Vehicle collision detection method, apparatus, and vehicle

Abstract

The application discloses a vehicle collision detection method and device and a vehicle, and applies to the technical field of vehicles.The method comprises the following steps: acquiring a first time sequence and a second time sequence of a target vehicle, the first time sequence and the second time sequence being obtained by sampling through different time sequence windows, each time sequence window comprising a to-be-detected moment, and each time sequence comprising attribute data in the sampled time sequence window, the attribute data comprising state data, driving behavior data, collision attribute data and radar data; performing feature change detection on the first time sequence to obtain attribute change features of the target vehicle; performing feature extraction on the second time sequence to obtain collision-related features of the target vehicle; inputting the attribute change features and the collision-related features into a pre-trained collision detection model to obtain a collision detection result of the target vehicle, the collision detection result indicating whether the target vehicle collides at the to-be-detected moment. The application can effectively improve the precision of vehicle collision detection.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, and in particular to a vehicle collision detection method, device, and vehicle. Background Technology

[0002] Vehicle collisions are a common type of accident in road traffic and a crucial component of vehicle driver assistance technology. Related technologies collect the vehicle's state signals while in motion and input them into a pre-trained collision detection model, which then identifies whether a collision has occurred, thus achieving vehicle collision detection. However, these technologies only consider the vehicle's state signals while in motion, resulting in an overly simplistic and one-sided collision detection logic, leading to low accuracy. Summary of the Invention

[0003] This application provides a vehicle collision detection method, apparatus, and vehicle to improve the accuracy of collision detection.

[0004] On the one hand, embodiments of this application provide a vehicle collision detection method, including the following steps:

[0005] A first time series and a second time series of the target vehicle are obtained; wherein the first time series and the second time series are obtained by sampling through different time series windows, each time series window includes the time to be measured, and each time series includes attribute data within the sampled time series window, the attribute data including state data, driving behavior data, collision attribute data and radar data;

[0006] Feature change detection is performed on the first time series to obtain the attribute change features of the target vehicle;

[0007] Feature extraction is performed on the second time series to obtain the collision-related features of the target vehicle;

[0008] The attribute change features and the collision-related features are input into a pre-trained collision detection model to obtain the collision detection result of the target vehicle; wherein, the collision detection result indicates whether the target vehicle has collided at the time to be tested.

[0009] On the other hand, embodiments of this application provide a vehicle collision detection device, including:

[0010] The acquisition module is used to acquire a first time series and a second time series of the target vehicle; wherein the first time series and the second time series are obtained by sampling through different time series windows, each time series window includes the time to be measured, and each time series includes attribute data within the sampled time series window, the attribute data including state data, driving behavior data, collision attribute data and radar data;

[0011] The first processing module is used to perform feature change detection on the first time series to obtain the attribute change features of the target vehicle;

[0012] The second processing module is used to extract features from the second time series to obtain the collision-related features of the target vehicle.

[0013] The detection module is used to input the attribute change features and the collision-related features into a pre-trained collision detection model to obtain the collision detection result of the target vehicle; wherein, the collision detection result indicates whether the target vehicle has collided at the time to be tested.

[0014] In another aspect, embodiments of this application provide a vehicle, including:

[0015] At least one processor;

[0016] At least one memory for storing at least one program;

[0017] When the at least one program is executed by the at least one processor, the at least one processor implements the above-described vehicle collision detection method.

[0018] According to an embodiment of this application, a vehicle collision detection method, apparatus, and vehicle are provided. The method acquires a first time series and a second time series of a target vehicle. The first and second time series are sampled through different time windows, each time window including the time to be measured. Each time series includes attribute data within the sampled time window, including state data, driving behavior data, collision attribute data, and radar data. Feature change detection is performed on the first time series to obtain attribute change features of the target vehicle. Feature extraction is performed on the second time series to obtain collision-related features of the target vehicle. The attribute change features and collision-related features are input into a pre-trained collision detection model to obtain a collision detection result for the target vehicle, indicating whether a collision occurred at the time to be measured. The technical solution according to this embodiment can effectively improve the accuracy of vehicle collision detection.

[0019] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0020] Figure 1 This is a flowchart of a vehicle collision detection method provided in this application;

[0021] Figure 2 This is a structural diagram of a vehicle collision detection device provided in this application;

[0022] Figure 3 This is an example drawing of a vehicle provided in this application;

[0023] Figure 4 This is a flowchart illustrating the implementation process of a vehicle collision detection method provided in this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0026] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0028] It should be noted that in all specific embodiments of this application, when processing is required based on data such as insurance records and maintenance records, permission or consent from the subject is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require the acquisition of data such as insurance records and maintenance records, separate permission or consent from the subject is obtained through pop-ups or redirection to a confirmation page. Only after obtaining the subject's separate permission or consent is the necessary data, such as insurance records and maintenance records, required for the proper functioning of embodiments of this application is acquired.

[0029] Vehicle collisions are a common type of accident in road traffic and a crucial component of vehicle driver assistance technology. Related technologies collect the vehicle's state signals while in motion and input them into a pre-trained collision detection model, which then identifies whether a collision has occurred, thus achieving vehicle collision detection. However, these technologies only consider the vehicle's state signals while in motion, making their collision detection logic overly simplistic and one-sided, leading to low accuracy. Furthermore, the vehicle data samples and their labels used for model training in these technologies often rely on manually set filtering rules. This results in phenomena such as misclassifying many samples with heights very close to actual collisions as collisions, further contributing to low collision detection accuracy.

[0030] Therefore, embodiments of this application provide a vehicle collision detection method, apparatus, and vehicle to improve the accuracy of vehicle collision detection.

[0031] First, the implementation steps of a vehicle collision detection method provided in this application embodiment will be described in detail below with reference to the accompanying drawings.

[0032] This application provides a vehicle collision detection method that can be applied to a terminal, a server, or software running on either a terminal or a server. The terminal can be a tablet, laptop, desktop computer, etc., but is not limited to these. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Furthermore, the server can be a node server in a blockchain network, but is not limited to these. Blockchain is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0033] Reference Figure 1 The vehicle collision detection method may include the following steps S101-S104.

[0034] S101, Obtain the first and second time series of the target vehicle.

[0035] It should be noted that both the first and second time series can be obtained through sampling using different time series windows. Each time series window can include the time point to be measured, and each time series can include attribute data within the sampled time series window. Attribute data can include state data, driving behavior data, collision attribute data, and radar data, etc. Specifically, state data refers to the state signals of the target vehicle while it is in motion; driving behavior data refers to data associated with the driver's behavior of the target vehicle; collision attribute data refers to the relevant signals of the target vehicle when a collision occurs; and radar data refers to obstacle signals detected by the target vehicle's onboard radar.

[0036] It is understood that the target vehicle refers to a vehicle for which a vehicle collision detection method is applicable in the embodiments of this application.

[0037] Specifically, this step sets two time series windows of different lengths based on the time to be measured, and obtains the corresponding attribute data through these two time series windows, thereby obtaining the first time series and the second time series of the target vehicle, so as to extract feature information that is highly correlated with the vehicle collision event from these two time series.

[0038] It is worth noting that both the time-series window used to collect the first time series and the time-series window used to collect the second time series must meet certain lengths, and the time-series window used to collect the first time series must be located within the time-series window used to collect the second time series. The purpose of this is twofold: firstly, by setting two time-series windows of different lengths for data sampling, the risk of data distortion due to an overly singular time scale can be effectively reduced, ensuring data quality; secondly, the first time series is used in subsequent steps to capture instantaneous changes in the target vehicle at the time of the test, while the second time series is used in subsequent steps to explore the actual condition of the target vehicle before and after the time of the test. By limiting the relationship between the time-series windows used to collect the first and second time series, the first time series is better at capturing short-term feature changes, and the second time series is better at capturing long-term feature trends, thus providing effective feature references for vehicle collision detection.

[0039] S102, perform feature change detection on the first time series to obtain the attribute change features of the target vehicle.

[0040] Specifically, there are often significant differences between the attribute data of a vehicle during normal driving and the attribute data of a vehicle at or after a collision. This difference can serve as a direct indication of a vehicle collision event; for example, the greater the difference, the greater the likelihood of a collision. Therefore, this step performs feature change detection on the first time series to obtain the attribute change features of the target vehicle. These attribute change features indicate the sudden changes and disturbances that occur to the target vehicle at the time of the test. This allows for precise localization of the difference between the target vehicle's condition at the time of the test and its condition during normal driving, thereby fully capturing the sudden changes and disturbances that occur to the target vehicle at the time of the test, i.e., instantaneous abrupt changes, such as sharp steering wheel turns or sudden speed drops. This enables a precise characterization of the vehicle's instantaneous dynamics, providing an effective reference for instantaneous dynamic features in vehicle collision detection.

[0041] S103, perform feature extraction on the second time series to obtain the collision-related features of the target vehicle.

[0042] Specifically, the vehicle's attribute data before and after the test time can serve as an indirect indication of a vehicle collision event. Accordingly, this step extracts features from the second time series to obtain collision-related features of the target vehicle. These collision-related features indicate key characteristics of the target vehicle before and after the test time that are highly correlated with the vehicle collision event. This effectively uncovers key features of the target vehicle before and after the test time that are highly correlated with the vehicle collision event, such as state features, driving behavior features, radar features, and collision warning features, thereby providing a valid contextual temporal feature reference for vehicle collision detection.

[0043] S104: Input the attribute change features and collision-related features into the pre-trained collision detection model to obtain the collision detection results of the target vehicle.

[0044] Specifically, this step inputs attribute change features and collision-related features into a pre-trained collision detection model. This model then identifies whether a collision has occurred with the target vehicle at the time of the test, obtaining the collision detection result for the target vehicle, thus achieving vehicle collision detection. The collision detection result indicates whether a collision has occurred with the target vehicle at the time of the test; that is, the collision detection result includes either "collision occurred" or "no collision occurred."

[0045] Optionally, the type of collision detection model can be set according to the actual situation, and this application embodiment does not limit it. For example, the collision detection model can be a traditional machine learning model such as XGBoost or Support Vector Machine (SVM); or, for example, the collision detection model can be a deep learning model such as Convolutional Neural Network (CNN), but it is not limited to these.

[0046] Therefore, the embodiments of this application not only fully consider the baseline data of the vehicle's state while driving, but also fully consider other multi-dimensional data such as driving behavior data, radar data, and vehicle collision-related signal data. These data together constitute attribute data. By extracting features from the attribute data of different time windows, the instantaneous changes of the vehicle at the time of the test can be accurately captured, while fully mining the key features of the vehicle before and after the time of the test that are highly correlated with the vehicle collision event. Then, vehicle collision detection is achieved based on this, thus effectively improving the accuracy of vehicle collision detection.

[0047] In some implementations, step S102 may include steps S201-S203.

[0048] S201, preprocess and extract features from the first time series to obtain the attribute autoregressive features and attribute lag features of various types of data in the first time series.

[0049] Specifically, the first time series, i.e., the true values ​​of attribute data at each moment in the first time series window, can be obtained through the aforementioned step S101. These data are all continuous value data. Therefore, in this step, the first time series is first preprocessed to ensure data quality. Then, based on the autoregressive lag order corresponding to each type of data in the first time series, the corresponding attribute autoregressive features are extracted from each type of data in the first time series. Similarly, based on the distributed lag order corresponding to each type of data in the first time series, the corresponding attribute lag features are extracted from each type of data in the first time series. This facilitates the use of the Autoregressive Distributed Lag (ARDL) model in subsequent steps to locate the target vehicle's condition under normal driving conditions. Here, the attribute autoregressive features indicate the lag period of the data, while the attribute lag features indicate the current value and lag period of the exogenous variables. Exogenous variables refer to other attribute data associated with the current class of attribute data. These can be flexibly set according to actual conditions; for example, other attribute data related to vehicle speed may include longitudinal acceleration, but are not limited to this.

[0050] Optionally, the preprocessing method can be set according to the actual situation, and this embodiment does not limit it. For example, during preprocessing, data cleaning is performed first, such as using interpolation or nearest-neighbor filling to handle missing values, filtering outliers based on the 3σ rule, etc. Then, assuming that all data are recorded in the same time step, if the sampling rates are different, signal frequency compensation processing is performed. That is, data with a period of one second is used as the standard, which is defined as a standard periodic signal. For data with a period of more than one second, the signal value closest to the standard periodic signal is taken and concatenated into the same row, but it is not limited to this.

[0051] Alternatively, the calibration method for the autoregressive lag order and distributed lag order of various types of data in the first time series can be set according to the actual situation, and this embodiment does not impose any restrictions on this. For example, during model training, information criteria such as the Bayesian information criterion, the Akaike information criterion, or cross-validation can be used to determine the autoregressive lag order and distributed lag order of various types of data in the first time series, but it is not limited to these methods.

[0052] Furthermore, in some optional implementations, the first time series can be obtained by sampling through a first time series window, and the implementation process of step S201 above may include:

[0053] The autoregressive window corresponding to the target class attribute data is extracted from the first time series window; wherein, the autoregressive window includes multiple time points from the p time points before the time to be tested to the time point before the time to be tested, where p represents the autoregressive lag order of the target class attribute data; the true value of the target class attribute data in the autoregressive window is determined as the attribute autoregressive feature of the target class attribute data; wherein, the target class attribute data represents any type of data in the first time series.

[0054] Here, for various attribute data in the first time series, we have: extract the autoregressive window corresponding to the current class attribute data from the first time series window. The true value of the current class attribute data in the autoregressive window is the attribute autoregressive feature of the current class attribute data. The autoregressive window includes multiple time points from the p time points before the test time to the time point before the test time, where p represents the autoregressive lag order of the current class attribute data. Based on this, the attribute autoregressive feature can be further expressed as Y. t-1 ,Y t-2 ,…,Y t-pFor example, assuming the longitudinal acceleration is represented as SRS_VehLongAccel, and its autoregressive lag order can be 5, the autoregressive features of the longitudinal acceleration can be represented as SRS_VehLongAccel_{t-1}, SRS_VehLongAccel_{t-2}, SRS_VehLongAccel_{t-3}, SRS_VehLongAccel_{t-4}, and SRS_VehLongAccel_{t-5}, in numerical form, such as -1.5 m / s. Thus, for each type of data in the first time series, the corresponding autoregressive window is extracted from the first time series window by using the autoregressive lag order of the current attribute data, and the attribute autoregressive features of the current attribute data are identified accordingly. This allows for precise location of the lag period of the current attribute data, thereby ensuring the accuracy of the attribute autoregressive features of the current attribute data.

[0055] Furthermore, in some optional implementations, the first time series can be obtained by sampling through a first time series window, and the implementation process of step S201 above may include:

[0056] The lag window corresponding to the target class attribute data is extracted from the first time series window; wherein, the lag window includes multiple times from the q times before the time to be measured to the time to be measured, and q represents the distribution lag order of the target class attribute data; the true value of the exogenous variable of the target class attribute data in the lag window is determined as the attribute lag feature of the target class attribute data; wherein, the target class attribute data represents any type of data in the first time series.

[0057] Here, for various attribute data in the first time series, we have: extracting the lag window corresponding to the current class attribute data from the first time series window, and determining the exogenous variables associated with the current class attribute data. Then, the true values ​​of the exogenous variables of the current class attribute data within the lag window are determined as the attribute lag features of the current class attribute data. The lag window includes multiple times from the q times before the test time to the test time itself, where q represents the distribution lag order of the current class attribute data. Based on this, the attribute lag features can be further expressed as X. t ,X t-1 ,…,X t-qFor example, assuming other attribute data related to vehicle speed may include longitudinal acceleration, which can be represented as SRS_VehLongAccel, and the distributed lag order of vehicle speed can be 2, then the attribute lag features of vehicle speed can be represented as SRS_VehLongAccel_{t}, SRS_VehLongAccel_{t-1}, and SRS_VehLongAccel_{t-2}, in numerical form, such as -1.5 m / s. Thus, for each type of data in the first time series, the corresponding lag window is extracted from the first time series window by using the distributed lag order of the current class attribute data, and the attribute lag features of the current class attribute data are identified accordingly. This allows for precise positioning of the current value and lag period of the exogenous variable of the current class attribute data, thereby ensuring the accuracy of the attribute lag features of the current class attribute data.

[0058] In one example, the first time series may include state data, driving behavior data, collision attribute data, and radar data within the first time series window. State data may include vehicle speed, longitudinal acceleration, average wheel speed (i.e., the average wheel speed of all four wheels), and lateral acceleration. Driving behavior data may include brake pedal position, accelerator pedal position, and steering wheel angular velocity. Collision attribute data may include Advanced Driver Assistance Systems (ADAS) signals, which may include Autonomous Emergency Braking (AEB) target deceleration signals, AEB target distance acceleration signals, and AEB target relative speed signals. Radar data may include obstacle distances from the vehicle's radar. The location of the vehicle's radar is not specifically limited; for example, the vehicle's radar may be located in multiple positions such as the front right, front left, rear right, and rear left of the target vehicle, but is not limited to these. Further assuming a distribution lag order of 2 and an autoregressive lag order of 5 for each type of data, the attribute autoregressive features and attribute lag features shown in Table 1 below can be obtained.

[0059] Table 1: Autoregressive and Lagative Features of Attributes

[0060]

[0061]

[0062]

[0063] S202, input the autoregressive and lag features of the attributes of various types of data in the first time series into the corresponding autoregressive distributed lag model to obtain the predicted values ​​of various types of data in the first time series at the time to be measured.

[0064] Specifically, the various types of data in the first time series often possess different dynamic characteristics, such as autocorrelation and causal relationships between exogenous variables. For instance, changes in vehicle speed may be primarily influenced by longitudinal acceleration and brake pedal position, while changes in longitudinal acceleration may be more significantly affected by factors such as accelerator pedal position. Therefore, this step first pre-trains corresponding ARDL models for each type of data in the first time series. This effectively uncovers the dynamic characteristics of different data, and the independent models allow for the selection of different autoregressive and distributed lag orders for each data point, thereby more accurately locating the target vehicle's condition under normal driving conditions.

[0065] Then, for each type of attribute data in the first time series, the autoregressive features and lag features of the current type of attribute data are input into the ARDL model corresponding to the current type of attribute data, so as to predict the predicted value of the current type of attribute data at the time to be tested by using the past data of the current type of attribute data and the current and past data of other exogenous variables. Among them, the predicted value of each type of data in the first time series at the time to be tested indicates the expected value of each type of data under the condition that the vehicle does not collide (i.e., normal driving), that is, it indicates a non-vehicle collision event (i.e., normal driving event). The predicted value satisfies the following formula (1):

[0066]

[0067] In equation (1), Y represents the predicted value of the current class attribute data at time t; c represents the preset constant term; t-i X represents the actual value of the current class attribute data at time ti; t-j φ represents the true value of the exogenous variable of the current class attribute data at time tj. i θ represents the regression coefficient of the current class attribute data at autoregressive lag order i, indicating the influence of each past period's value on time t; j This represents the lag coefficient of the exogenous variable of the current class attribute data at the distribution lag order j. It indicates the impact of the current class attribute data's exogenous variable in the current period and at each lag period on the current class attribute data.

[0068] S203, perform residual calculations on various types of data in the first time series and their predicted values ​​at the time to be measured, and obtain the residual feature values ​​of various types of data in the first time series as attribute change features.

[0069] Specifically, the true values ​​of various attribute data in the first time series can be obtained through the aforementioned step S101, which indicate the condition of the target vehicle at the time to be measured. The expected values ​​of various attribute data in the first time series can be obtained through the aforementioned step S202, which indicate the condition of the target vehicle under normal driving conditions. For various attribute data in the first time series, residual calculation is performed on the true value and expected value of the current class attribute data, that is, the difference between the true value and the expected value is calculated to obtain the residual feature value of the current class attribute data as the attribute change feature, as shown in the following formula (2):

[0070]

[0071] In equation (2), e t Y represents the residual characteristic value of the current class attribute data at time t; t This represents the actual value of the current class attribute data at time t.

[0072] In one example, using the attribute autoregressive features and attribute lag features shown in Table 1 above as a benchmark for residual calculation, the attribute change features shown in Table 2 below can be obtained. The field type of these attribute change features is Double.

[0073] Table 2: Characteristics of Attribute Changes

[0074]

[0075]

[0076] Therefore, this implementation method uses the ARDL model to predict various types of data in the first time series based on the attribute autoregressive features and attribute lag features of the current data type, obtaining the expected values ​​of various attribute data in the first time series. These values ​​indicate the state of the target vehicle under normal driving conditions. Subsequently, the expected value and the true value of the current data type are used to perform residual calculations to obtain the residual feature values ​​of the current data type. The residual feature values ​​of all data types together constitute the attribute change features of the target vehicle. This allows for accurate positioning of the degree of difference between the state of the target vehicle at the time of the test and its state under normal driving conditions. Based on this, it can fully capture the sudden changes and disturbances that occur to the target vehicle at the time of the test, i.e., instantaneous changes, such as sudden steering wheel turns or sudden speed drops, thereby achieving accurate characterization of the vehicle's instantaneous dynamics and providing an effective reference for instantaneous dynamic features for vehicle collision detection.

[0077] In some implementations, the second time series can be obtained by sampling through a second time window; step S103 may include the following steps S301-S303:

[0078] S301, preprocess the second time series;

[0079] S302, the second timing window is shortened to obtain a timing segment of a preset length;

[0080] S303, feature extraction is performed on the attribute data of the second time series within the time segment to obtain the target vehicle's state features, driving behavior features, collision attribute features, and radar features as collision-related features.

[0081] Specifically, this implementation method comprehensively utilizes multi-dimensional data, namely the second time series, in the selection of features. First, various types of data in the second time series are preprocessed to obtain a preprocessed second time series as a new second time series to ensure data quality. Then, time series segments of a preset length are extracted from the second time series window to facilitate the selection of the most representative data from a large amount of data. Finally, considering both the collision process and post-collision phenomena, such as collision process features like sudden speed changes, large acceleration, sudden braking, sudden steering wheel turns, nearby obstacles, airbag deployment, and automatic emergency braking system intervention, as well as post-collision features like seatbelt unfastening, door unlocking, door opening, and prolonged parking, feature extraction is performed on the attribute data within the time series segments of the second time series to obtain the target vehicle's state features, driving behavior features, collision attribute features, and radar features, which are used as collision-related features.

[0082] Optionally, the preprocessing method can be set according to the actual situation, and this embodiment does not limit it. For example, during preprocessing, data cleaning is performed first, such as using interpolation or nearest-neighbor filling to handle missing values, filtering outliers based on the 3σ rule, etc. Then, binary data is encoded using methods such as one-hot encoding, and multi-category data is encoded using methods such as label encoding. Finally, assuming that all data are recorded in the same time step, if the sampling rates are different, signal frequency compensation processing is performed. That is, data with a period of one second is used as the standard, which is defined as a standard periodic signal. For data with a period of more than one second, the signal value closest to the standard periodic signal is taken and concatenated into the same row, but it is not limited to this.

[0083] Alternatively, the preset length of the time segment can be flexibly set according to the actual situation. For example, a center point is selected from the second time window, such as the time to be measured as the center point, and then a time segment from N seconds before the center point to M seconds after the center point is extracted from the second time window. The values ​​of N and M can be flexibly set according to the actual situation, such as N being 6 and M being 5, but not limited to this.

[0084] In one example, the second time series may include state data, driving behavior data, collision attribute data, and radar data within the second time series window. The state data may include vehicle speed, positive longitudinal acceleration, negative longitudinal acceleration, average wheel speed (i.e., the average wheel speed of all four wheels), and lateral acceleration. The driving behavior data may include brake pedal position, accelerator pedal position, steering wheel angular velocity, gear position, door opening / closing status, and seatbelt tension. The collision attribute data may include ADAS signals and collision-related warning signals. ADAS signals may include AEB target deceleration signals, AEB target distance acceleration signals, AEB target relative speed signals, AEB automatic deceleration request signals, AEB steering suppression signals, AEB target confidence signals, and AEB target motion state signals. Collision-related warning signals may include rearview mirror LED collision warning signals, seat vibration warning signals, airbag collision signals, and close-range collision signals. The radar data may include obstacle distance data from the vehicle's radar. The location of the vehicle-mounted radar is not specifically limited. For example, it can be a vehicle-mounted radar located in multiple directions such as the front right, front left, rear right, and rear left of the target vehicle, but it is not limited to these.

[0085] Further, the time series segment is set as a time series segment from 6 seconds before the center point to 6 seconds after the center point. Based on this, feature extraction is performed on the above second time series, and the collision-related features shown in Table 3 below can be obtained.

[0086] Table 3: Collision-related features

[0087]

[0088]

[0089]

[0090]

[0091] Therefore, this implementation method first further subdivides the temporal segments of feature extraction to narrow the feature range and ensure the accuracy of the features. Then, feature extraction is performed based on full consideration of multi-dimensional data. This can effectively mine key features of the target vehicle before and after the time of the test that are highly correlated with the vehicle collision event, such as state features, driving behavior features, radar features and collision warning features, thereby providing an effective reference for contextual temporal features for vehicle collision detection.

[0092] In some embodiments, prior to step S104, the method further includes the following steps S401-S402:

[0093] S401, perform one-hot encoding on the Boolean features in the collision-related features;

[0094] S402 performs max-min normalization on numerical features in attribute change features and collision-related features.

[0095] Specifically, before inputting attribute change features and collision-related features into the pre-trained collision detection model, this implementation method requires one-hot encoding of all Boolean features and min-max normalization of all numerical features to ensure numerical quality and thus ensure the accuracy of vehicle collision detection.

[0096] Optionally, after the above processing, collision-related features and attribute change features are fused using a co-attention mechanism to obtain fused features, which serve as the final input to the model. Specifically, the cosine similarity between each feature value in the collision-related features and the attribute change features is calculated as the change attention weight of each feature value in the collision-related features. This assigns higher weights to collision-related features associated with transient mutations and reduces the weight ratio of collision-related features unrelated to transient mutations, mitigating the interference of irrelevant features. Subsequently, the product of each feature value in the collision-related features and its change attention weight is calculated as the enhanced feature value, thereby obtaining the enhanced collision-related features. This enables collision-related features to perceive key features hidden in transient mutations and thereby compensate for their lack of dynamic information. Simultaneously, the cosine similarity between each feature value in the attribute change features and the collision-related features is calculated as the collision attention weight for each feature value in the attribute change features. This assigns higher weights to attribute change features associated with vehicle collisions and reduces the weight of attribute change features unrelated to vehicle collisions, mitigating interference from irrelevant features. Then, the product of each feature value in the attribute change features and its collision attention weight is calculated as the enhanced feature value, resulting in the enhanced attribute change features. This allows the attribute change features to perceive key features hidden in the actual situation and compensate for their lack of contextual information. Finally, the enhanced collision-related features and the enhanced attribute change features are concatenated to form a fusion feature, which serves as the final input to the model. This process achieves feature complementarity and interaction, effectively enhancing the expression of key features associated with vehicle collision events, thereby improving the accuracy of vehicle collision detection.

[0097] In some embodiments, the above method may further include the following steps S501-S505.

[0098] S501, Obtain the first vehicle dataset.

[0099] Specifically, this step uses big data technology to collect multiple positive sample attribute data and multiple negative sample attribute data, and assigns corresponding label information to these data to form the first vehicle dataset, which is used for model training in subsequent steps. The first vehicle dataset can include a positive sample dataset and a negative sample dataset. The positive sample dataset can include multiple positive sample attribute data and their corresponding label information, and the negative sample dataset can include multiple negative sample attribute data and their corresponding label information. It should be understood that positive sample attribute data refers to attribute data corresponding to vehicle collision events, with the label information indicating a collision has occurred; negative sample attribute data refers to attribute data corresponding to non-vehicle collision events (i.e., normal driving events), with the label information indicating no collision has occurred.

[0100] Furthermore, in some optional implementations, the implementation process of step S501 above may include:

[0101] Identify several vehicles involved in a collision and several normal vehicles. Based on insurance and maintenance records, determine the positive sample time series window for each vehicle involved in a collision. The positive sample time series window includes the collision time, multiple past sample times preceding the collision time, and multiple subsequent sample times following the collision time. The collision time is the moment the collision occurs. Based on the target state time of each normal vehicle, determine the negative sample time series window for each normal vehicle. The negative sample time series window includes the target state time, multiple past sample times preceding the target state time, and multiple subsequent sample times following the target state time. The target state time includes at least one of the following: vehicle start time, vehicle travel time, vehicle deceleration to a stop time, or vehicle stop time. Extract the attribute data of each vehicle involved in a collision within its corresponding positive sample time series window as positive sample attribute data and assign corresponding label information to each positive sample attribute data. Extract the attribute data of each normal vehicle within its corresponding negative sample time series window as negative sample attribute data and assign corresponding label information to each negative sample attribute data.

[0102] Here, the collision dates of each vehicle are extracted from insurance and maintenance records of multiple vehicles involved in the collision. Based on this, the collision times of each vehicle are identified and labeled as positive sample timestamps. Simultaneously, at least one of the four normal driving scenarios—vehicle start-up, vehicle movement, vehicle deceleration to a stop, and vehicle parking—is selected as negative sample timestamps. For positive samples, the time point of the positive sample is used as the center point of the time series window. A positive sample time series window is constructed based on this, including the collision time, multiple past sample timestamps before the collision time, and multiple subsequent sample timestamps after the collision time. Then, the attribute data of each vehicle within its corresponding positive sample time series window is extracted as positive sample attribute data, and each positive sample attribute data is assigned corresponding label information, thereby constructing a positive sample dataset. For negative samples, the current time point is used as the center point of the time series window for each time point of the negative sample. Based on this, a negative sample time series window for the current time point is constructed, which includes the current time point, multiple past sample times before the current time point, and multiple subsequent sample times after the current time point. Then, the attribute data of each normal vehicle in the corresponding negative sample time series window is extracted as negative sample attribute data, and corresponding label information is assigned to each negative sample attribute data, thereby constructing a negative sample dataset.

[0103] In one example, the time points of the negative samples are selected as follows: (1) Vehicle start time: The vehicle speed changes from 0 km / h for 5 consecutive seconds to greater than 0 km / h for 5 consecutive seconds. The first moment when the vehicle speed changes from 0 to greater than 0 is taken as the vehicle start time; (2) Vehicle travel time: The vehicle speed is greater than 0 km / h for 11 consecutive seconds. The 6th second of this time window is taken as the vehicle travel time; (3) Vehicle deceleration to stop time: The vehicle speed changes from greater than 0 km / h for 5 consecutive seconds to 0 km / h for 5 consecutive seconds. The first moment when the vehicle speed changes from greater than 0 to 0 is taken as the vehicle deceleration to stop time; (4) Vehicle stop time: The vehicle speed is 0 km / h for 11 consecutive seconds. The 6th second of this time window is taken as the vehicle stop time.

[0104] In one example, the time points of the positive and negative samples are used as center points, and the time windows from 5 seconds before to 300 seconds after the center point are used to construct the positive and negative sample time windows. Here, the 5 seconds before the center point are used to examine abnormal vehicle states and driver operation before the collision, while the 300 seconds after the center point are used to examine whether the vehicle has phenomena such as prolonged parking or getting out to check the vehicle's condition after the collision. This ensures the quality of the training data and enables the collision detection model to better learn the mapping relationship between the input data and vehicle collision events.

[0105] S502, using the first vehicle dataset to perform a one-stage training and testing of the preset model to obtain the pre-trained model.

[0106] Specifically, in this step, firstly, the first vehicle dataset is preprocessed, including handling missing and outlier values, data encoding, and temporal alignment. The data samples in the preprocessed first vehicle dataset are randomly shuffled, and the dataset is split into training and testing sets according to a preset ratio. This preset ratio can be set according to actual conditions; for example, it could be 8:2, meaning the dataset is split into training and testing sets in an 8:2 ratio, but it is not limited to this. Secondly, to address the imbalance between positive and negative samples, the training set is resampled to obtain a new training set. For example, the SMOTE algorithm can be used to oversample the minority class, and the TomekLinks algorithm can be used to undersample the excessive number of samples, but this is not limited to this. Then, feature change detection and feature extraction are performed on the training set to obtain attribute change features and collision-related features. Boolean features in the collision-related features are one-hot encoded, and numerical features in the attribute change and collision-related features are subjected to min-max normalization to ensure data quality. It should be understood that the description of feature change detection and feature extraction here can adopt the description of feature change detection and feature extraction in the previous embodiments, and will not be repeated here. Next, the attribute change features and collision-related features of the training set are input into a preset model and the preset model is trained to obtain the trained preset model. Then, the trained preset model is tested using a test set to obtain the performance data of the trained preset model. Finally, if the performance data of the trained preset model meets the preset conditions, the training is considered complete, and the trained preset model is determined as the pre-trained model; otherwise, the training is considered incomplete, the hyperparameters are adjusted, and the process returns to the step of inputting the attribute change features and collision-related features of the training set into the preset model to achieve iterative training. Thus, through the training process of the first training stage, the pre-trained model is enabled to initially learn and explore the relationship between the input data and vehicle collision events, giving it a preliminary ability to identify vehicle collisions.

[0107] In one example, the default model is an XGBoost model, with the following parameters pre-set. Each parameter is a specific value that can be flexibly set according to the actual situation: the parameter range of the XGBoost model, the learning rate search range is lr1~lr2, the maximum tree depth search range is td1~td2, the sample weights of the smallest leaf node search range is mcw1~mcw2, the gamma value search range is γ1~γ2, the lambda value search range is λ1~λ2, and the alpha value search range is α1~α2. Subsequently, a Bayesian optimization method is used to search for optimal parameters within the parameter space of the XGBoost model. A loss function such as cross-entropy loss is used, and k-fold cross-validation is employed to avoid model overfitting.

[0108] In another example, after obtaining the pre-trained model, feature importance analysis and feature selection are performed to determine the types of attribute change features and collision-related features actually used for real-time detection, in order to ensure feature accuracy. Simply put, the feature importance and confusion matrix of the pre-trained model are calculated, and the importance of various features, the diversity of data samples, and the contribution of data samples to the prediction conclusions of the pre-trained model are evaluated by combining the visualization of the pre-trained model and attribute data. Unimportant features are removed, and the composition of data samples is supplemented or adjusted. This is existing technology and will not be elaborated further.

[0109] S503 uses a pre-trained model to perform collision detection on multiple test vehicles within a preset time period, and obtains the collision detection results for each test vehicle.

[0110] S504. The first vehicle dataset is updated based on the collision detection results of each test vehicle to obtain the second vehicle dataset.

[0111] S505 uses the second vehicle dataset to perform two-stage training and testing on the pre-trained model to obtain a pre-trained collision detection model.

[0112] Specifically, the first training phase enables the pre-trained model to distinguish between positive and negative samples. However, manually set negative sample selection rules cannot ultimately cover all non-collision scenarios. First, a preset time period is set, such as 24 hours (but not limited to this), and the pre-trained model is used to perform collision detection on multiple test vehicles within this time period, obtaining the collision detection results for each test vehicle. These results can be divided into two categories: correctly detected collisions and incorrectly detected collisions. Then, these results are used to update (expand) the first vehicle dataset to add a large number of high-quality negative samples, thus obtaining the second vehicle dataset. Finally, the pre-trained model is used to perform the second phase of training and testing using the new second vehicle dataset, resulting in a pre-trained collision detection model, thereby strengthening the model's ability to distinguish between real and false collisions. It should be understood that the description of the second phase of training and testing can follow the description of the first phase of training and testing, and will not be repeated here.

[0113] Furthermore, in some optional implementations, the specific implementation of step S504 above may include:

[0114] Test vehicles whose collision detection results indicate a collision occurred at the time to be tested are identified as suspected vehicles. Based on insurance and maintenance records, vehicles that did not collide within a preset time period are selected from the suspected vehicles as false positives. A negative sample time series window is determined for each false positive based on its target state time. This window includes the target state time, multiple past sample times preceding the target state time, and multiple subsequent sample times following the target state time. The target state time includes at least one of the following: vehicle start time, vehicle travel time, vehicle deceleration to a stop time, or vehicle stop time. Attribute data of each false positive vehicle within its corresponding negative sample time series window is extracted as auxiliary negative sample data, and each auxiliary negative sample is assigned a corresponding label. All auxiliary negative sample data and their corresponding labels are added to the first vehicle dataset to obtain the second vehicle dataset.

[0115] Here, to obtain a sufficiently large number of high-quality negative samples, collision detection results for each test vehicle are obtained by performing collision detection on multiple test vehicles using a pre-trained model within a preset time period. First, these test vehicles are screened. Specifically, test vehicles whose collision detection results indicate a collision occurred at the time of the test are identified as suspicious vehicles, denoted as set R. Suspicious vehicles are those detected by the model as having a collision, but it is uncertain whether a collision actually occurred. Then, based on the insurance and maintenance records of each suspicious vehicle, it can be determined whether a collision actually occurred. Based on this, vehicles that did not collide within the preset time period are selected from the suspicious vehicles as false positives. False positives are those vehicles that the model mistakenly detected as having a collision, but which did not actually occur, denoted as set R_FP. Since the data samples of falsely detected vehicles exhibit highly similar characteristics to those of real collisions, they are valuable for refining the model's distinction between true and false collisions. Therefore, for each falsely detected vehicle, a negative sample time series window is constructed using the target state time of the current falsely detected vehicle as the center point. The attribute data of the current falsely detected vehicle within its negative sample time series window is extracted as auxiliary negative sample data, and corresponding label information is assigned to this auxiliary negative sample data, with the label indicating that no collision occurred. It should be understood that the descriptions of time series window construction, data extraction, and labeling operations here can all be derived from the descriptions of time series window construction, data extraction, and labeling operations in the preceding steps, and will not be repeated here. Finally, each auxiliary negative sample data and its corresponding label information are added to the first vehicle dataset to obtain the second vehicle dataset, thereby updating the first vehicle dataset. In subsequent steps, the second vehicle dataset will be used to perform the second phase of training and testing, enhancing the model's ability to distinguish between real and fake collisions. This will enable the collision detection model to further learn and explore the relationship between data and vehicle collision events that are highly similar to real collisions, thereby effectively improving the performance of the collision detection model on vehicle collision detection tasks and thus effectively improving the accuracy of vehicle collision detection.

[0116] In addition, refer to Figure 2This application also provides a vehicle collision detection device, which may include: an acquisition module 601, a first processing module 602, a second processing module 603, and a detection module 604. The acquisition module 601 is used to acquire a first time series and a second time series of a target vehicle; wherein the first time series and the second time series are sampled through different time series windows, each time series window includes the time to be measured, and each time series includes attribute data within the sampled time series window, the attribute data including state data, driving behavior data, collision attribute data, and radar data; the first processing module 602 is used to perform feature change detection on the first time series to obtain the attribute change features of the target vehicle; the second processing module 603 is used to perform feature extraction on the second time series to obtain the collision-related features of the target vehicle; the detection module 604 is used to input the attribute change features and the collision-related features into a pre-trained collision detection model to obtain the collision detection result of the target vehicle; wherein the collision detection result indicates whether the target vehicle has collided at the time to be measured.

[0117] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0118] Finally, refer to Figure 3 This application also provides a vehicle, which includes:

[0119] At least one processor 701;

[0120] At least one memory 702 is used to store at least one program;

[0121] When at least one program is executed by at least one processor 701, the at least one processor 701 implements the above-described vehicle collision detection method.

[0122] The aforementioned vehicles can be private cars, such as sedans, sport utility vehicles (SUVs), multi-purpose vehicles (MPVs), or pickup trucks, or commercial vehicles, such as vans, buses, small trucks, or large trailers, or gasoline vehicles or new energy vehicles such as hybrid or pure electric vehicles.

[0123] The aforementioned memory 702, as a non-transitory network system, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory 702 may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory 702 may optionally include memory 702 remotely located relative to processor 701, and these remote memories 702 can be connected to processor 701 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0124] The aforementioned memory 702 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 702 and is called and executed by the processor 701.

[0125] The processor 701 described above can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0126] In some embodiments, the vehicle may further include:

[0127] Input / output interfaces are used to implement information input and output;

[0128] The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0129] The bus transmits information between various components of the device (such as processor 701, memory 702, input / output interface and communication interface);

[0130] The processor 701, memory 702, input / output interface, and communication interface can communicate with each other within the device via a bus.

[0131] The content of the above method embodiments is applicable to this vehicle embodiment. The specific functions implemented in this vehicle embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0132] The following will use an application scenario to illustrate the specific implementation process of the embodiments of this application.

[0133] In practical applications, OEMs currently face two major pain points: First, the first 10-15 minutes after a collision are the golden rescue time. OEMs need to confirm injuries and whether repairs are necessary within this period. Most users contact their insurance companies directly after an accident, relying on them for compensation and repair locations. This often results in vehicles not being returned to official after-sales service centers, leading to customer loss. Second, for minor or moderate accidents, some users choose not to repair their vehicles or delay repairs, posing safety risks. If OEMs could identify collision scenarios immediately, they could proactively provide repair services, improving the user experience. This application's embodiments can be used by vehicle OEMs to identify and determine collision accidents. These embodiments help OEMs identify and determine collision accidents, pinpointing the time of occurrence. This allows OEMs to immediately know about vehicles involved in collisions and proactively provide corresponding services, thereby expanding their user base and improving the user experience.

[0134] Specifically, refer to Figure 4 The specific implementation process of this application embodiment is as follows:

[0135] S01, Acquisition of training data: Extract the collision date of each vehicle from the insurance record data and maintenance record data of multiple collision vehicles, and find the collision time of each vehicle and mark it as the time point of the positive sample; at the same time, select at least one of the time points of the four normal driving scenarios of vehicle starting, vehicle driving, vehicle deceleration to stop and vehicle stopping as the time point of the negative sample. The selection of negative sample time points is as follows: (1) Vehicle start time: The vehicle speed changes from 0 km / h for 5 consecutive seconds to greater than 0 km / h for 5 consecutive seconds. The first moment when the vehicle speed changes from 0 to greater than 0 is taken as the vehicle start time; (2) Vehicle travel time: The vehicle speed is greater than 0 km / h for 11 consecutive seconds. The 6th second of this time window is taken as the vehicle travel time; (3) Vehicle deceleration to stop time: The vehicle speed changes from greater than 0 km / h for 5 consecutive seconds to 0 km / h for 5 consecutive seconds. The first moment when the vehicle speed changes from greater than 0 to 0 is taken as the vehicle deceleration to stop time; (4) Vehicle stop time: The vehicle speed is 0 km / h for 11 consecutive seconds. The 6th second of this time window is taken as the vehicle stop time. The positive sample time point and the negative sample time point are taken as the center point, and the time window is from 5 seconds before the center point to 300 seconds after the center point. Based on this, the positive sample time window and the negative sample time window are constructed. Attribute data for each collision vehicle within its corresponding positive sample time window are extracted as positive sample attribute data, and each positive sample attribute data is assigned a corresponding label, thus constructing a positive sample dataset. Attribute data for each normal vehicle within its corresponding negative sample time window are extracted as negative sample attribute data, and each negative sample attribute data is assigned a corresponding label, thus constructing a negative sample dataset. This yields the first vehicle dataset.

[0136] S02, Phase 1 Model Training and Testing: Using the first vehicle dataset and Bayesian optimization methods, the pre-set XGBoost model is trained and tested to obtain the pre-trained model.

[0137] S03, Training Data Update: First, the pre-trained model is used to perform collision detection on multiple test vehicles throughout the day to obtain the collision detection results for each test vehicle. Second, test vehicles whose collision detection results indicate a collision at the time to be tested are identified as suspected vehicles, denoted as set R. Suspected vehicles are those that the model detects as having collided but are unsure whether a collision actually occurred. Then, based on the insurance and maintenance records of each suspected vehicle, vehicles that have not collided within a preset time period are selected from the suspected vehicles as false detection vehicles. False detection vehicles are those that the model mistakenly detects as having collided but have not actually collided, denoted as set R_FP. Next, for each false detection vehicle, a negative sample time series window is constructed with the target state time of the current false detection vehicle as the center point, and the attribute data of the current false detection vehicle within its negative sample time series window is extracted as auxiliary negative sample data. The auxiliary negative sample data is assigned corresponding label information, which is "no collision occurred". Finally, each auxiliary negative sample data and its corresponding label information are added to the first vehicle dataset to obtain the second vehicle dataset.

[0138] S04, Second-stage model training and testing: Using the second vehicle dataset and Bayesian optimization method, the pre-trained model is trained and tested to obtain the pre-trained collision detection model.

[0139] S11, Acquisition of the test data: A first time series is acquired through a first time series window, and a second time series is acquired through a second time series window. The first and second time series windows must meet the corresponding lengths, and the first time series window must be located within the second time series window. The first time series may include state data, driving behavior data, collision attribute data, and radar data located within the first time series window. Specifically: State data may include vehicle speed, longitudinal acceleration, average wheel speed (i.e., the average wheel speed of all four wheels), and lateral acceleration, etc. Driving behavior data may include brake pedal position, accelerator pedal position, and steering wheel angular velocity, etc. Collision attribute data may include ADAS signals, such as AEB target deceleration signals, AEB target distance acceleration signals, and AEB target relative velocity signals, etc. Radar data may include obstacle distance from the vehicle's radar, etc. In addition, the second time series may include state data, driving behavior data, collision attribute data, and radar data within the second time series window. Specifically, the state data may include vehicle speed, positive longitudinal acceleration, negative longitudinal acceleration, average wheel speed (i.e., the average wheel speed of the four wheels), and lateral acceleration. The driving behavior data may include brake pedal position, accelerator pedal position, steering wheel angular velocity, gear status, door opening / closing status, and seat belt tension. The collision attribute data may include ADAS signals and collision-related warning signals. The ADAS signals may include AEB target deceleration signals, AEB target distance acceleration signals, AEB target relative speed signals, AEB automatic deceleration request signals, AEB steering suppression signals, AEB target confidence signals, and AEB target motion state signals, etc. The collision-related warning signals may include rearview mirror LED collision warning signals, seat vibration warning signals, airbag collision signals, and close-range collision signals, etc. The radar data may include obstacle distance from the vehicle radar, etc.

[0140] S12, Feature Change Detection: First, set the distribution lag order of each type of data in the first time series to 2 and the autoregressive lag order to 5. Based on this, the attribute autoregressive features and attribute lag features shown in Table 1 above can be obtained. Then, for each type of attribute data in the first time series, input the attribute autoregressive features and attribute lag features of the current type of attribute data into the ARDL model corresponding to the current type of attribute data to obtain the predicted value of the current type of attribute data at the time to be tested, as shown in formula (1) above. Then, perform residual calculation on the true value and expected value of the current type of attribute data, that is, calculate the difference between the true value and the expected value to obtain the residual feature value of the current type of attribute data as the attribute change feature, as shown in formula (2) above. Based on this, the attribute change features shown in Table 2 above can be obtained.

[0141] S13, Feature extraction: The time series segment is set as a time series segment from 6 seconds before the center point to 6 seconds after the center point. Based on this, feature extraction is performed on the second time series to obtain the collision-related features shown in Table 3 above.

[0142] S14, Pre-prediction processing: Perform one-hot encoding on all Boolean features and perform max-min normalization on all numerical features.

[0143] S15, Model Prediction: Input all processed collision-related features and attribute change features into a pre-trained collision detection model to obtain collision detection results, which are used to indicate whether a collision has occurred at the time of the test.

[0144] In summary, the embodiments of this application comprehensively consider multi-dimensional features, triggering collisions from three aspects: vehicle dynamics (such as vehicle collision posture), vehicle signals (such as ADAS signals), and human reactions before and after a collision (such as driver behavior data). This provides more evidence for accurate collision identification, including considering driving behavior information such as sudden braking, sudden steering wheel turning, vehicle AEB, radar obstacle detection, and vehicle system collision warnings, as well as post-collision behavior information such as prolonged parking, unfastening seat belts, and opening vehicle doors. Thus, by using richer information to assist in collision identification decisions, the accuracy of vehicle collision detection can be effectively improved.

[0145] On the other hand, this application proposes a two-stage training mechanism to address the problem of imbalanced positive and negative samples and excessive signal differences. In the first stage, signal segments from collision scenarios and signal segments from normal driving scenarios such as vehicle starting, vehicle driving, vehicle deceleration to a stop, and vehicle parking are used as training data. After calculating features, the model is trained to obtain a model with weak discrimination ability. In the second stage, samples that are misjudged as collisions by the model are added to the negative sample dataset to further enhance the model's ability to distinguish between real and false collisions, thereby effectively improving the model's performance on collision detection tasks.

[0146] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0147] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A vehicle collision detection method, characterized in that, include: A first time series and a second time series of the target vehicle are obtained; wherein the first time series and the second time series are obtained by sampling through different time series windows, each time series window includes the time to be measured, and each time series includes attribute data within the sampled time series window, the attribute data including state data, driving behavior data, collision attribute data and radar data; Feature change detection is performed on the first time series to obtain the attribute change features of the target vehicle; Feature extraction is performed on the second time series to obtain the collision-related features of the target vehicle; The attribute change features and the collision-related features are input into a pre-trained collision detection model to obtain the collision detection result of the target vehicle; wherein, the collision detection result indicates whether the target vehicle has collided at the time of the test. The step of detecting feature changes in the first time series to obtain the attribute change features of the target vehicle includes: The first time series is preprocessed and features are extracted to obtain the attribute autoregressive features and attribute lag features of various types of data in the first time series. The autoregressive and lag features of the attributes of various types of data in the first time series are input into the corresponding autoregressive distributed lag model to obtain the predicted values ​​of various types of data in the first time series at the time to be measured. Residual calculations are performed on various types of data in the first time series and their predicted values ​​at the time to be measured to obtain residual feature values ​​of various types of data in the first time series as the attribute change features.

2. The method according to claim 1, characterized in that, The first time series is obtained by sampling through a first time series window; the preprocessing and feature extraction of the first time series to obtain the attribute autoregressive features and attribute lag features of various types of data in the first time series includes: An autoregressive window corresponding to the target class attribute data is extracted from the first time series window; wherein, the autoregressive window includes data from the time series before the time to be measured. Multiple times from the time to the time before the time to be measured This represents the autoregressive lag order of the target class attribute data; The true value of the target class attribute data in the autoregressive window is determined as the attribute autoregressive feature of the target class attribute data; The target class attribute data refers to any one type of data in the first time series.

3. The method according to claim 1, characterized in that, The first time series is obtained by sampling through a first time series window; the preprocessing and feature extraction of the first time series to obtain the attribute autoregressive features and attribute lag features of various types of data in the first time series includes: The lag window corresponding to the target class attribute data is extracted from the first time series window; wherein, the lag window includes the data from the time before the time to be measured. Multiple times from the time to the time to be measured, This indicates the distribution lag order of the target class attribute data; The true value of the exogenous variable of the target class attribute data in the lag window is determined as the attribute lag feature of the target class attribute data; The target class attribute data refers to any one type of data in the first time series.

4. The method according to claim 1, characterized in that, The second time series is obtained by sampling through a second time window; the feature extraction of the second time series to obtain the collision-related features of the target vehicle includes: The second time series is preprocessed; The second timing window is shortened to obtain a timing segment of a preset length; Feature extraction is performed on the attribute data of the second time series within the time segment to obtain the state features, driving behavior features, collision attribute features, and radar features of the target vehicle as the collision-related features.

5. The method according to claim 1, characterized in that, The method further includes: Obtain a first vehicle dataset; wherein the first vehicle dataset includes a positive sample dataset and a negative sample dataset, the positive sample dataset includes multiple positive sample attribute data and label information corresponding to each positive sample attribute data, and the negative sample dataset includes multiple negative sample attribute data and label information corresponding to each negative sample attribute data; The first vehicle dataset is used to train and test the preset model for one stage to obtain the pre-trained model. During a preset time period, the pre-trained model is used to perform collision detection on multiple test vehicles to obtain the collision detection results for each test vehicle. The first vehicle dataset is updated based on the collision detection results of each test vehicle to obtain the second vehicle dataset; The pre-trained model is trained and tested in two stages using the second vehicle dataset to obtain the pre-trained collision detection model.

6. The method according to claim 5, characterized in that, The process of obtaining the first vehicle dataset includes: Identify a number of vehicles involved in the collision and a number of vehicles that were otherwise in good condition; Based on insurance record data and maintenance record data, a positive sample time sequence window is determined for each of the aforementioned collision vehicles; wherein, the positive sample time sequence window includes the collision time, multiple past sample times located before the collision time, and multiple subsequent sample times located after the collision time, wherein the collision time is the moment when the collision occurs; Based on the target state time of each of the normal vehicles, a negative sample time series window is determined for each of the normal vehicles; wherein, the negative sample time series window includes the target state time, a plurality of past sample times located before the target state time and a plurality of subsequent sample times located after the target state time, and the target state time includes at least one of the vehicle start time, vehicle driving time, vehicle deceleration to stop time or vehicle stopping time. Extract the attribute data of each of the collision vehicles in the corresponding positive sample time window as the positive sample attribute data, and assign corresponding label information to each of the positive sample attribute data. The attribute data of each normal vehicle within the corresponding negative sample time window are extracted as the negative sample attribute data, and corresponding label information is assigned to each negative sample attribute data.

7. The method according to claim 5, characterized in that, The step of updating the first vehicle dataset based on the collision detection results of each of the test vehicles to obtain a second vehicle dataset includes: The test vehicle whose collision detection result indicates that a collision occurred at the time to be tested is identified as a suspected vehicle. Based on insurance record data and maintenance record data, vehicles that did not have a collision during the preset time period are selected from the suspected vehicles as false detection vehicles. Based on the target state time of each of the falsely detected vehicles, a negative sample time series window is determined for each of the falsely detected vehicles; wherein, the negative sample time series window includes the target state time, a plurality of past sample times located before the target state time and a plurality of subsequent sample times located after the target state time, and the target state time includes at least one of the following: vehicle start time, vehicle driving time, vehicle deceleration to stop time, or vehicle stop time. Extract the attribute data of each falsely detected vehicle within the corresponding negative sample time window as auxiliary negative sample data, and assign corresponding label information to each of the auxiliary negative sample data. The auxiliary negative sample data and their corresponding label information are added to the first vehicle dataset to obtain the second vehicle dataset.

8. A vehicle collision detection device, characterized in that, include: The acquisition module is used to acquire a first time series and a second time series of the target vehicle; wherein the first time series and the second time series are obtained by sampling through different time series windows, each time series window includes the time to be measured, and each time series includes attribute data within the sampled time series window, the attribute data including state data, driving behavior data, collision attribute data and radar data; The first processing module is used to perform feature change detection on the first time series to obtain the attribute change features of the target vehicle; The second processing module is used to extract features from the second time series to obtain the collision-related features of the target vehicle. The detection module is used to input the attribute change features and the collision-related features into a pre-trained collision detection model to obtain the collision detection result of the target vehicle; wherein, the collision detection result indicates whether the target vehicle has collided at the time of the test. The step of detecting feature changes in the first time series to obtain the attribute change features of the target vehicle includes: The first time series is preprocessed and features are extracted to obtain the attribute autoregressive features and attribute lag features of various types of data in the first time series. The autoregressive and lag features of the attributes of various types of data in the first time series are input into the corresponding autoregressive distributed lag model to obtain the predicted values ​​of various types of data in the first time series at the time to be measured. Residual calculations are performed on various types of data in the first time series and their predicted values ​​at the time to be measured to obtain residual feature values ​​of various types of data in the first time series as the attribute change features.

9. A vehicle, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the vehicle collision detection method as described in any one of claims 1-7.

Images