An emergency scene driving behavior data processing method based on deep learning

Abstract

The application discloses an emergency scene driving behavior data processing method based on deep learning, and relates to the technical field of driving behavior data processing methods.The method comprises the following steps: setting at least two data simulation modes, wherein the two data simulation modes are a single working condition simulation scene and a comprehensive working condition simulation scene; collecting different driving data generated by different drivers in the two working condition simulation scenes; intelligently removing abnormal samples based on data correlation in each driving data and neural network anomaly detection; and obtaining first target driving data and second target driving data for establishing a database of a driver behavior model.The data processing method can effectively improve the reaction speed and safety of an automatic driving system, and realizes decision optimization and safety performance improvement of vehicle AI.

Description

TECHNICAL FIELD

[0001] The present application generally relates to the technical field of driving behavior data processing methods, and in particular to an emergency scene driving behavior data processing method based on deep learning. BACKGROUND

[0002] At present, the automatic driving system lacks comprehensive and accurate behavior data support for the steering and braking operations of human drivers in emergency scenes such as emergency braking, cutting in and cutting out, resulting in insufficient decision-making and control capabilities, specifically in terms of reaction time, vehicle behavior and safety performance that need to be improved. At the same time, existing data collection has problems such as poor hardware configuration synergy, poor data synchronization, unclear key parameter extraction and calibration, and abnormal data interference, making it difficult to effectively optimize the automatic driving algorithm to deal with complex emergency driving situations, directly reducing the safety performance of the automatic driving system and failing to meet the increasingly stringent requirements of automatic driving technology for vehicle response performance in emergency scenarios. SUMMARY

[0003] In view of the above-mentioned defects or shortcomings in the prior art, it is desirable to provide an emergency scene driving behavior data processing method based on deep learning.

[0004] The present application provides an emergency scene driving behavior data processing method based on deep learning, comprising the following steps:

[0005] At least two data simulation modes are set, and the two data simulation modes are single working condition simulation scenes and comprehensive working condition simulation scenes, respectively. Both working condition simulation scenes include emergency braking working condition, cutting in working condition and cutting out working condition.

[0006] Different driving data generated by different drivers in single working condition simulation scenes and comprehensive working condition simulation scenes are collected, and based on the data correlation within each driving data, combined with a neural network anomaly detection method, the two groups of driving data are subjected to abnormal sample elimination to obtain first target driving data and second target driving data.

[0007] Based on the first target driving data and the second target driving data, a training database is constructed; the training database is used to train a neural network-based driving behavior prediction model of the automatic driving system in emergency scenes; wherein the first target driving data and the second target driving data respectively include a plurality of key parameters of the target vehicle when generating braking behavior and steering behavior in the single working condition simulation scene and the comprehensive working condition simulation scene.

[0008] According to the technical scheme provided by the present application, at least two data simulation modes are set, including:

[0009] Select a plurality of drivers with different driving styles, and connect a data acquisition device for simulation;

[0010] Select test environment conditions, and input preset road adhesion coefficients and preset maximum set braking decelerations in the data acquisition device;

[0011] At least independently set an emergency braking condition, a cut-in condition and a cut-out condition in a simulation platform to form the single-condition simulation scene; and at least set a cut-out-braking condition and a cut-in-braking condition in the simulation platform to form the comprehensive-condition simulation scene.

[0012] According to the technical scheme provided by the application, different driving data generated by different drivers in a single-condition simulation scene are collected, including:

[0013] In the emergency braking condition,

[0014] Based on the braking instruction triggering time corresponding to each driver and the target vehicle response time, a personal average decision response time length corresponding to each driver is calculated, so as to calculate a comprehensive average decision response time length according to the personal average decision response time lengths;

[0015] Based on the braking pedal pressing time corresponding to each driver and the time at which the braking efficiency reaches the maximum braking deceleration, an average braking efficiency improvement time length corresponding to each driver is calculated, so as to calculate a comprehensive average braking efficiency improvement time length according to the average braking efficiency improvement time lengths;

[0016] Based on the maximum braking deceleration reached by each driver, a comprehensive average maximum braking deceleration is calculated.

[0017] According to the technical scheme provided by the application, different driving data generated by different drivers in a single-condition simulation scene are collected, including:

[0018] In the cut-in condition or the cut-out condition,

[0019] Confirm the risk occurrence time corresponding to each driver, which is used to mark the triggering time of the cut-in condition or the cut-out condition;

[0020] Obtain a steering wheel angle data set in a preset effective time length, and obtain a maximum steering wheel turning angle, the preset effective time length being set based on the starting time corresponding to the risk occurrence point;

[0021] A turning angle-time length relationship model is established based on actual data of the steering wheel angle data set changing with time length, and a steering wheel return time length is obtained based on the turning angle-time length relationship model and the maximum steering wheel turning angle.

[0022] According to the technical scheme provided by the application, the steering angle-time length relationship model is established, comprising:

[0023] The obtained steering wheel steering angle data set is subjected to polynomial fitting to obtain a fitted steering angle model expression and polynomial coefficient values, the polynomial coefficient values being each coefficient value fitted based on the steering angle model expression by the least square method;

[0024] Based on the steering angle model expression, the polynomial coefficient values and a preset steering wheel return relationship expression, an expression for representing a steering wheel return time is obtained, and based on the return time expression and a return start time corresponding to the maximum steering wheel steering angle, the steering angle-time length relationship model for outputting the steering wheel return time length is established.

[0025] According to the technical scheme provided by the application, different driving data generated by different drivers in a comprehensive working condition simulation scene is collected, comprising:

[0026] In the cut-out-braking working condition or the cut-in-braking working condition,

[0027] A working condition time sequence is set, the working condition time sequence being a risk trigger sequence of the cut-out working condition, the cut-in working condition and the braking working condition;

[0028] The first risk occurrence time and the second risk occurrence time of different drivers in the cut-out-braking working condition and the cut-in-braking working condition are obtained, the first risk occurrence time corresponding to the cut-out-braking working condition and the second risk occurrence time corresponding to the cut-in-braking working condition;

[0029] Based on the first risk occurrence time, the action instruction trigger time of different drivers in the cut-out-braking working condition, the time when the braking efficiency reaches the maximum braking deceleration, the maximum braking deceleration that can be reached, the maximum steering wheel steering angle and the steering wheel return time length, and combined with a deep learning model, the first target average decision response time length, the first target average braking efficiency improvement time length, the first target average maximum braking deceleration, the first target maximum steering wheel steering angle and the first target steering wheel return time length are respectively calculated;

[0030] Based on the second risk occurrence time, the action instruction trigger time of different drivers in the cut-in-braking working condition, the time when the braking efficiency reaches the maximum braking deceleration, the maximum braking deceleration that can be reached, the maximum steering wheel steering angle and the steering wheel return time length, and combined with a deep learning model, the second target average decision response time length, the second target average braking efficiency improvement time length, the second target average maximum braking deceleration, the second target maximum steering wheel steering angle and the second target steering wheel return time length are respectively calculated.

[0031] According to the technical scheme provided by the application, based on the data correlation inside each driving data, combined with the neural network anomaly detection method, the two groups of driving data are subjected to abnormal sample elimination to obtain first target driving data and second target driving data, which comprises:

[0032] In the emergency braking condition,

[0033] By using AI anomaly detection, the longitudinal acceleration of the target vehicle is obtained, and if the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold, the data sample at the corresponding time is eliminated.

[0034] In addition, the pedal stroke and the corresponding deceleration when the driver steps on the brake pedal are obtained, and if they do not match, the data sample in the time period is identified and eliminated by using the AI model.

[0035] According to the technical scheme provided by the application, based on the data correlation inside each driving data, combined with the neural network anomaly detection method, the two groups of driving data are subjected to abnormal sample elimination to obtain first target driving data and second target driving data, which comprises:

[0036] Based on the set condition time sequence, by using AI anomaly detection, the data sample that does not match the set condition time sequence or does not complete the complete cut-out-braking condition and cut-in-braking condition is eliminated.

[0037] In addition, the longitudinal acceleration of the target vehicle is obtained, and if the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold, the data sample at the time when the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold is eliminated.

[0038] In addition, the pedal stroke and the corresponding deceleration when the driver steps on the brake pedal are obtained, and if they do not match, the data sample in the time period is identified and eliminated by using the AI model.

[0039] In summary, the technical scheme specifically discloses an emergency scene driving behavior data processing method based on deep learning, which comprises the following steps: setting at least two data simulation modes, wherein the two data simulation modes are a single working condition simulation scene and a comprehensive working condition simulation scene, and the two working condition simulation scenes each comprise an emergency braking working condition, a cut-in working condition and a cut-out working condition; collecting different driving data generated by different drivers in the single working condition simulation scene and the comprehensive working condition simulation scene, and based on the data correlation in each driving data, combining a neural network anomaly detection method, performing abnormal sample elimination on the two groups of driving data to obtain first target driving data and second target driving data; based on the first target driving data and the second target driving data, constructing a training database; and the training database is used for training a neural network-based driving behavior prediction model of an automatic driving system in an emergency scene; wherein the first target driving data and the second target driving data each comprise a plurality of key parameters of a target vehicle when the target vehicle generates braking behavior and steering behavior in the single working condition simulation scene and a plurality of key parameters of the target vehicle when the target vehicle generates braking behavior and steering behavior in the comprehensive working condition simulation scene.

[0040] At present, the automatic driving system directly reduces the safety performance in emergency scenes such as emergency braking, cut-in and cut-out due to the lack of comprehensive and accurate behavior data of human drivers in these scenes. In the present application, the single working condition simulation scene and the comprehensive working condition simulation scene are set to collect comprehensive driving data, and the artificial intelligence model is used to realize strict abnormal sample elimination and behavior mode extraction, so as to construct a training database for optimizing the driving behavior model of the automatic driving system. The method can improve the reaction speed and safety of the automatic driving system, and realize the optimization and safety improvement of the automatic driving system. BRIEF DESCRIPTION OF DRAWINGS

[0041] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings:

[0042] Figure 1 It is a flowchart of an emergency scene driving behavior data processing method for an automatic driving system.

[0043] Figure 2 It is a flowchart of step S100 in the present application.

[0044] Figure 3 It is a flowchart of the braking working condition in step S200 in the present application.

[0045] Figure 4 It is a flowchart of the steering working condition in step S200 in the present application.

[0046] Figure 5 Figure 2 is a flowchart illustrating a process of step S206 in the present application.

[0047] Figure 6 Figure 3 is a flowchart illustrating a process of step S200 in the present application. DETAILED DESCRIPTION

[0048] The application will be further described below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the application, and not to limit the application. In addition, it should be noted that only parts related to the application are shown in the drawings for ease of description.

[0049] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0050] Embodiment 1

[0051] In order to make the technical solutions of the embodiments of the present application more clear and easy to understand, the application background of the embodiments of the present application is introduced below.

[0052] With the rapid development of artificial intelligence and intelligent transportation technology, autonomous driving systems are gradually moving from the laboratory to practical road applications. The core goal is to achieve a safer and more efficient travel experience. However, the reliability of autonomous driving systems in complex traffic environments, especially in emergency scenarios such as sudden obstacles, sudden vehicle cut-in / cut-out, etc., the decision-making and control capabilities are still the key bottlenecks restricting their large-scale landing. In such scenarios, the vehicle needs to complete braking or steering operations within a very short time, and any reaction delay or decision deviation can lead to serious accidents, so there are very high requirements for the real-time, accuracy and safety of autonomous driving systems.

[0053] The existing autonomous driving system is difficult to accurately simulate the emergency operation logic of human drivers in sudden scenarios such as emergency braking, vehicle cut-in, vehicle cut-out, etc., so it leads to a reaction time lag, a lack of vehicle behavior rationality and a lack of safety performance. Specifically, for example, in emergency braking, the system cannot achieve efficient braking within the shortest time, which may cause collision risks due to improper braking timing or force; or when facing other vehicles cutting in or needing to cut out to avoid obstacles, the control parameters such as steering angle and return timing are not reasonable, which easily leads to trajectory deviation or collision problems. This series of problems is closely related to the fact that the key operation parameters (such as the core features of steering and braking) of human drivers in emergency scenarios are not fully collected and deeply analyzed in the prior art, ultimately making it difficult for autonomous driving systems to optimize algorithms based on real driving behaviors.

[0054] In view of this, the present application provides an emergency scene driving behavior data processing method for vehicle artificial intelligence based on deep learning, which comprises the following steps: setting at least two data simulation modes, the two data simulation modes being a single working condition simulation scene and a comprehensive working condition simulation scene respectively, the two working condition simulation scenes both comprising an emergency braking working condition, a cut-in working condition and a cut-out working condition; collecting different driving data generated by different drivers in the single working condition simulation scene and the comprehensive working condition simulation scene, and based on the data correlation in each driving data, and using a deep learning anomaly detection model to remove abnormal samples from the two groups of driving data to obtain first target driving data and second target driving data; constructing a training database based on the first target driving data and the second target driving data; the training database is used to train a neural network-based driving behavior prediction model of an automatic driving system in an emergency scene; wherein the first target driving data and the second target driving data respectively comprise a plurality of key parameters of a target vehicle when generating braking behavior and steering behavior in the single working condition simulation scene, and a plurality of key parameters of the target vehicle when generating braking behavior and steering behavior in the comprehensive working condition simulation scene.

[0055] It can be seen that, by setting the single working condition simulation scene and the comprehensive working condition simulation scene, the method collects rich dynamic behavior data of the vehicle in different driving scenes, extracts driving features and removes abnormal samples in combination with an artificial intelligence modeling method to ensure the accuracy of the data, and finally guides the decision-making and control ability of the optimized automatic driving algorithm based on the established training database, especially in the driving situations of emergency braking, cut-in and cut-out, the reaction time, vehicle behavior and overall safety performance of the automatic driving system are evaluated and improved through the neural network model.

[0056] Please refer to Figure 1 The automatic driving system emergency scene driving behavior data processing method provided by the present embodiment has the structure as shown in the structural schematic diagram of the automatic driving system emergency scene driving behavior data processing method.

[0057] S100, setting at least two data simulation modes, the two data simulation modes being a single working condition simulation scene and a comprehensive working condition simulation scene respectively, the two working condition simulation scenes both comprising an emergency braking working condition, a cut-in working condition and a cut-out working condition;

[0058] In the automatic driving system emergency scene driving behavior data processing method, single working condition simulation scene and comprehensive working condition simulation scene (both containing emergency braking, cutting in and cutting out working condition) are set at the same time, which can improve the effectiveness and practicality of driving behavior data from multiple dimensions. For example, the single working condition simulation scene can accurately focus on the core operation logic of the driver in a single sudden scene by independently setting emergency braking, cutting in and cutting out working condition. The comprehensive working condition simulation scene simulates complex scenes (such as sudden braking when cutting out to avoid) in actual roads by setting combined working conditions such as cutting out-braking and cutting in-braking, which can capture the behavior rules of the driver in the coordinated operation of steering and braking, and supplement the multi-action connection logic data that cannot be covered in single working condition.

[0059] Specifically, referring to Figure 2 The at least two data simulation modes in step S100 are single working condition simulation and comprehensive working condition simulation, which specifically include the following steps:

[0060] S101, select a plurality of drivers with different driving styles and connect the simulation data collection device;

[0061] First, personnel with diversified driving styles are selected from different driving groups, and the driving age is more than 10 years to ensure that the driver has rich driving experience. The driving mileage is more than 30,000 kilometers to ensure that the performance of the driver in different driving conditions is stable. The age, occupation and marital status of the driver can be diversified, including personnel of different ages and occupations, to reflect a wider range of driving styles.

[0062] For example, the driving style covers cautious type (such as slow driving speed, gentle braking and steering operation), aggressive type (such as rapid response, large operation amplitude), and conventional type (between the above two types), to avoid data deviation caused by single style and ensure the comprehensiveness of the calibration result, so that the subsequently constructed database is more universal. The specific definition of the driving style is not specially limited in the embodiment of the application, as long as the difference in driving habits of the driver is ensured.

[0063] Then, the simulation data collection device is provided for these drivers. In this embodiment, the joint simulation environment of the Prescan simulation platform and the MATLAB data analysis tool is used for data collection and analysis in combination with the data enhancement and feature extraction method driven by deep learning. The specific hardware configuration is as follows:

[0064] Prescan simulation platform: simulates the behavior of a vehicle in different driving scenes, and can generate various driving working conditions and environmental factors (such as the appearance of obstacles, sudden braking of the front vehicle, etc.);

[0065] MATLAB analysis tool: used to extract data from the Prescan simulation platform and call the neural network modeling module for real-time feature extraction and pattern recognition of driving behavior;

[0066] Logitech steering wheel simulator: provides real-time feedback to the driver, simulating the real-time response when the driver operates the steering wheel, ensuring the authenticity of the driving behavior.

[0067] In the specific data collection phase, multiple drivers conduct driving tests under different conditions in the simulation environment, and multiple sets of driving data are collected in real time, including vehicle speed, steering angle, acceleration, brake pedal, accelerator pedal, and other signal data. The collection frequency can be set to 100Hz, and the data is stored based on the relative time axis; during the entire data collection process, all hardware devices are connected and synchronized through a high-speed data bus. Prescan and MATLAB realize real-time data interaction through a dedicated interface, and the deep learning model can preliminarily detect and eliminate abnormal signals during the collection process, serving as an AI anomaly detection means to accelerate data processing speed; the Logitech steering wheel simulator establishes a real-time feedback mechanism with the Prescan simulation environment, and the data collection in each scene includes real-time synchronization of multiple parameters, ensuring that all data reflects the accurate moment of driver operation and vehicle response.

[0068] S102、Select test environment conditions and input preset road adhesion coefficient and preset maximum set brake deceleration in data collection equipment;

[0069] In order to ensure the accuracy and reliability of the experimental results, the following parameters are set:

[0070] Environmental conditions: choose good road conditions on a sunny day without wind for testing;

[0071] Road adhesion coefficient: according to relevant experiments and regulations, set to 1.0, which meets the good road conditions of most cities;

[0072] Maximum brake deceleration of vehicle performance: refer to the test results of existing vehicle models on the market, select the maximum brake deceleration of ordinary vehicle models for setting, to ensure the universality of the experimental results. In addition, the test parameters are input into the AI training module inside the simulation data collection equipment, which is used to establish the nonlinear mapping relationship between environmental conditions and driving behavior later.

[0073] As can be seen, by setting standard test environment conditions, environmental interference factors can be effectively eliminated, making the operation data of different drivers comparable and ensuring the stability of the experimental results.

[0074] S103, at least independently set emergency braking condition, cut-in condition and cut-out condition in the simulation platform, to form a single condition simulation scene;

[0075] Regarding the establishment of a single condition simulation scene, mainly including three conditions independent operation, and each scene is defined according to specific driving environment and traffic conditions, including the following:

[0076] Scene 1, emergency braking scene: simulate the reaction when encountering sudden obstacles or emergency situations in high-speed driving state, and require to complete braking in the shortest time.

[0077] Scene 2, cut-in scene: simulate the cut-in of other vehicles from the side lane, and the driver needs to avoid collision to ensure no collision.

[0078] Scene 3, cut-out scene: simulate the appearance of obstacles in front of the vehicle, and the driver needs to change lanes according to the environment to avoid obstacles.

[0079] Through independent condition design, the core features of single operation (such as braking response time, steering angle change) can be isolated, which facilitates precise analysis of driving personnel's behavior habits under single task; Provide pure data source for subsequent extraction of basic parameters (such as comprehensive average decision response time, maximum steering wheel steering angle), avoid parameter confusion caused by multiple operation interference, at the same time, since the condition simulation scene comes from many drivers with different driving styles, therefore, the subsequent collection of driving data will be classified by neural network classifier, to provide training samples for individual driving behavior modeling.

[0080] S104, at least set cut-out-braking condition and cut-in-braking condition in the simulation platform, to form a comprehensive condition simulation scene.

[0081] Each scene under the comprehensive condition simulation scene is combined by two single conditions, to simulate complex and continuous emergency scenes, which also needs to be defined according to specific driving environment and traffic conditions, including the following:

[0082] Condition 1, cut-out-braking condition: first trigger the scene that the vehicle needs to cut out the lane (such as obstacles in front), and the driver needs to steer and cut out the process, the simulation platform suddenly increases new obstacles (such as static vehicles in the cut-out target lane), requiring the driver to perform steering and braking operations at the same time.

[0083] Condition 2, cut-in-braking condition: first trigger the side vehicle to cut into the lane of the vehicle, and the driver needs to steer to avoid, and the simulation platform triggers the sudden obstacle in front, requiring the driver to perform steering and braking operations at the same time.

[0084] It should be noted that the foregoing case is the working condition sequence of triggering the target vehicle to cut out or cut in first and then triggering the brake, and in the design, it can also be triggered first brake and then cut out or cut in, and the specific is not specially limited, and in the data processing link, the multi-task operation data in the comprehensive working condition is modeled through a deep learning time sequence model (such as LSTM or Transformer), so as to be closer to the complex scene of continuous occurrence of multiple emergency situations in the real road, and provide data reference for multi-task processing of the automatic driving system, so as to optimize the operation coordination logic in the complex scene.

[0085] In the embodiment of the application, the input of the neural network-based driving behavior prediction model includes multi-dimensional signal features obtained by driving data collection, specifically:

[0086] Brake-related parameters: brake instruction triggering time, brake pedal stroke, brake efficiency improvement duration, maximum brake deceleration; steering-related parameters: steering wheel angle sequence, maximum steering wheel turning angle, steering wheel return duration; vehicle dynamic parameters: vehicle longitudinal speed, longitudinal acceleration; risk triggering reference: risk occurrence point.

[0087] The output of the prediction model is the prediction result of the driving behavior under the corresponding emergency scene, specifically including:

[0088] Action type: braking, steering (cutting in or cutting out), or both; action timing: predicted brake start time, steering start time; action amplitude: predicted target brake deceleration, predicted target steering angle; action duration feature: predicted brake efficiency improvement process, steering wheel return process.

[0089] Through the input and output relationship of the above prediction model, the model can learn and predict the typical emergency operation logic of the driver under single working condition and comprehensive working condition, thereby providing reference for real-time decision and control of the automatic driving system in the emergency scene.

[0090] S200, collect different driving data generated by different drivers in single working condition simulation scenes and comprehensive working condition simulation scenes, and based on the data correlation inside each driving data, combine a neural network anomaly detection method to remove abnormal samples from the two groups of driving data, to obtain first target driving data and second target driving data;

[0091] The step S200 is the core link of the data processing method proposed in the embodiment of the application, mainly including driving data collection and abnormal sample elimination, and the purpose is to accurately extract driving behavior data in a single working condition simulation scene and a comprehensive working condition simulation scene, thereby laying a foundation for subsequent database construction. In the data collection and analysis process, the artificial intelligence modeling method is introduced, the deep learning feature extraction, time series modeling and abnormal detection mechanism are used, and the automation and accuracy of data processing are improved.

[0092] Specifically, the collection of different driving data generated by different drivers in the single working condition simulation scene and the comprehensive working condition simulation scene in the step S200 can be divided into the following two cases:

[0093] The first case is to collect different driving data generated by different drivers in the single working condition simulation scene, including the following steps:

[0094] I. Refer to Figure 3 In the emergency braking condition,

[0095] S201, based on the braking instruction triggering time corresponding to each driver and the target vehicle response time, the personal average decision response time corresponding to each driver is calculated, so as to calculate the comprehensive average decision response time according to the personal average decision response time.

[0096] In the traditional decision response time calculation, the artificial defined index (such as “brake pedal displacement more than 5% is regarded as instruction triggering” and “vehicle speed decreases by 0.5 m / s is regarded as vehicle response”) is often relied on, but the artificial index is easily affected by subjective experience and is difficult to adapt to complex driving scenes. In the embodiment of the application, the multi-dimensional signals of brake pedal, vehicle speed and acceleration are automatically encoded by using the feature extraction method based on convolutional neural network (CNN), and the redundant noise is eliminated, so as to reduce the deviation caused by artificial defined index and ensure that the calculation result is more stable.

[0097] Specifically, the brake behavior refers to the deceleration or stop of the vehicle controlled by the driver through the brake pedal; the analysis and calibration of the brake signal can optimize the brake control strategy, so that the automatic driving system can quickly and smoothly respond in the face of emergency; at the same time, the driving data collected by the brake behavior in the emergency braking condition can calculate the key parameters such as decision response time, brake efficiency improvement time and maximum brake deceleration.

[0098] Among them, the so-called feature extraction refers to extracting core parameters capable of representing the emergency operation behavior of the driver from the original driving data. These features mainly include: decision response time (calculated from the brake instruction trigger time and the vehicle response time), brake efficiency improvement time (calculated from the brake pedal down time and the maximum deceleration time), maximum brake deceleration, maximum steering wheel steering angle and steering wheel return time. These calculated features can be processed by a deep learning model (such as convolutional neural network CNN, long short-term memory network LSTM) later: finally, the input of the deep learning model is the above multi-dimensional time sequence signal, and the output is the prediction parameter or classification result of the driving behavior, such as action type, timing and amplitude. In this way, the original continuous signal is converted into structured behavior features, which are used to build a training database and guide the decision optimization of the autonomous driving system in emergency scenarios.

[0099] Specifically, the decision response time refers to the time required from the time when the driver makes a decision to the time when the vehicle executes the decision. In order to obtain the average decision response time of the driver under different working conditions, for each driver, the decision response time can be calculated by recording the time difference between the brake instruction trigger time (for example, brake pedal action) and the target vehicle response time (for example, deceleration or acceleration) of each driver. Here, the decision response time of the first i driver is denoted as ; specifically, the decision response time of the first i driver can be calculated by the following formula (1):

[0100] Formula (1);

[0101] In formula (1), is the target vehicle response time; is the brake instruction trigger time; is the decision response time of the first i driver.

[0102] Further, in order to ensure the accuracy of the test results, each driver can perform emergency braking conditions under multiple experimental conditions to obtain multiple data samples, and then calculate the personal average decision response time of each driver ;

[0103] The personal average decision response time of the first driver is calculated by the following formula (2):

[0104] Formula (2);

[0105] In formula (2), For the first The total number of data samples for each driver; For the first The driver's first Decision response time for a number of data samples.

[0106] Ultimately, the average individual decision response time for all drivers. The average decision response time of all drivers can be calculated using the following formula (3). :

[0107] Formula (3);

[0108] In formula (3), Total number of drivers; This is a comprehensive average decision response time; This represents the average decision response time for an individual.

[0109] S202. Based on the time when the brake pedal is pressed and the time when the braking efficiency reaches the maximum braking deceleration for each driver, the average braking efficiency improvement time for each driver is calculated, and the comprehensive average braking efficiency improvement time is calculated based on the average braking efficiency improvement time.

[0110] Next, the braking performance enhancement time refers to the time required from when the driver first presses the brake pedal until the vehicle's braking performance reaches its maximum value, and the maximum braking deceleration refers to the maximum deceleration achieved by the target vehicle during braking. For each driver, this can be achieved by recording the moment the driver presses the brake pedal. The time value at which the target vehicle's braking performance reaches its maximum braking deceleration The duration difference between them is used to calculate the braking performance improvement time. Here, the first... The duration of braking performance improvement for each driver is denoted as... Specifically, the duration of braking performance improvement for each driver. It is calculated using the following formula (4):

[0111] Formula (4);

[0112] in, For the first The duration of improved braking performance for each driver; The moment the brake pedal is depressed; This refers to the moment when the braking efficiency reaches its maximum braking deceleration.

[0113] Further, for each driver, the average braking performance improvement duration under multiple experimental conditions The average braking performance improvement duration under multiple experimental conditions is calculated by the following formula (5):

[0114] Formula (5);

[0115] In formula (5), is the braking performance improvement duration of the i-th data sample of the j-th driver; is the total number of data samples of the j-th driver; is the total number of data samples of the j-th driver; is the average braking performance improvement duration of the j-th driver.

[0116] Finally, the average braking performance improvement duration of all drivers is calculated by averaging the average braking performance improvement duration of all drivers, i.e., the comprehensive average braking performance improvement duration of all drivers is calculated by the following formula (6):

[0117] Formula (6);

[0118] wherein, is the comprehensive average braking performance improvement duration; is the total number of drivers; is the average braking performance improvement duration of the i-th driver.

[0119] S203, based on the maximum braking deceleration reached by each driver, the comprehensive average maximum braking deceleration is calculated.

[0120] For each driver, the maximum braking deceleration that each person can reach under different experimental conditions can be determined by recording the maximum deceleration reached by the vehicle under the entire emergency braking condition (without taking the average but taking the maximum value), and the maximum braking deceleration can be expressed by the following formula (7):

[0121] Formula (7);

[0122] is the maximum deceleration reached in the i-th data sample of the j-th driver. j Finally, the comprehensive average maximum braking deceleration of all drivers is calculated by the following formula (8):

[0123] is calculated by the following formula (8):​​​​​​

[0124] Equation (8);

[0125] wherein, is the comprehensive average maximum braking deceleration of all drivers; is the maximum braking deceleration of the i-th driver; is the maximum braking deceleration of the i-th driver; is the total number of drivers.

[0126] It should be noted that emergency braking refers to the situation where the driver must immediately brake the vehicle due to changes in road conditions, sudden braking of the vehicle in front, etc., in order to avoid a collision or reduce the severity of an accident. In this emergency braking condition, the time point at which the risk occurs can be when the braking deceleration of the vehicle first reaches -7 m / s², and the time at which the risk occurs is recorded as the risk time. Collecting driving data starting from the risk time is more accurate.

[0127] II. Referring to Figure 4 , in the cut-in condition or the cut-out condition,

[0128] S204, confirming the risk occurrence time corresponding to each driver, the risk occurrence time is used to determine the trigger time of the cut-in condition or the cut-out condition;

[0129] Here, it is first necessary to clarify that the steering behavior refers to the driver controlling the steering angle of the vehicle through the steering wheel, thereby affecting the trajectory and direction of travel of the vehicle. At the same time, the calibration of the steering behavior can help optimize the steering control of the automatic driving system, so that it can more accurately simulate and respond to the behavior of the driver. Here, two cases of cut-in and cut-out are considered.

[0130] In order to improve the modeling accuracy of the steering behavior, the embodiment introduces a time series analysis method based on long short-term memory network (LSTM) to model the steering wheel angle sequence, which can better depict the dynamic process of the driver's steering response before and after the risk occurrence point.

[0131] Then, for the cut-in condition or the cut-out condition in the emergency steering risk avoidance scenario, based on the driving data derived from the driving simulator in the embodiment of the present application, the following three key parameters need to be analyzed: risk occurrence time, maximum steering wheel steering angle, and steering wheel return duration.

[0132] In the embodiment of the present application, the emergency cut-in refers to the situation that the target vehicle must quickly cut in or change lanes to avoid collision with other vehicles or encounter obstacles during driving. In this working condition, the time point of risk occurrence can be when the right front wheel of the target vehicle is first laterally offset by more than 0.580 m, and the time point is recorded as the risk occurrence point; the cut-out and encounter with a stationary front vehicle refers to the situation that a stationary obstacle suddenly appears in front of the vehicle on the road after the vehicle cuts out or changes lanes, and the driver needs to take timely measures to avoid collision. In this working condition, the time point of risk occurrence is when the lateral displacement of the edge of the left rear bumper of GVT1 exceeds 1 / 4 of the width of GVT2, and the time point is recorded as the risk occurrence point.

[0133] S205, obtaining a steering wheel angle data set in a preset effective time length period, obtaining a maximum steering wheel turning angle, and the preset effective time length period is set based on a starting time corresponding to the risk occurrence point;

[0134] The maximum steering wheel angle refers to the maximum steering wheel turning angle reached by the driver during emergency turning. In the actual maximum steering wheel turning angle acquisition process, data 5 seconds before and after the risk occurrence point can be obtained as an effective data segment, and the steering wheel angle data in the effective data segment forms a steering wheel angle data set. The preset effective time length period in this embodiment is an effective time length period of 5 seconds before and after the risk occurrence point, and the length of the effective time length period is not limited in the present application.

[0135] In actual application, the data in the steering wheel angle data set of the driver's vehicle is smoothed and filtered to remove abnormal data points caused by collision or system deviation, and the maximum steering wheel angle is finally obtained by selecting the maximum steering wheel angle of 10 frames of slice data and taking the average value as the maximum steering wheel turning angle.

[0136] Finally, the maximum steering wheel turning angle is calculated by the following formula (9) :

[0137] Formula (9);

[0138] Wherein, is the steering wheel angle in each frame of data; is the maximum steering wheel turning angle.

[0139] S206, establishing a turning angle-time length relationship model based on the actual data of the steering wheel angle data set changing with time length, and obtaining the steering wheel return time length based on the turning angle-time length relationship model and the maximum steering wheel turning angle.

[0140] After obtaining the steering wheel angle data set, in order to accurately obtain the steering wheel return time, this invention uses the steering wheel angle data and the actual data of time change to fit and establish a steering angle-time relationship model, thereby accurately extracting the key indicator of steering wheel return time, making the autonomous driving system decision closer to human driving behavior.

[0141] In a preferred embodiment, see Figure 5 The process of establishing the steering angle-duration relationship model in step S206 specifically includes the following steps:

[0142] S2061. Perform polynomial fitting on the obtained steering wheel angle data set to obtain the fitted steering angle model expression and polynomial coefficient values. The polynomial coefficient values ​​are the coefficient values ​​of each term obtained by fitting the steering angle model expression using the least squares method.

[0143] The collected steering wheel angle data set is fitted with a polynomial. Here, a quadratic polynomial is selected for fitting. The final fitted steering angle model expression is as follows (10).

[0144] Formula (10);

[0145] in, Duration t The corresponding steering wheel angle; , and These are the polynomial coefficients.

[0146] Furthermore, the polynomial coefficients are obtained by fitting using the least squares method through the following formula (11).

[0147] Formula (11);

[0148] in, For the first The duration of each moment, The steering wheel angle was recorded from the actual test; ultimately, it was minimized. You can get , and The actual value of is the coefficient value of the polynomial coefficient.

[0149] S2062. Based on the steering angle model expression, polynomial coefficient values, and a pre-defined steering wheel return-to-center relationship, an expression is obtained to characterize the steering wheel return-to-center time. Based on the expression for the return-to-center time and the return-to-center start time corresponding to the maximum steering wheel steering angle, a steering angle-duration relationship model for outputting the steering wheel return-to-center time is established.

[0150] Further, the steering wheel turning angle is set to reach a percentage of the maximum turning angle, and the steering wheel turning back time is calculated, and the obtained steering wheel turning back relationship is set as formula (12):

[0151] Formula (12);

[0152] wherein, is the steering wheel turning angle at the steering wheel turning back time; is the maximum steering wheel turning angle; is a proportional coefficient of reaching the maximum steering wheel turning angle.

[0153] Combining formula (10) and formula (12), the equation of formula (13) can be obtained;

[0154] Formula (13);

[0155] By solving the quadratic equation of formula (13), the final expression for representing the steering wheel turning back time can be obtained, and the specific formula (14) is as follows:

[0156] Formula (14);

[0157] wherein, is the steering wheel turning back time; is the maximum steering wheel turning angle; is a proportional coefficient of reaching the maximum steering wheel turning angle.

[0158] Finally, when the expression for representing the steering wheel turning back time is calculated, and the data acquisition device also collects the steering wheel turning start time , the steering wheel turning back time can be obtained by formula (15) as follows:

[0159] Formula (15).

[0160] As described above, the steering angle-time relationship model can be established by formula (15) and formula (14), and in the actual application scenario, by collecting the braking instruction triggering time, the target vehicle response time, the braking pedal stepping time, the time when the braking efficiency reaches the maximum braking deceleration, the maximum braking deceleration reached by the driver, the steering wheel turning angle data and other driving data, the comprehensive average decision response time, the comprehensive average braking efficiency improvement time, the comprehensive average maximum braking deceleration, the maximum steering wheel turning angle and the steering wheel turning back time and other key parameters can be analyzed.

[0161] ​On this basis, a neural network regression model can also be used instead of polynomial fitting to establish a nonlinear steering angle-time relationship model, making the calculation of steering wheel return time more close to the real driving behavior characteristics. Specifically, the process of using a neural network regression model to replace polynomial fitting to establish a nonlinear steering angle-time relationship model includes:

[0162] Data preparation: using the preprocessed steering wheel angle data set as a sample, where the input features are time t , and the output label is the steering wheel angle at the corresponding time, the training set, validation set and test set are divided.

[0163] Model construction: select a neural network structure suitable for regression tasks (such as fully connected neural network, LSTM recurrent neural network, etc.), set the corresponding parameters (for example, the number of hidden layers, the number of neurons, the activation function).

[0164] Model training and optimization: train the model with the mean square error (MSE) of the predicted angle and the actual angle as the loss function, adjust the hyperparameters through the validation set to ensure that the fitting error of the model on the test set is minimized, and finally obtain the nonlinear steering angle-time relationship model.

[0165] Next, based on the data processing method of the single working condition simulation scene, the second case also includes: collecting different driving data generated by different drivers in a comprehensive working condition simulation scene, including the following steps:

[0166] III. See Figure 6 In the cut-out-braking condition or the cut-in-braking condition,

[0167] S207, set the working condition timing, which is the risk trigger sequence of the cut-out condition, the cut-in condition and the braking condition;

[0168] After setting the working condition timing, further use a deep learning algorithm based on sequence modeling to model the operation link of the driver in the multi-risk superposition scene, so as to quantify the operation logic and delay characteristics of the driver in the "multi-task conflict".

[0169] Specifically, for example, in the "cut-out-braking" condition, the driver first performs a steering action, and then triggers a braking operation in the new lane; a sequence modeling network (such as a long short-term memory network LSTM or Transformer) can capture the time dependence between the start of steering and the trigger of braking, thereby quantifying the reaction delay time of the driver in continuous task switching.

[0170] For example, in the "cut-in-braking" scenario, the driver first performs a steering maneuver to avoid a vehicle cutting in from the side, and then triggers braking due to an obstacle in front. In this scenario, the model not only extracts the timing characteristics of the steering angle change and brake pedal stroke, but also outputs the transition delay and action overlap between the two, to describe the driver's operation logic under multi-task conflict conditions.

[0171] Based on the foregoing, it is known that the comprehensive scenario simulation scenario includes comprehensive scenarios, i.e., simulating multi-risk superposition scenarios in real driving, so the timing of the comprehensive scenario and the interval length between the two scenarios (here, the interval length is the time interval between the risk occurrence points of the cut-out / cut-in scenario and the braking scenario) need to be set, i.e., whether to trigger cut-in or cut-out first or to trigger braking first. In the actual data processing and collection and processing process, various superposition scenarios can be tested, and here a cut-out-braking scenario or a cut-in-braking scenario is described, so the trigger sequence needs to be clarified in order to facilitate subsequent data screening and calculation.

[0172] S208, obtaining the first risk occurrence time and the second risk occurrence time of different drivers in the cut-out-braking scenario and the cut-in-braking scenario; the first risk occurrence time corresponds to the cut-out-braking scenario, and the second risk occurrence time corresponds to the cut-in-braking scenario;

[0173] Since the principle of the cut-out-braking scenario or the cut-in-braking scenario in data processing cannot be separated from the foregoing content description, a unified description is provided here. First, the risk occurrence time in the cut-out-braking scenario and the cut-in-braking scenario, i.e., the first risk occurrence time and the second risk occurrence time, needs to be obtained, and the first risk occurrence time and the second risk occurrence time should theoretically include the two risk occurrence times of the cut-out and braking scenarios and the two risk occurrence times of the cut-in and braking scenarios, to establish a time length reference for quantifying the driver's perception and response timing of different risks.

[0174] It should be noted that the risk occurrence point mentioned in the foregoing can be understood as the occurrence time of the obstacle test condition set in the simulation environment that causes the driver to perform braking and steering operations, and the risk trigger time is the decision time of the driver to perform braking and steering operations when encountering an obstacle based on the driving data (for example, the time when the driver steps on the brake pedal or starts to turn the steering wheel), and the first risk occurrence time and the second risk occurrence time here can follow the judgment method described above or be adjusted according to the comprehensive superposition scenario, which is not specially limited here.

[0175] S209、based on the first risk occurrence moment and the different driver's action instruction trigger moment in the cut-out-braking working condition, the moment when the braking efficiency reaches the maximum braking deceleration, the maximum braking deceleration that can be reached, the maximum steering wheel steering angle and the steering wheel return time, combined with the deep learning model, the first target average decision response time, the first target average braking efficiency improvement time, the first target average maximum braking deceleration, the first target maximum steering wheel steering angle and the first target steering wheel return time are calculated respectively;

[0176] S210、based on the second risk occurrence moment and the different driver's action instruction trigger moment in the cut-in-braking working condition, the moment when the braking efficiency reaches the maximum braking deceleration, the maximum braking deceleration that can be reached, the maximum steering wheel steering angle and the steering wheel return time, combined with the deep learning model, the second target average decision response time, the second target average braking efficiency improvement time, the second target average maximum braking deceleration, the second target maximum steering wheel steering angle and the second target steering wheel return time are calculated respectively.

[0177] Based on the foregoing formulas (1) to (15), the first target average decision response time, the first target average braking efficiency improvement time, the first target average maximum braking deceleration, the first target maximum steering wheel steering angle and the first target steering wheel return time in the cut-out-braking working condition are obtained; Similarly, based on the foregoing formulas (1) to (15), the second target average decision response time, the second target average braking efficiency improvement time, the second target average maximum braking deceleration, the second target maximum steering wheel steering angle and the second target steering wheel return time can also be obtained; Since the above data calculation is based on the same principle as shown in the foregoing content, it can be calculated by the deep learning model built in the Prescan simulation platform, so here it is not described in detail.

[0178] It should be noted that for the cut-out-braking working condition and the cut-in-braking working condition, the conversion delay time from steering to braking and the intervention time of braking during steering need to be calculated, which directly reflect the operation logic and operation delay of the driver under multi-task conflict; Among them, the conversion delay time is the time interval from the start of the previous operation (for example, cut-in or cut-out) to the start of the subsequent operation (for example, braking), which is a key parameter for reflecting the decision switching speed of the driver between two operations; The intervention time of braking during steering is the time interval from the second risk occurrence moment to the start of the subsequent operation, which is a key parameter reflecting the response speed of the driver to the new risk.

[0179] Specifically, taking the cut-out-braking working condition as an example, its trigger sequence is to cut out first and then brake, and for such a case, its conversion delay time can be calculated by the steering wheel steering start time and the braking start time is obtained by difference calculation; the intervention timing of braking in the turning process is the risk occurrence point of the set braking operation and the braking start time is obtained by difference calculation.

[0180] After a variety of data is collected, invalid data caused by collision, operation error or system abnormality needs to be excluded to ensure that the final target driving data can truly reflect the normal braking behavior of the driver, so that the "based on the data correlation of each driving data, the abnormal samples of the two groups of driving data are excluded, and the first target driving data and the second target driving data are obtained" in step S200, specifically includes the following steps: in the emergency braking condition,

[0181] Step A1, using AI anomaly detection, obtaining the longitudinal acceleration of the target vehicle, and determining that if the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold, the data sample at the corresponding time is excluded;

[0182] This is because when analyzing the braking efficiency improvement time and the maximum braking deceleration, collision samples need to be excluded to ensure the accuracy of the calibration data. Because when a collision occurs, the longitudinal acceleration of the vehicle may be greater than a certain threshold, so in the actual data processing process, the longitudinal acceleration of the target vehicle under different experimental conditions needs to be monitored, and the AI model is used to determine if the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold, then it is determined as a collision sample and needs to be excluded. Generally, the longitudinal acceleration threshold is usually set based on real vehicle crash test data (for example, it is set to ), that is, the absolute value of the longitudinal acceleration is greater than or equal to .

[0183] Step A2, obtaining the pedal stroke and corresponding deceleration when the driver steps on the brake pedal, and if the two do not match, using the AI model to identify and exclude the data samples in the time period.

[0184] Here, the correlation between brake pedal stroke and deceleration needs to be verified to exclude operation-response mismatch data caused by driver misoperation or system failure to ensure that the data can reflect the true braking control relationship; for example, under normal circumstances, a pedal stroke of 50% corresponds to a deceleration of , if a pedal stroke of 50% corresponds to a deceleration of at a certain time, it is determined as mismatch, and the AI model needs to exclude the data sample.

[0185] In addition, a deep learning anomaly detection method based on an autoencoder is introduced to analyze the reconstruction error of high-dimensional driving data and automatically identify potential abnormal samples, thereby further improving the accuracy of the elimination.

[0186] In another case, the "elimination of abnormal samples based on the data correlation between the two groups of driving data in step S200" includes the following steps:

[0187] Step B1: Based on the set working condition time sequence, use AI anomaly detection to eliminate data samples that do not match the set working condition time sequence or do not complete the complete cut-out-brake working condition and cut-in-brake working condition;

[0188] As can be seen from the content shown in step S207, in the comprehensive working condition simulation scenario, the working condition time sequence needs to be set. If the AI anomaly detection finds that the collected data does not match the time sequence, the data sample will be eliminated. For example, if the working condition time sequence is to cut in first and then turn, and the data sample shows that the brake operation is earlier than the cut-in operation, it is determined that the time sequence is chaotic, and the data sample is eliminated. Of course, there is also a case where, for example, the data sample only contains cut-out / cut-in operation and does not contain subsequent brake operation (such as turning without stepping on the brake or brake operation not reaching the effective pedal stroke), which is also determined as incomplete working condition, and the data sample also needs to be eliminated.

[0189] Step B2: Obtain the longitudinal acceleration of the target vehicle, and if the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold, eliminate the data sample at the time when the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold;

[0190] Step B3: Obtain the pedal stroke when the driver steps on the brake pedal and the deceleration corresponding to the pedal stroke. If the two do not match, the data sample in the period of the mismatch is eliminated.

[0191] The above steps B2 and B3 have the same purpose as the aforementioned steps A1 and A2, and the data sample elimination principle in the single working condition simulation scenario is consistent. The longitudinal acceleration threshold can be adjusted, and here it is not described in detail.

[0192] Finally, the obtained key parameters such as the comprehensive average decision response time, the comprehensive average brake efficiency improvement time, the comprehensive average maximum brake deceleration, the maximum steering wheel steering angle, and the steering wheel return time are further used as training samples to input the neural network driving behavior prediction model, which is used to build an intelligent decision database for the automatic driving system in emergency scenarios.

[0193] S300, based on the first target driving data and the second target driving data, a training database is constructed; the training database is used to train a driving behavior model of the autonomous driving system in an emergency scene;

[0194] The first target driving data and the second target driving data respectively include a plurality of key parameters of the target vehicle when generating braking behavior and steering behavior in a single working condition simulation scene and a comprehensive working condition simulation scene.

[0195] Specifically, the database contains a large amount of effective operation data of different drivers in various emergency scenes, which reflects the general rules (such as average decision response time, reasonable steering angle range) and diversified strategies (such as operation preferences of different style drivers) of human drivers in response to unexpected situations. In this embodiment, the above-mentioned key parameters are converted into structured feature vectors and input into a deep learning model (such as a convolutional neural network CNN and a long short-term memory network LSTM) for training. The prediction model of the driving behavior built in the autonomous driving system can learn these data to establish an operation benchmark consistent with the human driving logic, thereby reducing the safety hazards of the autonomous driving system.

[0196] For example, the single working condition data can help the model master the basic operation precision in simple emergency scenes. For example, by learning the maximum braking deceleration and efficiency improvement time in the emergency braking condition, the braking control algorithm of the system is optimized, so that the vehicle can achieve the optimal braking effect in the shortest time. In this process, the model constantly fits the mapping relationship between the braking signal and the vehicle response through supervised learning, so as to gradually approach the best braking logic of human drivers. The comprehensive working condition data can help the model improve the cooperative operation ability in complex emergency scenes. For example, by learning the timing coordination rule of steering and braking in the cut-in-braking condition, the model can avoid operation conflicts (such as blind braking when steering is too much, which leads to vehicle out of control) of the system in continuous emergency situations. In order to achieve this goal, the multi-dimensional time series data in the training database is input into the sequence modeling network to extract the time series dependence relationship across operations, thereby improving the prediction and decision-making ability of the model under the condition of multi-task superposition. At the same time, the key parameters (such as decision response time and steering return time) in the database also provide quantitative operation constraints for the model. For example, the model can set its maximum allowed response delay based on the average decision response time in the database, so as to ensure that there is no collision due to slow reaction in an emergency scene. At the same time, by learning the reasonable range of steering wheel return time, the model can avoid too rapid or slow steering operation to ensure the stability of the vehicle trajectory.

[0197] In addition, in the model training process, this embodiment also introduces an abnormal data filtering and reinforcement learning mechanism based on an autoencoder, so that the model not only can imitate the average behavior of human drivers, but also can continuously optimize the strategy in iterative training to improve the generalization ability to unknown emergency scenes.

[0198] It can be seen that, by constructing a structured and associated training database, the first target driving data and the second target driving data collected in the early stage are converted into human driving experience that can be learned by the automatic driving system. The key braking and steering parameters contained in the training database provide comprehensive training materials for the model from a single scene to a comprehensive scene, and finally help the automatic driving system to make safer and more human-logic driving decisions and controls in emergency scenes, thereby improving the reaction speed and safety of the automatic driving system, providing an accurate driver behavior model for the system in complex traffic scenes such as emergency braking, cutting in and cutting out, and realizing optimization and safety improvement of the automatic driving system.

[0199] Specifically, in a single working condition, for example, when a stationary obstacle suddenly appears in front of the vehicle during high-speed straight driving, the system can call the comprehensive average decision response time and the comprehensive average maximum braking deceleration in the training database as constraints to predict the typical reaction delay and maximum braking intensity of the driver, thereby optimizing the braking trigger logic and braking control parameters of the automatic driving system; in a comprehensive working condition, for example, when an obstacle appears again in the target lane during the process of cutting out to avoid the obstacle, the system can use the steering angle-time relationship and the braking intervention timing during steering learned by the model to predict the cooperative operation mode of the driver on the continuous risk, and reasonably allocate the timing and amplitude of steering and braking actions, thereby avoiding operation conflicts or trajectory instability of the automatic driving system in the case of multiple risks superimposed.

[0200] The above description is only the preferred embodiment of the present application and the explanation of the applied technical principles. It should be understood by those skilled in the art that the scope of the application involved in the present application is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with the technical features disclosed in the present application (but not limited to) having similar functions to form a technical solution.

Claims

1. A method for processing driving behavior data in emergency scenarios based on deep learning, characterized in that, The method includes the following steps: At least two data simulation modes are set, namely a single working condition simulation scenario and a comprehensive working condition simulation scenario. Both working condition simulation scenarios include: emergency braking condition, cut-in condition and cut-out condition. Different driving data generated by different drivers in single-condition simulation scenarios and comprehensive-condition simulation scenarios are collected. Based on the data correlation within each driving data, combined with the neural network anomaly detection method, abnormal samples are removed from the two sets of driving data to obtain the first target driving data and the second target driving data. Based on the first target driving data and the second target driving data, a training database is constructed; the training database is used to train the autonomous driving system to perform a neural network-based driving behavior prediction model in emergency scenarios; wherein, the first target driving data and the second target driving data respectively include: multiple key parameters of the target vehicle when it generates braking and steering behavior in the single working condition simulation scenario and multiple key parameters of the target vehicle when it generates braking and steering behavior in the comprehensive working condition simulation scenario. This includes collecting different driving data generated by different drivers under comprehensive driving condition simulation scenarios, including: In the cut-out-braking or cut-in-braking conditions, a working condition sequence is set, which is the risk triggering order of the cut-out condition, the cut-in condition, and the braking condition; Obtain the first and second risk occurrence times for different drivers under the cut-out-brake and cut-in-brake conditions; the first risk occurrence time corresponds to the cut-out-brake condition, and the second risk occurrence time corresponds to the cut-in-brake condition; Based on the moment when the first risk occurs and the moment when the action command of different drivers is triggered in the cut-out-brake condition, the moment when the braking efficiency reaches the maximum braking deceleration, the maximum achievable braking deceleration, the maximum steering wheel angle, and the steering wheel return time, the average decision response time of the first target, the average braking efficiency improvement time of the first target, the average maximum braking deceleration of the first target, the maximum steering wheel angle of the first target, and the steering wheel return time of the first target are calculated by combining a deep learning model. Based on the timing of the second risk occurrence and the timing of the action commands triggered by different drivers under the engagement-braking condition, the timing when the braking efficiency reaches the maximum braking deceleration, the maximum achievable braking deceleration, the maximum steering wheel angle, and the steering wheel return time, combined with a deep learning model, the average decision response time of the second target, the average braking efficiency improvement time of the second target, the average maximum braking deceleration of the second target, the maximum steering wheel angle of the second target, and the steering wheel return time of the second target are calculated respectively.

2. The method for processing emergency driving behavior data based on deep learning according to claim 1, characterized in that, Set at least two data simulation modes, including: Select multiple drivers with different driving styles and connect them to simulation data acquisition equipment; Select the test environment conditions, and input the preset road surface adhesion coefficient and the preset maximum braking deceleration into the data acquisition device; The simulation platform shall independently set up at least three operating conditions: emergency braking, cutting in, and cutting out, to constitute the single operating condition simulation scenario; and at least two operating conditions, cutting out-braking and cutting in-braking, shall be set up in the simulation platform to constitute the comprehensive operating condition simulation scenario.

3. The method for processing emergency driving behavior data based on deep learning according to claim 2, characterized in that, Collect different driving data generated by different drivers in a single simulated driving scenario, including: Under the aforementioned emergency braking condition, Based on the braking command trigger time and the target vehicle response time for each driver, the individual average decision response time for each driver is calculated, and the comprehensive average decision response time is calculated based on the individual average decision response time. Based on the time when the brake pedal is pressed and the time when the braking efficiency reaches the maximum braking deceleration for each driver, the average braking efficiency improvement time for each driver is calculated, and the comprehensive average braking efficiency improvement time is calculated based on the average braking efficiency improvement time for each driver. The combined average maximum braking deceleration is calculated based on the maximum braking deceleration achieved by each driver.

4. The method for processing emergency driving behavior data based on deep learning according to claim 3, characterized in that, Collecting different driving data generated by different drivers in a single simulated driving scenario also includes: Under the aforementioned cutting-in or cutting-out conditions, The risk occurrence point time corresponding to each driver is identified, and the risk occurrence point time is used to mark the trigger time of the cut-in or cut-out condition; Obtain the steering wheel angle data set within a preset effective time period to obtain the maximum steering wheel angle. The preset effective time period is set based on the start time corresponding to the risk occurrence point. By using actual data on the steering wheel angle data set that changes with time, a steering angle-time relationship model is established, and based on the steering angle-time relationship model and the maximum steering wheel angle, the steering wheel return-to-center time is obtained.

5. The method for processing emergency driving behavior data based on deep learning according to claim 4, characterized in that, Establish a steering angle-duration relationship model, including: The obtained steering wheel angle data set is subjected to polynomial fitting to obtain the fitted steering angle model expression and polynomial coefficient values. The polynomial coefficient values ​​are the coefficient values ​​of each term obtained by fitting the steering angle model expression using the least squares method. Based on the steering angle model expression, polynomial coefficient values, and a preset steering wheel return-to-center relationship, an expression for characterizing the steering wheel return-to-center time is obtained. Based on the expression for the return-to-center time and the return-to-center start time corresponding to the maximum steering wheel steering angle, a steering angle-time relationship model for outputting the steering wheel return-to-center duration is established.

6. The method for processing emergency driving behavior data based on deep learning according to claim 4, characterized in that, Based on the data correlation within each set of driving data, and combined with a neural network anomaly detection method, abnormal samples are removed from the two sets of driving data to obtain the first target driving data and the second target driving data, including: Under the emergency braking condition, AI anomaly detection is used to obtain the longitudinal acceleration of the target vehicle. If the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold, the data sample at the corresponding time point is removed. In addition, the system obtains the pedal travel and corresponding deceleration when the driver presses the brake pedal. If the two do not match, the system uses an AI model to identify and remove data samples within that time period.

7. The method for processing emergency driving behavior data based on deep learning according to claim 1, characterized in that, Based on the data correlation within each set of driving data, and combined with a neural network anomaly detection method, abnormal samples are removed from the two sets of driving data to obtain the first target driving data and the second target driving data, including: Based on the set operating condition timing sequence, AI anomaly detection is used to remove data samples that do not conform to the set operating condition timing sequence or have not completed the cut-out-braking and cut-in-braking operating conditions. In addition, the longitudinal acceleration of the target vehicle is obtained, and if the longitudinal acceleration is greater than or equal to a longitudinal acceleration threshold, the data samples at the time when the longitudinal acceleration is greater than or equal to the longitudinal acceleration threshold are removed. In addition, the system obtains the pedal travel and the corresponding deceleration when the driver presses the brake pedal. If the two do not match, the data samples within the time period in which the mismatch occurs will be removed.

Images