Vehicle emergency control method and system under abnormal behavior of driver

Abstract

The invention relates to the technical field of driving safety, in particular to a vehicle emergency control method and system under driver abnormal behaviors, and the method comprises the steps: firstly obtaining source domain monitoring data collected by a target vehicle, and then generating target domain monitoring data matched with the source domain monitoring data based on sample data, inputting the source domain monitoring data and the target domain monitoring data into a preset behavior judgment model to obtain driver behavior data output by the preset behavior judgment model, and finally performing emergency control on the target vehicle based on the driver behavior data. Through the cross-domain data generation and unified model training mechanism, the matched target domain data is generated based on the source domain monitoring data, and the unified preset behavior judgment model is used for behavior recognition, so that the dependence on sensors such as a camera and a radar is eliminated, and the behavior recognition accuracy is improved. Driver abnormity monitoring and emergency control can be achieved for old vehicle models without adding hardware, and the problems that a traditional scheme is high in hardware requirement and poor in compatibility are effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of driving safety technology, and in particular to a vehicle emergency control method and system under abnormal driver behavior. Background Art

[0002] With the continuous improvement of automobile intelligence, driver abnormal behavior monitoring and emergency control technology has become an important research direction to improve driving safety. Drivers may make operational errors due to distraction (such as using mobile phones), fatigue driving, sudden illness or health problems, which may lead to traffic accidents. Therefore, real-time monitoring of the driver's status and taking active control measures (such as automatic deceleration, lane keeping, emergency braking, etc.) when abnormalities occur are of vital importance to prevent accidents and protect the safety of life and property.

[0003] However, the existing driver abnormal behavior monitoring and emergency control systems usually need to integrate multiple types of hardware devices such as cameras, millimeter wave radars, inertial measurement units (IMUs), steering wheel torque sensors, etc. to collect relevant data. Although this multi-sensor fusion architecture can improve monitoring accuracy, it places high demands on vehicle hardware configuration, severely limiting the large-scale application of driver abnormal behavior monitoring technology, making a large number of old vehicles unable to enjoy the protection of advanced safety technologies.

[0004] Therefore, people need a driver abnormal behavior emergency control method and system that does not require large-scale hardware modification and has strong compatibility. Summary of the invention

[0005] Therefore, the present invention provides a vehicle emergency control method and system under abnormal driver behavior, so as to solve the problems of high hardware requirements and poor compatibility of safety systems in the prior art.

[0006] The present invention provides a vehicle emergency control method under abnormal driver behavior, comprising:

[0007] Obtain source domain monitoring data collected by the target vehicle;

[0008] Generate target domain monitoring data that matches the source domain monitoring data based on the sample data, wherein the target domain monitoring data is a type of data that cannot be collected by the target vehicle;

[0009] Inputting the source domain monitoring data and the target domain monitoring data into a preset behavior judgment model to obtain the driver behavior data output by the preset behavior judgment model;

[0010] Based on the driver behavior data, emergency control is performed on the target vehicle.

[0011] In a preferred implementation, the source domain monitoring data includes source domain time series data, and the target domain monitoring data includes target domain time series data; generating target domain monitoring data matching the source domain monitoring data based on the sample data includes:

[0012] The source domain time series data is input into the preset knowledge transfer model to obtain the target domain time series data output by the preset knowledge transfer model, wherein the preset knowledge transfer model is trained based on the sample data and is used to generate time series data of a different type that matches the input time series data according to the time series characteristics of the time series data.

[0013] In a preferred implementation, the preset knowledge transfer model includes an encoder, a domain adaptation layer and a decoder connected in sequence, wherein:

[0014] The encoder is used to extract the time series features of the source domain time series data and generate a first context vector;

[0015] The domain adaptation layer includes a plurality of sequentially connected feedforward neural network layers, and is used to adjust the feature distribution of the first context vector to obtain a second context vector;

[0016] The decoder is used to generate target domain time series data according to the second context vector.

[0017] In a preferred implementation, before the step of performing emergency control on the target vehicle based on the driver behavior data, the method further comprises:

[0018] The source domain monitoring data is input into a preset lightweight model to obtain the driver behavior data output by the preset lightweight model, wherein the preset lightweight model is obtained based on knowledge distillation training of a preset behavior judgment model.

[0019] In a preferred implementation, the sample data includes source domain sample data and sample behavior data corresponding to the source domain sample data; before the step of inputting the source domain monitoring data into a preset lightweight model to obtain the driver behavior data output by the preset lightweight model, the method further includes training the preset lightweight model, specifically including:

[0020] Obtaining an initial lightweight model;

[0021] Acquire historical source domain monitoring data as first training data;

[0022] Obtaining historical driver behavior data corresponding to the first training data as a teacher model output result corresponding to the first training data;

[0023] Calculate the similarity between the first training data and the source domain sample data, and select the sample behavior data corresponding to the source domain sample data with the highest similarity as the true label corresponding to the first training data;

[0024] Inputting the first training data into the initial lightweight model to obtain a student model output result corresponding to the first training data;

[0025] Establishing a first loss function according to the true label corresponding to the first training data, the output result of the teacher model and the output result of the student model;

[0026] The initial lightweight model is optimized according to the first loss function to obtain a trained preset lightweight model.

[0027] In a preferred implementation, the sample data further includes target domain sample data corresponding to the source domain sample data; and training the preset lightweight model further includes:

[0028] Selecting source domain sample data with a similarity higher than a preset threshold as second training data;

[0029] Using the sample behavior data corresponding to the second training data as the true label corresponding to the second training data;

[0030] Inputting the second training data and the target domain sample data corresponding to the second training data into the preset behavior judgment model to obtain the teacher model output result corresponding to the second training data;

[0031] Inputting the second training data into the initial lightweight model to obtain the student model output result corresponding to the second training data;

[0032] Establishing a second loss function according to the true label corresponding to the second training data, the output result of the teacher model and the output result of the student model;

[0033] The initial lightweight model is optimized according to the second loss function to obtain a trained preset lightweight model.

[0034] In a preferred implementation, training a preset lightweight model further includes:

[0035] Selecting source domain sample data representing abnormal driver behavior as the third training data;

[0036] Using the sample behavior data corresponding to the third training data as the true label corresponding to the third training data;

[0037] Inputting the third training data and the target domain sample data corresponding to the third training data into the preset behavior judgment model to obtain the teacher model output result corresponding to the third training data;

[0038] Inputting the third training data into the initial lightweight model to obtain a student model output result corresponding to the third training data;

[0039] Establishing a third loss function according to the true label corresponding to the third training data, the output result of the teacher model and the output result of the student model;

[0040] The initial lightweight model is optimized according to the third loss function to obtain a trained preset lightweight model.

[0041] In a preferred implementation, training a preset lightweight model further includes:

[0042] Perform weighted summation on the first loss function, the second loss function and the third loss function to obtain a total loss function;

[0043] The initial lightweight model is optimized according to the total loss function to obtain a trained preset lightweight model.

[0044] In a preferred implementation, emergency control of a target vehicle is performed based on driver behavior data, including:

[0045] Determine the behavior risk level based on driver behavior data;

[0046] Emergency control is carried out on the target vehicle based on the behavioral risk level.

[0047] The present invention also provides a vehicle emergency control system under abnormal driver behavior, comprising:

[0048] A data acquisition module, used to obtain source domain monitoring data collected by the target vehicle;

[0049] A data filling module is used to generate target domain monitoring data that matches the source domain monitoring data based on the sample data, wherein the target domain monitoring data is a type of data that cannot be collected by the target vehicle;

[0050] A behavior judgment module, used to input source domain monitoring data and target domain monitoring data into a preset behavior judgment model to obtain driver behavior data output by the preset behavior judgment model;

[0051] The emergency control module is used to perform emergency control on the target vehicle based on the driver's behavior data.

[0052] The beneficial effects of adopting the above embodiment are:

[0053] The present invention provides a vehicle emergency control method and system under abnormal driver behavior, which first obtains source domain monitoring data collected by a target vehicle, then generates target domain monitoring data matching the source domain monitoring data based on sample data, then inputs the source domain monitoring data and the target domain monitoring data into a preset behavior judgment model, obtains driver behavior data output by the preset behavior judgment model, and finally performs emergency control on the target vehicle based on the driver behavior data. The present invention generates matching target domain data based on source domain monitoring data through cross-domain data generation and unified model training mechanism, and uses a unified preset behavior judgment model for behavior recognition, thereby getting rid of dependence on sensors such as cameras and radars, so that old models can achieve driver abnormality monitoring and emergency control without adding hardware, effectively solving the problems of high hardware requirements and poor compatibility of traditional solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A method flow chart of a vehicle emergency control method under abnormal driver behavior provided by the present invention;

[0055] Figure 2 for Figure 1 Specific step diagram of step S106;

[0056] Figure 3 for Figure 1 Another specific step diagram of step S106;

[0057] Figure 4 for Figure 1 Another specific step diagram of step S106;

[0058] Figure 5 This is a system structure diagram of the vehicle emergency control system under abnormal driver behavior provided by the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Combination Figure 1 As shown, a specific embodiment of the present invention discloses a vehicle emergency control method under abnormal driver behavior, comprising:

[0061] S101, obtaining source domain monitoring data collected by the target vehicle;

[0062] S102, generating target domain monitoring data matching the source domain monitoring data based on the sample data, wherein the target domain monitoring data is a type of data that cannot be collected by the target vehicle;

[0063] S103, inputting the source domain monitoring data and the target domain monitoring data into a preset behavior judgment model to obtain driver behavior data output by the preset behavior judgment model;

[0064] S104: Perform emergency control on the target vehicle based on the driver behavior data.

[0065] In the above process, source domain monitoring data refers to the type of data that can be collected by the target vehicle, target domain monitoring data refers to the type of data that cannot be collected by the target vehicle, and sample data is accurate data prepared in advance. In the invention, when the source domain monitoring data is known, the law represented by the sample data can be used to complete the target domain monitoring data. The preset behavior judgment model is a model for uniformly judging the driver's behavior. Through the above scheme, when facing the safety control of some old models, the model can be allowed to collect only the data that it can collect (ie, source domain target data), and then automatically complete the target domain feature data for these data to generate complete and reasonable data, so as to use the data of the preset behavior judgment model for behavior monitoring. It can be understood that the above steps S102 and S103 can be executed in the cloud.

[0066] This embodiment generates matching target domain data based on source domain monitoring data through cross-domain data generation and a unified model training mechanism, and uses a unified preset behavior judgment model for behavior recognition, thereby getting rid of dependence on sensors such as cameras and radars. This allows old car models to achieve driver abnormality monitoring and emergency control without the need for new hardware, effectively solving the problems of high hardware requirements and poor compatibility of traditional solutions.

[0067] Specifically, the specific method of monitoring the target domain data can be implemented by any existing technology. For example, the largest number of values ​​in the sample data can be counted as the completed target domain monitoring data, or the source domain monitoring data and the data part of the sample data that is of the same type as the source domain monitoring data can be matched, and the data entry that is most similar to the source domain monitoring data can be selected, and the data part of the data entry that is different from the source domain monitoring data can be used as the target domain monitoring data.

[0068] The present invention also provides a more preferred solution for completing the target domain monitoring data. In a new embodiment, the source domain monitoring data includes source domain time series data, and the target domain monitoring data includes target domain time series data. On this basis, the above step S102, generating target domain monitoring data matching the source domain monitoring data based on the sample data, specifically includes:

[0069] The source domain time series data is input into the preset knowledge transfer model to obtain the target domain time series data output by the preset knowledge transfer model, wherein the preset knowledge transfer model is trained based on the sample data and is used to generate time series data of a different type that matches the input time series data according to the time series characteristics of the time series data.

[0070] In the above process, the preset knowledge transfer model uses the labeled sample data (complete data, including source domain and target domain) from other models or scenes to train the model, and transfers the learned time series features to the target domain (old models with missing data). The preset knowledge transfer model captures the long-term dependency between the source domain and the target domain time series data (such as the driver's stepping on the accelerator must be accompanied by a forward line of sight scan, and the lane deviation rate is determined by the steering angle and vehicle speed). The generated data has coherence in the time dimension, avoiding the "behavioral fault" caused by traditional random completion. The model enforces the physical consistency of the source domain and target domain data through comparative learning (such as the covariance relationship between the acceleration change rate and the steering angle change), so that the generated target domain data conforms to the laws of vehicle dynamics without relying on a large amount of real scene labeled data.

[0071] It is understandable that abnormal driver behavior (such as sudden deceleration caused by sudden illness) often has time evolution characteristics. Time series data can fully capture the start, development and end of the action, while discrete data (such as single-frame heart rate values) are prone to lose key dynamic information. By analyzing the temporal correlation between source domain time series data (such as steering wheel angle sequence) and target domain time series data (such as eye movement trajectory), an implicit logical chain of driver intention and vehicle control can be established (such as "frequently looking at the phone - steering wheel deviating from the lane - emergency braking"), improving the causal reasoning ability of behavior recognition.

[0072] Compared with other solutions, this embodiment reuses limited data through generative models to improve data utilization efficiency, and the time series knowledge migration supports cross-domain generalization, while ensuring the physical rationality of the generated data and maximizing the value of existing data. It not only solves the behavior monitoring blind spot problem caused by the lack of sensors in traditional solutions, but also avoids the logical paradox risks that may be caused by pure data generation methods, providing a reusable technical paradigm for the low-cost and high-precision implementation of driver status monitoring systems.

[0073] It is conceivable that the above-mentioned preset knowledge transfer model can be implemented by using any existing machine learning model, and the present invention further provides a feasible embodiment. In a new embodiment, the preset knowledge transfer model includes an encoder, a domain adaptation layer and a decoder connected in sequence, wherein:

[0074] The encoder is used to extract the time series features of the source domain time series data and generate a first context vector;

[0075] The domain adaptation layer includes a plurality of sequentially connected feedforward neural network layers, and is used to adjust the feature distribution of the first context vector to obtain a second context vector;

[0076] The decoder is used to generate target domain time series data according to the second context vector.

[0077] In the above content, the implementation of the encoder is preferably based on the Transformer architecture (such as Temporal FusionTransformer) or LSTM network, and the long-term dependencies of the source domain time series data (such as the acceleration change trend within 10 consecutive seconds) are captured through the multi-head attention mechanism. The input source domain time series data (such as the throttle opening sequence) is mapped to the first context vector, and the time series dynamic features (such as action acceleration) and implicit behavior patterns (such as driver predictive operations) are retained during the encoding process. The domain adaptation layer adopts a multi-layer feedforward neural network stacking structure to enforce the consistency of the feature space mapping between the target domain and the source domain, such as eliminating cross-domain data distribution differences (such as the domain offset of driver behavior data and vehicle dynamics parameters), and ensuring that the generated target domain time series data is strongly associated with the source domain in terms of physical laws (such as the temporal coupling of sudden acceleration actions and line of sight shifts). The decoder module can also be implemented based on the inverse process of Transformer or the time series prediction capability of LSTM, reconstructing the target domain time series data (such as the driver's pupil diameter change curve) from the adjusted second context vector, and generating a time series sequence that conforms to the target domain data distribution (such as the biological signal sampling rate) while maintaining time series continuity, supporting multimodal output (such as generating eye movement trajectories and gesture actions at the same time).

[0078] It can be imagined that the above embodiment is most effective when it comes to compensating time series data, but in practice the target domain data often also includes some other non-time series data. In this case, the data can be completed by mixing the multiple completion methods mentioned above.

[0079] Furthermore, if the above method is deployed and executed locally, the vehicle system needs to have sufficiently high hardware computing power. If it is executed in the cloud, the vehicle communication environment needs to be good. In practice, the above two conditions may not be met at the same time when the vehicle is driving. Therefore, in order to solve the contradiction between insufficient local deployment computing power and reliance on cloud communication in the traditional solution, the invention also provides a preferred embodiment, which further establishes a lightweight model based on the preset behavior judgment model to facilitate local deployment, so as to reduce the hardware requirements and environmental requirements for safety detection.

[0080] Specifically, in a preferred embodiment, before step S104, emergency control of the target vehicle is performed based on the driver behavior data, step S105 is also included, inputting the source domain monitoring data into a preset lightweight model to obtain the driver behavior data output by the preset lightweight model, wherein the preset lightweight model is obtained based on knowledge distillation training of a preset behavior judgment model.

[0081] The above method uses knowledge distillation technology to train a lightweight model (i.e., student model) based on a preset behavior judgment model (teacher model), compressing the model size and reducing inference latency without losing the performance of the original model, thus meeting the real-time operation requirements of automotive-grade chips (such as MCUs). In addition, compared to the preset behavior judgment model, the lightweight model only needs to input source domain monitoring data to realize behavior judgment, which is more in line with the vehicle's own situation and has a faster inference speed.

[0082] Knowledge distillation is a technology that achieves model compression through knowledge transfer between models. Its essence is to transfer the knowledge of a complex teacher model (preset behavior judgment model) to the student model (lightweight model) in an implicit form. Unlike traditional fine-tuning, knowledge distillation does not rely on a large amount of labeled data, but guides the learning of the student model by mining the feature representation of the intermediate layer of the teacher model.

[0083] The general process of knowledge distillation is: input the same data into the teacher model and the student model at the same time, and according to the output of the teacher model, the output of the student model and the true label corresponding to the data itself, a loss function can be established to complete the training. Specifically, after obtaining the output of the teacher model, it is necessary to perform softmax normalization on the output of the teacher model based on a preset temperature to obtain a soft label. Then, based on the same preset temperature, the output of the student model is softmax normalized to obtain a soft prediction. A soft loss function is calculated based on the soft label and the soft prediction. Then, the output of the student model itself is used as a hard prediction, and the true label is used as a hard label, and a hard loss function is calculated based on the hard label and the hard prediction. Then, by fusing the soft loss function and the hard loss function based on a specific ratio, a final loss function can be obtained to train the student model and complete the knowledge distillation from the teacher model to the student model.

[0084] In this embodiment, if you want to achieve knowledge distillation of the lightweight model through the predictive behavior judgment model, you need to face two problems. First, in the prior art, the teacher model and the student model will input the same data, while in this embodiment, the preset behavior judgment model and the lightweight model input data are different. Another problem is that when building a lightweight model, it is necessary to use the user's source domain monitoring data to train it, so that the lightweight model can be meaningful and realize personalized local deployment. However, as can be seen from the previous content, this system can only collect real source domain monitoring data, and its corresponding target domain monitoring data and driver status data are all based on algorithm inference prediction. Therefore, when performing knowledge distillation, it is impossible to know the real label corresponding to the source domain monitoring data.

[0085] Therefore, the present invention further provides a feasible knowledge distillation method that can solve the above problems. Specifically, in a preferred embodiment, the sample data includes source domain sample data and sample behavior data corresponding to the source domain sample data. The source domain sample data is sample data of the same type as the source domain monitoring data. Then the sample behavior data is the real driver behavior data corresponding to each sample data. In step S105, the source domain monitoring data is input into the preset lightweight model to obtain the driver behavior data output by the preset lightweight model. The vehicle emergency control method under abnormal driver behavior also includes step S106, training the preset lightweight model, combining Figure 2 As shown, this step specifically includes:

[0086] S201, obtaining an initial lightweight model;

[0087] S202, obtaining historical source domain monitoring data as first training data;

[0088] S203, obtaining historical driver behavior data corresponding to the first training data as a teacher model output result corresponding to the first training data;

[0089] S204, calculating the similarity between the first training data and the source domain sample data, and selecting the sample behavior data corresponding to the source domain sample data with the highest similarity as the true label corresponding to the first training data;

[0090] S205, inputting the first training data into the initial lightweight model to obtain a student model output result corresponding to the first training data;

[0091] S206, establishing a first loss function according to the true label corresponding to the first training data, the output result of the teacher model and the output result of the student model;

[0092] S207. Optimize the initial lightweight model according to the first loss function to obtain a trained preset lightweight model.

[0093] In the above process, the historical source domain monitoring data and historical driver behavior data are all existing data collected or predicted in the past. The lightweight model is trained using the historical source domain monitoring data to achieve personalization.

[0094] After the first training data is input into the initial lightweight model, the corresponding student model output result is naturally obtained. The historical driver behavior data corresponding to the first training data is generated by the predictive behavior model itself, so it can be directly used as the output result of the teacher model corresponding to the first training data, solving the problem of inconsistent input between the teacher model and the student model.

[0095] In addition, in most cases, the driver's abnormal behavior data is similar, and abnormal behavior accounts for a minority of all behavior types. For most normal behaviors, as long as the lightweight model determines that it is still within the scope of normal behavior, even if the specific behavior type is evaluated incorrectly, it will not affect the final control effect. Therefore, this embodiment also matches the source domain sample data that is most similar to the first training data, and approximates its sample behavior data as the true label, which solves the problem of missing true labels corresponding to the first training data and realizes knowledge distillation.

[0096] Furthermore, the historical behavior data itself is inferred again based on the inferred target domain monitoring data, which means that the output results of the teacher model corresponding to the first training data may not be accurate enough. Therefore, it is better to use more realistic data for further training to improve authenticity.

[0097] Combination Figure 3 As shown, in a preferred embodiment, the sample data also includes target domain sample data corresponding to the source domain sample data (the target domain sample data is sample data of the same type as the target domain monitoring data), and the above step S106, training the preset lightweight model, also includes:

[0098] S301, selecting source domain sample data with a similarity higher than a preset threshold as second training data;

[0099] S302, using the sample behavior data corresponding to the second training data as the true label corresponding to the second training data;

[0100] S303, inputting the second training data and the target domain sample data corresponding to the second training data into a preset behavior judgment model to obtain the teacher model output result corresponding to the second training data;

[0101] S304, inputting the second training data into the initial lightweight model to obtain a student model output result corresponding to the second training data;

[0102] S305, establishing a second loss function according to the true label corresponding to the second training data, the output result of the teacher model and the output result of the student model;

[0103] S306. Optimize the initial lightweight model according to the second loss function to obtain a trained preset lightweight model.

[0104] Similarly, based on the similarity of different drivers' behaviors and the tolerance of the present invention to deviations in normal behavior judgments, the present embodiment further uses sample data as second training data to provide more realistic training data for the preset lightweight model to improve model accuracy.

[0105] The sample behavior data corresponding to the second training data is naturally its corresponding real label. After the second training data is input into the initial lightweight model, the corresponding student model output result is naturally obtained. In addition, the second training data is essentially the source domain sample data, which itself has the corresponding real target domain sample data. Therefore, it can be input into the preset behavior judgment model to obtain the teacher model output result, thereby realizing knowledge distillation.

[0106] Furthermore, based on the same reason, it is conceivable that in practice, driver behavior data that characterize abnormal behavior is ultimately a minority of data. The historical source domain monitoring data mentioned above as the first training data (because in most cases users have normal driving behavior, the collected historical source domain monitoring data may all be data that characterizes normal conditions) and the source domain sample data (sample data similar to normal condition data) as the second training data may not include data that characterizes abnormal behavior. Therefore, in order to improve the generalization of the lightweight model, the present invention also provides an embodiment. In combination with Figure 4 As shown, in a new embodiment, the above step S106, training the preset lightweight model, specifically also includes:

[0107] S401, selecting source domain sample data representing abnormal driver behavior as third training data;

[0108] S402, using the sample behavior data corresponding to the third training data as the true label corresponding to the third training data;

[0109] S403, inputting the third training data and the target domain sample data corresponding to the third training data into a preset behavior judgment model to obtain a teacher model output result corresponding to the third training data;

[0110] S404, inputting the third training data into the initial lightweight model to obtain a student model output result corresponding to the third training data;

[0111] S405, establishing a third loss function according to the true label corresponding to the third training data, the output result of the teacher model and the output result of the student model;

[0112] S406. Optimize the initial lightweight model according to the third loss function to obtain a trained preset lightweight model.

[0113] In this embodiment, source domain sample data characterizing abnormal driver behavior is selected as the third training data, providing the initial lightweight model with more diverse, realistic and reasonable training samples, so that the lightweight model can learn the judgment logic of abnormal behavior, thereby improving its generalization.

[0114] Among them, the sample behavior data corresponding to the third training data is naturally its corresponding true label. After the third training data is input into the initial lightweight model, the corresponding student model output result is naturally obtained. In addition, the third training data is essentially the source domain sample data, which itself has the corresponding real target domain sample data. Therefore, it can be input into the preset behavior judgment model to obtain the teacher model output result, thereby realizing knowledge distillation.

[0115] It can be understood that the three optimization processes described in the above three embodiments can be performed separately, that is, the initial lightweight model can be back-propagated and optimized separately using the first loss function, the second loss function and the third loss function, or the three loss functions can be fused and back-propagated and optimized simultaneously to improve training efficiency.

[0116] Specifically, in a preferred embodiment, the above step S106, training the preset lightweight model, further includes:

[0117] Perform weighted summation on the first loss function, the second loss function and the third loss function to obtain a total loss function;

[0118] The initial lightweight model is optimized according to the total loss function to obtain a trained preset lightweight model.

[0119] Among them, the weight ratio of the first loss function, the second loss function and the third loss function can be reasonably adjusted according to the proportion of the target domain monitoring data.

[0120] Furthermore, in a preferred embodiment, the above step S104, performing emergency control on the target vehicle based on the driver behavior data, specifically includes:

[0121] Determine the behavior risk level based on driver behavior data;

[0122] Emergency control is carried out on the target vehicle based on the behavioral risk level.

[0123] This embodiment achieves efficient coordination of abnormal driver behavior monitoring and emergency control by building a dynamic hierarchical response mechanism. For example, in the risk level determination link, based on the driver behavior data (such as distraction index and fatigue level score) output by the preset behavior judgment model, combined with the dynamic threshold algorithm, the risk can be divided into three levels: low (0-30 points), medium (31-70 points), and high (71-100 points), and the first-level warning (such as voice prompt), second-level intervention (such as power reduction) and third-level emergency control (such as emergency braking) are triggered respectively to ensure that the intervention intensity is accurately matched with the risk level.

[0124] Combination Figure 5 As shown, the present invention also provides a vehicle emergency control system under abnormal driver behavior, comprising:

[0125] The data acquisition module 510 is used to acquire source domain monitoring data collected by the target vehicle;

[0126] A data filling module 520 is used to generate target domain monitoring data matching the source domain monitoring data based on the sample data, wherein the target domain monitoring data is a type of data that cannot be collected by the target vehicle;

[0127] The behavior judgment module 530 is used to input the source domain monitoring data and the target domain monitoring data into a preset behavior judgment model to obtain the driver behavior data output by the preset behavior judgment model;

[0128] The emergency control module 540 is used to perform emergency control on the target vehicle based on the driver behavior data.

[0129] Furthermore, in one embodiment, the vehicle emergency control system under the above-mentioned abnormal driver behavior further includes:

[0130] Data learning module, used to train preset lightweight models;

[0131] The quick judgment module is used to input the source domain monitoring data into a preset lightweight model to obtain the driver behavior data output by the preset lightweight model, wherein the preset lightweight model is obtained based on knowledge distillation training of a preset behavior judgment model.

[0132] It should be noted here that the corresponding system provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be found in the corresponding contents in the above method embodiments, which will not be repeated here.

[0133] The present invention provides a vehicle emergency control method and system under abnormal driver behavior, which first obtains source domain monitoring data collected by a target vehicle, then generates target domain monitoring data matching the source domain monitoring data based on sample data, then inputs the source domain monitoring data and the target domain monitoring data into a preset behavior judgment model, obtains driver behavior data output by the preset behavior judgment model, and finally performs emergency control on the target vehicle based on the driver behavior data. The present invention generates matching target domain data based on source domain monitoring data through cross-domain data generation and unified model training mechanism, and uses a unified preset behavior judgment model for behavior recognition, thereby getting rid of dependence on sensors such as cameras and radars, so that old models can achieve driver abnormality monitoring and emergency control without adding hardware, effectively solving the problems of high hardware requirements and poor compatibility of traditional solutions.

[0134] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0135] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A vehicle emergency control method under abnormal driver behavior, characterized in that: include: Obtain source domain monitoring data collected by the target vehicle; Generate target domain monitoring data that matches the source domain monitoring data based on the sample data, wherein the target domain monitoring data is a type of data that cannot be collected by the target vehicle; Inputting the source domain monitoring data and the target domain monitoring data into a preset behavior judgment model to obtain the driver behavior data output by the preset behavior judgment model; Based on the driver behavior data, emergency control is performed on the target vehicle.

2. The vehicle emergency control method under abnormal driver behavior according to claim 1, characterized in that: The source domain monitoring data includes source domain time series data, and the target domain monitoring data includes target domain time series data. The target domain monitoring data matching the source domain monitoring data is generated based on the sample data, including: The source domain time series data is input into the preset knowledge transfer model to obtain the target domain time series data output by the preset knowledge transfer model, wherein the preset knowledge transfer model is trained based on the sample data and is used to generate time series data of a different type that matches the input time series data according to the time series characteristics of the time series data.

3. The vehicle emergency control method under abnormal driver behavior according to claim 2, characterized in that: The preset knowledge transfer model includes an encoder, a domain adaptation layer, and a decoder connected in sequence, where: The encoder is used to extract the time series features of the source domain time series data and generate a first context vector; The domain adaptation layer includes a plurality of sequentially connected feedforward neural network layers, and is used to adjust the feature distribution of the first context vector to obtain a second context vector; The decoder is used to generate target domain time series data according to the second context vector.

4. The vehicle emergency control method under abnormal driver behavior according to claim 1, characterized in that: Before the step of performing emergency control on the target vehicle based on the driver behavior data, the method further includes: The source domain monitoring data is input into a preset lightweight model to obtain the driver behavior data output by the preset lightweight model, wherein the preset lightweight model is obtained based on knowledge distillation training of a preset behavior judgment model.

5. The vehicle emergency control method under abnormal driver behavior according to claim 4, characterized in that: The sample data includes source domain sample data and sample behavior data corresponding to the source domain sample data; before the step of inputting the source domain monitoring data into a preset lightweight model to obtain the driver behavior data output by the preset lightweight model, the method further includes training the preset lightweight model, specifically including: Obtaining an initial lightweight model; Acquire historical source domain monitoring data as first training data; Acquire historical driver behavior data corresponding to the first training data as a teacher model output result corresponding to the first training data; Calculate the similarity between the first training data and the source domain sample data, and select the sample behavior data corresponding to the source domain sample data with the highest similarity as the true label corresponding to the first training data; Inputting the first training data into the initial lightweight model to obtain a student model output result corresponding to the first training data; Establishing a first loss function according to the true label corresponding to the first training data, the output result of the teacher model and the output result of the student model; The initial lightweight model is optimized according to the first loss function to obtain a trained preset lightweight model.

6. The vehicle emergency control method under abnormal driver behavior according to claim 5, characterized in that: The sample data also includes target domain sample data corresponding to the source domain sample data; Training preset lightweight models also includes: Selecting source domain sample data with a similarity higher than a preset threshold as second training data; Using the sample behavior data corresponding to the second training data as the true label corresponding to the second training data; Inputting the second training data and the target domain sample data corresponding to the second training data into the preset behavior judgment model to obtain the teacher model output result corresponding to the second training data; Inputting the second training data into the initial lightweight model to obtain the student model output result corresponding to the second training data; Establishing a second loss function according to the true label corresponding to the second training data, the output result of the teacher model and the output result of the student model; The initial lightweight model is optimized according to the second loss function to obtain a trained preset lightweight model.

7. The vehicle emergency control method under abnormal driver behavior according to claim 6, characterized in that: Training preset lightweight models also includes: Selecting source domain sample data representing abnormal driver behavior as the third training data; Using the sample behavior data corresponding to the third training data as the true label corresponding to the third training data; Inputting the third training data and the target domain sample data corresponding to the third training data into the preset behavior judgment model to obtain the teacher model output result corresponding to the third training data; Inputting the third training data into the initial lightweight model to obtain a student model output result corresponding to the third training data; Establishing a third loss function according to the true label corresponding to the third training data, the output result of the teacher model and the output result of the student model; The initial lightweight model is optimized according to the third loss function to obtain a trained preset lightweight model.

8. The vehicle emergency control method under abnormal driver behavior according to claim 7, characterized in that: Training preset lightweight models also includes: Perform weighted summation on the first loss function, the second loss function and the third loss function to obtain a total loss function; The initial lightweight model is optimized according to the total loss function to obtain a trained preset lightweight model.

9. The vehicle emergency control method under abnormal driver behavior according to claim 1, characterized in that: Based on the driver behavior data, emergency control of the target vehicle is performed, including: Determine the behavior risk level based on driver behavior data; Emergency control is carried out on the target vehicle based on the behavioral risk level.

10. A vehicle emergency control system under abnormal driver behavior, characterized in that: include: A data acquisition module, used to obtain source domain monitoring data collected by the target vehicle; A data filling module is used to generate target domain monitoring data that matches the source domain monitoring data based on the sample data, wherein the target domain monitoring data is a type of data that cannot be collected by the target vehicle; A behavior judgment module, used to input source domain monitoring data and target domain monitoring data into a preset behavior judgment model to obtain driver behavior data output by the preset behavior judgment model; The emergency control module is used to perform emergency control on the target vehicle based on the driver's behavior data.

Images