a Bmax needs a period of time; the third stage is the stage where the braking acceleration remains unchanged, corresponding to τ3; the fourth stage is the brake recovery process, corresponding to τ4, which refers to the process of gradually recovering to zero after the brake pedal is released.

[0080] Based on the above process, the following scene warning threshold D W and the following scene emergency braking warning threshold D B are calculated:

[0081] D W=D B +v ego ·t w

[0082]

[0083] Where v ego This refers to the vehicle's speed; a Bmax The maximum braking deceleration of this vehicle is determined by filtering within the dataset; f0 represents the minimum detection frequency of the intelligent vehicle target detection module and the ranging module; t w The warning time set for the system.

[0084] Therefore, when quantifying the criticality of the following scenario, for L≤D W The corresponding time t fol-beg As the initial moment in the quantization model of the following scenario, the distance between the vehicle and the vehicle in front gradually decreases from this moment, and the relative speed also decreases until the relative speed reaches 0. Therefore, the moment t corresponding to the relative speed of the two vehicles reaching 0 is defined as follows. fol-end This marks the end of the scenario-critical quantification model.

[0085] 3.2 Behavioral Analysis of Lane Changing Intrusion Scenarios

[0086] Lane intrusion by a preceding vehicle is a common scenario. In such situations, intelligent vehicles not only exhibit longitudinal control behavior but also, when necessary, lateral control to prevent traffic accidents. Therefore, lane intrusion scenarios pose a significant challenge for intelligent vehicles in testing. Based on driver data and research into actual driving experience, experienced drivers, from their perspective, judge the relative position of the preceding vehicle's front steering wheels to the lane lines and the lateral speed at which the preceding vehicle intrudes into their lane. Furthermore, the preceding vehicle may brake during the lane change, requiring the tested vehicle to brake and, in specific lane intrusion scenarios, make necessary steering and lane changes to avoid accidents. This invention employs a decoupled longitudinal and lateral motion analysis model to define the upper and lower bounds of a key quantitative model for lane-changing scenarios involving lane intrusion.

[0087] For lateral decoupling in scenarios where a vehicle intrudes into another vehicle's lane, this invention defines the lane line crossed by the intruding lane when changing lanes to the vehicle's own lane as the intruding lane line, such as... Figure 8 As shown, the lateral intrusion distance q of the preceding vehicle is used as an important parameter to determine the start of a lane change, specifically expressed as:

[0088]

[0089]

[0090] In the formula, The lateral offset distance of the vehicle i axis and the lane axis of the vehicle can be directly obtained in the data set; Width lane is the highway lane line width, which is a constant of 3.75 meters; d is the lateral distance between the vehicle axis and the invading lane line; wherein y i and y j are the horizontal coordinate values of IDs i and j in the data set, respectively; height j is the width of vehicle j; is the heading angle of vehicle j, which is calculated by the angle parameter and the orientation parameter provided in the data set, Since the value is very small in actual calculation, it can be replaced by and ; w b is the wheelbase of the vehicle, and the wheelbase of a small vehicle is defined as 2.5m and the wheelbase of a large truck is defined as 4m in the present application. When q is 0, it is defined as the start time of the front vehicle invasion scene, and when the front vehicle is offset from the lane line center line of the vehicle by less than or equal to 0.1m, it is defined as the end time of the front vehicle invasion scene.

[0091] 4. Key scene extraction:

[0092] The open source data sets commonly used for intelligent vehicle testing and verification are the HighD data set and the China Highway and Expressway Congestion Scene Data Set (Aerial Dataset for China Congested Highway and Expressway, AD4CHE) proposed by DJI. The method proposed in the present patent can process these two data sets to obtain key test scenes. The specific extraction process is as follows: the total trajectory statistical information in the aerial data set is used to determine two important trajectory characteristics, the number of lane changes (numLaneChanges) and the minimum time to collision (minTTC), which are used for the preliminary screening of the front vehicle invasion scene and the following vehicle scene, respectively; for the front vehicle invasion scene, when the number of lane changes is not 0 and there is a rear vehicle after the lane line (LaneId) occupied by the invading vehicle changes, it can be defined as a front vehicle invasion scene, after determining the invading scene, the vehicle trajectory before the LaneId changes is analyzed, the time corresponding to the front vehicle lateral invasion distance q is calculated as the start time of the key quantitative model, and the time corresponding to the invading vehicle and the lane line offset less than or equal to 0.1m is found as the end time of the key quantitative model, and then the key of the front vehicle invasion scene is quantified; for the following vehicle scene, TTC is a commonly used index to express the key of the following vehicle scene, first, scenes where there is no front vehicle are filtered out, and then the upper and lower limits of the key quantitative model are determined using the method in section 3.1 to calculate the key of the following vehicle scene.

[0093] This section mainly describes how to use the method proposed in the application in aerial data set to extract key scenes, and use the method of the application to quantify the key scenes extracted, and compare with the commonly used index TTC to prove the effectiveness of the method proposed in the application. The main steps of extraction are as shown in Figure 9

[0094] In the specific extraction process, the front vehicle lane change intrusion scene and the following vehicle scene are screened by means of the numLaneChanges and minTTC parameters in the data set, and the corresponding scene start time and end time are determined by the method of the previous step; after the scene is extracted, the scene key is quantified by the method of the application, and the output result is as shown in Figure 10 The left coordinate axis in the figure is the scale of the method proposed in the application for scene quantization, and the right coordinate axis is the scale of the traditional TTC-based scene key quantization. When the speed of the host vehicle and the front vehicle is the same, TTC tends to infinity. From the result graph, it can be seen that:

[0095] 1. The method proposed in the application determines the front cut-in scene start point as the most critical scene key moment, and at the same time, the TTC is the smallest, which represents that the scene at this moment is the most dangerous;

[0096] 2. The larger the TTC, the safer the scene, and at the same time, the method proposed in the application represents the safety of the scene when the value is small, and the trend of the two can be compared to be the same;

[0097] 3. At the same time, TTC is not sensitive to lateral motion objects, while the method proposed in the application indicates the key scene in the time dimension before TTC, which shows that the method of the application is more sensitive to lateral motion objects.​

Claims

1. A method for aerial data set oriented intelligent vehicle key test scene extraction, characterized in that: The steps are as follows: (1) The key analysis of the position distribution of vehicles around the ego vehicle determines the boundary, i.e. the range domain [θ1, θ2] of the test scene concerned: n ] Eight kinds of position information relative to the vehicle are extracted in the aerial data set, namely left front, front, right front, left side, right side, left rear, front and right rear; In order to describe the scene key, the polar coordinate system is introduced, the bird's eye view two-dimensional perspective in the aerial data set is converted to one-dimensional feature space, the vehicle effective contour distance curve is used to describe the position distribution of the possible collision part of the surrounding vehicles relative to the vehicle, the centroid of the vehicle is taken as the origin of the curve, the road direction of the vehicle forward movement is taken as the x axis, and the clockwise direction is taken as positive, the contour information of the surrounding vehicles within 360° of the vehicle is mapped to the polar coordinate system; The coordinate transformation is performed on the position information, taking the left front end of the circumscribed rectangle frame of vehicle j as an example: In the formula, s j is the horizontal coordinate after rotation transformation during the lane-changing process of vehicle j; t j is the vertical coordinate after rotation transformation during the lane-changing process of vehicle j; θ1 is the angle between the line connecting the centroid of the host vehicle i and the left front end of the target vehicle j and the forward direction of the host vehicle; d1 is the distance between the centroid of the host vehicle i and the left front end of the target vehicle j; The outer rectangle frame of the vehicle j is transformed to other position information coordinates; all the angles and distances of the possible collision parts relative to the main vehicle i jointly constitute θ ij and r ij , and is input as the field intensity formula in the second section; The part relative to the vehicle that may collide is selected as the collection area of CDCE, and the area where the collision may occur is scanned from the centroid of vehicle i, and the intersection of the scanning ray and the circumscribed rectangle frame of the other vehicle is recorded in the polar coordinate system as the collection point; (2) Scene key quantification based on field strength theory: Based on the risk field theory of complex traffic conditions, the key quantitative model of vehicle behavior in the scene is established, and the field strength n formed by the ego vehicle i and other vehicles j at time t is defined as: where M i and M j are equivalent masses, whose values are related to the vehicle type, mass, and are set as specific constants; R is a road condition parameter, set as a specific constant; k1, k2, and G are all constants greater than 0; θ ij is the angle between the line connecting the mass center of the host vehicle i and each point on the possible collision profile of the target vehicle j and the forward direction of the host vehicle, and is calculated in the same way as θ defined in the previous step; v ij is the relative speed between the two vehicles, and the specific expression is The relative speeds of each point on the same vehicle outer frame line are equal; it can be seen that the greater the field strength, the higher the scene criticality, and at the t n moment, the scene criticality is represented by the integral within the range of θ1and θ n . Since R i , M j , and M i are all constants with respect to θ at the same instant, the above equation simplifies to: Then, from t1 to t N The scene significance at the moment is determined by summing up each timestamp, and the specific expression is: (3) Scene key behavior analysis, determine the scene key quantity model of time domain [t1, t N ]: The two scenes of following a car and the front car invading the lane line are analyzed respectively, and the time domain [t1, t N ] of the key scene quantification model is determined, while the range domain [θ1, θ n ] needs to be determined by the position distribution in the first step, and the corresponding scenes in the aerial data set are quantified, and then the scenes with test value are screened out for application; (3.1) Analysis of following vehicle scene behavior: A follow-up scene classification key importance quantification model is designed, and a follow-up scene warning threshold D is set W and a follow-up scene emergency braking warning threshold D B The follow-up scene is screened into three states, and the key importance of the follow-up scene is quantified; When the distance L between the host vehicle and the preceding vehicle is greater than D W , such a scenario belongs to a safe scenario; when L is greater than D B and less than or equal to D W , a pre-warning signal is triggered to remind the driver that a vehicle is present in front, and such a scenario has a certain potential risk; when L is less than D B , the host vehicle should brake at the maximum braking deceleration as the control target, and such a scenario has a high risk; Calculating a following scenario early warning threshold value D based on a vehicle braking process W and an emergency braking following scenario early warning threshold value D B : D W = D B + v ego · t w where v ego is the vehicle speed; a Nmax is the maximum braking deceleration of the vehicle, selected from the data set; f0is the minimum detection frequency of the intelligent vehicle target detection module and the distance measurement module; t w is the warning time set by the system; Therefore, when the scene key is quantified for the following vehicle scene, L≤D W The time t fol-beg As the initial time of the following vehicle scene quantification model, the distance between the host vehicle and the preceding vehicle will gradually decrease, and the relative speed will also decrease until the relative speed is 0. Therefore, the time t fol-end As the end time of the scene key quantification model; (3.2) Analysis of front vehicle lane changing intrusion scene behavior The upper and lower bounds of the lane changing scene key quantification model are analyzed by decoupling the front vehicle intrusion lane changing scene; For the lateral decoupling of the front vehicle intrusion lane changing scene, the lane line that is pressed by the intrusion lane when changing to the vehicle lane becomes the intrusion lane line, and the front vehicle lateral intrusion distance q is an important parameter for determining the start of lane changing, which is specifically expressed as: wherein, is the lateral offset distance between the vehicle i's centerline and the lane centerline, which is directly obtained from the dataset; Width lane is the lane line width; d is the lateral distance between the vehicle i's centerline and the invading lane line; wherein, y i and y j are the x-coordinate values of the data set with ID i and j, respectively; height j is the width of the vehicle j; is the heading angle of the vehicle j, which is calculated by the angle parameter and the orientation parameter provided in the data set, w b is the wheelbase of the vehicle; q is defined as the start time of the front vehicle invasion scene when it is 0, and the end time of the front vehicle invasion scene when the offset between the current vehicle and the lane line centerline is less than or equal to 0.1 m. (4) Key scene extraction: The extraction process is as follows: two important trajectory feature quantities, vehicle lane changing frequency and minimum collision time, are determined by the overall trajectory statistical information in the aerial data set, which are used for the preliminary screening of the front vehicle intrusion scene and the following vehicle scene; For the front vehicle intrusion scene, when the lane changing frequency is not 0 and there is a rear vehicle after the change of the lane line occupied by the intrusion vehicle, it can be defined as a front vehicle intrusion scene, the time corresponding to the front vehicle lateral intrusion distance q is calculated as the starting time of the key quantification model after analyzing the vehicle trajectory before the change of the lane line, the time corresponding to the lane line offset of the intrusion vehicle less than or equal to 0.1 m is found as the ending time of the key quantification model, and then the key of the front vehicle intrusion scene is quantified; For the following vehicle scene, TTC is a commonly used index to express the key of the following vehicle scene, first, the scene where the front vehicle does not exist is filtered out, and then the upper and lower limits of the key quantification model are determined by the method in section 3.1 to calculate the key of the following vehicle scene. 2.The aerial data set oriented intelligent vehicle key test scene extraction method according to claim 1, characterized in that: The aerial photography data refers to collecting and extracting feature data of road vehicles from an overlooking angle through a UAV device; the aerial photography data set extracts each frame of motion feature of the vehicle through image recognition to obtain basic vehicle motion features, including position, shape feature, transverse and longitudinal speed, yaw angular velocity, offset of the center point from the current lane center line, distance from the front vehicle and position vehicle serial number information in eight directions, and at the same time, abstract representation parameters are calculated, including headway and collision time.

3. The aerial data set oriented intelligent vehicle key test scene extraction method according to claim 1, characterized in that: In step 3.1, the deceleration and braking time relationship of the braking process includes: the first stage is the intelligent vehicle perception system sampling stage, corresponding to τ1; the second stage is the brake braking process, corresponding to τ2, and as the brake pedal moves, the brake gap is generated. The vehicle produces a braking acceleration to the maximum braking deceleration a Bmax It takes a period of time; the third stage is the stage where the braking acceleration remains unchanged, corresponding to τ3; the fourth stage is the brake recovery process, corresponding to τ4, which refers to the process of gradually recovering to zero after the brake pedal is released.

Images