Vehicle attitude control method, electronic equipment and vehicle

Abstract

The invention relates to the technical field of vehicle control, and provides a vehicle attitude control method, electronic equipment and a vehicle. The method comprises the following steps: acquiring state parameters of a vehicle, and acquiring image data of an object placing area outside the vehicle; determining an object type according to the image data by using a pre-trained image recognition model; determining predicted attitude parameters of the vehicle according to the state parameters and the object type by using a pre-trained attitude prediction model; and receiving historical driving parameters of the vehicle in a preset time period, determining a target risk level of the vehicle attitude by using a pre-trained decision model according to the predicted attitude parameters, the historical driving parameters and the object type, and controlling the vehicle based on the target risk level. In this way, the target risk level comprehensively considers the predicted attitude parameter, the historical driving parameter and the object type of the object placed outside the vehicle, and the attitude instability risk caused by the object placed outside the vehicle can be reduced.

Description

Technical Field

[0001] This disclosure relates to the field of vehicle control technology, and in particular to a method for controlling vehicle attitude, electronic equipment, and vehicle. Background Technology

[0002] During vehicle operation, the system typically identifies potential instability risks based on set parameters. However, it does not consider the impact of externally placed objects on the vehicle's attitude, making it impossible to accurately predict instability risks.

[0003] In view of this, how to accurately predict the instability risk of vehicle attitude has become an urgent technical problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of this disclosure is to provide a vehicle attitude control method, electronic device and vehicle to solve the problem of the inability to accurately predict the instability risk of vehicle attitude in the prior art.

[0005] To achieve the above objectives, the first aspect of this disclosure provides a method for controlling vehicle attitude, the method comprising: Obtain the vehicle's status parameters and image data of the area where objects are placed outside the vehicle; The object type is determined based on the image data using a pre-trained image recognition model; The predicted attitude parameters of the vehicle are determined using a pre-trained attitude prediction model based on the state parameters and the object type. The system receives historical driving parameters of the vehicle over a preset time period, uses a pre-trained decision model to determine the target risk level of the vehicle's posture based on the predicted posture parameters, the historical driving parameters, and the object type, and controls the vehicle based on the target risk level.

[0006] Based on the same inventive concept, a second aspect of this disclosure proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0007] Based on the same inventive concept, a third aspect of this disclosure proposes a vehicle that includes the electronic equipment described in the second aspect.

[0008] As described above, the vehicle attitude control method, electronic device, and vehicle provided in this disclosure are as follows: The method acquires the vehicle's state parameters and image data of the area where an object is placed outside the vehicle. A pre-trained image recognition model is used to determine the object type based on the image data. A pre-trained attitude prediction model is used to determine the predicted attitude parameters of the vehicle based on the state parameters and the object type, taking into account the influence of the object type on the vehicle's attitude. Historical driving parameters of the vehicle over a preset time period are received. A pre-trained decision model is used to determine the target risk level of the vehicle attitude based on the predicted attitude parameters, historical driving parameters, and object type. The target risk level of the vehicle attitude comprehensively considers the predicted attitude parameters, historical driving parameters, and the object type of the object placed outside the vehicle, taking into account the influence of the object type on the risk of vehicle attitude instability. This makes the target risk level of the vehicle attitude more accurate. Controlling the vehicle based on the target risk level can reduce the risk of attitude instability caused by objects placed outside the vehicle. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in this disclosure or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a vehicle attitude control method according to an embodiment of the present disclosure; Figure 2 This is a flowchart of the vehicle load sensing and dynamic attitude control method according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram of the structure of the vehicle attitude control device according to an embodiment of the present disclosure; Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0012] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0013] Based on the background description, in actual off-road vehicle use, drivers often attach loads such as tents, buckets, and trailers to the roof or rear of the vehicle. These loads not only increase the overall vehicle weight but also significantly alter the vehicle's center of gravity and attitude response characteristics. Current technologies typically rely on fixed center of gravity models or manually set parameters, failing to dynamically identify the current load state based on varying load conditions, and unable to accurately predict attitude instability risks under load changes, especially during off-road maneuvers such as slopes, sharp turns, and side tilts, which pose a high risk of rollover. Furthermore, existing systems cannot adaptively adjust driving behavior restriction strategies based on real-time load and dynamic response, leading to off-road safety relying on human experience and posing significant risks. Therefore, there is an urgent need for an intelligent compensation mechanism that combines load identification with vehicle dynamic modeling to reduce the risk of attitude instability caused by load changes.

[0014] As mentioned above, how to accurately predict the instability risk of vehicle attitude has become an important research question.

[0015] Based on the above description, such as Figure 1 As shown, the vehicle attitude control method proposed in this embodiment includes: Step 101: Obtain the vehicle's status parameters and image data of the area where the object is placed outside the vehicle.

[0016] In practice, the vehicle's state parameters are the vehicle's driving state data at the current moment. These state parameters include at least one of the following: attitude angle, wheel speed, yaw rate, and suspension height. Specifically, the vehicle's attitude angle is collected using an Inertial Measurement Unit (IMU); the wheel speed is collected using wheel speed sensors; the yaw rate is collected using a yaw rate sensor; and the suspension height is collected using four-wheel suspension height sensors.

[0017] The area where objects are placed externally on the vehicle can be the roof area or the rear suspension area. By acquiring image data of this area, it's possible to accurately determine whether an object is placed there and precisely identify the type of object. Specifically, this is achieved using a visual camera with adjustable angles in all directions (front, rear, left, and right) to capture image data of the area where objects are placed externally.

[0018] In addition, a roof pressure sensor can be installed on the vehicle roof. The roof pressure sensor collects roof pressure parameters, and based on these parameters, it determines whether an object is placed in the designated area on the roof, and verifies the type of object detected.

[0019] The state parameters are preprocessed to obtain preprocessed state parameters. Preprocessing includes: timestamp synchronization, multi-sensor alignment, and missing data imputation. Image data is also preprocessed to obtain preprocessed image data. Preprocessing includes: image size standardization, image brightness normalization, and noise filtering. This ensures high consistency of various data types within the same time window.

[0020] Step 102: Determine the object type based on the image data using a pre-trained image recognition model.

[0021] In practice, the image recognition model is a pre-trained model used to identify the types of objects in image data. Specifically, the image data is input into the pre-trained image recognition model, which then uses the model to determine the object type and object position parameters from the image data.

[0022] Step 103: Determine the predicted attitude parameters of the vehicle based on the state parameters and the object type using a pre-trained attitude prediction model.

[0023] In practice, the attitude prediction model is a pre-trained model used to predict vehicle attitude parameters. Specifically, the state parameters and object type are input into the pre-trained image recognition model, and the pre-trained attitude prediction model is used to determine the vehicle's predicted attitude parameters based on the state parameters and object type. Thus, when determining the vehicle's predicted attitude parameters using the pre-trained attitude prediction model, both the vehicle's state parameters and the object type placed outside the vehicle are considered, taking into account the influence of the object type on the vehicle's attitude.

[0024] Step 104: Receive the vehicle's historical driving parameters over a preset time period, use a pre-trained decision model to determine the target risk level of the vehicle's posture based on the predicted posture parameters, the historical driving parameters, and the object type, and control the vehicle based on the target risk level.

[0025] In practice, historical driving parameters refer to the vehicle's driving parameters within a preset time period prior to the current moment. For example, if the preset time period is 5 seconds, the historical driving parameters are the vehicle's driving parameters within 5 seconds prior to the current moment. Historical driving parameters can be the vehicle's off-road maneuvers, including at least one of the following: uphill / downhill, side tilt, sharp turns, and obstacle crossing.

[0026] The decision model is pre-trained to predict the risk level of attitude instability when a vehicle is subjected to an externally placed object and operates according to historical driving parameters. Specifically, the predicted attitude parameters, historical driving parameters, and object type are input into the pre-trained decision model, which then determines the target risk level of the vehicle's attitude based on these parameters. Thus, when determining the target risk level of the vehicle's attitude using the pre-trained decision model, the predicted attitude parameters, historical driving parameters, and the object type placed externally are comprehensively considered, taking into account the impact of the object type on the risk of vehicle attitude instability.

[0027] Through the above embodiments, vehicle state parameters and image data of the area where objects are placed outside the vehicle are acquired. A pre-trained image recognition model is used to determine the object type based on the image data. A pre-trained attitude prediction model is used to determine the predicted attitude parameters of the vehicle based on the state parameters and object type, taking into account the impact of the object type on the vehicle's attitude. Historical driving parameters of the vehicle over a preset time period are received. A pre-trained decision model is used to determine the target risk level of the vehicle's attitude based on the predicted attitude parameters, historical driving parameters, and object type. The target risk level of the vehicle's attitude comprehensively considers the predicted attitude parameters, historical driving parameters, and the object type of the objects placed outside the vehicle, taking into account the impact of the object type on the risk of vehicle attitude instability. This makes the target risk level of the vehicle's attitude more accurate. Controlling the vehicle based on the target risk level can reduce the risk of attitude instability caused by objects placed outside the vehicle.

[0028] In some embodiments, step 102 includes: Step 1021: Use a pre-trained image recognition model to extract and process the image data to obtain object features, and determine key features from the object features.

[0029] In practice, the pre-trained image recognition model uses a multi-layer residual block stacking method, which has a powerful feature extraction capability and can stably identify object features in complex off-road environments such as occlusion, lighting changes, and foreign object interference.

[0030] The attention mechanism module in a pre-trained image recognition model can enhance the focus on local targets and identify key features from object features. For example, key features could be the edge features of a car roof tent or the geometric features of an irregularly shaped bucket.

[0031] Step 1022: Determine the bounding box of the object region based on the key features, and determine the object position parameters based on the bounding box.

[0032] In practice, the object features are divided into multiple grids, and the bounding box of the object region is determined based on these grids. The center coordinates and size parameters of the bounding box are determined using an object detection algorithm, and the object's position parameters are determined based on these center coordinates and size parameters.

[0033] Step 1023: Determine the object type corresponding to the key feature based on the pre-stored correspondence; wherein, the correspondence is the correspondence between the features and types of an object.

[0034] In practice, the pre-stored correspondences include the correspondences between the features and types of various objects. The object type corresponding to the key features is determined based on the pre-stored correspondences.

[0035] For example, the pre-stored correspondences include: tent edge features corresponding to tent types, and bucket geometric features corresponding to bucket types. When the key feature of an object is the tent edge feature, the object type is determined to be tent type. When the key feature of an object is the bucket geometric feature, the object type is determined to be bucket type.

[0036] In addition, the object features are divided into multiple grids, and the predicted type and probability distribution within each grid are determined. The predicted type with the highest probability distribution is then taken as the object type. For example, the object features are divided into three grids: the first grid is predicted as "clothing" with a probability distribution of 0.5, the second grid is predicted as "tent" with a probability distribution of 0.8, and the third grid is predicted as "headscarf" with a probability distribution of 0.2. In the above scenario, the predicted type of the second grid is taken as the object type, i.e., the object type is determined to be "tent".

[0037] The above scheme utilizes a pre-trained image recognition model to extract object features from image data, and identifies key features from these features, enabling accurate extraction of key object characteristics. Based on these key features, bounding boxes are determined for the object regions, and object position parameters are determined from the bounding boxes, ensuring accurate object position parameter identification. Finally, the object type corresponding to the key features is determined based on a pre-stored correspondence; where the correspondence is between object features and types. This allows for the rapid and accurate identification of object types placed outside vehicles.

[0038] In some embodiments, the pre-training process of the image recognition model includes: Step 102A: Obtain the sample image and the corresponding type label and location label of the sample image.

[0039] In practice, sample images and their corresponding type and location labels are acquired. The image recognition model is pre-trained based on the sample images and their corresponding type and location labels.

[0040] A label dataset is established using manual annotation and a semi-automatic calibration system. Sample images in the dataset are pre-associated with their corresponding type and location labels. The dataset includes various object combinations; for example, the type labels corresponding to sample images include at least one of the following: empty, rooftop tent, bucket, towed equipment, and combined load. This allows the trained image recognition model to generalize across multiple usage scenarios.

[0041] Step 102B: Input the sample image into the first neural network and use the first neural network to determine the prediction type and prediction location of the sample image.

[0042] In practice, the first neural network can be an improved residual network (ResNet), which is used as the backbone network of the image recognition model.

[0043] The sample image is input into the first neural network, and the first neural network is used to determine the prediction type and prediction location of the sample image.

[0044] Step 102C: Determine the type loss function by comparing the predicted type and the type label, and determine the location loss function by comparing the predicted location and the location label.

[0045] In practice, the type label is the actual type corresponding to the pre-labeled sample image. The type loss function is determined by comparing the predicted type and the type label. The type loss function reflects the deviation of the first neural network in predicting the object type.

[0046] Location labels are the actual locations corresponding to pre-labeled sample images. The location loss function is determined by comparing the predicted location with the location labels. The location loss function reflects the deviation of the first neural network in predicting the object's location.

[0047] Step 102D: Weight the type loss function and the location loss function to obtain the joint loss function.

[0048] In practice, the first weight corresponding to the type loss function and the second weight corresponding to the position loss function are retrieved from the pre-stored database. The type loss function and the position loss function are then weighted based on the first and second weights to obtain the joint loss function. In this way, the joint loss function can comprehensively reflect the deviation of the first neural network in predicting object type and object position, thus enabling faster and more accurate training of the first neural network based on the joint loss function.

[0049] Step 102E: Adjust the model parameters of the first neural network based on the joint loss function to obtain an image recognition model.

[0050] In practice, the model parameters of the first neural network are adjusted based on the joint loss function to obtain an updated recognition model, and the loss function of the updated recognition model is determined. If the loss function of the updated recognition model satisfies a preset convergence condition, the updated recognition model is used as the image recognition model. If the loss function of the updated recognition model does not satisfy the preset convergence condition, the model parameters of the updated recognition model are further adjusted to obtain the image recognition model.

[0051] During the training process of image recognition models, transfer learning strategies can be adopted to introduce pre-trained parameters of general object detection models to accelerate convergence. At the same time, fine-tuning can be performed based on image data of historical mass-produced models of car manufacturers to improve recognition accuracy.

[0052] During the inference phase, the image recognition model can dynamically adjust the input frame rate based on the current vehicle speed to optimize edge computing efficiency. The image recognition model is ultimately deployed in an in-vehicle high-performance artificial intelligence (AI) controller, working with middleware to interact with other sensor systems while maintaining high frame rate and low latency image recognition capabilities. This provides high-confidence load information input for subsequent dynamic response and motion constraint models.

[0053] The above method obtains sample images along with their corresponding type and location labels. These sample images are then input into a first neural network, which determines the predicted type and location of the sample images. A type loss function is determined by comparing the predicted type and type label, and a location loss function is determined by comparing the predicted location and location label. The type and location loss functions are then weighted to obtain a joint loss function. This joint loss function comprehensively reflects the bias of the first neural network in predicting object type and location, allowing for faster and more accurate training of the first neural network. The model parameters of the first neural network are then adjusted based on the joint loss function to obtain the image recognition model.

[0054] In some embodiments, step 103 includes: Step 1031: Convert the object type into a type parameter vector and convert the state parameter into a state time sequence vector of a preset duration.

[0055] In practice, the object type is converted into a type parameter vector, and the state parameters are converted into a state time sequence vector of a preset duration. By converting both the object type and the state parameters into vectors, the attitude prediction model can process the type parameter vector and the state time sequence vector.

[0056] Step 1032: The type parameter vector and the state time sequence vector are concatenated to obtain a fused vector.

[0057] In practice, to facilitate the processing of type parameter vectors and state time sequence vectors by the attitude prediction model, the type parameter vectors and state time sequence vectors are concatenated to obtain a fusion vector.

[0058] Step 1033: Using the pre-trained attitude prediction model, determine the predicted pitch angle and predicted roll angle of the vehicle based on the fusion vector.

[0059] In practice, in order to accurately model the vehicle attitude parameters under different object types, a pre-trained attitude prediction model is used to determine the vehicle's predicted pitch angle and predicted roll angle based on the fusion vector.

[0060] The attitude prediction model includes multi-layer stacked GRU units and combines type parameter vectors and state temporal vectors as fusion vectors, which can improve the accuracy of the attitude prediction model in modeling complex nonlinear dynamic features.

[0061] Meanwhile, the pose prediction model incorporates Dropout and LayerNorm mechanisms to enhance robustness and prevent overfitting. Dropout is a regularization method designed to reduce overfitting in neural networks. By randomly discarding a portion of neurons and their connections, it retains only some activation paths in each training iteration, allowing the remaining neurons to function more independently rather than relying on specific combination patterns. LayerNorm's robustness stems from its dynamic normalization mechanism, effectively adapting to changes in sequence features by normalizing each sample using a Z-score in the feature dimension.

[0062] Step 1034: Determine the stability risk score based on the predicted pitch angle and the predicted roll angle.

[0063] In practice, the attitude prediction model can identify unstable dynamic trends caused by sudden changes in object type, road condition disturbances, or improper operation, and determine a stability risk score based on the predicted pitch angle and predicted roll angle, which is used to guide the subsequent decision-making model to determine whether historical driving parameters should be restricted.

[0064] Step 1035: Use the predicted pitch angle, the predicted roll angle, and the stability risk score as the predicted attitude parameters of the vehicle.

[0065] In practice, the predicted attitude parameters include: predicted pitch angle, predicted roll angle, and stability risk score. This allows for a comprehensive and accurate prediction of the vehicle's attitude parameters.

[0066] The above scheme converts object types into type parameter vectors and state parameters into state time-series vectors of preset duration. The type parameter vectors and state time-series vectors are concatenated to obtain a fusion vector, facilitating processing of these vectors by the attitude prediction model. Using a pre-trained attitude prediction model, the predicted pitch and roll angles of the vehicle are determined based on the fusion vector. A stability risk score is then determined based on the predicted pitch and roll angles. Using the predicted pitch, roll, and stability risk score as the vehicle's predicted attitude parameters enables comprehensive and accurate prediction of the vehicle's attitude parameters.

[0067] In some embodiments, the pre-training process of the pose prediction model includes: Step 103A: Obtain the sample object type and sample state parameters, and collect the actual attitude angle of the vehicle.

[0068] In practice, the sample object type can be the type label used when training the image recognition model. The sample object type includes at least one of the following: unloaded, rooftop tent, bucket, towed equipment, and combined load.

[0069] The sample state parameters are time-series data segments collected under multiple off-road action scenarios. The sample state data within each time window includes sensor data within the time before and after the current moment, and the number of samples is increased and the data coverage is expanded by sliding the window.

[0070] The actual attitude angle of the vehicle is collected. After the second neural network outputs the predicted attitude angle, the first loss function is determined by comparing the predicted attitude angle with the actual attitude angle.

[0071] Step 103B: Input the sample object type and the sample state parameters into the second neural network, and use the second neural network to determine the predicted posture angle and the predicted risk score.

[0072] In practice, the second neural network can be a Gated Recurrent Unit (GRU), which serves as the backbone network of the attitude prediction model. The GRU possesses efficient temporal memory capabilities, enabling it to learn vehicle attitude parameters for different object types during training.

[0073] The sample object type and sample state parameters are input into the second neural network, which is then used to determine the predicted pose angle and the predicted risk score.

[0074] Step 103C: The root mean square error between the predicted attitude angle and the actual attitude angle is used as the first loss function. A second loss function is determined based on the predicted risk score. The first loss function and the second loss function are weighted to obtain a joint loss function.

[0075] In practice, the root mean square error between the predicted attitude angle and the actual attitude angle is used as the first loss function. The first loss function can reflect the deviation of the second neural network in predicting the vehicle attitude angle. The training objective of the first loss function is to minimize the error between the predicted attitude angle and the actual attitude angle.

[0076] The actual risk score is obtained, and the second loss function is determined by comparing the predicted risk score and the actual risk score. The second loss function can reflect the deviation of the second neural network in predicting the risk score.

[0077] The first weight corresponding to the first loss function and the second weight corresponding to the second loss function are retrieved from pre-stored data. The first and second loss functions are then weighted together to obtain a joint loss function. This joint loss function comprehensively reflects the deviation of the second neural network in predicting pose angles and risk scores, thus enabling faster and more accurate training of the second neural network.

[0078] Step 103D: Adjust the model parameters of the second neural network based on the joint loss function to obtain the pose prediction model.

[0079] In practice, the model parameters of the second neural network are adjusted based on the joint loss function to obtain an updated prediction model, and the loss function of the updated prediction model is determined. If the loss function of the updated prediction model satisfies the preset convergence condition, the updated prediction model is used as the pose prediction model. If the loss function of the updated prediction model does not satisfy the preset convergence condition, the model parameters of the updated prediction model are further adjusted to obtain the pose prediction model.

[0080] The attitude prediction model supports fine-tuning through online incremental learning to adapt to different driving styles and driver habits. Deployed in the intermediate layer computing module of the vehicle controller, the attitude prediction model has real-time computing capabilities, ensuring that attitude trend prediction is completed at the millisecond level during vehicle operation, thereby providing timely warnings to the decision-making model.

[0081] The above scheme obtains the sample object type and sample state parameters, and collects the actual attitude angle of the vehicle. The sample object type and sample state parameters are input into a second neural network, which determines the predicted attitude angle and predicted risk score. The root mean square error between the predicted and actual attitude angles is used as the first loss function. A second loss function is determined based on the predicted risk score, and the first and second loss functions are weighted to obtain a joint loss function. This joint loss function comprehensively reflects the deviation of the second neural network in predicting attitude angles and risk scores, thus enabling faster and more accurate training of the second neural network. The model parameters of the second neural network are adjusted based on the joint loss function to obtain the attitude prediction model.

[0082] In some embodiments, step 104 includes: Step 1041: Using a pre-trained decision model, determine the probability distribution of the vehicle attitude risk level based on the predicted attitude parameters, the historical driving parameters, and the object type; wherein the probability distribution is a probability distribution corresponding to multiple risk levels.

[0083] In practice, the decision model is used to determine whether a vehicle is allowed to perform a high-risk action based on the object type and predicted attitude parameters. The decision model includes a state space and an action space.

[0084] Specifically, predicted attitude parameters, historical driving parameters, and object type are input into a pre-trained decision model. The state space receives the predicted attitude parameters, historical driving parameters, and object type. Using the pre-trained decision model, the action space determines the probability distribution of the vehicle attitude risk level based on the predicted attitude parameters, historical driving parameters, and object type. The probability distribution consists of multiple probability distributions corresponding to different risk levels.

[0085] For example, the probability distribution for low-risk level output by the decision model is 0.2, the probability distribution for medium-risk level is 0.5, and the probability distribution for high-risk level is 0.8.

[0086] Step 1042: Determine the target risk level with the highest probability distribution from the multiple risk levels, determine the target control method corresponding to the target risk level, and control the vehicle based on the target control method.

[0087] In practice, the action space determines the target risk level with the highest probability distribution from multiple risk levels. For example, if the probability distribution corresponding to the low risk level is 0.2, the probability distribution corresponding to the medium risk level is 0.5, and the probability distribution corresponding to the high risk level is 0.8, then the target risk level is determined to be the high risk level.

[0088] Different risk levels correspond to different vehicle control methods. Specifically, a target control method is determined for the target risk level, and the vehicle is controlled based on this target control method. For example, a low risk level corresponds to the first control method, a medium risk level corresponds to the second control method, and a high risk level corresponds to the third control method. When the target risk level is high, the target control method is determined to be the third control method, and the vehicle is controlled based on this third control method.

[0089] The above scheme utilizes a pre-trained decision-making model to determine the probability distribution of vehicle attitude risk levels based on predicted attitude parameters, historical driving parameters, and object types. This probability distribution comprises multiple risk levels. The target risk level with the highest probability distribution is then determined from these risk levels. The corresponding target control method is then identified, and the vehicle is controlled based on this target control method. This approach accurately predicts the target risk level at which the vehicle will execute historical driving parameters under predicted attitude parameters and the type of object placed outside the vehicle. Based on this target risk level, it can then determine whether to restrict historical driving parameters.

[0090] In some embodiments, step 1042 includes: Step 10421: In response to determining that the target risk level is low risk level, the target control mode is determined to be prompt control mode, and prompt information indicating that there is a risk in the vehicle posture is displayed on the in-vehicle screen.

[0091] In practice, the decision-making model pre-sets allowance strategies, warning strategies, and restriction strategies. Each strategy corresponds to a different level of target control.

[0092] When the target risk level is low, the permission policy is triggered, allowing the execution of historical driving parameters. In the above scenario, the target control method is determined to be a prompt control method, and a prompt message indicating a risk to the vehicle's posture is displayed on the in-vehicle screen.

[0093] In some scenarios, if the vehicle's historical driving parameters indicate a sharp turn, and the decision model determines that the vehicle is at a low risk level when making a sharp turn under the object type and predicted attitude parameters, then the vehicle is allowed to make the sharp turn.

[0094] Step 10422: In response to determining that the target risk level is a medium risk level, the target control mode is determined to be an early warning control mode, and an early warning message indicating that there is a risk in the vehicle's posture is given through a buzzer.

[0095] In practice, when the target risk level is medium risk, an early warning strategy is triggered, and a warning is issued based on historical driving parameters. In the above scenario, the target control method is determined to be the early warning control method, and a buzzer is used to issue a warning indicating that there is a risk to the vehicle's posture.

[0096] In some scenarios, if the vehicle's historical driving parameters indicate a sharp turn, and the decision model determines that the vehicle is at a medium risk level when making a sharp turn under the object type and predicted attitude parameters, then an early warning will be issued for the vehicle making a sharp turn.

[0097] Step 10423: In response to determining that the target risk level is a high risk level, the target control mode is determined to be a restriction control mode. An alarm is sounded by a buzzer indicating that there is a risk in the vehicle's posture, and the historical driving parameters are restricted.

[0098] In practice, when the target risk level is high, a restriction strategy is triggered to limit the execution of historical driving parameters. In the above scenario, the target control method is determined to be a restriction control method. An alarm is triggered via a buzzer to indicate a risk to the vehicle's posture, and historical driving parameters are restricted.

[0099] In some scenarios, if the vehicle's historical driving parameters indicate a sharp turn, and the decision model determines that the vehicle is at a high risk level when making a sharp turn under the object type and predicted attitude parameters, then the vehicle is restricted from making sharp turns.

[0100] The above scheme determines the control method as follows: When the target risk level is low, the control method is a prompt control method, displaying a risk warning on the vehicle screen. Historical driving parameters are allowed to be executed at low risk levels. When the target risk level is medium, the control method is a warning control method, using a buzzer to warn of a risk to the vehicle's posture. Historical driving parameters are also warn of a risk at medium risk levels. When the target risk level is high, the control method is a restriction control method, using a buzzer to warn of a risk to the vehicle's posture and restricting historical driving parameters. Historical driving parameters are restricted at high risk levels.

[0101] In some embodiments, the pre-training process of the decision model includes: Step 104A: Obtain sample posture parameters, sample driving parameters, and sample object type.

[0102] In practice, sample attitude parameters can include various sample attitude angles and sample stability risk levels.

[0103] The sample driving parameters can be a variety of sample off-road actions. The sample driving parameters include at least one of the following: uphill / downhill, side roll, sharp turn, and obstacle crossing.

[0104] The sample object type can be a type label used when training the image recognition model, wherein the sample object type includes at least one of the following: unloaded, rooftop tent, bucket, towed equipment, and combined load.

[0105] Step 104B: Input the sample posture parameters, the sample driving parameters, and the sample object type into the third neural network.

[0106] In practice, the third neural network can be a reinforcement learning (RL) model, which serves as the backbone network of the decision-making model.

[0107] Step 104C: Determine the predicted risk level using the policy network in the third neural network, and determine the reward function corresponding to the predicted risk level using the value network in the third neural network.

[0108] In practice, the sample posture parameters, sample driving parameters, and sample object type are input into the third neural network. This third neural network includes a policy network and a value network. The policy network within the third neural network is used to determine the predicted risk level, and the value network within the third neural network is used to determine the reward function corresponding to the predicted risk level.

[0109] The reward function is based on vehicle attitude stability, while also incorporating object center offset, dynamic attitude change rate, and historical action success rate as reference indicators.

[0110] Step 104D: Adjust the model parameters of the policy network based on the reward function to obtain the decision model.

[0111] In practice, the model parameters of the policy network in the third neural network are adjusted based on the reward function to obtain an updated policy model, and the loss function of the updated policy model is determined. If the loss function of the updated policy model satisfies the preset convergence condition, the updated policy model is used as the decision model. If the loss function of the updated policy model does not satisfy the preset convergence condition, the model parameters of the updated policy model are further adjusted to obtain the decision model.

[0112] The decision-making model is trained using a Deep Q-Network (DQN) structure from deep reinforcement learning, which combines a policy network and a value network to evaluate the target risk level.

[0113] During the training of the decision-making model, a hybrid training approach combining simulation and real-vehicle testing is employed. Initial strategy training is first completed on the automaker's in-house off-road simulation platform, followed by fine-tuning using real-world road sampling data to ensure the model's high real-world adaptability. The decision model outputs decision commands, which, in conjunction with the vehicle controller, trigger attitude compensation or action restriction logic. It also supports a dynamic strategy update mechanism, allowing for strategy shifts based on different driver styles and usage frequencies. The RL model also features an interpretable output mechanism, recording the logic chains that trigger action restrictions for later backtracking analysis and over-the-air (OTA) strategy optimization. The entire RL module operates at the control system's decision-making layer, exhibiting high real-time performance and stability. This ensures personalized dynamic action control strategies are implemented in complex off-road scenarios, minimizing the risk of attitude instability or rollover caused by externally placed objects.

[0114] The above method obtains sample posture parameters, sample driving parameters, and sample object types. These parameters are then input into a third neural network. The policy network within the third neural network determines the predicted risk level, and the value network determines the reward function corresponding to that risk level. This makes the reward function more accurate, allowing for faster and more precise training of the third neural network. The decision model is then obtained by adjusting the model parameters of the policy network based on the reward function.

[0115] In some embodiments, to achieve efficient collaboration between the image recognition model, the pose prediction model, and the decision model, a multi-model fusion strategy and joint training mechanism are used to ensure an optimal balance in information transmission, feature sharing, and decision consistency among the multiple models. The fusion strategy employs a multi-stage cascaded structure. First, the object type and position parameters output by the image recognition model are used as static environment parameters. These static environment parameters are injected into the time-series input structure of the pose prediction model in the form of encoded vectors, enabling the pose prediction model to recognize the current object type during dynamic response modeling, thereby enhancing the modeling accuracy of the pose change trend. The pose prediction parameters output by the pose prediction model are further input into the decision model as state parameters, assisting the decision model in predicting the risk of the current action and selecting limiting strategies. During model fusion, feature vector transmission is the primary method, without direct parameter fusion, ensuring that each model can be updated independently without interference, facilitating vehicle deployment and subsequent maintenance. In terms of the training mechanism, a staged joint training approach is adopted. First, the image recognition model and the pose prediction model are trained separately on a large-scale labeled dataset to achieve basic recognition and prediction capabilities. Then, an intermediate data channel is constructed to combine the outputs of the two models into a state vector for the decision model to learn strategies. To overcome the fragmentation of training processes between models, a joint loss function is implemented, comprising a recognition accuracy loss function, a pose prediction error loss function, and a policy reward maximization objective function. These three are combined with weights to form a joint training objective. During policy training, the weight ratios of each model are dynamically adjusted to achieve overall system synergistic optimization. Furthermore, in the reinforcement learning phase, a model self-feedback mechanism is introduced. After policy execution, the decision-making model feeds back the results to the pose prediction model, marking whether the predicted pose parameters are of a high-risk level. This guides the pose prediction model to perform weighted learning on high-risk pose parameters during training, improving its sensitivity to these parameters. The entire training process is executed in parallel on a high-performance distributed server, and a data replay buffer is used to prioritize training samples, improving training efficiency and convergence speed. The key to model fusion lies in feature dimension alignment and time window synchronization mechanisms. In each input, the temporal correspondence between image data, state parameters, and historical driving parameters is ensured, avoiding semantic shifts between models. This achieves a truly collaborative decision-making system, providing a powerful algorithmic foundation for vehicle pose compensation and motion restriction strategies.

[0116] In some embodiments, to ensure the stability and high accuracy of the fusion model in real-world in-vehicle environments, this disclosure constructs a systematic model tuning and accuracy optimization process, employing multi-dimensional tuning mechanisms for the image recognition model, pose prediction model, and decision model. Key tuning parameters for the image recognition model include: the number of residual blocks, feature channel width, convolution kernel size, and the activation position of the attention mechanism. Automated Machine Learning (AutoML) is used to optimize the structure on labeled image datasets, ultimately obtaining a model structure that balances performance and accuracy under edge computing conditions. Image enhancement parameters (e.g., brightness normalization amplitude, affine transformation range, random occlusion probability) are also incorporated into the optimization space to improve the generalization ability of the image recognition model in harsh off-road environments. Core tuning parameters for the pose prediction model include: the number of GRU unit layers, hidden state dimension, time window length, and learning rate strategy. To improve the robustness of pose prediction, a Bayesian optimization algorithm is introduced to automatically adjust the time step size and dropout rate of the GRU. After multiple validations on the test set, the optimal structural combination is selected for deployment. Simultaneously, feature clustering is performed on different driving behavior modes, and corresponding weights are trained for each. During vehicle operation, personalized model weights are loaded based on driver labels to achieve dynamic model adaptation. Parameter tuning of the decision model focuses on the weight design of the reward function and the optimization of the convergence speed of the policy network. By introducing an automatic adjustment mechanism for the attitude instability penalty coefficient, action coherence reward factor, and safety boundary weight ratio, the system can effectively improve the decision stability of the RL policy under high-risk action boundary conditions. Throughout the parameter tuning process, a model accuracy tracking module and an error backtracking mechanism are integrated. Recognition deviations, attitude prediction errors, and action selection results are logged for each decision cycle, and an accuracy evolution curve is generated for parameter tuning feedback. Furthermore, in the simulation testing phase before system deployment, a virtual test bench is used to simulate complex off-road routes and various load combinations. The system automatically retrains sample segments with high misclassification rates, effectively reducing the recognition error rate and action miscontrol rate. Through this complete parameter tuning and accuracy optimization system, the robustness and accuracy of each sub-model under different off-road scenarios are ultimately unified, laying a solid foundation for actual vehicle deployment.

[0117] In some embodiments, a highly integrated online inference architecture enables real-time model inference and attitude compensation control during vehicle operation, maintaining millisecond-level response time throughout the entire process from data acquisition to action limitation decision-making. Cameras capture real-time image data of the roof and external mounting areas. This image data first enters an image recognition model for object type identification. The object type is immediately encoded as a load vector and input to the attitude prediction module, triggering the acquisition of pressure sensor and IMU data in parallel. All sensor data and load vectors together form a unified state sequence, which is input to the attitude prediction model for dynamic response trend prediction. The output predicted attitude parameters are input to the decision model, which performs a risk assessment by combining recent historical driving parameters and current action intent. The decision model outputs a decision on whether the current off-road action is permitted and sends it to the vehicle's underlying control unit via the Controller Area Network (CAN) bus, triggering corresponding attitude compensation strategies or action limitation logic. If there is a risk of attitude instability under the current object type conditions, such as a rooftop tent combined with a large-angle turn on a side slope, the system will send an action limitation command. Simultaneously, the instrument panel displays a risk warning and triggers compensation measures such as chassis speed limiting, steer-by-wire intervention, or active suspension height adjustment. To ensure stability and reliability, a data integrity detection mechanism is introduced during the inference process. When critical sensor data is missing, a redundant estimation module is activated to ensure process continuity. Furthermore, the online inference module uses an asynchronous pipeline architecture to divide image processing, sensor reading, and model computation into parallel sub-threads, significantly improving overall throughput efficiency. The system supports a configurable threshold adaptive mechanism, automatically adjusting the sensitivity of action limits based on historical successful / failed off-road action samples to accommodate different terrains and user driving preferences. Throughout operation, the system maintains a high-frequency inference cycle and low-latency control response, ensuring that the entire process from load recognition to action decision-making does not exceed 100 milliseconds. This effectively enables rapid perception and timely response to sudden attitude changes under complex off-road conditions, thereby significantly improving the vehicle's dynamic safety performance.

[0118] Through the above embodiments, vehicle state parameters and image data of the area where objects are placed outside the vehicle are acquired. A pre-trained image recognition model is used to determine the object type based on the image data. A pre-trained attitude prediction model is used to determine the predicted attitude parameters of the vehicle based on the state parameters and object type, taking into account the impact of the object type on the vehicle's attitude. Historical driving parameters of the vehicle over a preset time period are received. A pre-trained decision model is used to determine the target risk level of the vehicle's attitude based on the predicted attitude parameters, historical driving parameters, and object type. The target risk level of the vehicle's attitude comprehensively considers the predicted attitude parameters, historical driving parameters, and the object type of the objects placed outside the vehicle, taking into account the impact of the object type on the risk of vehicle attitude instability. This makes the target risk level of the vehicle's attitude more accurate. Controlling the vehicle based on the target risk level can reduce the risk of attitude instability caused by objects placed outside the vehicle.

[0119] It should be noted that the embodiments of this disclosure can also be further described in the following ways: Figure 2 This is a flowchart of a vehicle load sensing and dynamic attitude control method according to an embodiment of this disclosure. Figure 2 As shown, the vehicle load perception and dynamic attitude control method includes: Step 1: Vehicle data acquisition and preprocessing.

[0120] The sensors include: a vision camera with adjustable front, rear, left, and right angles; an inertial measurement unit (IMU); wheel speed sensors; yaw rate sensors; four-wheel suspension height sensors; and a roof pressure sensor module. The vision camera collects image data of the roof and rear of the vehicle; the IMU collects the vehicle's attitude angles; the wheel speed sensors collect wheel speeds; the yaw rate sensors collect yaw rate; the four-wheel suspension height sensors collect suspension heights; and the roof pressure sensors collect roof pressure parameters.

[0121] Step 2: Load the image recognition model _ResNet.

[0122] An improved residual network (ResNet) serves as the backbone structure of the image recognition model. It utilizes a pre-trained image recognition model to determine object types based on image data. The model employs a multi-layered stacked residual block approach, possessing powerful feature extraction capabilities and enabling stable object feature recognition even in complex off-road environments with occlusion, lighting variations, and foreign object interference.

[0123] Step 3, Dynamic Response Modeling - GRU.

[0124] To accurately model the predicted attitude parameters of the vehicle under different object types, a gated recurrent unit (GRU) is introduced as the core structure of the attitude prediction model. The pre-trained attitude prediction model is used to determine the predicted attitude parameters of the vehicle based on the state parameters and object type.

[0125] Step 4, Dynamic Adaptive Decision Making (RL).

[0126] The decision model is used to assess whether a specific high-risk action is permissible based on object type and predicted attitude parameters. A pre-trained decision model determines the target risk level of the vehicle's attitude based on predicted attitude parameters, historical driving parameters, and object type, and then controls the vehicle based on this target risk level.

[0127] Step 5: Multi-model fusion and joint training.

[0128] To achieve efficient collaboration among image recognition models, pose prediction models, and decision-making models, a multi-model fusion strategy and joint training mechanism are used to ensure that multiple models achieve the optimal balance in information transmission, feature sharing, and decision consistency.

[0129] Step 6: Model tuning and accuracy optimization.

[0130] To ensure the stability and high accuracy of the fusion model in a real-world vehicle environment, this disclosure presents a systematic model tuning and accuracy optimization process, employing multi-dimensional tuning mechanisms for the image recognition model, attitude prediction model, and decision model.

[0131] Step 7: Online inference and real-time compensation control.

[0132] During vehicle operation, a highly integrated online inference architecture enables real-time model inference and attitude compensation control, maintaining millisecond-level response time throughout the entire process from data acquisition to action constraint decision-making.

[0133] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0134] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0135] Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides a vehicle attitude control device.

[0136] refer to Figure 3 The vehicle attitude control device includes: The acquisition module 301 is configured to acquire the vehicle's status parameters and acquire image data of the area where objects are placed outside the vehicle. The physical type determination module 302 is configured to determine the object type based on the image data using a pre-trained image recognition model; The predicted attitude parameter determination module 303 is configured to determine the predicted attitude parameters of the vehicle based on the state parameters and the object type using a pre-trained attitude prediction model. The control module 304 is configured to receive historical driving parameters of the vehicle over a preset time period, determine the target risk level of the vehicle's posture based on the predicted posture parameters, the historical driving parameters, and the object type using a pre-trained decision model, and control the vehicle based on the target risk level.

[0137] In some embodiments, the physical type determination module 302 includes: The feature extraction unit is configured to extract object features from the image data using a pre-trained image recognition model, and to determine key features from the object features. The object position parameter determination unit is configured to determine the bounding box of the object region based on the key features, and to determine the object position parameters based on the bounding box. The object type determination unit is configured to determine the object type corresponding to the key feature based on a pre-stored correspondence; wherein the correspondence is the correspondence between the object's features and its type.

[0138] In some embodiments, the apparatus further includes a first training module, the first training module comprising: The first sample acquisition unit is configured to acquire a sample image and the type label and location label corresponding to the sample image; The first prediction unit is configured to input the sample image into a first neural network and use the first neural network to determine the prediction type and prediction location of the sample image. The comparison processing unit is configured to determine a type loss function by comparing the predicted type and the type label, and to determine a location loss function by comparing the predicted location and the location label. The weighted processing unit is configured to perform weighted processing on the type loss function and the location loss function to obtain a joint loss function; The image recognition model training unit is configured to adjust the model parameters of the first neural network based on the joint loss function to obtain an image recognition model.

[0139] In some embodiments, the predicted attitude parameter determination module 303 includes: The conversion unit is configured to convert the object type into a type parameter vector and the state parameter into a state timing vector of a preset duration; The splicing processing unit is configured to splice the type parameter vector and the state time sequence vector to obtain a fusion vector; The attitude angle prediction unit is configured to use a pre-trained attitude prediction model to determine the vehicle's predicted pitch angle and predicted roll angle based on the fusion vector. A stability risk score determination unit is configured to determine a stability risk score based on the predicted pitch angle and the predicted roll angle. The predicted attitude parameter determination unit is configured to use the predicted pitch angle, the predicted roll angle, and the stability risk score as the predicted attitude parameters of the vehicle.

[0140] In some embodiments, the apparatus further includes a second training module, the second training module comprising: The second sample acquisition unit is configured to acquire the sample object type and sample state parameters, and to collect the actual attitude angle of the vehicle. The second prediction unit is configured to input the sample object type and the sample state parameters into a second neural network, and use the second neural network to determine the predicted posture angle and the predicted risk score. The joint loss function determination unit is configured to use the root mean square error between the predicted attitude angle and the actual attitude angle as a first loss function, determine a second loss function based on the predicted risk score, and perform weighted processing on the first loss function and the second loss function to obtain a joint loss function. The pose prediction model training unit is configured to adjust the model parameters of the second neural network based on the joint loss function to obtain the pose prediction model.

[0141] In some embodiments, the control module 304 includes: The probability distribution determination unit is configured to use a pre-trained decision model to determine the probability distribution of the vehicle posture risk level based on the predicted posture parameters, the historical driving parameters, and the object type; wherein the probability distribution is a probability distribution corresponding to multiple risk levels respectively; The control unit is configured to determine the target risk level with the highest probability distribution from the plurality of risk levels, determine the target control mode corresponding to the target risk level, and control the vehicle based on the target control mode.

[0142] In some embodiments, the control unit includes: The first control subunit is configured to, in response to determining that the target risk level is low risk level, determine the target control mode as prompt control mode and display prompt information indicating that there is a risk to the vehicle posture on the in-vehicle screen; The second control subunit is configured to, in response to determining that the target risk level is a medium risk level, determine that the target control mode is a warning control mode, and issue a warning through a buzzer to indicate that there is a risk in the vehicle's posture. The third control subunit is configured to, in response to determining that the target risk level is high risk, determine that the target control mode is a restrictive control mode, issue an alarm via a buzzer indicating a risk to the vehicle's posture, and restrict the historical driving parameters.

[0143] In some embodiments, the apparatus further includes a third training module, the third training module comprising: The third sample acquisition unit is configured to acquire sample posture parameters, sample driving parameters, and sample object type; The input unit is configured to input the sample posture parameters, the sample driving parameters, and the sample object type into a third neural network; The reward function determination unit is configured to determine the predicted risk level using the policy network in the third neural network, and to determine the reward function corresponding to the predicted risk level using the value network in the third neural network. The decision model training unit is configured to adjust the model parameters of the policy network based on the reward function to obtain a decision model.

[0144] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0145] The apparatus of the above embodiments is used to implement the corresponding vehicle attitude control method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0146] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the vehicle attitude control method described in any of the above embodiments.

[0147] Figure 4 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0148] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), 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 specification.

[0149] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0150] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0151] The communication interface 1040 is used to connect the communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB (Universal Serial Bus), network cable, etc.) or wireless means (such as mobile network, WIFI (Wireless Fidelity), Bluetooth, etc.).

[0152] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0153] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0154] The electronic devices described above are used to implement the corresponding vehicle attitude control methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0155] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform vehicle attitude control as described in any of the above embodiments. The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0156] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the vehicle attitude control method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0157] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a vehicle, including the vehicle posture control device, or electronic device, or storage medium in the above embodiments, wherein the vehicle device implements the vehicle posture control method described in any of the above embodiments.

[0158] The vehicle described in the above embodiments is used to implement the vehicle attitude control method described in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0159] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a computer program product, including computer program instructions. When the computer program instructions are run on a computer, the computer causes the computer to execute the vehicle posture control method as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0160] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0161] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0162] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0163] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0164] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0165] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0166] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0167] This disclosure is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this disclosure. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the protection scope of this disclosure.

Claims

1. A method for controlling vehicle attitude, characterized in that, The method includes: Obtain the vehicle's status parameters and image data of the area where objects are placed outside the vehicle; The object type is determined based on the image data using a pre-trained image recognition model; The predicted attitude parameters of the vehicle are determined using a pre-trained attitude prediction model based on the state parameters and the object type. The system receives historical driving parameters of the vehicle over a preset time period, uses a pre-trained decision model to determine the target risk level of the vehicle's posture based on the predicted posture parameters, the historical driving parameters, and the object type, and controls the vehicle based on the target risk level.

2. The method according to claim 1, characterized in that, The step of determining the object type based on the image data using a pre-trained image recognition model includes: The image data is processed using a pre-trained image recognition model to obtain object features, and key features are determined from the object features. Based on the key features, the bounding box of the object region is determined, and the object position parameters are determined based on the bounding box. The object type corresponding to the key feature is determined based on the pre-stored correspondence; wherein, the correspondence is the correspondence between the features and types of the object.

3. The method according to claim 1, characterized in that, The pre-training process of the image recognition model includes: Obtain the sample image and the corresponding type label and location label of the sample image; The sample image is input into a first neural network, and the first neural network is used to determine the prediction type and prediction location of the sample image; The type loss function is determined by comparing the predicted type and the type label, and the location loss function is determined by comparing the predicted location and the location label. The joint loss function is obtained by weighting the type loss function and the location loss function. The image recognition model is obtained by adjusting the model parameters of the first neural network based on the joint loss function.

4. The method according to claim 1, characterized in that, The step of determining the predicted attitude parameters of the vehicle using a pre-trained attitude prediction model based on the state parameters and the object type includes: The object type is converted into a type parameter vector, and the state parameter is converted into a state time sequence vector of a preset duration. The type parameter vector and the state time sequence vector are concatenated to obtain a fusion vector; Using a pre-trained attitude prediction model, the predicted pitch angle and predicted roll angle of the vehicle are determined based on the fusion vector; A stability risk score is determined based on the predicted pitch angle and the predicted roll angle. The predicted pitch angle, the predicted roll angle, and the stability risk score are used as the vehicle's predicted attitude parameters.

5. The method according to claim 1, characterized in that, The pre-training process of the pose prediction model includes: Obtain the sample object type and sample state parameters, and collect the actual attitude angle of the vehicle; The sample object type and the sample state parameters are input into the second neural network, and the second neural network is used to determine the predicted posture angle and the predicted risk score. The root mean square error between the predicted attitude angle and the actual attitude angle is used as the first loss function. A second loss function is determined based on the predicted risk score. The first loss function and the second loss function are weighted to obtain a joint loss function. The pose prediction model is obtained by adjusting the model parameters of the second neural network based on the joint loss function.

6. The method according to claim 1, characterized in that, The step of determining the target risk level of the vehicle attitude using a pre-trained decision model based on the predicted attitude parameters, the historical driving parameters, and the object type, and controlling the vehicle based on the target risk level, includes: Using a pre-trained decision model, a probability distribution of the vehicle attitude risk level is determined based on the predicted attitude parameters, the historical driving parameters, and the object type; wherein, the probability distribution is a probability distribution corresponding to multiple risk levels respectively; The target risk level with the highest probability distribution is determined from the multiple risk levels, the target control method corresponding to the target risk level is determined, and the vehicle is controlled based on the target control method.

7. The method according to claim 6, characterized in that, The step of determining the target control method corresponding to the target risk level and controlling the vehicle based on the target control method includes: In response to determining that the target risk level is low risk, the target control mode is determined to be a prompt control mode, and a prompt message indicating that there is a risk in the vehicle's posture is displayed on the in-vehicle screen; In response to determining that the target risk level is a medium risk level, the target control mode is determined to be an early warning control mode, and an early warning is given by means of a buzzer indicating that there is a risk in the vehicle's posture; In response to determining that the target risk level is a high risk level, the target control mode is determined to be a restrictive control mode. An alarm is sounded to indicate that there is a risk in the vehicle's posture, and the historical driving parameters are restricted.

8. The method according to claim 1, characterized in that, The pre-training process of the decision model includes: Obtain sample posture parameters, sample driving parameters, and sample object type; The sample posture parameters, sample driving parameters, and sample object type are input into the third neural network; The predicted risk level is determined using the policy network in the third neural network, and the reward function corresponding to the predicted risk level is determined using the value network in the third neural network. The decision model is obtained by adjusting the model parameters of the policy network based on the reward function.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1 to 8.

10. A vehicle, characterized in that, Includes the electronic device as described in claim 9.

Images