Vehicle yaw identification method, device, equipment, storage medium and program product

Abstract

The application provides a vehicle yaw identification method, device, equipment, storage medium and program product. The method comprises the following steps: a causal diagram model is constructed to represent the causal relationship among a driving intention signal, an environmental interference factor and a real yaw result; based on the causal diagram model and observation data pre-labeled with the real yaw result, an intervention distribution can be predicted by using a backdoor adjustment algorithm; and then, whether a yaw event detected by a vehicle in real time is a real yaw can be judged by using the intervention distribution. Based on the observation data and the intervention distribution, the influence of the environmental interference factor on the causal relationship between the driving intention signal and the real yaw result can be excluded. Through this causal discrimination method, the yaw misidentification caused by the environmental factor interference and the misjudgment of temporary behaviors such as obstacle avoidance is reduced, and the accuracy of yaw identification and the reliability of navigation are improved.

Description

Technical Field

[0001] This application relates to the field of vehicle navigation, and in particular to a vehicle yaw recognition method, device, equipment, storage medium, and program product. Background Technology

[0002] Navigation software helps vehicles reach their destinations by planning routes and providing real-time guidance, and is widely used in various vehicle travel scenarios. Accurately identifying whether the vehicle has deviated from its course during navigation is crucial. This not only indicates whether the vehicle can trigger navigation replanning in a timely manner, but also whether it can filter out false alarms caused by temporary disturbances, thus improving the reliability of the navigation system.

[0003] Currently, vehicle yaw detection primarily relies on data such as satellite positioning and visual perception. This type of data is susceptible to environmental interference; for example, heavy rain and strong light can cause sensor signals to drift, leading to false yaw detection. Furthermore, actions such as avoiding obstacles while driving are also frequently misidentified as yaw. Such false yaw detection not only triggers unnecessary navigation replanning but also generates incorrect error messages, severely impacting navigation reliability. Summary of the Invention

[0004] This application provides a vehicle yaw recognition method, device, equipment, storage medium, and program product to improve the accuracy of yaw recognition.

[0005] In a first aspect, embodiments of this application provide a vehicle yaw detection method, including:

[0006] Construct a cause-effect graph model, which includes outcome variables, processing variables, and confounding variables. Outcome variables represent whether the vehicle actually veers off course, processing variables represent driving intention signals, and confounding variables represent environmental interference factors that affect driving intention signals and / or vehicle veergence detection.

[0007] Based on a causal graph model and observational data with pre-labeled true yaw results, a backdoor adjustment algorithm is used to predict the intervention distribution; the intervention distribution represents the causal relationship between the driving intention signal and the true yaw result after excluding the influence of confounding variables.

[0008] The intervention distribution is used to determine whether the yaw events detected by the vehicle in real time are true yaws.

[0009] Among some possible implementations, based on causal graphical models and observational data pre-labeled with true yaw results, a backdoor adjustment algorithm is used to predict the intervention distribution, including:

[0010] Based on observational data, the conditional probability of actual yaw occurring under different combinations of driving intention signals and confounding variables is determined;

[0011] Based on the observed data, determine the marginal probability of each confounding variable.

[0012] The intervention distribution is obtained by weighted summation of the conditional probability and the marginal probability.

[0013] In the above method, for each driving intention signal, the conditional probability under different environments is weighted by the natural distribution of the confounding variables, so that the distribution of the driving intention signal under various environments is consistent with the distribution of the confounding variables in the natural environment. This can cut off the backdoor path of the confounding variables affecting the generation of the driving intention signal, eliminate the false associations caused by environmental interference factors, and obtain the pure causal effect of driving intention on the real deviation.

[0014] In some possible implementations, the vehicle includes an ultraviolet polarization sensor, and before predicting the intervention distribution using a backdoor adjustment algorithm based on causal graphical models and pre-labeled observation data with actual yaw results, it also includes:

[0015] If the confidence level of the vehicle's yaw detection data is lower than a preset threshold, the absolute heading angle output by the ultraviolet polarization sensor is obtained, as well as the real yaw sample under the condition that the absolute heading angle affects the driving intention signal.

[0016] The absolute heading angle is introduced as an instrumental variable into the causal graph model. The instrumental variable and the confounding variable are independent of each other, and the instrumental variable indirectly affects the judgment of the true yaw by influencing the driving intention signal.

[0017] Update observational data using real yaw samples.

[0018] In the above method, by pre-acquiring yaw events detected when the driver can use the ultraviolet polarization sensor and labeling whether the yaw event is a true yaw, true yaw samples containing instrumental variables can be generated. The newly added true yaw samples can provide training data under adverse conditions for the prediction of causal graphical models and intervention distributions, ensuring the accuracy of instrumental variable estimation. Furthermore, the instrumental variables can be used to identify the causal effect of driving intention on true yaw, maintaining the robustness of yaw determination.

[0019] In some possible implementations, intervention distribution is used to determine whether a yaw event detected by the vehicle in real time is a true yaw, including:

[0020] In response to the vehicle detecting a yaw event, acquire the confusion variables and driving intention signals at the current moment;

[0021] Based on the confounding variables and driving intention signals at the current moment, the corresponding probability values ​​are matched from the intervention distribution as the causal confidence of the yaw event;

[0022] If the causal confidence level is greater than or equal to the first threshold, the yaw event is determined to be a true yaw.

[0023] If the causal confidence level is less than or equal to the second threshold, the yaw event is determined to be a misidentification.

[0024] In the above method, based on the causal relationship between the driving intention signal and the actual deviation, as well as the intervention distribution, the causal confidence of the deviation event can be determined. Combined with the dual threshold judgment mechanism, the causal confidence can be transformed into clear navigation control commands, effectively reducing the deviation misjudgment rate.

[0025] In some possible implementations, the obfuscation variables at the current moment are obtained, including:

[0026] Real-time acquisition of vehicle perception data, including at least one of visual perception data, satellite positioning data, and inertial measurement data;

[0027] Based on the perceived data, identify the interfering factors in the current environment;

[0028] The interference level is determined based on the degree of influence of interference factors on yaw detection, and the interference level is used as a confusion variable.

[0029] In the above approach, by collecting multimodal sensing data and identifying specific interference factors, complex environmental states can be mapped into discrete interference levels and used as the values ​​of the confusion variable Z. Thus, the confusion variable can be used to calculate the intervention distribution and dynamically adapt to the current environment in subsequent decision-making.

[0030] Among the possible implementations are:

[0031] If the causal confidence is greater than the second threshold and less than the first threshold, then the missing information of the perception data is generated through a spatiotemporal generative adversarial network, and the missing information is used to correct the perception data.

[0032] Based on the corrected perception data, identify the semantics of the current scene;

[0033] Based on the corrected perception data and the semantics of the current scene, counterfactual simulations are performed on the potential trajectory of the vehicle, and the simulation results are used to determine whether the yaw event is a real yaw.

[0034] In the above approach, for ambiguous scenarios where the causal confidence level is between the first and second thresholds, a second verification can be performed after the perceptual data is completed using a spatiotemporal generative adversarial network. This completion method has a higher completion effect than traditional linear interpolation and other methods, which can improve the data completion rate, reduce spatial errors, increase the correctness of counterfactual simulation results, and improve the yaw recognition accuracy.

[0035] Among the possible implementations are:

[0036] Based on the vehicle's historical trajectory and current scene semantics, predict the vehicle's potential future veergence path and use the potential veergence path to output a warning; and / or,

[0037] When the vehicle's current real road does not match the navigation map topology path, a reasonable driving path is determined based on the semantics of the current scene, and the deviation event is judged as a real deviation according to the reasonable driving path.

[0038] In the above approach, predicting potential risks based on trajectory dynamics and changing the passive response of navigation software when veering off course to active warning can provide drivers with more time to correct course and achieve risk prediction and alerts.

[0039] In some possible implementations, after using the intervention distribution to determine whether the yaw event detected by the vehicle in real time is a true yaw, the following is also included:

[0040] Acquire perception data from roadside units surrounding the vehicle when a yaw event occurs;

[0041] Based on the perception data from the roadside unit, determine whether the yaw event detected by the vehicle in real time is a real yaw, and obtain the first judgment result;

[0042] If the first judgment result is inconsistent with the second judgment result, the first threshold or the second threshold shall be corrected according to the first judgment result. The second judgment result is the judgment result of whether the yaw event is a real yaw using the intervention distribution.

[0043] In the above approach, the roadside unit is not limited by the vehicle's sensor field of view, and has a wider field of view than the vehicle, thus reflecting the real road conditions more objectively. Utilizing more reliable roadside data to dynamically adjust the threshold allows the vehicle's subsequent judgments, based on causal graph models and interventions distributed in similar scenarios, to more closely align with the semantics of the real road, thereby improving the accuracy of yaw detection.

[0044] Among the possible implementations are:

[0045] Obtain the causal relationship between yaw detection and multiple preset driving scenarios;

[0046] When a vehicle is performing yaw detection, the vehicle's dependence weight on different perception data is dynamically adjusted based on the real-time driving scenario and causal relationships.

[0047] In the above approach, when fusing multimodal sensor data to detect yaw events, the fusion weights of each sensor can be dynamically adjusted according to the real-time driving scenario and the causal relationship between the driving scenario and yaw detection, rather than relying on static rules or fixed weights. This can reduce false alarms of yaw events caused by correlation fitting from the root.

[0048] In some possible implementations, multiple vehicle nodes store their respective observation data locally; other methods include:

[0049] For any vehicle node, the intervention distribution and / or the pre-parameters used to predict the intervention distribution are encrypted and / or de-identified, and the processed data is uploaded to a distributed node network containing multiple vehicle nodes; and the local intervention distribution and / or pre-parameters are updated using the data uploaded by other vehicle nodes to the distributed node network.

[0050] The intervention distribution and pre-parameters are obtained based on the observation data stored locally by the vehicle nodes.

[0051] In the above approach, a distributed node network composed of multiple vehicle nodes replaces the traditional central server, which avoids the impact of single-point failures on the yaw recognition function. Furthermore, by uploading only the encrypted and anonymized intervention distribution and its pre-parameters, the model can be deployed locally on the vehicle, and the vehicle's original trajectory data does not leave the vehicle's local storage. This not only improves the model's generalization ability through data sharing but also reduces the risk of privacy leaks.

[0052] Secondly, embodiments of this application provide a vehicle yaw detection device, comprising:

[0053] The causal construction module is used to build a causal graph model. The causal graph model includes outcome variables, processing variables, and confounding variables. The outcome variables represent whether the vehicle actually veers off course, the processing variables represent the driving intention signal, and the confounding variables represent environmental interference factors that affect the driving intention signal and / or vehicle veergence detection.

[0054] The prediction module is used to predict the intervention distribution based on a causal graph model and observation data with pre-labeled true yaw results, using a backdoor adjustment algorithm. The intervention distribution represents the causal relationship between the driving intention signal and the true yaw result after excluding the influence of confounding variables.

[0055] The judgment module is used to determine whether the yaw event detected by the vehicle in real time is a real yaw by using the intervention distribution.

[0056] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0057] The memory stores the instructions that the computer executes;

[0058] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0059] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0060] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0061] The vehicle yaw recognition method, apparatus, device, storage medium, and program product provided in this application construct a causal graph model to represent the causal relationship between driving intention signals, environmental interference factors, and actual yaw results. Based on the causal graph model and observation data pre-labeled with actual yaw results, a backdoor adjustment algorithm can be used to predict the intervention distribution. The intervention distribution can then be used to determine whether the yaw event detected by the vehicle in real time is a true yaw. Based on the observation data and the intervention distribution, the influence of environmental interference factors on the causal relationship between driving intention signals and actual yaw results can be eliminated. Through this causal discrimination method, misidentification of yaw caused by environmental interference and misjudgment of temporary actions such as obstacle avoidance is reduced, thereby improving the accuracy of yaw recognition and the reliability of navigation. Attached Figure Description

[0062] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0063] Figure 1 A flowchart illustrating a vehicle yaw recognition method provided as an example in this application;

[0064] Figure 2 A causal relationship illustration of a causal graph model provided as an example in this application Figure 1 ;

[0065] Figure 3 A causal relationship illustration of a causal graph model provided as an example in this application Figure 2 ;

[0066] Figure 4 This application provides an exemplary schematic diagram of the architecture of a yaw recognition system.

[0067] Figure 5 This application provides a schematic diagram of the structure of a vehicle yaw detection device as an example.

[0068] Figure 6 This is a schematic diagram of the structure of an electronic device provided as an example in this application.

[0069] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0070] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0071] With the widespread use of in-vehicle navigation, the ability to accurately identify vehicle deviation (i.e., the vehicle deviating from the planned navigation path) during navigation is extremely important. Once deviation is detected, not only should the route be replanned, but the user should also be alerted through voice prompts (such as "You have deviated from the navigation path and will be replanning the route soon").

[0072] Currently, vehicles primarily rely on satellite signals such as GNSS (Global Navigation Satellite System) for positioning. During navigation, vehicles can detect whether they are veering off course based on GNSS data, IMU (Inertial Measurement Unit) data, and data from sensors such as cameras and radar.

[0073] However, vehicle yaw recognition has several limitations: GNSS signals are easily interfered with by thick clouds and heavy rain, resulting in positioning delays and position drift. This positioning deviation manifests as the vehicle being on the planned navigation path but being misidentified as deviating from the planned path, leading to misidentification of yaw. Cameras are prone to problems such as rain and fog obscuring the lens, and poor lighting, making it difficult to clearly capture key visual features such as lane lines, resulting in a lack of basis for yaw judgment. Relying solely on IMU data will accumulate errors over time, and the longer the driving time, the worse the reliability of yaw recognition. In addition, vehicles often need to avoid pedestrians or temporarily detour through construction zones, causing temporary deviations in the vehicle's trajectory. Existing yaw recognition algorithms rely on data correlation analysis, which can easily misjudge these non-active trajectory changes as real yaws, affecting the accuracy of navigation judgments.

[0074] The aforementioned limitations in yaw detection often lead to misidentification of yaw during vehicle navigation, which not only triggers unnecessary navigation replanning but also generates false error messages, seriously affecting the reliability of navigation.

[0075] Based on this, a technical concept is proposed: when a vehicle detects a veergence based on data such as satellite signals and visual perception, the system uses driving intention signals such as turn signals and steering wheel angles to determine whether this veergence is a genuine veergence requiring navigation replanning. Thus, when the vehicle experiences positioning inaccuracies due to signal drift, driving intention signals such as the turn signals not being activated and the steering wheel angle not changing abruptly can be used to determine that the vehicle is likely not veerging, thereby reducing false identifications caused by positioning errors.

[0076] However, temporary actions such as avoiding obstacles may prompt the driver to activate turn signals or turn the steering wheel. Environmental factors such as obstacles can also interfere with driving intention signals, potentially leading to misidentification of yaw if relying solely on these signals. Therefore, based on the aforementioned technical concept, a causal graph model is constructed to determine the net causal effect of the driving intention signal on the actual yaw result after excluding the influence of environmental factors such as obstacles and heavy rain. This effect is then used to determine whether the vehicle has actually yawed. This causal discrimination method reduces the interference of environmental factors, avoids misidentifying temporary actions such as avoiding obstacles as yaw, and improves the accuracy of yaw identification and the reliability of navigation.

[0077] The vehicle yaw recognition method based on the technical concept of this application can be applied to intelligent connected vehicles, high-precision navigation, and autonomous driving perception and decision-making systems. Specific application scenarios can include regular urban roads and highways, as well as extreme driving scenarios such as rainstorms, foggy weather, mining areas, off-road roads, tunnel groups, underground garages, construction sections, and urban canyons.

[0078] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0079] Figure 1 This application provides an exemplary flowchart of a vehicle yaw recognition method, as shown below. Figure 1 As shown, the method includes:

[0080] Step S101: Construct a causal graph model.

[0081] The causal graphical model includes outcome variables, treatment variables, and confounding variables. Outcome variables represent whether the vehicle actually veers off course. Treatment variables represent the driving intention signal. Confounding variables represent environmental interference factors that affect the driving intention signal and / or vehicle veergence detection.

[0082] Figure 2 A causal relationship illustration of a causal graph model provided as an example in this application Figure 1.like Figure 2 As shown: Y is the outcome variable, which can indicate whether the vehicle actually veers off course; X is the driving intention signal, which can include the turn signal being turned on, voice turn commands (such as "turn right"), sudden changes in steering wheel angle, etc.; Z is the confounding variable, which can include various environmental interference factors, such as visual perception failure caused by rain or fog, obstacles that need to be avoided such as construction cones or pedestrians crossing the road, and satellite positioning offset caused by GNSS multipath effects, etc.

[0083] It is understandable that the environmental interference factors represented by the confusion variables may affect changes in driving intention signals and may also cause inaccurate vehicle positioning, leading to the misidentification of a vehicle that is in the correct position (i.e., the position on the navigation route) as being in a veergency.

[0084] Step S102: Based on the causal graph model and observation data with pre-labeled real yaw results, the backdoor adjustment algorithm is used to predict the intervention distribution.

[0085] The intervention distribution represents the causal relationship between the driving intention signal and the actual yaw result after excluding the influence of confounding variables.

[0086] In this embodiment of the application, based on the constructed causal graph model, learning can be performed using a large amount of historical observation data (including driving intention signals, environmental interference factors, and manually labeled real yaw results). For example... Figure 2 The causal relationship shown indicates that Z simultaneously affects both X and Y, forming a backdoor path. Without intervention, this would lead to a false association between X and Y. Therefore, a backdoor adjustment algorithm can be used to de-obfuscate the causal relationship between X and Y, thereby blocking the influence of Z on the causal relationship between X and Y and calculating an intervention distribution that reflects the true probability of deviation caused purely by driving intention.

[0087] Step S103: Use the intervention distribution to determine whether the yaw event detected by the vehicle in real time is a real yaw.

[0088] Among these, deviation events are detected based on vehicle satellite positioning data, visual perception data (such as camera data), and IMU data. For example, navigation software can monitor vehicle satellite positioning and camera footage. When it detects that the vehicle's satellite positioning is not on the planned navigation path or that the vehicle has left its current lane based on camera footage, the navigation software can classify this as a deviation event.

[0089] In this embodiment, the intervention distribution is actually a confidence lookup table for determining velocities based on the causal relationship between driving intention signals and actual velocities. It can record the probability of actual velocities occurring under various combinations of driving intention signals and environmental interference factors. For example, for a detected velocity event, when the driver activates the turn signal and the environment is clear and undisturbed, the causal confidence is as high as 95%, indicating that it is highly likely to be a real velocity. However, when the driver does not activate the turn signal and encounters heavy rain, the causal confidence is only 10%, indicating that it is likely a misidentification. Furthermore, when the turn signal is activated but there is an obstacle ahead that needs to be temporarily avoided, the causal confidence may be less than 40%.

[0090] For example, when navigation software detects a yaw event, it can match the confidence level corresponding to the current driving scenario from the intervention distribution and compare the confidence level with a preset threshold. If the confidence level is higher than the threshold, it is determined to be a real yaw and triggers route replanning; otherwise, it is regarded as a temporary disturbance or sensor false alarm and is ignored.

[0091] In the above embodiments, by constructing a causal graph model, the causal relationship between driving intention signals, environmental interference factors, and actual yaw results can be represented. Based on the causal graph model and observation data pre-labeled with actual yaw results, a backdoor adjustment algorithm can be used to predict the intervention distribution. The intervention distribution can then be used to determine whether the yaw event detected by the vehicle in real time is a true yaw. Based on the observation data and the intervention distribution, the influence of environmental interference factors on the causal relationship between driving intention signals and actual yaw results can be eliminated. Through this causal discrimination method, yaw misidentification caused by environmental interference and misjudgment of temporary actions such as obstacle avoidance is reduced, improving the accuracy of yaw identification and the reliability of navigation.

[0092] In one embodiment, based on a causal graph model and observational data pre-labeled with actual yaw results, a backdoor adjustment algorithm is used to predict the intervention distribution, including:

[0093] Based on observational data, the conditional probability of actual veergence under different combinations of driving intention signals and confounding variables is determined; based on observational data, the marginal probability of each confounding variable is determined; the intervention distribution is obtained by weighted summation of the conditional probability and the marginal probability.

[0094] In the embodiments of this application, the backdoor adjustment algorithm can be expressed as the following formula (1).

[0095] (1)

[0096] Wherein, P(Y|do(X=x)) represents the intervention distribution, which in this embodiment can be understood as "if a forced driving intention signal X=x (such as forced activation of the turn signal), what is the probability that the vehicle will actually veer (Y=1)?" Without forcing an experiment, this formula can be used to predict the probability distribution of forced intervention using existing observational data.

[0097] On the right side of formula (1), P(Y | X=x, Z=z) is the conditional probability, representing the frequency of the actual velocity Y occurring in historical data when the driver's intention is X and the confounding variable is Z. For example, it can include information such as "the probability of velocity when using lights in heavy rain" and "the probability of velocity when using lights in clear weather". P(Z) is the marginal probability, which can represent the actual proportion of various environmental disturbances Z (such as clear weather, heavy rain, construction, etc.) in the natural environment.

[0098] In the above embodiments, for each driving intention signal x, the conditional probability under different environments is weighted by the natural distribution P(Z) of the confusion variable Z, so that the distribution of X under various environments is consistent with the distribution of Z in the natural environment. This can cut off the backdoor path generated by Z influencing X, eliminate the false associations generated by environmental interference factors, and obtain the pure causal effect of driving intention on real deviation.

[0099] In one embodiment, the vehicle includes an ultraviolet polarization sensor and, prior to predicting the intervention distribution using a backdoor adjustment algorithm based on a causal graphical model and observation data pre-labeled with actual yaw results, may further include:

[0100] If the confidence level of the vehicle's perception data is lower than a preset threshold, the absolute heading angle output by the ultraviolet polarization sensor is obtained, as well as the actual yaw sample when the absolute heading angle affects the driving intention signal; the absolute heading angle is introduced as an instrumental variable into the causal graph model; and the observation data is updated using the actual yaw sample.

[0101] The vehicle yaw detection data can include satellite positioning data, IMU data, and visual perception data.

[0102] Figure 3 A causal relationship illustration of a causal graph model provided as an example in this application Figure 2 .like Figure 3 As shown, the instrumental variable I and the confounding variable Z are independent of each other. The instrumental variable I can indirectly affect the judgment of the true yaw by influencing the driving intention signal.

[0103] Navigation software primarily relies on yaw detection data collected by the vehicle's own sensors (such as GNSS modules and cameras) to detect vehicle yaw events. These sensors are susceptible to environmental interference and may fail to varying degrees, leading to unreliable data. In extreme environments, these sensors may even fail completely; for example, in heavy rain, not only are satellite signals inaccurate, but cameras also fail to identify the correct lane. To address this, this application employs an ultraviolet polarization sensor instead of the failed traditional sensor. The absolute heading angle output by the ultraviolet polarization sensor is unaffected by rain or fog and can serve as a stable heading source.

[0104] For example, such as Figure 3 As shown, the absolute heading angle is introduced as an instrumental variable I into the causal graphical model. I is independent of confounding variables (such as rain, fog, and multipath effects), and I only indirectly affects the true yaw judgment by influencing the driver's intention. Yaw events detected when the driver can use the ultraviolet polarization sensor are acquired in advance, and these events are labeled as true yaws, generating true yaw samples containing the instrumental variable. The newly added true yaw samples can provide training data under adverse conditions for the causal graphical model and the prediction of the intervention distribution, ensuring the accuracy of the instrumental variable estimation.

[0105] In real-time yaw determination, by introducing instrumental variables, even in the event of GNSS or camera failure, an ultraviolet polarization sensor can be used to replace the failed sensor. This allows the instrumental variables to identify the causal effect of driving intention on the actual yaw, thus maintaining the robustness of yaw determination.

[0106] In one embodiment, determining whether a yaw event detected by the vehicle in real time is a true yaw using intervention distribution includes:

[0107] In response to the vehicle detecting a yaw event, the system acquires the confusion variables and driving intention signals at the current moment; based on the confusion variables and driving intention signals at the current moment, it matches the corresponding probability values ​​from the intervention distribution as the causal confidence of the yaw event; if the causal confidence is greater than or equal to the first threshold, the yaw event is determined to be a real yaw; if the causal confidence is less than or equal to the second threshold, the yaw event is determined to be a misidentification.

[0108] The first and second thresholds can be set custom-defined or calibrated through prior testing. The first and second thresholds can be equal or unequal.

[0109] For example, causal graph models, observation data, and intervention distributions can be pre-deployed locally on the vehicle. When a yaw event is detected during navigation, vehicle sensors can be used to acquire current environmental interference factors (such as heavy rain or obstacles) and driving intention signals (such as turn signal status). The acquired data is then used as an index to match the corresponding probability value from the vehicle's local intervention distribution as the causal confidence score. This causal confidence score can represent the probability that the yaw event reflects the user's yaw intention under the current environment and driving intention signals.

[0110] For example, the first threshold is 90% and the second threshold is 40%. If the causal confidence is greater than or equal to 90%, the deviation event can be determined to be a systematic deviation (such as positioning drift), which is a real deviation that requires a response from the navigation software and triggers route replanning. If the causal confidence is less than or equal to 40%, it is determined to be a driver's active behavior (such as avoiding obstacles or taking a temporary detour), which is a deviation misidentification and does not require replanning or outputting deviation prompts.

[0111] In the above embodiments, based on the causal relationship between the driving intention signal and the actual deviation, as well as the intervention distribution, the causal confidence of the deviation event can be determined. Combined with the dual threshold judgment mechanism, the causal confidence can be transformed into explicit navigation control commands, effectively reducing the deviation misjudgment rate.

[0112] In some possible implementations, the obfuscation variables at the current moment are obtained, including:

[0113] Real-time acquisition of vehicle perception data; identification of interference factors in the current environment based on the perception data; determination of interference level based on the degree of influence of interference factors on yaw detection, and use the interference level as a confusion variable.

[0114] The perception data includes at least one of visual perception data, satellite positioning data, and inertial measurement data.

[0115] For example, when a yaw event is detected during navigation, perception data can be collected in real time by the vehicle's multimodal sensors. The perception data may include visual information captured by the camera (such as lane line clarity and obstacle recognition), satellite positioning signals received by GNSS (such as signal-to-noise ratio and positioning accuracy factor), and inertial measurement data output by the IMU (such as acceleration and angular velocity).

[0116] Based on this sensory data, specific interference factors can be identified: for example, visual image analysis can determine whether there is rain or fog obstruction, construction cones, or pedestrians crossing the road; GNSS signal quality can assess whether there is multipath effect or signal attenuation; and IMU data fluctuations can determine the degree of road bumpiness. The identification results of this step can also be used to determine the current driving scenario of the vehicle. For example, poor GNSS signal quality may indicate that the vehicle is in a weak signal environment such as a tunnel, and a high degree of road bumpiness may indicate that the vehicle is in off-road terrain such as mountains.

[0117] After identifying the interfering factors, their impact on yaw detection can be transformed into discrete interference levels, which can then be used as a confusion variable Z. For example, the interference level can include two levels: no interference and interference, corresponding to Z=0 and Z=1 respectively.

[0118] In the above embodiments, by collecting multimodal sensing data and identifying specific interference factors, complex environmental states can be mapped into discrete interference levels and used as the values ​​of confusion variables Z. Thus, confusion variables Z can be used to calculate the intervention distribution and dynamically adapt to the current environment in subsequent decision-making.

[0119] In one embodiment, after matching the causal confidence level, the following may also be included:

[0120] If the causal confidence level is greater than the second threshold and less than the first threshold, then the missing information of the perception data is generated through a spatiotemporal generative adversarial network, and the missing information is used to correct the perception data; based on the corrected perception data, the semantics of the current scene are identified; based on the corrected perception data and the semantics of the current scene, counterfactual simulation of the vehicle's potential trajectory is performed, and the simulation results are used to determine whether the yaw event is a real yaw.

[0121] In this embodiment, if the causal confidence level corresponding to a veergence event is between a first threshold and a second threshold, meaning it cannot be determined with a high probability that it is a genuine veergence or a misidentification of veergence, then ST-GAN (SpatialTransformer Generative Adversarial Networks) can be used to complete the vehicle perception data, and the completed data can then be used for verification. For example, if there is an obstacle in front of the vehicle's lane causing the lane line to be obscured, and the vehicle bypasses the obstacle, resulting in missing lane line information collected by the onboard camera, ST-GAN can be used to complete the data. After completion, a counterfactual simulation is performed on the potential trajectory to determine whether the vehicle would veer if there were no obstacle. The simulation results are used to determine whether it is a genuine veergence (for example, if the simulation results show that the vehicle trajectory returns to the original lane after the construction cone is removed, then the current veergence is determined to be a temporary avoidance, not a genuine veergence).

[0122] In some possible implementations, when performing data completion using ST-GAN, heading angle data output from an ultraviolet polarization sensor can be introduced to further enhance the data completion effect. Experimental results show that using the implementation method of this application to complete missing vehicle perception data achieves a completion rate of 95% and a spatial error of less than 0.3 meters, significantly outperforming traditional completion methods such as linear interpolation and Kalman prediction.

[0123] In the above embodiments, for ambiguous scenarios where causal confidence is between the first and second thresholds, secondary verification can be performed after completing the perceived data using a spatiotemporal generative adversarial network. This completion method is more effective than traditional linear interpolation, improving data completion rate, reducing spatial error, increasing the correctness of counterfactual simulation results, and improving yaw recognition accuracy. Furthermore, this yaw judgment method based on quantified causal confidence and threshold partitions avoids the confusion between "active detour (causally controllable)" and "system misjudgment (false association)" in traditional binary judgments. It distinguishes between active detour and misjudgment by constructing a logical chain of "yaw cause-effect." For example, in active detour, the driver's turning intention is clear, and the causal confidence is ≤40%, which is judged as "false association"; in system misjudgment, there is no clear intention variable, and the causal confidence is ≥90%, which is judged as "real yaw." The intermediate interval can be further verified using a spatiotemporal generative adversarial network, significantly improving yaw recognition accuracy.

[0124] In one embodiment, the method may further include:

[0125] Based on the vehicle's historical trajectory and current scene semantics, predict the vehicle's potential future deviation path and use the potential deviation path to output a warning.

[0126] In this embodiment, proactive early warning can be achieved based on the vehicle's historical trajectory and scene semantics. For example, the historical trajectory of a period prior to a veergency (e.g., the first 3 seconds) can be acquired to capture the vehicle's movement trend and driving intention. Combined with current scene semantics (e.g., blurred lane lines ahead, the presence of construction cones, or pedestrians crossing), the potential future veergency path of the vehicle can be simulated and a warning issued. For instance, when lane lines are blurred due to wear, the trajectory of the vehicle potentially deviating to the left can be predicted based on the steering wheel angle change and IMU heading angle in the previous 3 seconds, and a warning can be issued to the driver 2-3 seconds before the actual deviation occurs. Similarly, when an obstacle is detected in front of the vehicle, the avoidance route can be predicted based on the steering wheel angle change and IMU heading angle in the previous 3 seconds and output to the driver.

[0127] In the above embodiments, by predicting potential risks based on trajectory dynamics, the passive response of navigation software when deviating from the course is changed to an active warning, which can provide the driver with more time to correct the course and realize risk prediction and warning.

[0128] In one embodiment, the method may further include:

[0129] When the vehicle's current real road does not match the navigation map topology path, a reasonable driving path is determined based on the semantics of the current scene, and the deviation event is judged as a real deviation according to the reasonable driving path.

[0130] Traditional yaw detection algorithms mostly rely on high-definition map data, which in turn depend on fixed road topology. In temporary scenarios such as road construction detours or off-road terrain without lane markings, where standardized lane lines are absent and road boundaries frequently change, traditional solutions relying on high-definition map topology matching completely fail. This leads to false associations due to a "map-to-real-road mismatch" (e.g., a vehicle traveling on a temporary road is flagged as veerging from its course by the map). Furthermore, vehicles often encounter potholes, steep slopes, and other uneven terrain, resulting in significant trajectory fluctuations and further complicating yaw detection.

[0131] In this embodiment of the application, the current scene semantics may include the terrain semantics around the vehicle. The combination of "terrain semantics + driver intention + trajectory causal logic" can also determine whether a real deviation has occurred (for example, when off-roading, the temporary road boundary generated by the LiDAR can be used as a causal variable, and combined with the driver's steering intention, it can be determined whether the trajectory is actively selected).

[0132] For example, when off-roading in a mining area, the temporary road boundaries generated by LiDAR serve as causal variables. If the driver's steering intention aligns with the direction of the temporary road, it is determined to be an active choice rather than a deviation. Furthermore, ST-GAN can simulate and generate temporary road scene data, supplementing the prior information missing from the high-precision map, enabling the decision-making logic to adapt to the surrounding road environment. In this way, the dependence on fixed map topology can be reduced, false deviations caused by map-road mismatches can be avoided, and the robustness of deviation recognition and the accuracy of deviation recognition in special terrain scenarios such as mining areas or off-road driving can be improved.

[0133] In one embodiment, after determining whether the yaw event detected by the vehicle in real time is a true yaw using the intervention distribution, the method further includes:

[0134] Acquire perception data from roadside units around the vehicle when a yaw event occurs; based on the perception data from the roadside units, determine whether the yaw event detected by the vehicle in real time is a real yaw, and obtain a first judgment result; if the first judgment result is inconsistent with the second judgment result, then correct the first threshold or the second threshold based on the first judgment result, and the second judgment result is the judgment result of whether the yaw event is a real yaw using the intervention distribution.

[0135] V2I (Vehicle-to-Infrastructure) communication refers to wireless communication technology between vehicles and roadside infrastructure (i.e., roadside units). Roadside units can integrate sensors such as LiDAR, cameras, and millimeter-wave radar, and can generate semantic information from a bird's-eye view based on the data collected by these sensors. This information can represent the position and motion status of road boundaries, lane lines, obstacles, and vehicles, providing a description of the road environment from a perspective other than the vehicle's viewpoint.

[0136] In this embodiment, a dual-channel yaw recognition system for vehicles and roadside units can be formed based on the V2I mechanism. For the vehicle channel, the method described in the above embodiments can be used to determine whether the yaw event is a genuine yaw, thus obtaining a second determination result. For the roadside unit channel, information about the road environment where the vehicle is located at the time of the yaw event can be obtained based on V2I, and this information can be used to determine whether the yaw event is a genuine yaw, thus obtaining a first determination result.

[0137] In some possible implementations, for areas with weak signal strength such as tunnel groups or underground parking garages, the topological features inside the tunnel or underground parking garage (such as ventilation openings, lighting layout, etc.) can be pre-stored locally on the vehicle. Alternatively, the topological features inside the tunnel or underground parking garage can be obtained through the V2I mechanism of this application embodiment. Based on these topological features, the vehicle positioning data and surrounding environment data can be supplemented by combining the ST-GAN in the above embodiments, and can be matched and verified with LiDAR point cloud data.

[0138] In the above embodiments, the roadside unit is not limited by the vehicle sensor's field of view and typically has a wider field of view than the vehicle, enabling it to more objectively reflect the real road conditions (such as confirming whether the current lane is suitable for lane changing, whether there is a construction area, etc.). If the roadside unit's judgment result is inconsistent with the vehicle's causal judgment result, it indicates that there may be factors in the current environment that the vehicle sensors have not observed (such as the degree of tunnel signal attenuation or temporary traffic control, etc.). More reliable roadside data can be used to dynamically correct the decision threshold (such as 90% or 40%), making the vehicle's subsequent judgments based on the causal graph model and intervention distribution in similar scenarios closer to the real road semantics, thereby improving the accuracy of yaw recognition.

[0139] In one embodiment, the method may further include:

[0140] Obtain the causal relationship between yaw detection and multiple preset driving scenarios; when the vehicle performs yaw detection, dynamically adjust the dependence weight of different perception data when detecting whether a yaw event exists, based on the real-time driving scenario and causal relationship of the vehicle.

[0141] Currently, navigation software relies on multimodal sensor data (such as IMU data and camera data) from the vehicle to detect veer-off events. When running detection algorithms, data from each sensor is typically fused with fixed weights before determining whether a veer-off has occurred. This approach has significant limitations. For example, GNSS signals are highly reliable in clear weather but less reliable in heavy rain. However, the weight of GNSS signals is the same during multimodal data fusion in both weather conditions. This leads to reliance on satellite positioning even when GNSS signals weaken in heavy rain, or on inertial measurement systems when IMU noise increases in off-road conditions. This method cannot adapt to changes in driving scenarios, resulting in highly unstable accuracy in initial veer-off event detection.

[0142] In this embodiment, historical data can be used to learn the causal relationship between yaw detection and various driving scenarios. For example, analysis shows that in heavy rain scenarios, the heading angle data from the ultraviolet polarization sensor has the strongest causal correlation with the actual yaw, while the causal correlation of GNSS signals is significantly weakened due to multipath effects. In off-road scenarios, the causal correlation between IMU+LiDAR terrain matching data and the actual yaw is stronger than that of IMU data alone. Based on these causal relationships, when detecting yaw events in real time, the dependence weight of different sensing data can be dynamically adjusted according to the specific scenario in which the vehicle is currently located (such as identifying heavy rain or off-roading through sensing data). For example, in heavy rain, the weight of ultraviolet polarization data can be increased, and the weight of GNSS data can be decreased.

[0143] In some possible implementations, the roadside unit can sense the vehicle's driving scenario and generate a corresponding scenario weight matrix, which is then transmitted to the vehicle via V2I. When the vehicle navigation software detects a yaw event, it allocates the fusion weights of each sensor according to this matrix.

[0144] In the above embodiments, when fusing multimodal sensor data to detect yaw events, the fusion weights of each sensor can be dynamically adjusted according to the real-time driving scenario and the causal relationship between the driving scenario and yaw detection, rather than relying on static rules or fixed weights. This can reduce false alarms of yaw events caused by correlation fitting from the root.

[0145] In one embodiment, multiple vehicle nodes each store their own observation data locally, and the method further includes:

[0146] For any vehicle node, the intervention distribution and / or the pre-parameters used to predict the intervention distribution are encrypted and / or anonymized, and the processed data is uploaded to a distributed node network containing multiple vehicle nodes; for any vehicle node, the local intervention distribution and / or pre-parameters are updated using the data uploaded by other vehicle nodes in the distributed node network.

[0147] The intervention distribution and preconditions are obtained based on the observation data stored locally by the vehicle nodes. The preconditions may include conditional probabilities and marginal probabilities as described in the above embodiments. The methods for determining the preconditions and intervention distribution can be referred to the relevant embodiments described above, and will not be repeated here.

[0148] Currently, vehicle yaw detection applications typically involve setting up a central server in the cloud. Vehicles act as clients, uploading their driving trajectories and other data to the cloud server. The cloud server then trains a unified yaw detection model based on the large amount of trajectory data uploaded by multiple vehicles. After training, any vehicle can upload its real-time trajectory to the cloud, which then inputs the real-time trajectory into the model and sends the model's yaw detection result back to the corresponding vehicle. This method carries a single point of failure risk; if the cloud server fails, vehicles cannot perform yaw detection. Furthermore, uploading trajectory data to the cloud poses a privacy risk, but without uploading trajectory data, the cloud model lacks sufficient training samples, leading to poor model generalization ability and difficulty in adapting to diverse yaw detection needs.

[0149] In this application embodiment, a privacy-preserving collaborative training mechanism based on a distributed node network is proposed to address the aforementioned issues. Under this mechanism, even if multiple vehicles need to share data for training, each vehicle can train a yaw recognition model based on causality locally. Observation data and trajectory data containing user privacy can be stored locally on the vehicle, and only non-sensitive data such as pre-parameters and intervention distributions are encrypted and anonymized before being shared with the cloud and other vehicles. For example, vehicles that have not experienced rainstorm scenarios can refer to the conditional probability statistics of other vehicle nodes under rainstorm conditions to improve their yaw recognition capabilities in similar scenarios.

[0150] The yaw detection model based on causal relationships can be used to implement the relevant steps in the above method embodiments, such as constructing a causal graph model and predicting the intervention distribution based on the causal graph model. Encryption processing may include adding differential privacy noise, and desensitization processing may include removing vehicle identification tags to ensure that it is impossible to trace back to a specific vehicle or trip based on shared data.

[0151] In some possible implementations, vehicle nodes can be nodes within the same consortium blockchain. A consortium blockchain is a type of blockchain that requires authorized participation; only certified vehicle nodes can participate in data sharing and consensus verification. It not only possesses the basic characteristics of a blockchain but also offers enhanced privacy.

[0152] In the above embodiments, a distributed node network composed of multiple vehicle nodes replaces the traditional central server, which avoids the impact of single-point failures on the yaw recognition function. Furthermore, by employing a consortium blockchain to share data and limiting the shared data to encrypted and anonymized intervention distributions and their pre-parameters, the model is deployed locally on the vehicle, and the vehicle's original trajectory data does not leave the vehicle. This approach not only improves the model's generalization ability through shared data but also reduces the risk of privacy leaks, achieving dual protection of model co-evolution and user privacy.

[0153] Figure 4 This is a schematic diagram of the architecture of a yaw recognition system provided as an example in this application. Figure 4 As shown, the system may include a cloud layer and a vehicle layer. The vehicle layer may refer to the vehicle nodes in the above embodiments, and the cloud layer may be a cloud server used to coordinate the various vehicle nodes.

[0154] The vehicle layer can include sensors such as GNSS, IMU, cameras, wheel speedometers and ultraviolet polarization sensors for sensing the vehicle's surrounding environment and its own state. It can also include software modules such as a lightweight BEV (bird's-eye view) feature extraction network, a causal inference engine, an ST-GAN data completion module and an intent confidence estimator for yaw recognition.

[0155] The lightweight BEV feature extraction network utilizes deep learning to stitch together camera footage, generating a bird's-eye view of the road environment surrounding the vehicle based on data from multiple onboard cameras. The ST-GAN data completion module uses ST-GAN to complete data based on camera data, combined with heading angle data output from an ultraviolet polarization sensor. The causal inference engine can determine the causal relationship between driving intention signals and actual yaw and predict intervention distribution. The intention confidence estimator can match the causal confidence of the current yaw event to the intervention distribution and generate a yaw determination result based on a preset first or second threshold. The onboard layer can also encrypt model parameters such as the intervention distribution locally before uploading them to the cloud layer.

[0156] The cloud layer can include modules such as a federated learning coordinator, a regional scene parameter library, and a user anonymization behavior profiling engine. The regional scene parameter library can store scene parameters related to multiple regions. The federated learning coordinator can maintain global model parameters based on model parameters uploaded by at least one vehicle node, generating a global fusion model update package. The user anonymization behavior profiling engine can be used to analyze data such as deviation and avoidance behaviors of user groups after anonymizing user privacy; this data can be used to support the learning and updating of global model parameters.

[0157] In some possible implementations, the vehicle layer can also include an extreme scenario adaptation module, which can customize perception strategies for extreme driving scenarios such as heavy rain, mining areas, and tunnels. For example, for scenarios where vehicle perception is difficult, such as rain or heavy snow, a stable heading source unaffected by light can be used, i.e., the absolute heading angle output by an ultraviolet polarization sensor can be used to detect yaw. In scenarios such as mining areas where the boundaries of roads change frequently and map topology matching is lacking, terrain semantics can be identified through camera data, IMU data, etc., to determine temporary road boundaries and, based on this, to determine a reasonable driving route to determine whether yaw has occurred, thus eliminating the reliance on high-precision maps. In scenarios with weak communication signals, such as tunnels or underground parking garages, scene topology data can be pre-stored, or data from roadside units can be obtained through the aforementioned V2I mechanism, combined with the pre-stored topology data or V2I data for yaw identification, improving the problem of GNSS signal failure.

[0158] The yaw recognition system proposed in this application is based on "causal logic to remove false information, generation technology to complete the system, and extreme scenario adaptation". It can build a three-level collaborative system of "vehicle-road-cloud" and break through the traditional "data association + fixed rules" paradigm of yaw recognition. It solves the problems of false association, data missing, map dependence and privacy leakage in yaw recognition from the root.

[0159] The yaw detection system of this application constructs a causal graph model based on driving intent and yaw detection. By using Do-Calculus (Do calculus, also known as intervention calculus) to block confounding variables and retain true causal features, it effectively reduces false yaw associations. Experiments have verified that the false yaw misjudgment rate can be reduced by 70%. Furthermore, it utilizes spatiotemporal consistent priors to generate high-fidelity data and predicts potential risks based on trajectory dynamics, maintaining stable detection even in scenarios with missing data, thus possessing proactive warning capabilities. In addition, customized perception strategies (such as ultraviolet polarization, terrain semantics, and road feature matching) are implemented for different extreme scenarios to achieve accurate adaptation. Experiments have verified that the yaw detection accuracy across all scenarios is ≥98%.

[0160] Figure 5 This application provides an exemplary structural diagram of a vehicle yaw detection device, as shown below. Figure 5 As shown, the vehicle yaw recognition device 500 provided in this embodiment includes:

[0161] The causal construction module 501 is used to construct a causal graph model. The causal graph model includes outcome variables, processing variables, and confounding variables. The outcome variables represent whether the vehicle has a real yaw, the processing variables represent the driving intention signal, and the confounding variables represent environmental interference factors that affect the driving intention signal and / or vehicle yaw detection.

[0162] The prediction module 502 is used to predict the intervention distribution based on the causal graph model and observation data with pre-labeled true yaw results using a backdoor adjustment algorithm; the intervention distribution represents the causal relationship between the driving intention signal and the true yaw result after excluding the influence of confounding variables.

[0163] The judgment module 503 is used to determine whether the yaw event detected by the vehicle in real time is a real yaw by using the intervention distribution.

[0164] In some possible implementations, the prediction module 502 can also be used to: determine the conditional probability of a real yaw occurring under different combinations of driving intention signals and confounding variables based on observation data; determine the marginal probability of each confounding variable occurring based on observation data; and obtain the intervention distribution by weighted summation of the conditional probability and the marginal probability.

[0165] In some possible implementations, the causal construction module 501 can also be used to: if the confidence level of the vehicle's yaw detection data is lower than a preset threshold, obtain the absolute heading angle output by the ultraviolet polarization sensor, and obtain the real yaw sample when the absolute heading angle affects the driving intention signal; introduce the absolute heading angle as an instrumental variable into the causal graph model, where the instrumental variable and the confounding variable are independent of each other, and the instrumental variable indirectly affects the judgment of the real yaw by influencing the driving intention signal; and update the observation data using the real yaw sample.

[0166] In some possible implementations, the judgment module 503 can also be used to: in response to the vehicle detecting a yaw event, obtain the confusion variable and driving intention signal at the current moment; based on the confusion variable and driving intention signal at the current moment, match the corresponding probability value from the intervention distribution as the causal confidence of the yaw event; if the causal confidence is greater than or equal to a first threshold, determine that the yaw event is a real yaw; if the causal confidence is less than or equal to a second threshold, determine that the yaw event is a misidentification.

[0167] In some possible implementations, the judgment module 503 can also be used to: collect vehicle perception data in real time, including at least one of visual perception data, satellite positioning data and inertial measurement data; identify interference factors in the current environment based on the perception data; determine the interference level based on the degree of influence of the interference factors on yaw detection, and use the interference level as a confusion variable.

[0168] In some possible implementations, the judgment module 503 can also be used to: identify the semantics of the current scene based on the corrected perception data; perform counterfactual simulation of the vehicle's potential trajectory based on the corrected perception data and the semantics of the current scene, and determine whether the yaw event is a real yaw based on the simulation results.

[0169] In some possible implementations, the judgment module 503 can also be used to: predict the vehicle's potential future deviation path based on the vehicle's historical trajectory and the semantics of the current scene, and use the potential deviation path to output a warning.

[0170] In some possible implementations, the judgment module 503 can also be used to: determine a reasonable driving path based on the semantics of the current scene when the vehicle's current real road does not match the navigation map topology path, and determine whether the deviation event is a real deviation according to the reasonable driving path.

[0171] In some possible implementations, the judgment module 503 can also be used to: acquire the perception data of the roadside units around the vehicle when the yaw event occurs; determine whether the yaw event detected by the vehicle in real time is a real yaw based on the perception data of the roadside units, and obtain a first judgment result; if the first judgment result is inconsistent with the second judgment result, then the first threshold or the second threshold is corrected according to the first judgment result, and the second judgment result is the judgment result of whether the yaw event is a real yaw using the intervention distribution.

[0172] Some possible implementations also include a yaw detection module, which can be used to: obtain the causal relationship between yaw detection and multiple preset driving scenarios; and dynamically adjust the vehicle's dependence weight on different perception data when detecting the presence of a yaw event, based on the real-time driving scenario and causal relationship of the vehicle.

[0173] In some possible implementations, a shared collaboration module is also included, which can be used to: for any vehicle node, encrypt and / or de-identify the intervention distribution and / or the pre-parameters used to predict the intervention distribution, and upload the processed data to a distributed node network containing multiple vehicle nodes, wherein the intervention distribution and pre-parameters are obtained based on the observation data stored locally by the vehicle nodes; for any vehicle node, update the local intervention distribution and / or pre-parameters using the data uploaded by other vehicle nodes in the distributed node network.

[0174] The vehicle yaw recognition device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0175] Figure 6 This is a schematic diagram of the structure of an electronic device provided in this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus 604.

[0176] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0177] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0178] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0179] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0180] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0181] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0182] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0183] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0184] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0185] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0186] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0187] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0188] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0189] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0190] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for vehicle yaw detection, characterized in that, include: A cause-effect graph model is constructed, which includes outcome variables, processing variables, and confounding variables. The outcome variables represent whether the vehicle actually veers off course, the processing variables represent the driving intention signal, and the confounding variables represent environmental interference factors that affect the driving intention signal and / or vehicle veergence detection. Based on the causal graph model and observation data pre-labeled with actual yaw results, the intervention distribution is predicted using a backdoor adjustment algorithm; the intervention distribution represents the causal relationship between the driving intention signal and the actual yaw result after excluding the influence of the confounding variables. The intervention distribution is used to determine whether the yaw event detected by the vehicle in real time is a real yaw.

2. The method according to claim 1, characterized in that, The method of predicting the intervention distribution based on the causal graph model and pre-labeled observation data with true yaw results, using a backdoor adjustment algorithm, includes: Based on the observed data, the conditional probability of a true yaw occurring under different combinations of driving intention signals and confusion variables is determined; Based on the observed data, determine the marginal probability of each confounding variable. The intervention distribution is obtained by weighted summation of the conditional probability and the marginal probability.

3. The method according to claim 1 or 2, characterized in that, The vehicle includes an ultraviolet polarization sensor, and before predicting the intervention distribution using a backdoor adjustment algorithm based on the causal graph model and observation data pre-labeled with real yaw results, it also includes: If the confidence level of the vehicle's yaw detection data is lower than a preset threshold, the absolute heading angle output by the ultraviolet polarization sensor is obtained, and the actual yaw sample is obtained when the absolute heading angle affects the driving intention signal. The absolute heading angle is introduced as an instrumental variable into the causal graph model. The instrumental variable is independent of the confounding variable, and the instrumental variable indirectly affects the judgment of the true yaw by influencing the driving intention signal. The observation data are updated using the actual yaw samples.

4. The method according to claim 1 or 2, characterized in that, The step of using the intervention distribution to determine whether the yaw event detected by the vehicle in real time is a true yaw includes: In response to the vehicle detecting a yaw event, acquire the confusion variables and driving intention signals at the current moment; Based on the confusion variables and driving intention signals at the current moment, the corresponding probability values ​​are matched from the intervention distribution as the causal confidence of the yaw event; If the causal confidence level is greater than or equal to the first threshold, then the yaw event is determined to be a true yaw. If the causal confidence level is less than or equal to the second threshold, the yaw event is determined to be a misidentification.

5. The method according to claim 4, characterized in that, The process of obtaining the confusion variables at the current moment includes: Real-time acquisition of vehicle perception data, including at least one of visual perception data, satellite positioning data, and inertial measurement data; Based on the perceived data, identify the interference factors in the current environment; Based on the degree of influence of the aforementioned interference factors on yaw detection, the interference level is determined, and the interference level is used as a confusion variable.

6. The method according to claim 5, characterized in that, Also includes: If the causal confidence is greater than the second threshold and less than the first threshold, then the missing information of the perceived data is generated by a spatiotemporal generative adversarial network, and the missing information is used to correct the perceived data. Based on the corrected perception data, identify the semantics of the current scene; Based on the corrected perception data and the semantics of the current scene, a counterfactual simulation of the vehicle's potential trajectory is performed, and the simulation results are used to determine whether the yaw event is a real yaw.

7. The method according to claim 6, characterized in that, Also includes: Based on the vehicle's historical trajectory and the semantics of the current scene, predict the vehicle's potential future deviation path, and use the potential deviation path to output an early warning. And / or, When the vehicle's current real road does not match the navigation map topology path, a reasonable driving path is determined based on the current scene semantics, and the deviation event is judged as a real deviation according to the reasonable driving path.

8. The method according to claim 4, characterized in that, After using the intervention distribution to determine whether the yaw event detected by the vehicle in real time is a real yaw, the process also includes: Acquire the perception data of the roadside units around the vehicle when the yaw event occurs; Based on the perception data of the roadside unit, it is determined whether the yaw event detected by the vehicle in real time is a real yaw, and a first judgment result is obtained; If the first judgment result is inconsistent with the second judgment result, the first threshold or the second threshold shall be corrected according to the first judgment result. The second judgment result is the judgment result of whether the deviation event is a real deviation using the intervention distribution.

9. The method according to claim 1 or 2, characterized in that, Also includes: Obtain the causal relationship between yaw detection and multiple preset driving scenarios; When the vehicle is performing yaw detection, the vehicle's dependence weight on different perception data is dynamically adjusted according to the real-time driving scenario and the causal relationship. Alternatively, multiple vehicle nodes may store their respective observation data locally, and the method may further include: For any vehicle node, the intervention distribution and / or the pre-parameters used to predict the intervention distribution are encrypted and / or de-identified, and the processed data is uploaded to a distributed node network containing the multiple vehicle nodes; and the local intervention distribution and / or pre-parameters are updated using the data uploaded by other vehicle nodes to the distributed node network. The intervention distribution and the precondition parameters are obtained based on the observation data stored locally by the vehicle nodes.

10. A vehicle yaw detection device, characterized in that, include: The causal construction module is used to construct a causal graph model, which includes outcome variables, processing variables, and confounding variables. The outcome variables represent whether the vehicle has actually veered off course. The processing variables represent the driving intention signal. The confounding variables represent environmental interference factors that affect the driving intention signal and / or vehicle veergence detection. The prediction module is used to predict the intervention distribution based on the causal graph model and observation data pre-labeled with the actual yaw results, using a backdoor adjustment algorithm; the intervention distribution represents the causal relationship between the driving intention signal and the actual yaw result after excluding the influence of the confounding variables. The judgment module is used to determine whether the yaw event detected by the vehicle in real time is a real yaw using the intervention distribution.

11. A computer-readable storage medium / computer program product / electronic device, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 9; and / or, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 9; and / or, The electronic device includes a processor and a memory communicatively connected to the processor; wherein the memory stores computer-executable instructions, and the processor executes the computer-executable instructions stored in the memory to implement the method as described in any one of claims 1 to 9.

Images