Automatic extraction method and device for vehicle dangerous driving scene

Abstract

This invention belongs to the field of intelligent driving technology and discloses an automatic extraction method and device for dangerous driving scenarios of vehicles. When the lateral acceleration or longitudinal deceleration of a vehicle exceeds a danger threshold, a pre-extraction request with a timestamp is generated. Subsequently, static road semantic elements are extracted by backtracking forward visual perception data, and a predicted wheel track is projected based on the vehicle's current speed and heading angle. Semantic elements that spatially overlap with this track are activated as valid environment masks. Next, the physical positional relationship between the actual movement trajectory of the vehicle that triggered the pre-extraction request and the valid environment mask is calculated to determine the spatiotemporal logical consistency. If a consistency is determined, the current situation is confirmed as a legitimate terrain avoidance operation and the extraction is canceled; otherwise, data from the corresponding time period is extracted in a package. This invention effectively filters out false dangerous scenarios caused by reasonable terrain avoidance, avoids the pollution of the autonomous driving training set by redundant data, and reduces the cost of manual cleaning and labeling.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, specifically to a method and apparatus for automatically extracting dangerous driving scenarios of vehicles. Background Technology

[0002] With the rapid development of autonomous driving technology and Advanced Driver Assistance Systems (ADAS), utilizing massive amounts of natural driving data for algorithm training and system closed-loop verification has become an industry consensus. Among these, dangerous driving scenarios (such as emergency braking and emergency avoidance) contain rich information on the interaction between humans, vehicles, and the environment, representing high-value data for improving the ability of autonomous driving systems to handle extreme situations. Therefore, how to automatically and accurately extract highly valuable dangerous driving scenario fragments from continuous daily vehicle driving data is a fundamental and critical technical requirement in the field of autonomous driving data.

[0003] Currently, the industry generally adopts a threshold triggering mechanism based on vehicle dynamics parameters to achieve automatic extraction of dangerous scenarios. Specifically, existing technologies typically rely on onboard data acquisition terminals to monitor the vehicle's CAN bus data in real time. When a sudden change in dynamic indicators such as longitudinal deceleration, lateral acceleration, or steering wheel angle change rate is detected and exceeds a pre-set dangerous physical threshold, it is determined that a dangerous driving behavior has occurred, thereby triggering a data interception command. The video and sensor data within a certain time window before and after the triggering time are packaged and saved as a dangerous scenario.

[0004] However, in complex real-world road environments, a vehicle's violent movements do not always signify a dangerous traffic situation or loss of control. For example, when drivers encounter potholes, damaged road surfaces, traffic cones placed for construction, or legally mandated speed bumps, they often take actions such as sudden braking or sharp steering to avoid the terrain. In these situations, the vehicle's dynamic parameters may momentarily exceed preset danger thresholds, but this is essentially a legal and normal operation taken by the driver to ensure smooth and safe driving. Existing extraction logic, lacking the perception and integration of the semantics of the external environment, cannot establish a connection between the vehicle's own motion state and external static road conditions. Consequently, it cannot distinguish between emergency actions to deal with sudden dangers and legitimate avoidance actions to deal with static terrain, leading to frequent misjudgments and extraction of the latter as dangerous scenarios.

[0005] The redundant data corresponding to such pseudo-dangerous scenarios will seriously pollute the training set for autonomous driving, significantly increasing the cost of manual data cleaning and scene labeling in the cloud, which urgently needs improvement. Summary of the Invention

[0006] In view of the above-mentioned shortcomings mentioned in the background art, the purpose of this invention is to provide an automatic extraction method and device for dangerous driving scenarios of vehicles.

[0007] This invention provides an automatic extraction method for dangerous driving scenarios of vehicles, the method comprising the following steps: Real-time acquisition of vehicle dynamics data; when the lateral acceleration or longitudinal deceleration in the dynamics data exceeds a preset danger threshold, a scene pre-extraction request with the current timestamp is generated. In response to the scene pre-extraction request, forward visual perception data within a preset time window before the current timestamp is retrieved to extract static road semantic elements, and a predicted wheel track is projected forward based on the vehicle's current speed and heading angle. The static road semantic elements that spatially overlap with the predicted wheel track are activated as valid environment masks. The actual motion trajectory of the vehicle that triggered the scene pre-extraction request is obtained and its physical positional relationship with the effective environment mask is calculated to determine whether the motion change characteristics of the vehicle at this time are consistent with the spatiotemporal logic of avoiding the effective environment mask. If the scenario is determined to be a match, the vehicle's current behavior is confirmed as a legitimate terrain avoidance operation, the scenario pre-extraction request is revoked, and data extraction is suppressed; if the scenario is determined to be a mismatch or the effective environment mask is not activated, the scenario pre-extraction request is executed, and data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario.

[0008] The present invention also provides an automatic extraction device for dangerous driving scenarios of vehicles, the device comprising: The pre-extraction request generation module is used to acquire vehicle dynamics data in real time. When the lateral acceleration or longitudinal deceleration in the dynamics data is detected to exceed a preset danger threshold, a scene pre-extraction request with the current timestamp is generated. The environment mask activation module is used to respond to the scene pre-extraction request, backtrack the forward visual perception data within a preset time window before the current timestamp to extract static road semantic elements, and project the predicted wheel track band forward based on the vehicle's current speed and heading angle, and activate the static road semantic elements that spatially overlap with the predicted wheel track band as valid environment masks. The consistency determination module is used to obtain the actual motion trajectory of the vehicle that triggers the scene pre-extraction request and calculate its physical positional relationship with the effective environment mask, so as to determine whether the motion change characteristics of the vehicle at this time are consistent with the spatiotemporal logic of avoiding the effective environment mask. The scene extraction and execution module is used to confirm that the vehicle's current behavior is a legal terrain avoidance operation if the scene is determined to match, cancel the scene pre-extraction request and suppress data extraction; if the scene is determined to not match or the effective environment mask is not activated, the scene pre-extraction request is executed, and data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scene.

[0009] After an abnormal vehicle dynamics index triggers a pre-extraction request, this invention constructs an effective environment mask based on static road semantic elements and performs spatiotemporal logic consistency verification in conjunction with the vehicle's actual movement trajectory. This enables accurate differentiation between emergency operations involving real sudden dangers and legitimate driving behaviors that avoid static terrain (such as speed bumps, potholes, etc.). Therefore, this invention effectively filters out and intercepts false triggers of dangerous scenarios, significantly saving extremely limited storage space on the vehicle and bandwidth for vehicle-to-cloud transmission. It also avoids the contamination of the autonomous driving algorithm training set by massive amounts of redundant data, thereby significantly reducing the cost of subsequent manual data cleaning and scene labeling in the cloud, and effectively improving the operational efficiency and data quality of the autonomous driving data closed-loop system. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an automatic extraction method for dangerous driving scenarios of a vehicle, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the process for determining whether the characteristics of vehicle motion changes match the spatiotemporal logic of avoiding an effective environmental mask, as disclosed in an embodiment of the present invention. Figure 3 This is a schematic diagram of the further processing flow for determining the spatiotemporal logic consistency as disclosed in the embodiments of the present invention; Figure 4 This is a schematic diagram of the structure of an automatic extraction device for dangerous driving scenarios of a vehicle, as disclosed in an embodiment of the present invention. Detailed Implementation

[0011] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0012] This embodiment can be applied to the process of a vehicle driving in a real road environment. During the driving process, the vehicle may actively take emergency braking or emergency steering actions due to the presence of static road semantic elements ahead, which may cause the lateral acceleration or longitudinal deceleration in the vehicle dynamics data to exceed a preset danger threshold. It should be understood that this scenario is only used to illustrate the pseudo-dangerous scenario background faced by this application, and does not limit the specific vehicle type, vehicle speed, road type, type and number of static road semantic elements, nor does it limit the specific network structure of the visual perception model.

[0013] It should be noted that the core of the technical solution of this invention lies in: after the dynamic index abnormally triggers the pre-extraction request, it further combines the static road semantic elements in the forward visual perception data, the vehicle's predicted wheel track and the vehicle's actual movement trajectory to verify whether the vehicle's movement change belongs to a legitimate operation to avoid static terrain, and retains high-value dangerous scenarios of excessive defense through avoidance intensity verification when necessary.

[0014] Reference Figure 1 As shown, this embodiment of the invention provides an automatic extraction method for dangerous driving scenarios of vehicles, the method including the following steps S1 to S4.

[0015] S1: Real-time acquisition of vehicle dynamics data; when the lateral acceleration or longitudinal deceleration in the dynamics data exceeds a preset danger threshold, a scene pre-extraction request with the current timestamp is generated. The onboard extraction terminal acquires vehicle dynamics data generated during vehicle operation in real time. This data can originate from at least one of the following: vehicle CAN bus, inertial measurement unit, chassis controller, steering system, or braking system. It includes at least one of the following: current vehicle speed, heading angle, lateral acceleration, longitudinal deceleration, steering wheel angle, and rate of change of steering wheel angle. Lateral acceleration characterizes the intensity of vehicle movement during steering or lateral motion, while longitudinal deceleration characterizes the intensity of vehicle braking or deceleration.

[0016] Pre-defined hazard thresholds can include lateral acceleration hazard thresholds and longitudinal deceleration hazard thresholds. These thresholds can be calibrated based on vehicle model, road conditions, sensor configuration, data acquisition requirements, and driving style, and can be the same or different. When the lateral acceleration in the vehicle dynamics data exceeds the preset lateral acceleration hazard threshold, or the longitudinal deceleration exceeds the preset longitudinal deceleration hazard threshold, it indicates that the vehicle may have experienced a sudden dynamic change such as emergency steering or emergency braking. In this case, instead of immediately packaging long data segments or directly identifying the current moment as a dangerous driving scenario, a scenario pre-extraction request with the current timestamp is generated.

[0017] The current timestamp is used to identify the moment when an anomaly occurs in the vehicle dynamics data and triggers the pre-extraction request; for example, it can be denoted as T0. It can be understood that this current timestamp can serve as a reference moment for subsequent backtracking of forward visual perception data, or as a reference moment for the final extraction of data for the corresponding time period. The scene pre-extraction request must at least carry this current timestamp.

[0018] As an example, the scene pre-extraction request may also carry at least one of the following: trigger type identifier, lateral acceleration or longitudinal deceleration at the time of triggering, current vehicle speed, heading angle, etc. The trigger type identifier is used to distinguish whether the scene pre-extraction request is triggered by lateral acceleration exceeding a danger threshold or by longitudinal deceleration exceeding a danger threshold.

[0019] S2: In response to the scene pre-extraction request, backtrack the forward visual perception data within the preset time window before the current timestamp to extract static road semantic elements, and project the predicted wheel track band forward based on the vehicle's current speed and heading angle, and activate the static road semantic elements that spatially overlap with the predicted wheel track band as valid environment masks. After generating a scene pre-extraction request with the current timestamp, respond to the scene pre-extraction request and perform environment semantic verification.

[0020] First, the forward visual perception data within a preset time window prior to the current timestamp is reviewed. This preset time window can be set based on the perception data frame rate, vehicle motion state, data caching capacity, and data acquisition requirements. Furthermore, the forward visual perception data can come from image sequences or video frames captured by the forward-facing camera, or from other visual perception devices that can provide information about the road environment ahead of the vehicle. By reviewing the forward visual perception data prior to the current timestamp, environmental information about the road area that the vehicle has already passed or is about to pass before the dynamic change occurred can be obtained, thus providing a basis for determining whether the dynamic change was caused by static terrain ahead.

[0021] When extracting static road semantic elements, these elements can include at least one of the following: speed bumps, potholes, water accumulation areas, and construction barriers. It is understood that these static road semantic elements are relatively fixed terrain features or obstacles in the road environment that do not move rapidly over time, and may prompt drivers to take evasive actions such as emergency braking or emergency steering.

[0022] As an example, forward visual perception data within a preset time window prior to the current timestamp is used to extract static road semantic elements, including: S21: Using a preset visual perception model, target detection and semantic segmentation are performed on the forward visual perception data within a preset time window before the current timestamp obtained by backtracking, to determine the category of the static road semantic element and the pixel coordinates of the static road semantic element in the forward visual perception data. In practice, object detection and semantic segmentation can be performed using a pre-built visual perception model. This model can employ a cascaded or jointly trained object detection and semantic segmentation network, or it can utilize other visual perception models capable of simultaneously outputting object categories and pixel-level regions. Object detection is used to locate potential static road semantic elements in the backtracked forward visual perception data and determine their category. Semantic segmentation is used to determine the pixel regions occupied by these elements in the image. The categories can include at least one of speed bumps, potholes, puddles, and construction barriers. Pixel regions can be represented by bounding boxes, center points, segmentation masks, etc., thereby obtaining the pixel coordinates of the static road semantic elements in the forward visual perception data.

[0023] In practical implementation, target detection and semantic segmentation can be performed on multiple frames of images within a preset time window, and the detection results of multiple frames can be temporally correlated, voted, or fused to suppress false detections and missed detections in a single frame and improve the stability of static road semantic element recognition and localization.

[0024] S22: Based on the intrinsic and extrinsic calibration parameters of the vehicle-mounted camera, the pixel coordinates are inversely projected onto the local coordinate system of the vehicle to obtain the actual physical distance and two-dimensional contour of the static road semantic element relative to the vehicle.

[0025] After obtaining the pixel coordinates of the static road semantic elements, the pixel coordinates in the image plane are further transformed to the vehicle local coordinate system using the intrinsic and extrinsic calibration parameters of the vehicle-mounted camera. These intrinsic and extrinsic calibration parameters may include at least one of the following: camera intrinsic parameters, distortion coefficients, camera extrinsic parameters, and the rotation and translation relationship of the camera relative to the vehicle coordinate system.

[0026] The transformation process can incorporate a ground plane assumption, that is, assuming that static road semantic elements are located near the ground where the vehicle is traveling, thereby mapping pixels in the 2D image to the local 2D plane or 3D space of the vehicle. Through this inverse projection, the actual physical distance and 2D contour of the static road semantic elements relative to the vehicle can be obtained. The actual physical distance can include at least one of longitudinal distance, lateral distance, or relative orientation; the 2D contour can include at least one of width, length, area, or boundary shape. Optionally, the inverse projection results of multiple frames can be fused or filtered to reduce single-frame perceptual errors.

[0027] Through the above processing, the position and size of static road semantic elements are unified under the vehicle local coordinate system for subsequent spatial overlap determination with the predicted wheel track zone.

[0028] Simultaneously or subsequently, based on the vehicle's current speed and heading angle, a predicted wheel track zone is projected forward. As an example, this projection of the predicted wheel track zone based on the vehicle's current speed and heading angle includes: S23: Obtain the vehicle's inherent track width parameter, and calculate the vehicle's predicted driving centerline based on the current vehicle speed and the preset prediction time length, starting from the vehicle's current position and extending along the direction of the heading angle. In practice, the inherent wheelbase parameter of the vehicle, which characterizes the wheelbase between the left and right wheels on the same axle, is obtained. This parameter can be used as the basis for subsequently determining the width of the predicted wheel track band. This inherent wheelbase parameter can be pre-stored in the vehicle parameter library or provided in real time by the chassis system.

[0029] Starting from the vehicle's current position and following the heading angle, the predicted driving centerline of the vehicle is calculated based on the current vehicle speed and a preset prediction time. The current position can be understood as the position corresponding to the origin of the vehicle's local coordinate system, and the heading angle can be derived from the aforementioned vehicle dynamics data. The preset prediction time can be set according to vehicle speed, sensing distance, data processing capabilities, and data acquisition requirements; it is used to limit the time range for forward prediction.

[0030] In practice, the current vehicle speed can be multiplied by a preset prediction time length to obtain the longitudinal distance the vehicle may travel within that prediction time length. This longitudinal distance is then extended along the heading angle to obtain the predicted driving centerline. If the vehicle maintains its current motion, the predicted driving centerline can be a straight line extending along the heading angle. If the vehicle exhibits a continuous turning trend, the predicted driving centerline can be corrected using a vehicle kinematics model to present a curve that matches the vehicle's future driving trend. Essentially, the predicted driving centerline represents the central path along which the vehicle is most likely to travel in the future.

[0031] S24: Using the predicted driving centerline as a reference, expand outwards to the left and right sides by half of the inherent wheel track parameter as boundaries, generate a strip-shaped polygonal region on a local two-dimensional plane, and use the polygonal region as the predicted wheel track band.

[0032] In practice, sampling can be performed along the predicted driving centerline at preset intervals to obtain multiple center points. At each center point, the normal direction of the predicted driving centerline is determined, and the centerline is offset to the left and right by half of the inherent wheelbase parameter, respectively, to obtain the corresponding left and right boundary points. The left boundary points are connected sequentially to form the left boundary line, and the right boundary points are connected sequentially to form the right boundary line. The left and right boundary lines are then closed at the start and end points of the predicted driving centerline, thus forming a strip-shaped polygonal region on a local two-dimensional plane, which is the predicted wheel track band. It can be understood that this predicted wheel track band is used to characterize the road surface area that the vehicle is most likely to traverse in the future, its width matching the vehicle's wheelbase, and its extension direction matching the vehicle's current heading angle.

[0033] After obtaining the predicted wheel track, the extracted static road semantic elements are spatially overlapped with the predicted wheel track. Specifically, the actual physical distance and two-dimensional contour of the static road semantic elements relative to the vehicle can be mapped onto the same local two-dimensional plane, and it is determined whether at least part of the contour of the static road semantic element falls within the predicted wheel track or intersects with the boundary of the predicted wheel track. If a static road semantic element falls within the predicted wheel track, or its two-dimensional contour partially overlaps with the predicted wheel track, it is determined that the static road semantic element has spatial overlap with the predicted wheel track and is activated as a valid environment mask. If a static road semantic element is only located in an adjacent lane, on the roadside, or in other areas that are not necessarily traversed by the vehicle and does not have spatial overlap with the predicted wheel track, it is not activated as a valid environment mask. If multiple static road semantic elements have spatial overlap with the predicted wheel track, they can be activated as valid environment masks respectively, or one or more of them that are closest to the vehicle, have the highest risk, or have the greatest overlap with the vehicle's predicted driving path can be selected as valid environment masks according to actual needs.

[0034] By employing the above method, static road semantic elements such as speed bumps, potholes, or construction barriers in adjacent lanes can be avoided from causing the vehicle to falsely trigger filtering, thereby improving the accuracy of environmental semantic verification. It can be understood that the effective environmental mask can be interpreted as a static road semantic element confirmed to be located within the vehicle's predicted wheel track zone in the forward visual perception data. This element is used in subsequent step S3 to determine whether the vehicle's actual trajectory matches the spatiotemporal logic of avoiding the static road semantic element.

[0035] It should be noted that if no valid environment mask is activated during the aforementioned process, it indicates that the abrupt change in vehicle dynamics is not caused by static road semantic elements located within the vehicle's predicted wheel track band. In this case, the subsequent step S4 will directly execute the scene pre-extraction request and extract data for the corresponding time period based on the current timestamp, packaging it into a dangerous driving scene.

[0036] S3: Obtain the actual motion trajectory of the vehicle that triggers the scene pre-extraction request and calculate its physical positional relationship with the effective environment mask, so as to determine whether the motion change characteristics of the vehicle at this time are consistent with the spatiotemporal logic of avoiding the effective environment mask; In this step, after activating a valid environment mask, a spatiotemporal logic matching determination is further performed. It should be noted that if no valid environment mask is activated, the matching determination in this step is unnecessary, and the process directly proceeds to step S4 to execute the scene pre-extraction request.

[0037] If a valid environment mask has been activated, the actual vehicle trajectory triggered by the pre-extraction request is obtained. This actual vehicle trajectory can be calculated based on at least one of the following: vehicle dynamics data, inertial measurement unit data, global positioning system data, visual odometry data, or chassis kinematics model. The actual vehicle trajectory may include at least one of the following: vehicle position sequence, heading angle change, longitudinal displacement, and lateral displacement over a period of time before and after the triggering moment.

[0038] After obtaining the vehicle's actual trajectory, the physical positional relationship between it and the effective environment mask is calculated. Specifically, the vehicle's actual trajectory and the effective environment mask can be mapped to the same vehicle local coordinate system or the same world coordinate system, and at least one of the following is calculated: longitudinal distance, lateral offset, relative orientation, and spatial overlap between the trajectory and the mask outline. Based on the above physical positional relationship, it is determined whether the vehicle's current motion change characteristics are consistent with the spatiotemporal logic of avoiding the effective environment mask. Specifically, spatiotemporal logic consistency can be understood as follows: the vehicle's motion change occurs in time before approaching the effective environment mask, spatially in the direction of avoiding the effective environment mask, and the avoidance magnitude is sufficient to avoid spatial overlap with the effective environment mask.

[0039] As an example, refer to Figure 2 As shown, determining whether the vehicle's current motion characteristics match the spatiotemporal logic of avoiding the effective environmental mask includes: S31: If the scene pre-extraction request is triggered by the longitudinal deceleration exceeding the danger threshold, calculate the longitudinal braking distance between the starting position of the vehicle when the longitudinal deceleration exceeds the danger threshold and the effective environment mask. When the longitudinal braking distance is within a reasonable braking range that matches the current vehicle speed, determine that the spatiotemporal logic is consistent. In specific implementation, for the aforementioned scenarios, the trigger type identifier carried by the scene pre-extraction request can be used to identify whether the pre-extraction request was triggered by longitudinal deceleration exceeding a danger threshold. In this case, the starting position of the vehicle when the longitudinal deceleration exceeds the danger threshold is obtained, and the longitudinal braking distance between this starting position and the effective environment mask is calculated. The longitudinal braking distance characterizes the longitudinal distance required for the vehicle to travel from the start of emergency braking to the location of the effective environment mask.

[0040] Next, the calculated longitudinal braking distance is compared with a reasonable braking range matching the current vehicle speed. This reasonable braking range can be pre-defined based on at least one of the following factors: current vehicle speed, road surface adhesion coefficient, driver's average reaction time, and vehicle braking performance. If the longitudinal braking distance falls within this reasonable braking range, it indicates that the driver's braking behavior matches the semantic elements of the static road ahead in both timing and distance, and it can be determined that the vehicle's motion change characteristics are consistent with the spatiotemporal logic of the effective environment mask for avoidance. If the longitudinal braking distance is not within this reasonable braking range, it is determined to be inconsistent.

[0041] S32: If the scene pre-extraction request is triggered by the lateral acceleration exceeding the danger threshold, calculate the maximum lateral offset of the vehicle's actual trajectory. When the sum of the maximum lateral offset and the width parameter of the effective environment mask is greater than or equal to the safety avoidance threshold, determine that the spatiotemporal logic is consistent.

[0042] In practical implementation, for the aforementioned scenarios, the trigger type identifier can also be used to identify that the pre-extraction request was triggered by lateral acceleration exceeding a danger threshold. In this case, the maximum lateral offset is calculated based on the vehicle's actual trajectory. The maximum lateral offset characterizes the maximum lateral displacement of the vehicle relative to its original direction of travel within a certain period before and after the trigger moment.

[0043] The width parameter of the effective environmental mask is further obtained, and the sum of the maximum lateral offset and this width parameter is compared with a safe avoidance threshold. The safe avoidance threshold can be calibrated based on at least one of the following factors: vehicle width, lane width, avoidance margin, and current vehicle speed. When the sum of the maximum lateral offset and the width parameter is greater than or equal to the safe avoidance threshold, it indicates that the vehicle's lateral avoidance range is sufficient to avoid the effective environmental mask, and it can be determined that the vehicle's motion change characteristics are consistent with the spatiotemporal logic of avoiding the effective environmental mask; otherwise, it is determined that they are inconsistent.

[0044] Understandably, this step can establish a correlation between the vehicle's actual trajectory and the effective environment mask in both time and space, thereby distinguishing whether a sudden change in vehicle dynamics is a genuine evasion maneuver against the effective environment mask or another abnormal action unrelated to the effective environment mask.

[0045] It should be noted that if a scene pre-extraction request is triggered simultaneously by both lateral acceleration and longitudinal deceleration exceeding the danger threshold, S31 or S32 can be executed according to the preset priority, or S31 and S32 can be executed separately and a comprehensive judgment can be made. This embodiment does not limit this. Furthermore, parameters such as reasonable braking range and safe avoidance threshold can be calibrated according to vehicle type, road conditions, sensor configuration, and data acquisition requirements.

[0046] S4: If the match is determined, the current behavior of the vehicle is confirmed as a legal terrain avoidance operation, the scene pre-extraction request is cancelled and data extraction is suppressed; if the match is determined or the effective environment mask is not activated, the scene pre-extraction request is executed, and the data of the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scene.

[0047] In this step, after determining whether the vehicle's motion change characteristics match the spatiotemporal logic of the effective environmental mask for avoidance, different data extraction decisions are further executed based on the determination results. Specifically: If the result is deemed inconsistent, or if no valid environment mask is activated in step S2, it indicates that the sudden change in vehicle dynamics is not caused by static road semantic elements located within the vehicle's predicted wheel track zone, and may belong to a real sudden danger, driver error, or other dangerous game scenario. In this case, the previously generated scenario pre-extraction request is executed, and data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario. The corresponding time period can be set to a certain number of seconds before and after the current timestamp according to data collection requirements; this embodiment does not impose such a limitation.

[0048] If the match is determined, the vehicle's current behavior is preliminarily confirmed as a legitimate terrain avoidance maneuver, i.e., an emergency braking or emergency steering maneuver taken by the driver to avoid a valid environmental mask ahead. However, to avoid incorrectly filtering out high-value dangerous data where the driver's excessive panic leads to an overly large avoidance maneuver and the vehicle is on the verge of losing control, this embodiment does not immediately and unconditionally cancel the pre-extraction request after a match is determined. Instead, it further performs an avoidance intensity check to distinguish between legitimate and restrained terrain avoidance maneuvers and dangerous avoidance maneuvers that are excessively defensive.

[0049] As an example, refer to Figure 3 As shown, if the match is determined, the vehicle's current behavior is confirmed as a legitimate terrain avoidance operation, the scene pre-extraction request is revoked, and data extraction is suppressed, including: S41: If the match is determined, obtain the physical size parameters of the effective environment mask in the forward visual perception data, and determine the minimum avoidance requirement corresponding to the effective environment mask based on the physical size parameters. In practice, after determining that the spatiotemporal logic matches, the size of the effective environment mask in the physical world is further determined. The physical size parameters may include at least one of the following: width, length, height, area, or outline size of the effective environment mask, which can be determined based on the two-dimensional outline obtained by the aforementioned inverse projection.

[0050] As an example, the minimum avoidance requirement can be determined through a pre-defined size-avoidance requirement mapping relationship. This mapping relationship can be pre-stored in the form of a calibration table or mapping curve. Its input is the physical size parameters of the effective environment mask, and its output is the minimum avoidance requirement corresponding to that effective environment mask. For example, for lateral avoidance scenarios, when the width of the effective environment mask is 0.3 meters, the corresponding minimum lateral offset can be set to 0.4 meters; when the width of the effective environment mask is 0.5 meters, the corresponding minimum lateral offset can be set to 0.6 meters; and when the width of the effective environment mask is 0.8 meters, the corresponding minimum lateral offset can be set to 0.9 meters. It should be understood that the above values ​​are only used to illustrate the positive correlation between width and minimum lateral offset; the specific values ​​can be determined based on vehicle width, tire width, safety margin, and calibration test results.

[0051] As another example, in longitudinal braking scenarios, when the effective environmental mask is a high speed bump, the vehicle may experience a significant vertical impact when passing over it. In this case, the minimum avoidance requirement can be reflected as the minimum longitudinal deceleration. For example, for a speed bump with a height of 0.05 meters, the corresponding minimum longitudinal deceleration can be set to 0.2g; for a speed bump with a height of 0.1 meters, the corresponding minimum longitudinal deceleration can be set to 0.3g. It should be understood that the above values ​​are only used to illustrate the positive correlation between height and minimum longitudinal deceleration; the specific values ​​can be determined based on suspension characteristics, tire characteristics, and ride comfort requirements.

[0052] In practical implementation, if the two-dimensional outline of the effective environment mask is irregular in shape, its circumscribed rectangle or equivalent outline size can be used as the physical size parameter. If multiple effective environment masks exist simultaneously, the minimum avoidance requirement corresponding to each effective environment mask can be determined separately, and the maximum value among them can be used as the basis for the reasonable upper limit of the subsequent mapping security response.

[0053] S42: Based on the minimum avoidance requirement and the current vehicle speed, map out a reasonable safety response upper limit for the effective environment mask, extract the action extreme value from the dynamic data that triggers the scene pre-extraction request, and compare the action extreme value with the reasonable safety response upper limit; In practice, after obtaining the minimum avoidance requirement, a reasonable safe response limit is determined based on the current vehicle speed and the effective environmental mask. This reasonable safe response limit characterizes the range of reasonable actions a driver must take to achieve a reasonable avoidance under the current vehicle speed and effective environmental mask conditions. It may include at least one of the following: lateral offset limit, steering wheel angle change rate limit, and longitudinal deceleration limit.

[0054] The correspondence between the upper limit of a reasonable safe response, the minimum avoidance requirement, and the current vehicle speed can be determined through a pre-calibrated mapping table, mapping curve, or lookup table interpolation. For example, under the same minimum avoidance requirement, the higher the current vehicle speed, the higher the allowable upper limit of a reasonable safe response can be adjusted accordingly; conversely, under the same current vehicle speed, the greater the minimum avoidance requirement, the higher the allowable upper limit of a reasonable safe response can also be adjusted accordingly. It is understood that the specific mapping relationship can be calibrated based on vehicle type, road conditions, sensor configuration, and data acquisition requirements; this embodiment does not impose such limitations.

[0055] As an example, the reasonable safe response upper limit can be determined through a pre-defined two-dimensional mapping table. This table takes the minimum avoidance requirement and the current vehicle speed as input and the reasonable safe response upper limit as output. For example, in lateral avoidance scenarios, when the minimum avoidance requirement is a minimum lateral offset of 0.6 meters and the current vehicle speed is 40 km / h, the corresponding lateral offset upper limit can be set to 1.2 meters; when the current vehicle speed is 60 km / h, the corresponding lateral offset upper limit can be set to 1.5 meters; and when the current vehicle speed is 80 km / h, the corresponding lateral offset upper limit can be set to 1.8 meters. It should be understood that the above values ​​are only used to illustrate the trend that, under the same minimum avoidance requirement, the higher the vehicle speed, the wider the allowable lateral offset upper limit. The specific values ​​can be determined based on the vehicle's handling characteristics, tire adhesion characteristics, and calibration test results.

[0056] As another example, for longitudinal braking scenarios, when the minimum avoidance requirement is a minimum longitudinal deceleration of 0.3g and the current vehicle speed is 40km / h, the corresponding upper limit of longitudinal deceleration can be set to 0.5g; when the current vehicle speed is 60km / h, the corresponding upper limit of longitudinal deceleration can be set to 0.6g. It should be understood that the above values ​​are only used to illustrate the corresponding trend between vehicle speed and the upper limit of reasonable safe response; the specific values ​​can be determined based on the braking system performance, road surface adhesion conditions, and ride comfort requirements.

[0057] In practice, if the mapping table does not directly contain the combination of the current minimum avoidance requirement and the current vehicle speed, the corresponding reasonable safety response upper limit can be obtained through linear interpolation or nearest neighbor interpolation. If the reasonable safety response upper limit includes upper limits in multiple dimensions, such as both lateral offset upper limit and steering wheel angle change rate upper limit, then the upper limit of each dimension can be compared with the corresponding extreme value of the action, and if any dimension exceeds the corresponding upper limit, it can be judged as excessive self-defense.

[0058] After determining a reasonable safety response upper limit, the extreme values ​​of actions are extracted from the dynamic data of the pre-extracted request for the triggering scenario. These extreme values ​​can include at least one of the following: maximum lateral offset, maximum steering wheel angle change rate, and maximum longitudinal deceleration. They characterize the maximum magnitude of the actions taken by the driver within a certain period before and after the triggering moment. For example, if the trigger type is lateral acceleration triggering, the extreme value can be the maximum lateral offset of the vehicle's actual trajectory within a preset time window before and after the triggering moment; if the trigger type is longitudinal deceleration triggering, the extreme value can be the maximum absolute value of the longitudinal deceleration within a preset time window before and after the triggering moment.

[0059] Furthermore, the extracted extreme values ​​of the actions are compared with the upper limit of a reasonable safe response.

[0060] S43: If the extreme value of the action does not exceed the reasonable safety response limit, then the current behavior of the vehicle is confirmed as a legal terrain avoidance operation that matches the environment, and the scene pre-extraction request is cancelled and data extraction is suppressed; In practice, when the comparison results show that the driver's evasive action does not exceed the aforementioned reasonable safety response limit, it indicates that the driver's evasive action is within a reasonable range in terms of magnitude and constitutes a legitimate terrain evasion operation matching the current environment. At this point, the scenario pre-extraction request is revoked and data extraction is suppressed. Suppressing data extraction may include at least one of the following: not packaging data for the current time period, not uploading data for the current time period, deleting cached pre-extracted data, or retaining only low-priority logs.

[0061] The above methods can effectively filter out false dangerous scenarios caused by legal terrain avoidance, prevent redundant data from polluting the autonomous driving training set, and save vehicle-side storage space and vehicle-to-cloud transmission bandwidth.

[0062] S44: If the extreme value of the action exceeds the upper limit of the reasonable safety response, the current behavior of the vehicle is determined to be an excessive defensive dangerous avoidance operation, the suppression of the data extraction is forcibly rejected, and the scenario pre-extraction request is executed instead. Data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario.

[0063] In practice, when the comparison results show that the driver's evasive maneuver exceeds the aforementioned reasonable safety response limit, it indicates that although the driver is indeed evading a valid environmental mask ahead, the magnitude of the evasive maneuver is too large and the response is too violent, which may lead to vehicle instability, crossing lane boundaries, or causing other traffic risks. This constitutes an excessive and dangerous evasive maneuver. At this point, the previously planned data suppression operation is forcibly rejected, and a scenario pre-extraction request is executed instead. Data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario.

[0064] The above method can adaptively retain high-value dangerous data such as driver overreaction and vehicles on the verge of losing control, and avoid such data being incorrectly filtered out because it is misjudged as legitimate terrain avoidance.

[0065] Understandably, this embodiment, after dynamic pre-triggering, environmental semantic verification, and spatiotemporal logic consistency determination, further implements avoidance intensity verification, thereby achieving refined control over data extraction decisions. On the one hand, for legitimate and restrained terrain avoidance operations, pre-extraction is revoked and data extraction is suppressed, thus filtering out false dangerous scenarios; on the other hand, for excessively defensive danger avoidance operations, the suppression operation is forcibly rejected and data extraction is executed, thus retaining high-value dangerous scenarios. Therefore, this embodiment can achieve a balance between filtering redundant data and retaining high-value data, improving the operational efficiency and data quality of the autonomous driving data closed-loop system.

[0066] Reference Figure 4 As shown, this embodiment of the invention also provides an automatic extraction device for dangerous driving scenarios of vehicles, the device comprising: The pre-extraction request generation module 100 is used to acquire vehicle dynamics data in real time. When the lateral acceleration or longitudinal deceleration in the dynamics data is detected to exceed a preset danger threshold, a scene pre-extraction request with the current timestamp is generated. The environment mask activation module 200 is used to respond to the scene pre-extraction request, backtrack the forward visual perception data within a preset time window before the current timestamp to extract static road semantic elements, and project the predicted wheel track band forward based on the vehicle's current speed and heading angle, and activate the static road semantic elements that spatially overlap with the predicted wheel track band as valid environment masks. The consistency determination module 300 is used to obtain the actual motion trajectory of the vehicle that triggers the scene pre-extraction request and calculate its physical positional relationship with the effective environment mask, so as to determine whether the motion change characteristics of the vehicle at this time are consistent with the spatiotemporal logic of avoiding the effective environment mask. The scene extraction execution module 400 is used to confirm that the vehicle's current behavior is a legal terrain avoidance operation if the scene is determined to match, cancel the scene pre-extraction request and suppress data extraction; if the scene is determined to not match or the effective environment mask is not activated, the scene pre-extraction request is executed, and data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scene.

[0067] As an example, the static road semantic elements include at least one of speed bumps, potholes, water accumulation areas, and construction barriers; the environment mask activation module 200 is specifically used for: Using a pre-set visual perception model, target detection and semantic segmentation are performed on the forward visual perception data within a preset time window before the current timestamp obtained from the retrospective process, to determine the category of the static road semantic element and the pixel coordinates of the static road semantic element in the forward visual perception data. Based on the intrinsic and extrinsic calibration parameters of the vehicle-mounted camera, the pixel coordinates are inversely projected onto the local coordinate system of the vehicle to obtain the actual physical distance and two-dimensional contour of the static road semantic element relative to the vehicle.

[0068] As an example, the environment mask activation module 200 is further used for: Obtain the vehicle's inherent track width parameters, and calculate the vehicle's predicted driving centerline based on the current vehicle speed and a preset prediction time length, starting from the vehicle's current position and extending along the direction of the heading angle. Using the predicted driving centerline as a reference, the boundary is extended outward by half of the inherent wheel track parameter to the left and right, respectively, to generate a strip-shaped polygonal region on a local two-dimensional plane, and the polygonal region is used as the predicted wheel track band.

[0069] As an example, the matching degree determination module 300 is specifically used for: If the scenario pre-extraction request is triggered by the longitudinal deceleration exceeding the danger threshold, calculate the longitudinal braking distance between the starting position of the vehicle when the longitudinal deceleration exceeds the danger threshold and the effective environment mask. When the longitudinal braking distance is within a reasonable braking range that matches the current vehicle speed, it is determined that the spatiotemporal logic is consistent. If the scenario pre-extraction request is triggered by the lateral acceleration exceeding the danger threshold, the maximum lateral offset of the vehicle's actual trajectory is calculated. When the sum of the maximum lateral offset and the width parameter of the effective environment mask is greater than or equal to the safety avoidance threshold, the spatiotemporal logic is determined to be consistent.

[0070] As an example, the scene extraction execution module 400 is specifically used for: If a match is determined, the physical size parameters of the effective environment mask in the forward visual perception data are obtained, and the minimum avoidance requirement corresponding to the effective environment mask is determined based on the physical size parameters. Based on the minimum avoidance requirement and the current vehicle speed, a reasonable safety response upper limit is mapped for the effective environment mask. The extreme values ​​of the actions in the dynamic data that trigger the scene pre-extraction request are extracted, and the extreme values ​​of the actions are compared with the reasonable safety response upper limit. If the extreme value of the action does not exceed the reasonable safety response limit, then the current behavior of the vehicle is confirmed as a legal terrain avoidance operation that matches the environment, and the scene pre-extraction request is cancelled and data extraction is suppressed. If the extreme value of the action exceeds the upper limit of the reasonable safety response, the current behavior of the vehicle is determined to be an excessive defensive evasive operation. The suppression of the data extraction is forcibly rejected, and the scenario pre-extraction request is executed instead. Data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario.

[0071] Given that the automatic extraction device (or system, readable storage medium) for dangerous driving scenarios provided in this application is an equivalent mapping of the aforementioned method embodiments on a virtual module architecture or underlying hardware entity, both are based on the completely consistent overall inventive concept. To keep the specification concise and avoid redundancy, any data flow details and logical judgment criteria not elaborated upon in the device embodiments should be directly referenced and uniformly understood by those skilled in the art in conjunction with the relevant descriptions of the aforementioned method embodiments.

[0072] It should be clear that as long as the core technical concept of this invention—using static environmental semantic masks to reverse verify spatiotemporal logic to filter out false dangers and supplementing it with action extreme value circuit breaking to capture excessive defense scenarios—is adhered to, those skilled in the art, when actually deploying autonomous driving data collection services, should fall within the legal protection scope defined by the claims of this application by making conventional numerical evolutions, equivalent substitutions, or adaptive fine-tunings to the aforementioned trajectory projection algorithm, threshold mapping model, or environmental verification logic, taking into account differences in chassis parameters of different vehicle models, the arrangement scheme of perception sensors, or the characteristics of computing power platforms.

Claims

1. An automatic extraction method for dangerous driving scenarios of vehicles, characterized in that, Includes the following steps: Real-time acquisition of vehicle dynamics data; when the lateral acceleration or longitudinal deceleration in the dynamics data exceeds a preset danger threshold, a scene pre-extraction request with the current timestamp is generated. In response to the scene pre-extraction request, forward visual perception data within a preset time window before the current timestamp is retrieved to extract static road semantic elements, and a predicted wheel track is projected forward based on the vehicle's current speed and heading angle. The static road semantic elements that spatially overlap with the predicted wheel track are activated as valid environment masks. The actual motion trajectory of the vehicle that triggered the scene pre-extraction request is obtained and its physical positional relationship with the effective environment mask is calculated to determine whether the motion change characteristics of the vehicle at this time are consistent with the spatiotemporal logic of avoiding the effective environment mask. If the scenario is determined to be a match, the vehicle's current behavior is confirmed as a legitimate terrain avoidance operation, the scenario pre-extraction request is revoked, and data extraction is suppressed; if the scenario is determined to be a mismatch or the effective environment mask is not activated, the scenario pre-extraction request is executed, and data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario.

2. The method for automatically extracting dangerous driving scenarios of vehicles according to claim 1, characterized in that, The static road semantic elements include at least one of speed bumps, potholes, water accumulation areas, and construction barriers; therefore, the forward visual perception data within a preset time window prior to the current timestamp is retrieved to extract the static road semantic elements, including: Using a pre-set visual perception model, target detection and semantic segmentation are performed on the forward visual perception data within a preset time window before the current timestamp obtained from the retrospective process, to determine the category of the static road semantic element and the pixel coordinates of the static road semantic element in the forward visual perception data. Based on the intrinsic and extrinsic calibration parameters of the vehicle-mounted camera, the pixel coordinates are inversely projected onto the local coordinate system of the vehicle to obtain the actual physical distance and two-dimensional contour of the static road semantic element relative to the vehicle.

3. The method for automatically extracting dangerous driving scenarios of vehicles according to claim 1, characterized in that, Based on the vehicle's current speed and heading angle, the predicted wheel track area is projected forward, including: Obtain the vehicle's inherent track width parameters, and calculate the vehicle's predicted driving centerline based on the current vehicle speed and a preset prediction time length, starting from the vehicle's current position and extending along the direction of the heading angle. Using the predicted driving centerline as a reference, the boundary is extended outward by half of the inherent wheel track parameter to the left and right, respectively, to generate a strip-shaped polygonal region on a local two-dimensional plane, and the polygonal region is used as the predicted wheel track band.

4. The method for automatically extracting dangerous driving scenarios of vehicles according to claim 3, characterized in that, Determining whether the vehicle's current motion characteristics match the spatiotemporal logic of avoiding the effective environmental mask includes: If the scenario pre-extraction request is triggered by the longitudinal deceleration exceeding the danger threshold, calculate the longitudinal braking distance between the starting position of the vehicle when the longitudinal deceleration exceeds the danger threshold and the effective environment mask. When the longitudinal braking distance is within a reasonable braking range that matches the current vehicle speed, it is determined that the spatiotemporal logic is consistent. If the scenario pre-extraction request is triggered by the lateral acceleration exceeding the danger threshold, the maximum lateral offset of the vehicle's actual trajectory is calculated. When the sum of the maximum lateral offset and the width parameter of the effective environment mask is greater than or equal to the safety avoidance threshold, the spatiotemporal logic is determined to be consistent.

5. The method for automatically extracting dangerous driving scenarios of vehicles according to claim 1, characterized in that, If the match is determined, the vehicle's current behavior is confirmed as a legitimate terrain avoidance maneuver. The scene pre-extraction request is revoked and data extraction is suppressed, including: If a match is determined, the physical size parameters of the effective environment mask in the forward visual perception data are obtained, and the minimum avoidance requirement corresponding to the effective environment mask is determined based on the physical size parameters. Based on the minimum avoidance requirement and the current vehicle speed, a reasonable safety response upper limit is mapped for the effective environment mask. The extreme values ​​of the actions in the dynamic data that trigger the scene pre-extraction request are extracted, and the extreme values ​​of the actions are compared with the reasonable safety response upper limit. If the extreme value of the action does not exceed the reasonable safety response limit, then the current behavior of the vehicle is confirmed as a legal terrain avoidance operation that matches the environment, and the scene pre-extraction request is cancelled and data extraction is suppressed. If the extreme value of the action exceeds the upper limit of the reasonable safety response, the current behavior of the vehicle is determined to be an excessive defensive evasive operation. The suppression of the data extraction is forcibly rejected, and the scenario pre-extraction request is executed instead. Data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario.

6. An automatic extraction device for dangerous driving scenarios of vehicles, characterized in that, The device includes: The pre-extraction request generation module is used to acquire vehicle dynamics data in real time. When the lateral acceleration or longitudinal deceleration in the dynamics data is detected to exceed a preset danger threshold, a scene pre-extraction request with the current timestamp is generated. The environment mask activation module is used to respond to the scene pre-extraction request, backtrack the forward visual perception data within a preset time window before the current timestamp to extract static road semantic elements, and project the predicted wheel track band forward based on the vehicle's current speed and heading angle, and activate the static road semantic elements that spatially overlap with the predicted wheel track band as valid environment masks. The consistency determination module is used to obtain the actual motion trajectory of the vehicle that triggers the scene pre-extraction request and calculate its physical positional relationship with the effective environment mask, so as to determine whether the motion change characteristics of the vehicle at this time are consistent with the spatiotemporal logic of avoiding the effective environment mask. The scene extraction and execution module is used to confirm that the vehicle's current behavior is a legal terrain avoidance operation if the scene is determined to match, cancel the scene pre-extraction request and suppress data extraction; if the scene is determined to not match or the effective environment mask is not activated, the scene pre-extraction request is executed, and data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scene.

7. The automatic extraction device for dangerous driving scenarios of a vehicle according to claim 6, characterized in that, The static road semantic elements include at least one of speed bumps, potholes, water accumulation areas, and construction barriers; the environment mask activation module is specifically used for: Using a pre-set visual perception model, target detection and semantic segmentation are performed on the forward visual perception data within a preset time window before the current timestamp obtained from the retrospective process, to determine the category of the static road semantic element and the pixel coordinates of the static road semantic element in the forward visual perception data. Based on the intrinsic and extrinsic calibration parameters of the vehicle-mounted camera, the pixel coordinates are inversely projected onto the local coordinate system of the vehicle to obtain the actual physical distance and two-dimensional contour of the static road semantic element relative to the vehicle.

8. The automatic extraction device for dangerous driving scenarios of a vehicle according to claim 6, characterized in that, The environment mask activation module is further used for: Obtain the vehicle's inherent track width parameters, and calculate the vehicle's predicted driving centerline based on the current vehicle speed and a preset prediction time length, starting from the vehicle's current position and extending along the direction of the heading angle. Using the predicted driving centerline as a reference, the boundary is extended outward by half of the inherent wheel track parameter to the left and right, respectively, to generate a strip-shaped polygonal region on a local two-dimensional plane, and the polygonal region is used as the predicted wheel track band.

9. The automatic extraction device for dangerous driving scenarios of a vehicle according to claim 6, characterized in that, The matching degree determination module is specifically used for: If the scenario pre-extraction request is triggered by the longitudinal deceleration exceeding the danger threshold, calculate the longitudinal braking distance between the starting position of the vehicle when the longitudinal deceleration exceeds the danger threshold and the effective environment mask. When the longitudinal braking distance is within a reasonable braking range that matches the current vehicle speed, it is determined that the spatiotemporal logic is consistent. If the scenario pre-extraction request is triggered by the lateral acceleration exceeding the danger threshold, the maximum lateral offset of the vehicle's actual trajectory is calculated. When the sum of the maximum lateral offset and the width parameter of the effective environment mask is greater than or equal to the safety avoidance threshold, the spatiotemporal logic is determined to be consistent.

10. An automatic extraction device for dangerous driving scenarios of a vehicle according to claim 6, characterized in that, The scene extraction and execution module is specifically used for: If a match is determined, the physical size parameters of the effective environment mask in the forward visual perception data are obtained, and the minimum avoidance requirement corresponding to the effective environment mask is determined based on the physical size parameters. Based on the minimum avoidance requirement and the current vehicle speed, a reasonable safety response upper limit is mapped for the effective environment mask. The extreme values ​​of the actions in the dynamic data that trigger the scene pre-extraction request are extracted, and the extreme values ​​of the actions are compared with the reasonable safety response upper limit. If the extreme value of the action does not exceed the reasonable safety response limit, then the current behavior of the vehicle is confirmed as a legal terrain avoidance operation that matches the environment, and the scene pre-extraction request is cancelled and data extraction is suppressed. If the extreme value of the action exceeds the upper limit of the reasonable safety response, the current behavior of the vehicle is determined to be an excessive defensive evasive operation. The suppression of the data extraction is forcibly rejected, and the scenario pre-extraction request is executed instead. Data for the corresponding time period is extracted based on the current timestamp and packaged into a dangerous driving scenario.

Images