An automatic identification and segmentation method for intelligent driving key interaction scene

Abstract

The application discloses an automatic identification and segmentation method for intelligent driving key interaction scenes, and relates to the technical field of intelligent driving and artificial intelligence information processing, and comprises the following steps: acquiring vehicle perception data in the intelligent driving process, analyzing the lateral position change trend of adjacent lane targets and the predicted collision time between the adjacent lane targets and the vehicle, and generating a potential cut-in scene mark when it is detected that the lateral position of the adjacent lane target continuously approaches the center line of the vehicle lane and the predicted collision time meets the preset cut-in risk condition. Through the multi-source information fusion and state relaxation double boundary segmentation method, the application realizes high-precision identification and complete time window extraction of intelligent driving key interaction scenes, effectively reduces false triggering, and improves data quality. At the same time, through the automatic processing process, manual labeling is not needed, large-scale and standardized driving scene dataset construction can be realized, and reliable data support is provided for intelligent driving analysis and optimization.

Description

Technical Field

[0001] This invention relates to the fields of intelligent driving and artificial intelligence information processing technology, specifically to an automatic identification and segmentation method for key interactive scenarios in intelligent driving. Background Technology

[0002] With the development of intelligent driving technology, the application of human-machine co-driving systems in real-world road environments is gradually increasing. During the evaluation of the human-machine co-driving experience, negative experiences of the driver (such as tension and panic) are not evenly distributed throughout the entire driving process, but rather mainly concentrated in specific key interaction scenarios, such as the intelligent driving system suddenly issuing a takeover request or an adjacent vehicle cutting into the current lane at close range. Therefore, to accurately evaluate the interaction effect between the intelligent driving system and the driver, it is necessary to precisely identify and automatically segment key interaction scenario fragments with analytical value from massive amounts of driving data, thereby providing an effective data foundation for subsequent human-machine co-driving experience analysis, system optimization, and the construction of driving datasets. Currently, the identification and analysis of key intelligent driving scenarios typically employ data analysis methods based on vehicle operating status, driving behavior parameters, or system event signals. This involves detecting changes in specific parameters in the driving log to identify abnormal events or specific driving scenarios.

[0003] Existing technologies have the following shortcomings: Firstly, current key interaction scenario recognition technologies still have certain limitations. On the one hand, existing technologies typically rely on a single threshold or simple rules for scenario judgment. For example, they might determine a following-to-close event by setting a headway (THW) less than a preset value, or identify emergency driver intervention by detecting sudden changes in steering wheel angular velocity. Because these methods primarily rely on a single data indicator, they are prone to false detections in complex and dynamic traffic environments. For instance, in congested urban roads, vehicles may maintain a relatively small headway for extended periods, but this does not necessarily represent a dangerous interaction event with analytical value, leading to the incorrect identification of numerous invalid scenarios. Secondly, existing methods typically employ post-event data analysis, importing driving logs and writing rule scripts to find data segments that meet certain conditions. This approach usually only identifies the event trigger point and cannot automatically determine complete and reasonable time boundaries based on the scene's changing process. It cannot accurately delineate the state changes before the event or the recovery process after the event, resulting in the generated analysis data window potentially containing a large amount of irrelevant information or missing key interaction information, making it difficult to meet the requirements for high-precision recognition and complete data segmentation of key interaction scenarios in intelligent driving.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic identification and segmentation method for key interactive scenarios in intelligent driving, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an automatic identification and segmentation method for key interactive scenarios in intelligent driving, comprising the following steps: Acquire vehicle perception data during intelligent driving, analyze the lateral position change trend of targets in adjacent lanes and the predicted collision time between targets in adjacent lanes and the vehicle. When it is detected that the lateral position of targets in adjacent lanes is continuously moving towards the center line of the vehicle, and the predicted collision time meets the preset entry risk conditions, generate a potential entry scenario flag. Based on the potential entry scenario marker, the lane change status of the target in the adjacent lane and the intelligent driving control behavior of the vehicle are detected. When the target in the adjacent lane is detected to continuously cross the lane line and enter the lane, and the intelligent driving control behavior of the vehicle generates a deceleration command, the potential entry scenario marker is identified as the entry interaction scenario, and the time corresponding to the deceleration command is identified as the core event time of the entry interaction scenario. Based on the core moment of the event, a time-reverse search is performed on the vehicle operation data and driver physiological data corresponding to the entry interaction scenario to obtain the time position when the vehicle's longitudinal motion state recovers to a stable state and the driver's psychological load recovers to the stable state before the trigger, and this time position is determined as the starting boundary of the entry interaction scenario. Based on the core moment of the event, a forward time search is performed on the vehicle operation data and driver physiological data corresponding to the entry interaction scenario to obtain the time position when the vehicle resumes a stable following state and the driver's psychological load is reduced to a preset stable state, and this time position is determined as the end boundary of the entry interaction scenario. Based on the start and end boundaries, continuous driving data is extracted to generate data segments corresponding to the interactive scenarios, thereby enabling automatic identification and complete time window segmentation of key interactive scenarios in intelligent driving.

[0007] The system acquires vehicle perception data during intelligent driving and analyzes the lateral position change trends of targets in adjacent lanes and the predicted collision time between the targets in adjacent lanes and the vehicle itself. This includes: continuously acquiring the position information of targets in adjacent lanes, continuously tracking the lateral movement of targets in adjacent lanes in the road coordinate system, and determining whether the target vehicle has a tendency to move towards the center line of the lane based on the lateral position changes of targets in adjacent lanes over a continuous period of time; simultaneously acquiring the longitudinal distance change information between targets in adjacent lanes and the vehicle itself, predicting the collision risk between targets in adjacent lanes and the vehicle itself based on the vehicle's motion state, and generating a corresponding potential entry scenario flag when the lateral position change trend and the predicted collision time simultaneously meet the potential entry scenario judgment conditions.

[0008] When a target vehicle in an adjacent lane is detected to continuously cross the lane line and enter the current lane, and the vehicle's intelligent driving control behavior generates a deceleration command, the potential entry scenario marker is identified as an entry interaction scenario, and the time corresponding to the deceleration command is identified as the core event moment of the entry interaction scenario. This includes: continuously detecting the trajectory of the target vehicle after the potential entry scenario marker is generated, and determining whether the target vehicle gradually enters the lane where the vehicle is located from the adjacent lane; when a target vehicle is detected to continuously cross the lane line and enter the current lane, and the vehicle's intelligent driving control behavior is detected to output a deceleration command based on the target vehicle's motion state, the current potential entry scenario marker is converted into a confirmed entry interaction scenario, and the time and location of the deceleration command are taken as the core event moment.

[0009] Based on the core moment of the event, a reverse time search is performed on the vehicle operation data and driver physiological data corresponding to the interaction scenario. This includes: using the core moment of the event as a time reference point, continuously searching along the time axis of the driving data in the direction before the event, simultaneously acquiring vehicle operation status data and driver EEG data, and analyzing changes in the vehicle's longitudinal motion state and the driver's psychological load; by determining whether the vehicle's longitudinal motion state has recovered to a stable state and whether the driver's psychological load has recovered to a stable state before the trigger, the time position that meets the recovery conditions of the preceding state is determined, and this time position is determined as the starting boundary of the interaction scenario.

[0010] The process of restoring the vehicle's longitudinal motion state to a stable state includes: scanning the vehicle's operating data backward from the core moment of the event, detecting changes in the vehicle's longitudinal acceleration, and determining that the vehicle's motion state meets the recovery conditions when the vehicle's longitudinal acceleration smoothly falls back to within the range of ±0.2 m / s² and remains stable. Here, ±0.2 m / s² represents the allowable range of variation in longitudinal acceleration when it returns to a stable state, and is used to determine whether the vehicle has recovered from the dynamic adjustment state during the interaction process to a stable operating state.

[0011] The process of restoring the driver's psychological load to a stable state before the trigger includes: analyzing the driver's EEG data, obtaining the decoded driver's psychological load value, comparing the current psychological load value with the average psychological load level in the 30 seconds before the scene trigger, and determining that the driver's psychological state has returned to a stable state when the driver's psychological load value returns to the average level in the 30 seconds before the scene trigger is within ±0.5 standard deviations. The corresponding time position is used as the basis for judgment in the process of determining the starting boundary.

[0012] Based on the core moment of the event, a forward time search is performed on the vehicle operation data and driver physiological data corresponding to the entry interaction scenario. This includes: using the core moment of the event as a time reference point, continuously searching along the driving data time axis in the direction after the event occurs, and synchronously analyzing the changes in vehicle operation status and driver psychological load; when it is detected that the vehicle has returned to a stable following state and the vehicle acceleration has stabilized again, and the driver's psychological load has decreased to a preset stable state, the corresponding time position is determined as the end boundary of the entry interaction scenario, so as to complete the automatic determination of the event end time position.

[0013] The process of reducing the driver's psychological load to a preset stable state includes: continuously analyzing the driver's EEG data after the core moment of the event; when the decoded driver's psychological load value is below 0.4 for 3 consecutive seconds, it is determined that the driver's psychological load has reached a stable state; where 0.4 is the calibration threshold in the process of judging the driver's psychological load, and 3 consecutive seconds are used to judge the continuous stability of the driver's psychological load recovery process, so as to avoid the deviation of the end boundary judgment caused by the change at a single time point.

[0014] Based on the start and end boundaries, continuous driving data is extracted to generate data segments corresponding to the entry interaction scenario. This includes: using a determined start boundary as the starting position for data extraction and a determined end boundary as the ending position for data extraction; extracting continuous driving data between the start and end boundaries; and extracting vehicle operation data and driver physiological data corresponding to the entry interaction scenario to form a complete entry interaction scenario data segment for subsequent analysis.

[0015] To achieve automatic identification and complete time window segmentation of key interactive scenarios in intelligent driving, the system continuously analyzes vehicle perception data, vehicle operation data, and driver physiological data. Without the need for manual post-event scenario screening and boundary labeling, it automatically completes the detection of potential entry scenarios, confirmation of entry interaction scenarios, determination of the core moment of the event, determination of the start boundary, and determination of the end boundary. Based on the determined time window, it generates standardized entry interaction scenario data fragments.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention identifies key interactive scenarios in intelligent driving through a multi-source information fusion mechanism. It no longer relies on single vehicle parameters or threshold conditions for judgment, but comprehensively analyzes multiple types of information, including vehicle motion state, changes in target vehicle behavior, and intelligent driving control responses, to achieve accurate judgment of key interactive behaviors. By fusing deterministic event signals with dynamic logical judgment conditions, it effectively reduces false triggering problems caused by non-critical factors such as road congestion and normal vehicle following changes, improving the accuracy and reliability of scene recognition results. This allows the recognition process to more accurately distinguish between ordinary congested following states and dangerous incursions, providing a more accurate scene data foundation for subsequent intelligent driving experience evaluation and data analysis.

[0017] This invention utilizes a state-relaxation-based dual-boundary segmentation algorithm to dynamically determine the time range of key interaction scenarios. Instead of using fixed time windows for data extraction, it automatically determines the start and end points of the scenario based on vehicle motion and driver state changes. By analyzing the state recovery process before and after the event, it can accurately eliminate smooth driving data irrelevant to the interaction event while retaining key data information from the event's formation, development, and recovery processes. This ensures that the generated data window has complete information content and high data quality. This approach effectively improves the data validity in subsequent causal relationship analysis, driving behavior analysis, and human-machine co-driving experience evaluation, and enhances the signal-to-noise ratio of the scene data.

[0018] This invention automates the entire data processing process, from triggering key interaction scenarios and confirming them to segmenting time windows, eliminating the need for manual post-event data review, event filtering, and scenario boundary annotation. By automatically analyzing continuous driving data, it can quickly extract valuable key interaction scenarios from massive amounts of driving data and automatically generate corresponding standardized scenario data fragments. This automated processing method reduces subjective differences in manual annotation, improves the consistency and efficiency of scenario data construction, meets the needs of large-scale intelligent driving test data processing, and enables the automated and standardized construction of driving scenario datasets, providing reliable data support for intelligent driving system optimization and human-machine interaction performance evaluation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1This is a flowchart of an automatic identification and segmentation method for key interactive scenarios in intelligent driving, according to the present invention. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0022] This invention provides, for example Figure 1 The method for automatic identification and segmentation of key interaction scenarios in intelligent driving, as shown, includes the following steps: During the initial screening and core triggering processes, environmental perception data is continuously acquired during the operation of intelligent driving, and the motion status of road-related targets is analyzed in real time to achieve preliminary screening of potential key interaction scenarios. Regarding the motion changes of target vehicles in adjacent lanes, the focus is on analyzing the lateral position change trend of target vehicles. By continuously collecting the position information of target vehicles in the road coordinate system, the lateral movement trajectory of target vehicles is determined. When a continuous change in the lateral position of a target vehicle in an adjacent lane is detected, and the lateral position of the target vehicle gradually moves towards the center line of the lane where the vehicle is located, it indicates that the target vehicle has a tendency to migrate laterally into the lane. This motion state is different from the normal adjacent lane driving state and can serve as an important basis for judging whether the target vehicle may cut into the lane.

[0023] While analyzing the lateral movement trend of the target vehicle, the system further assesses the potential interaction risks after the target vehicle enters the lane by considering the longitudinal distance changes between the target vehicle and the vehicle itself. By acquiring the real-time longitudinal distance between the target vehicle and the vehicle, and predicting the future collision risk based on their motion states, the predicted time of collision (TTC) between them is calculated. When the TTC is detected to be less than a preset threshold, it indicates that the target vehicle's current movement trend may cause the safe distance between them to decrease rapidly, posing a high interaction risk.

[0024] The information on the target vehicle's lateral position continuously moving towards the center line of its lane, together with the risk information that the predicted collision time (TTC) between the target vehicle and the current vehicle is less than a preset threshold, constitutes a comprehensive judgment condition for potential cutting-in behavior. By simultaneously analyzing the lateral movement trend of the target vehicle and the longitudinal interaction risk between vehicles, the misidentification problem caused by relying solely on a single vehicle parameter for scenario judgment can be avoided. For example, when judging solely based on changes in distance between vehicles, in normal traffic congestion, vehicles may maintain a small distance for a long time, but this does not necessarily mean that the target vehicle is cutting in; when judging solely based on changes in the target vehicle's lateral position, the vehicle may have a slight lateral deviation, but it has not entered an actual interaction risk state. Therefore, by jointly analyzing changes in lateral position and predicted collision time, the accuracy of potential cutting-in scenario screening can be improved.

[0025] During data analysis, intelligent driving operation data is continuously input into the scene recognition process. Perception data is processed sequentially over time to dynamically track the motion of target vehicles in adjacent lanes. When the lateral position change of a target vehicle shows a trend of continuously moving towards the centerline of the current lane, and the predicted time of collision (TTC) between the target vehicle and the current vehicle meets the risk assessment criteria, a corresponding potential entry scenario flag is generated. This potential entry scenario flag indicates the presence of interactive behaviors in the current environment that may affect the intelligent driving operation, providing a basis for further confirmation of specific interaction scenarios.

[0026] Potential entry scenario markers are not directly used as the final entry scenario result, but rather as a preliminary screening criterion to reduce the scope of invalid analysis in continuous driving data. In real road environments, vehicle operating states are affected by various factors such as road structure, traffic flow changes, and driving behavior. Some target vehicles may exhibit brief lateral movements without actually entering the lane. Therefore, by setting potential entry scenario markers, it is possible to ensure that key interactive behaviors are captured while avoiding a large number of invalid scenario records caused by single instantaneous state changes.

[0027] In response to continuously changing road traffic environments, by continuously analyzing the lateral movement of target vehicles and the longitudinal risk status between vehicles, potential intrusion interactions can be detected in advance, and key interaction scenarios can be initially triggered when preset judgment conditions are met. This process is based on the actual movement changes of the target vehicle and does not rely on a single fixed threshold for judgment. Instead, it filters interactive behaviors that may affect the operation of intelligent driving by analyzing the correlation between changes in the lateral trajectory of the target vehicle and the predicted collision time between vehicles.

[0028] In the specific implementation process, environmental perception data can include information on the target vehicle's position changes and the distance between the target vehicle and the current vehicle. By continuously acquiring and analyzing the above data, it is determined whether the target vehicle has a tendency to move from the adjacent lane towards the current lane. When the lateral position of the target vehicle continues to move towards the center line of the current lane, and the predicted collision time (TTC) between the two vehicles is less than a preset condition, the current target vehicle is considered to have potential cutting-in characteristics, and a potential cutting-in scenario flag is triggered.

[0029] Through the data analysis process described above, time zones with potential critical interaction risks can be quickly located from a large amount of continuous driving data, providing a data foundation for subsequent scene confirmation and complete time window segmentation. This approach reduces the false detection problem caused by relying solely on single vehicle parameters in traditional methods, improves the adaptability of the critical interaction scene identification process to complex dynamic driving environments, and provides accurate event triggering criteria for subsequent scene boundary determination based on changes in vehicle and driver states.

[0030] Once a potential lane-entry scenario is detected, the system continuously analyzes the target vehicle's motion state and the vehicle's intelligent driving control state to determine whether the target vehicle has already engaged in actual lane-entry interaction and to confirm the scenario for the initially screened potential lane-entry behaviors. Specifically, while continuously tracking the trajectory of target vehicles in adjacent lanes, the system assesses whether the target vehicle continuously crosses lane lines to enter the vehicle's lane. When the target vehicle only shows a lateral approach trend but does not actually change lanes, it remains in a potential lane-entry state; when the target vehicle's trajectory further develops, manifesting as continuously crossing lane lines and entering the vehicle's lane, it indicates that the target vehicle has transitioned from a potential lateral movement state to an actual lane-entry behavior.

[0031] As the target vehicle continuously crosses the lane lines to enter this lane, the control response status of the vehicle's intelligent driving system is further analyzed to determine whether the target vehicle's entry behavior has actually affected the vehicle's intelligent driving operation. When the vehicle's intelligent driving system detects that it generates a corresponding deceleration command based on the target vehicle's entry behavior, it indicates that the target vehicle's entry into this lane has triggered the intelligent driving system's control response. This interaction process has clear vehicle interaction characteristics and no longer belongs to the general vehicle motion changes in an ordinary road environment.

[0032] After completing the correlation analysis between the target vehicle's motion state and the vehicle's intelligent driving control behavior, the previously generated potential entry scenario flags are upgraded to confirmed entry scenarios. This confirmation process is jointly determined by the target vehicle's actual lane crossing behavior and the vehicle's intelligent driving system's deceleration response behavior, avoiding judgments based solely on changes in the target vehicle's lateral position or distance between vehicles, thereby reducing misjudgments caused by non-critical interactive behaviors such as road congestion or slight lateral vehicle deviation.

[0033] During the specific judgment process, the system continuously acquires information on the target vehicle's positional changes and tracks and analyzes whether the target vehicle continues to move towards the lane. Once the target vehicle's lateral movement continues and it crosses the lane line, entering the lane where the vehicle is located, the system further determines whether the vehicle's intelligent driving system should initiate deceleration control in response to the target vehicle's movement. If the target vehicle completes lane change and the vehicle's intelligent driving system simultaneously outputs a deceleration command, then the current target vehicle lane change is considered a complete vehicle interaction event, and this event is identified as a lane change scenario.

[0034] After confirming the scenario, mark the moment when the vehicle's intelligent driving system issues a deceleration command as the core point of the event. The core of the incident The temporal location used to characterize the key interaction between the target vehicle's approach behavior and the vehicle's control response is an important reference for subsequent scene data analysis and time window segmentation. By identifying the core points of the event... It can accurately locate the core time region where key interactions occur in continuous driving data, providing a unified time reference for subsequent forward analysis of vehicle and driver state changes before the event, and backward analysis of vehicle and driver state recovery process after the event.

[0035] Core of the incident The determination of the event core point is not simply based on the exact moment the target vehicle enters the lane, but rather on the correlation between the target vehicle's actual entry behavior and the vehicle's intelligent driving control response. When the target vehicle only exhibits a tendency to move closer to the center line of the lane, it is not directly considered the event core point because no actual vehicle interaction has yet occurred. Only when the target vehicle continuously crosses the lane line and enters the lane, further causing the vehicle's intelligent driving system to issue a deceleration command, is the corresponding time and location determined as the event core point. .

[0036] Through the aforementioned scenario establishment process, changes in vehicle movement within a continuous road environment can be transformed into key scenario events with clear interactive significance. This process first filters potential entry scenario markers to identify possible target vehicle entry behaviors, and then further confirms these by examining the target vehicle's actual lane change status and the vehicle's intelligent driving control response, resulting in more accurate and effective identification of the entry scenarios.

[0037] Compared to scene recognition methods that rely solely on a single parameter, this process comprehensively considers changes in the target vehicle's trajectory and the control behavior of the vehicle's intelligent driving system, enabling the scene recognition results to reflect the actual human-vehicle interaction process. In complex traffic environments, vehicles may experience brief lateral movements due to changes in road width, adjustments in driving behavior, or changes in traffic flow. If these movements are not considered in conjunction with the vehicle's intelligent driving control response, non-critical motion behaviors can easily be mistaken for intrusion events. By adding a deceleration command as a scene confirmation condition, non-influential target vehicle movements can be further filtered out, improving the accuracy of key interaction scene recognition.

[0038] At the same time, utilize the core points of the event Serving as a time benchmark for entering a scenario is beneficial for extracting valuable scenario segments from the complete driving data stream. After identifying the core point of the event, we can further analyze the vehicle's operating status, environmental changes, and driver status changes before the event, as well as the vehicle's recovery process and driver status changes after the event, thereby forming complete data on the interactive scenario.

[0039] Therefore, by continuously analyzing the motion state of target vehicles in potential entry scenarios, when a target vehicle is detected continuously crossing lane lines to enter the lane, and the vehicle's intelligent driving system issues a deceleration command accordingly, the potential entry scenario flag is upgraded to a confirmed entry scenario, and the moment the deceleration command is issued is marked as the core point of the event. This enables accurate conversion from potential interactive behaviors to identifying key interactive scenarios, providing a reliable foundation for subsequent full scene time window segmentation based on vehicle motion state and driver state.

[0040] From the core of the event At the moment of initiation, the driving data and driver's EEG data corresponding to the interaction scenario are scanned and analyzed in a forward direction to determine the reasonable data range before the interaction scenario occurs, and to obtain a precise starting point for analysis that reflects changes in the vehicle's state and the driver's state before the event is triggered. This process does not use a fixed time length as the scene pre-context window, but rather dynamically searches continuous data based on changes in vehicle motion and the driver's psychological load, thereby determining the starting boundary that can fully cover the event formation process.

[0041] When performing a search of preceding windows, focus on the core event point. As a temporal reference point, driving data and EEG data are continuously analyzed forward along the timeline. Driving data reflects the vehicle's motion state changes before the interaction occurred, while EEG data captures changes in the driver's psychological state. By simultaneously analyzing the vehicle's longitudinal motion and the driver's psychological load, it is determined whether the current time position has returned to the stable state before the scene was triggered.

[0042] Specifically, during vehicle state analysis, the longitudinal acceleration of the vehicle is monitored. By determining whether the longitudinal acceleration gradually recovers from its dynamic state before the interaction to a stable state, the stability of the vehicle's motion is assessed. When the algorithm detects that the vehicle's longitudinal acceleration smoothly returns to a preset range and remains stable for a continuous period, the vehicle's motion is considered to have recovered to a relatively stable state.

[0043] The smooth recovery state of vehicle longitudinal acceleration is determined by the following conditions: in, The unit representing longitudinal acceleration, indicating the change in acceleration in the longitudinal direction of a vehicle; The range represents the permissible interval of variation in the vehicle's longitudinal acceleration when it returns to a stable state, and is used to determine whether the vehicle's motion state has decreased to a stable level. This judgment condition allows identification of the time frame at which the vehicle gradually recovers from an interactive state to a stable operating state.

[0044] Simultaneously, during the driver state analysis, EEG data is processed, and the driver's psychological load value is obtained based on the decoded EEG data. The driver's state changes are analyzed by determining whether the psychological load has recovered to the stable level before the scenario was triggered. The driver's psychological load recovery status is determined by comparing it with the average psychological load level over a period of time before the scenario was triggered.

[0045] The specific judgment criteria are: the average level within ±0.5 standard deviation range of the 30 seconds prior to scene triggering; Among them, the 30 seconds before the scene is triggered represents the data time range used to calculate the reference state of the driver's stable psychological load. This time range is located before the intervention interaction event occurs and is used to reflect the driver's normal psychological state when not affected by the current interaction event; the average level represents the average state of the driver's psychological load value within this time range and is used as a reference benchmark for judging the recovery of psychological load. The standard deviation range indicates the allowable fluctuation of the driver's psychological load value around the reference average level. This range is used to determine whether the current driver's psychological load has recovered to the stable state before the incident.

[0046] During the search for the starting point of the preceding window, both the vehicle motion state recovery condition and the driver's psychological load recovery condition must be met simultaneously. When the algorithm starts from the core point of the event... During the forward scan, a certain time position is detected that simultaneously meets the following two state requirements: On the one hand, the longitudinal acceleration of this vehicle smoothly returns to: Within the range, and maintain a stable state; On the other hand, the driver's psychological load value obtained from decoding was restored to within the range of ±0.5 standard deviations of the average level in the 30 seconds before the scene was triggered.

[0047] This time position is then determined as the starting position of the preceding window and marked as follows: in, This indicates the precise starting point for analysis within the interactive scenario, used to determine the boundary position of valid data before the event occurs. This time position is not obtained by manually setting a fixed time length, but is dynamically determined based on the combined recovery status of the vehicle and driver states.

[0048] Determined through the above methods This ensures that the data entering the analysis scope has complete pre-event state information. Before the interaction scenario occurs, the vehicle's motion state and the driver's psychological state are usually in a relatively stable state. However, as the target vehicle gradually shows a tendency to engage, the vehicle state and driver state will change. Therefore, by searching forward to the point where the vehicle state and driver state both recover, we can avoid including too much irrelevant smooth driving data in the analysis window, while ensuring that important state changes before the engagement event are not missed.

[0049] In actual continuous driving data processing, driving data and EEG data are analyzed in a corresponding manner according to their temporal relationship. This is based on the core points of the event. As a baseline, the corresponding data is traversed forward gradually using a time-reverse search method. If both the vehicle longitudinal acceleration recovery condition and the driver's psychological load recovery condition are not simultaneously met, the search continues forward; when both conditions are met simultaneously, the search stops, and the corresponding time position is determined. .

[0050] This method utilizes the dual constraints of vehicle motion and driver psychological state to determine the boundaries of the preceding window, enabling the starting point of scene analysis to more accurately reflect the state changes prior to the formation of the interaction. Compared to a fixed-length truncation method, this approach can automatically adjust the analysis range based on the actual changes in different entry scenarios, improving the effectiveness of the generated scene data fragments.

[0051] Meanwhile, by introducing two judgment conditions—the recovery state of vehicle longitudinal acceleration and the recovery state of driver psychological load—the incompleteness caused by relying solely on vehicle state or driver state for boundary determination can be avoided. For example, determining the starting point based solely on changes in vehicle longitudinal acceleration may fail to reflect the process of changes in driver psychological state; similarly, determining the starting point based solely on changes in driver psychological load may fail to accurately reflect the recovery status of vehicle motion. Therefore, by considering both types of data simultaneously, the effective analysis area before the interaction scenario occurs can be more comprehensively determined.

[0052] Ultimately, by starting from the core of the event The system scans forward driving data and EEG data, searching for time points where both the vehicle's longitudinal acceleration recovery condition and the driver's psychological load recovery condition are simultaneously met, thus determining the precise starting point for analysis when entering the interaction scenario. This starting point can serve as an important boundary for constructing the time window of the subsequent complete scene, providing an accurate data range for the automatic segmentation of interactive scenes in continuous driving data.

[0053] From the core of the event Starting from a specific moment, continuous backward scanning analysis is performed on the driving data and driver's EEG data corresponding to the interaction scenario to determine the effective recovery process after the interaction scenario ends, and to obtain a precise end boundary that reflects the recovery of both the vehicle's state and the driver's psychological state. This process comprehensively judges the changes in the vehicle's operating state and the driver's psychological load after the event, and does not use a fixed time length as the end position of the scenario, but dynamically determines the end time position of the interaction scenario based on the recovery of both the vehicle's state and the driver's psychological state.

[0054] During the subsequent window search, the core event point is... As a time reference point, continuous data corresponding to the transition interaction scenario is scanned and analyzed along the time axis backward. During the scanning process, vehicle operation status data and driver EEG data are acquired simultaneously. By analyzing changes in vehicle motion status and driver psychological load, it is determined whether the transition interaction process has ended and whether the vehicle and driver have returned to a stable state.

[0055] The assessment of vehicle operating status primarily analyzes the changes in the vehicle's motion after the target vehicle enters the lane. Once the target vehicle has entered, the vehicle's intelligent driving control behavior adjusts according to the current traffic conditions, potentially resulting in deceleration, adjustments to following distance, and other motion changes. As the impact of the entry interaction gradually decreases, the vehicle's motion gradually stabilizes. Therefore, during the backward scan, it is necessary to determine whether the vehicle has returned to a stable following state and whether the vehicle's acceleration has stabilized again.

[0056] The vehicle status recovery condition is indicated by: the vehicle has returned to a stable following state, and the acceleration has stabilized again.

[0057] Here, "this vehicle" refers to the current vehicle participating in the intelligent driving interaction process; "stable following state" indicates that the vehicle's operating state has returned to normal following state after the target vehicle completes its engagement; and "acceleration stabilizes again" indicates that the vehicle's longitudinal motion changes have decreased, and the vehicle is no longer in a state of significant dynamic adjustment caused by the engagement interaction. This condition can be used to determine whether the interaction effect on the vehicle side has ended.

[0058] Simultaneously, the driver's EEG data is analyzed. The driver's psychological load value is obtained through EEG data decoding, and changes in psychological load are used to determine whether the driver has recovered to a lower psychological load state. During the approach of the target vehicle, the driver's psychological load may increase due to changes in the environment and vehicle control behavior. After the interaction process ends, the driver's psychological load gradually decreases. Therefore, by detecting the time during which the driver's psychological load remains below a preset threshold, it is possible to determine whether the driver's psychological state has recovered.

[0059] The condition for restoring driver's psychological load is: the driver's psychological load is below 0.4 for 3 consecutive seconds; Among them, driver psychological load refers to the parameter reflecting the changes in the driver's psychological state obtained by decoding the driver's EEG data; 3 consecutive seconds indicates the length of time that the driver's psychological load needs to be maintained to meet the conditions, in order to avoid misjudgment caused by a single instantaneous decrease in psychological load; 0.4 indicates the calibration threshold in the process of judging driver psychological load, used to judge whether the driver's psychological load has been reduced to a stable level.

[0060] During the boundary search at the end of the subsequent window, both the vehicle recovery condition and the driver's psychological load recovery condition must be met simultaneously. When the algorithm moves from the event's core point... During the forward scan, a certain time position was detected that simultaneously met the following conditions: On the one hand, the vehicle has resumed stable following, and the vehicle's acceleration has stabilized again; On the other hand, the driver's psychological load is below 0.4 (calibrated threshold) for 3 consecutive seconds.

[0061] This time position is then determined as the end position of the interaction scene and marked as follows: in, This indicates the precise end point of the analysis for the interaction scenario, used to determine the termination boundary of the data segment in that scenario on the timeline. This end point is not set at a fixed time interval, but rather determined jointly based on the recovery status of the vehicle's operational state and the driver's psychological state.

[0062] By starting from the core of the event A backward scan can fully capture the vehicle's adjustment process and the driver's psychological recovery process after the interaction scenario begins. Even after the target vehicle completes the transition, the vehicle's intelligent driving control behavior may continue to affect the vehicle's operating state, and the driver may still be under significant psychological stress. Directly capturing scene data at fixed intervals may not accurately cover the complete recovery process, resulting in the loss of some key state change information. Therefore, dynamically searching for the location where both the vehicle and driver states recover can more accurately determine the scene's end boundary.

[0063] In continuous driving data processing, driving data and EEG data are analyzed according to a unified temporal relationship. Starting from the core of the event... The system begins iterating through the data. If neither the vehicle stabilization condition nor the driver's psychological stress recovery condition is met simultaneously, the search continues. When both conditions are met, the search stops, and the corresponding time position is determined. .

[0064] Determined in this way This ensures that the final data segment of the interactive scenario includes the complete event occurrence process and the recovery process after the event ends. This termination boundary not only reflects the recovery status of the vehicle's motion state but also incorporates changes in the driver's psychological state, making the generated data window more in line with the needs of human-machine co-driving experience analysis.

[0065] After completing the initial boundary and the end boundary Once determined, the continuous driving data stream is automatically trimmed based on two time boundaries to form corresponding data packets for the interactive scenarios: in, This indicates the precise starting point for analysis when entering the interaction scenario, used to determine the range of valid data before the event occurs; This indicates the precise endpoint of the analysis when entering the interactive scenario, used to determine the range of valid data after the event occurs; It represents the complete time range of the cut-in interaction scene defined by the start boundary and the end boundary, and is used to extract high-quality scene data fragments from continuous driving data.

[0066] Using the above method, data can be automatically and completely extracted from a continuous data stream. This is a high-quality data package for the entry scenario. It contains key vehicle operation information and driver status changes during the entry interaction process, avoiding data redundancy or omissions of critical data caused by fixed time windows in traditional methods.

[0067] The entire subsequent window segmentation process is based on the core event point. Based on this, the method automatically determines the end boundary of the interactive scene by searching backwards for the recovery positions of the vehicle's state and the driver's psychological state. Compared to methods that determine the end time solely based on changes in vehicle movement, this approach further considers the driver's psychological load recovery process, resulting in a more complete determination of the scene's end position.

[0068] Simultaneously, by comprehensively judging three aspects—vehicle stability in following other vehicles, stable acceleration, and driver psychological load remaining below 0.4 for three consecutive seconds—the true recovery phase after the impact of the interaction can be effectively identified. When the vehicle has returned to normal operation and the driver's psychological load has also returned to a low level, it indicates that the interaction scenario has completed its entire process, and the corresponding time position can be used as the precise endpoint for analysis. .

[0069] Ultimately, through the core of the event Scan backwards and determine based on vehicle condition recovery conditions and driver psychological load recovery conditions. Combined with the previous window's determination It can automatically and completely extract from continuous driving data streams: The high-quality entry scenario data packages within the scope provide a complete and effective data foundation for subsequent analysis of key interactive scenarios in intelligent driving.

[0070] This invention identifies key interactive scenarios in intelligent driving through a multi-source information fusion mechanism. It no longer relies on single vehicle parameters or threshold conditions for judgment, but comprehensively analyzes multiple types of information, including vehicle motion state, changes in target vehicle behavior, and intelligent driving control responses, to achieve accurate judgment of key interactive behaviors. By fusing deterministic event signals with dynamic logical judgment conditions, it effectively reduces false triggering problems caused by non-critical factors such as road congestion and normal vehicle following changes, improving the accuracy and reliability of scene recognition results. This allows the recognition process to more accurately distinguish between ordinary congested following states and dangerous incursions, providing a more accurate scene data foundation for subsequent intelligent driving experience evaluation and data analysis.

[0071] This invention utilizes a state-relaxation-based dual-boundary segmentation algorithm to dynamically determine the time range of key interaction scenarios. Instead of using fixed time windows for data extraction, it automatically determines the start and end points of the scenario based on vehicle motion and driver state changes. By analyzing the state recovery process before and after the event, it can accurately eliminate smooth driving data irrelevant to the interaction event while retaining key data information from the event's formation, development, and recovery processes. This ensures that the generated data window has complete information content and high data quality. This approach effectively improves the data validity in subsequent causal relationship analysis, driving behavior analysis, and human-machine co-driving experience evaluation, and enhances the signal-to-noise ratio of the scene data.

[0072] This invention automates the entire data processing process, from triggering key interaction scenarios and confirming them to segmenting time windows, eliminating the need for manual post-event data review, event filtering, and scenario boundary annotation. By automatically analyzing continuous driving data, it can quickly extract valuable key interaction scenarios from massive amounts of driving data and automatically generate corresponding standardized scenario data fragments. This automated processing method reduces subjective differences in manual annotation, improves the consistency and efficiency of scenario data construction, meets the needs of large-scale intelligent driving test data processing, and enables the automated and standardized construction of driving scenario datasets, providing reliable data support for intelligent driving system optimization and human-machine interaction performance evaluation.

[0073] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An automatic identification and segmentation method for key interactive scenarios in intelligent driving, characterized in that, Includes the following steps: Acquire vehicle perception data during intelligent driving, analyze the lateral position change trend of targets in adjacent lanes and the predicted collision time between targets in adjacent lanes and the vehicle. When it is detected that the lateral position of targets in adjacent lanes is continuously moving towards the center line of the vehicle, and the predicted collision time meets the preset entry risk conditions, generate a potential entry scenario flag. Based on the potential entry scenario marker, the lane change status of the target in the adjacent lane and the intelligent driving control behavior of the vehicle are detected. When the target in the adjacent lane is detected to continuously cross the lane line and enter the lane, and the intelligent driving control behavior of the vehicle generates a deceleration command, the potential entry scenario marker is identified as the entry interaction scenario, and the time corresponding to the deceleration command is identified as the core event time of the entry interaction scenario. Based on the core moment of the event, a time-reverse search is performed on the vehicle operation data and driver physiological data corresponding to the entry interaction scenario to obtain the time position when the vehicle's longitudinal motion state recovers to a stable state and the driver's psychological load recovers to the stable state before the trigger, and this time position is determined as the starting boundary of the entry interaction scenario. Based on the core moment of the event, a forward time search is performed on the vehicle operation data and driver physiological data corresponding to the entry interaction scenario to obtain the time position when the vehicle resumes a stable following state and the driver's psychological load is reduced to a preset stable state, and this time position is determined as the end boundary of the entry interaction scenario. Based on the start and end boundaries, continuous driving data is extracted to generate data segments corresponding to the interactive scenarios, thereby enabling automatic identification and complete time window segmentation of key interactive scenarios in intelligent driving.

2. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 1, characterized in that, The system acquires vehicle perception data during intelligent driving and analyzes the lateral position change trends of targets in adjacent lanes and the predicted collision time between the targets in adjacent lanes and the vehicle itself. This includes: continuously acquiring the position information of targets in adjacent lanes, continuously tracking the lateral movement of targets in adjacent lanes in the road coordinate system, and determining whether the target vehicle has a tendency to move towards the center line of the lane based on the lateral position changes of targets in adjacent lanes over a continuous period of time; simultaneously acquiring the longitudinal distance change information between targets in adjacent lanes and the vehicle itself, predicting the collision risk between targets in adjacent lanes and the vehicle itself based on the vehicle's motion state, and generating a corresponding potential entry scenario flag when the lateral position change trend and the predicted collision time simultaneously meet the potential entry scenario judgment conditions.

3. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 1, characterized in that, When a target vehicle in an adjacent lane is detected to continuously cross the lane line and enter the current lane, and the vehicle's intelligent driving control behavior generates a deceleration command, the potential entry scenario marker is identified as an entry interaction scenario, and the time corresponding to the deceleration command is identified as the core event moment of the entry interaction scenario. This includes: continuously detecting the trajectory of the target vehicle after the potential entry scenario marker is generated, and determining whether the target vehicle gradually enters the lane where the vehicle is located from the adjacent lane; when a target vehicle is detected to continuously cross the lane line and enter the current lane, and the vehicle's intelligent driving control behavior is detected to output a deceleration command based on the target vehicle's motion state, the current potential entry scenario marker is converted into a confirmed entry interaction scenario, and the time and location of the deceleration command are taken as the core event moment.

4. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 1, characterized in that, Based on the core moment of the event, a reverse time search is performed on the vehicle operation data and driver physiological data corresponding to the interaction scenario. This includes: using the core moment of the event as a time reference point, continuously searching along the time axis of the driving data in the direction before the event, simultaneously acquiring vehicle operation status data and driver EEG data, and analyzing changes in the vehicle's longitudinal motion state and the driver's psychological load; by determining whether the vehicle's longitudinal motion state has recovered to a stable state and whether the driver's psychological load has recovered to a stable state before the trigger, the time position that meets the recovery conditions of the preceding state is determined, and this time position is determined as the starting boundary of the interaction scenario.

5. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 4, characterized in that, The process of restoring the vehicle's longitudinal motion state to a stable state includes: scanning the vehicle's operating data backward from the core moment of the event, detecting changes in the vehicle's longitudinal acceleration, and determining that the vehicle's motion state meets the recovery conditions when the vehicle's longitudinal acceleration smoothly falls back to within the range of ±0.2 m / s² and remains stable. Here, ±0.2 m / s² represents the allowable range of variation in longitudinal acceleration when it returns to a stable state, and is used to determine whether the vehicle has recovered from the dynamic adjustment state during the interaction process to a stable operating state.

6. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 4, characterized in that, The process of restoring the driver's psychological load to a stable state before the trigger includes: analyzing the driver's EEG data, obtaining the decoded driver's psychological load value, comparing the current psychological load value with the average psychological load level in the 30 seconds before the scene trigger, and determining that the driver's psychological state has returned to a stable state when the driver's psychological load value returns to the average level in the 30 seconds before the scene trigger is within ±0.5 standard deviations. The corresponding time position is used as the basis for judgment in the process of determining the starting boundary.

7. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 1, characterized in that, Based on the core moment of the event, a forward time search is performed on the vehicle operation data and driver physiological data corresponding to the entry interaction scenario. This includes: using the core moment of the event as a time reference point, continuously searching along the driving data time axis in the direction after the event occurs, and synchronously analyzing the changes in vehicle operation status and driver psychological load; when it is detected that the vehicle has returned to a stable following state and the vehicle acceleration has stabilized again, and the driver's psychological load has decreased to a preset stable state, the corresponding time position is determined as the end boundary of the entry interaction scenario, so as to complete the automatic determination of the event end time position.

8. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 7, characterized in that, The process of reducing the driver's psychological load to a preset stable state includes: continuously analyzing the driver's EEG data after the core moment of the event; when the decoded driver's psychological load value is below 0.4 for 3 consecutive seconds, it is determined that the driver's psychological load has reached a stable state; where 0.4 is the calibration threshold in the process of judging the driver's psychological load, and 3 consecutive seconds are used to judge the continuous stability of the driver's psychological load recovery process, so as to avoid the deviation of the end boundary judgment caused by the change at a single time point.

9. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 1, characterized in that, Based on the start and end boundaries, continuous driving data is extracted to generate data segments corresponding to the entry interaction scenario. This includes: using a determined start boundary as the starting position for data extraction and a determined end boundary as the ending position for data extraction; extracting continuous driving data between the start and end boundaries; and extracting vehicle operation data and driver physiological data corresponding to the entry interaction scenario to form a complete entry interaction scenario data segment for subsequent analysis.

10. The automatic identification and segmentation method for key interactive scenarios in intelligent driving according to claim 1, characterized in that, To achieve automatic identification and complete time window segmentation of key interactive scenarios in intelligent driving, the system continuously analyzes vehicle perception data, vehicle operation data, and driver physiological data. Without the need for manual post-event scenario screening and boundary labeling, it automatically completes the detection of potential entry scenarios, confirmation of entry interaction scenarios, determination of the core moment of the event, determination of the start boundary, and determination of the end boundary. Based on the determined time window, it generates standardized entry interaction scenario data fragments.

Images