Vehicle transverse speed prediction method and system, computer equipment and vehicle

Abstract

The embodiment of the invention relates to the technical field of automatic driving and vehicle positioning, in particular to a vehicle transverse speed prediction method and system, computer equipment and a vehicle. The method mainly comprises the following steps: collecting a motion sensor data set of a vehicle in a preset time period; the motion sensor dataset includes at least an angular velocity, an acceleration, and a wheel speed of the vehicle. And inputting the motion sensor data set into a pre-trained transverse speed prediction model to obtain the transverse speed of the vehicle at the current moment. According to the method, the transverse speed is directly predicted through the data of the motion sensor by adopting the mapping method based on the pre-training model, so that the dependence of a traditional tire model-based method on numerous physical parameters which are difficult to accurately calibrate or estimate on line is effectively avoided, and the robustness and practicability of transverse speed estimation in the actual operation of the vehicle are remarkably improved.

Description

Technical Field

[0001] This application relates to the fields of autonomous driving and vehicle positioning technology, and in particular to a method, system, computer device and vehicle for predicting vehicle lateral speed. Background Technology

[0002] With the development of autonomous driving technology, the requirements for high-precision and high-reliability vehicle positioning are becoming increasingly stringent. High-precision positioning technology for autonomous driving generally adopts multi-sensor fusion positioning, and relies particularly on inertial navigation algorithms based on IMU (Inertial Measurement Unit).

[0003] Since the accuracy of inertial navigation algorithms diverges over time, traditional methods often employ vehicle non-integrity constraint algorithms to improve the recursive accuracy of inertial navigation. This involves assuming that the vehicle's lateral and vertical velocities are zero, and then constraining the inertial navigation algorithm to suppress the divergence of inertial navigation recursive errors, thereby improving positioning accuracy.

[0004] However, during maneuvers such as turning and changing lanes, the wheels generate a sideslip angle, resulting in an actual lateral velocity that is not zero but a small non-zero value (e.g., 0.1 m / s). If the traditional vehicle non-integrity constraint algorithm is still applied in this scenario, imposing a strict lateral zero-velocity constraint, the accuracy of the inertial navigation algorithm cannot be effectively improved, thereby reducing the accuracy of vehicle positioning. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, computer device, and vehicle for predicting vehicle lateral speed in response to at least one of the above-mentioned technical problems.

[0006] In a first aspect, embodiments of this application provide a method for predicting the lateral speed of a vehicle, the method comprising: Collect motion sensor data of the vehicle within a preset time period. The motion sensor data set should include at least the vehicle's angular velocity, acceleration, and wheel speed.

[0007] The motion sensor dataset is input into a pre-trained lateral velocity prediction model to obtain the vehicle's lateral velocity at the current moment. The lateral velocity prediction model is used to characterize the mapping relationship between historical motion sensor datasets and historical lateral velocities.

[0008] In some embodiments, collecting motion sensor data of a vehicle within a preset time period further includes the following steps: The time interval is determined according to the preset sampling frequency, and the angular velocity, acceleration, wheel speed and steering wheel angle of the vehicle are collected within the preset time period. The component angular velocity of the rear wheel speed difference is determined based on the wheel speed difference between the left and right wheels of the rear axle and the wheel track. Determine the vehicle's wheel speed and angular velocity based on the steering wheel angle, the vehicle's wheelbase, and the wheel speeds of the left and right rear wheels. A motion sensor dataset is constructed based on the vehicle's angular velocity, acceleration, wheel speed, steering wheel angle, rear wheel speed differential angular velocity, and wheel speed rotation angle.

[0009] In some embodiments, inputting the motion sensor dataset into a pre-trained lateral velocity prediction model to obtain the lateral velocity of the vehicle at the current moment further includes the following steps: By using a pre-trained lateral velocity prediction model, local feature extraction is performed on a motion sensor dataset to obtain a local feature sequence. Establish temporal dependencies for local feature sequences, and determine the lateral speed of the vehicle at the current moment based on the temporal dependencies.

[0010] In some embodiments, the vehicle lateral speed prediction method further includes: For local features at different time steps in the local feature sequence, corresponding weights are assigned. The weights are used to measure the importance of the local features at each time step in predicting the lateral speed of the vehicle at the current moment.

[0011] In some embodiments, the vehicle lateral speed prediction method further includes: The vehicle's pose at the current moment is determined based on its lateral velocity and the corresponding observation noise variance; where the observation noise variance is used to characterize the confidence level of the vehicle's lateral velocity at the current moment.

[0012] In some embodiments, the vehicle lateral speed prediction method further includes: The historical motion sensor dataset of the vehicle under different driving conditions is input into the initial lateral velocity prediction model to obtain the predicted lateral velocity. Based on the error between the predicted lateral velocity and the standard lateral velocity, the model parameters of the initial lateral velocity prediction model are corrected until the error converges to a preset error threshold. Based on the corrected model parameters, the pre-trained lateral velocity prediction model is determined.

[0013] In some embodiments, the vehicle lateral speed prediction method further includes: Based on the vehicle's speed information in the navigation coordinate system and the pre-determined installation angle parameters of the inertial measurement unit relative to the vehicle, the vehicle's speed information in the vehicle body coordinate system is calculated through coordinate transformation; wherein, the installation angle parameters are used to correct the vehicle's speed information in the navigation coordinate system to the vehicle body coordinate system during coordinate transformation. The lateral velocity component of the vehicle's speed information in the vehicle body coordinate system is used as the standard lateral velocity.

[0014] In a second aspect, embodiments of this application provide a vehicle lateral speed prediction system, the system comprising: a sensor array and a controller; The sensor array is used to collect motion sensor datasets of the vehicle in response to data acquisition commands from the controller; The controller is used to execute the vehicle lateral speed prediction method as provided in any embodiment of the first aspect of this application.

[0015] In a third aspect, embodiments of this application provide a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the vehicle lateral speed prediction method provided in any embodiment of the first aspect of this application.

[0016] In a fourth aspect, embodiments of this application provide a vehicle that includes a vehicle lateral speed prediction system as provided in the second aspect of this application.

[0017] The aforementioned vehicle lateral velocity prediction method, system, computer equipment, and vehicle, by employing a mapping method based on a pre-trained model to directly predict lateral velocity from motion sensor data, effectively avoid the dependence of traditional tire-based methods on numerous physical parameters that are difficult to accurately calibrate or estimate online. This significantly improves the robustness and practicality of lateral velocity estimation in actual vehicle operation. By collecting and utilizing a time-series dataset of motion sensors over a preset time period as model input, the prediction process can consider the dynamic process and contextual information of vehicle motion, rather than just a single instantaneous state. This allows the model to more accurately capture and predict non-zero lateral velocities generated during dynamic maneuvers such as turning and lane changes, solving the key problem of the failure of traditional vehicle non-holonomic constraints in these scenarios. Attached Figure Description

[0018] Figure 1 This is a diagram illustrating the application environment of the vehicle lateral speed prediction method in some embodiments; Figure 2 This is a flowchart illustrating the vehicle lateral speed prediction method in some embodiments; Figure 3 This is a flowchart illustrating the steps involved in constructing a motion sensor dataset in some embodiments; Figure 4 This is a flowchart illustrating the feature extraction steps in some embodiments; Figure 5 This is a flowchart illustrating the prediction of lateral velocity in some embodiments; Figure 6This is a flowchart illustrating standard lateral speeds in some embodiments; Figure 7 Here are some block diagrams of the vehicle lateral velocity prediction system in some embodiments; Figure 8 This is a diagram showing the internal structure of a computer device in some embodiments. Detailed Implementation

[0019] To make the technical solutions and advantages of this application clearer, the embodiments and related technical content of this application will be further described in detail below with reference to the accompanying drawings and text description. It should be understood that the embodiments described below are only used to explain the technical solutions of the embodiments of this application and are not intended to limit more possible implementations of this application.

[0020] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0021] For ease of understanding, Figure 1 An application environment is illustrated, in which a controller 110 is built into a vehicle 120 to execute the steps of a vehicle lateral speed prediction method. During execution, the controller 110 can communicate with other devices or modules in the vehicle 120 via a network to obtain data related to vehicle lateral speed prediction sent by sensors 130 (sensor 130 refers to any sensor on the vehicle) or other devices or modules installed on the vehicle 120. The controller 110 can be implemented using a standalone controller or a controller cluster consisting of multiple controllers.

[0022] The controller can be implemented using at least one of the following hardware forms: programmable logic array (PLA), field-programmable gate array (FPGA), digital signal processor (DSP), application-specific integrated circuit (ASIC), general-purpose processor, or other programmable logic device.

[0023] Of course, the vehicle lateral speed prediction method provided in this application embodiment can also be applied to more scenarios not shown.

[0024] In a first aspect, embodiments of this application provide a method for predicting the lateral speed of a vehicle, which can be applied to... Figure 1 In the application environment shown, it can be applied to Figure 1 Taking controller 110 as an example, in some embodiments, such as Figure 2As shown, the vehicle lateral speed prediction method includes steps S210 and S220 that can be executed by the controller 110.

[0025] Step S210: Collect motion sensor data of the vehicle within a preset time period.

[0026] The motion sensor dataset includes at least the vehicle's angular velocity, acceleration, and wheel speed.

[0027] The preset time period refers to the time span used to construct the temporal input window. Its length needs to balance dynamic responsiveness and modeling stability, and can range from 1 second to 2.5 seconds. With a frequency of 50Hz, this corresponds to a sequence length of 50 to 125 time steps. The preset time period is not a fixed, absolute time window, but rather a sliding time window that traces backward from the current moment, ensuring that the model always infers based on the latest dynamic context.

[0028] A motion sensor dataset refers to a set of time-series data that reflects the motion state of a vehicle, collected by various motion sensors (IMU, wheel speed sensors, etc.) installed on the vehicle.

[0029] Among them, angular velocity refers to the instantaneous angular rate of the vehicle's rotation around its three axes (roll axis X, pitch axis Y, and yaw axis Z), which is output by the gyroscope in the IMU installed near the vehicle's center of gravity, and is measured in rad / s; its Z-axis component directly reflects the vehicle's steering strength, while the X / Y-axis components characterize the rate of change of the vehicle's attitude, together constituting the key observation of the vehicle's rotational dynamics.

[0030] Acceleration refers to the linear acceleration of a vehicle along its three axes, measured by accelerometers in the IMU, and is measured in m / s². The Y-axis component has a differential relationship with the lateral velocity, while the X-axis and Z-axis reflect the driving or braking intensity and road excitation, respectively.

[0031] Wheel speed refers to the instantaneous rotational linear velocity of each of the four wheels of a vehicle (left front wheel, right front wheel, left rear wheel, and right rear wheel), output by wheel speed sensors, and measured in m / s. The difference in wheel speeds among the four wheels implies the vehicle's steering geometry, drive or braking distribution, and tire lateral deviation behavior.

[0032] Data from various motion sensors can be collected synchronously and time-stamped to avoid phase errors introduced by asynchronous sampling. The data acquisition process can be achieved by the controller sending periodic trigger commands to each sensor via a CAN (Controller Area Network) bus or Ethernet, ensuring timing consistency. The motion sensor dataset is not limited to the three basic quantities mentioned above; it can also be expanded to include auxiliary variables such as steering wheel angle, yaw rate estimates, and IMU temperature compensation parameters, thereby enhancing the characterization of tire sideslip mechanisms.

[0033] Step S220: Input the motion sensor dataset into the pre-trained lateral velocity prediction model to obtain the lateral velocity of the vehicle at the current moment.

[0034] Among them, the lateral velocity prediction model is used to characterize the mapping relationship between the historical motion sensor dataset and the historical lateral velocity.

[0035] A pre-trained lateral velocity prediction model refers to a machine learning model (such as a deep learning model) that has learned the complex mapping relationship between specific inputs and outputs through a training process. This model is specifically trained to learn the mapping relationship between historical motion sensor datasets and historical lateral velocities. Its input is a motion sensor dataset containing time-series data such as angular velocity, acceleration, and wheel speed, and its output is lateral velocity.

[0036] The mapping relationship between historical motion sensor datasets and historical lateral speeds can be a nonlinear function relationship learned through massive amounts of real vehicle dynamic data. This mapping relationship has strong generalization ability and does not depend on physical parameters that are difficult to obtain accurately online, such as tire lateral stiffness, vertical load distribution, and road adhesion coefficient. Therefore, it can maintain stable performance on complex road surfaces such as different vehicle models, different load states, and rain, snow, and slippery surfaces.

[0037] By employing a mapping method based on a pre-trained model to directly predict lateral velocity from motion sensor data, this approach effectively avoids the reliance of traditional tire-based methods on numerous physical parameters that are difficult to calibrate or estimate online. This significantly improves the robustness and practicality of lateral velocity estimation in actual vehicle operation. By collecting and utilizing a time-series dataset of motion sensors over a pre-defined time period as model input, the prediction process considers the dynamic process and contextual information of vehicle motion, rather than just a single instantaneous state. This allows the model to more accurately capture and predict non-zero lateral velocities generated during dynamic maneuvers such as turning and lane changes, solving the key problem of the failure of traditional vehicle non-holonomic constraints in these scenarios.

[0038] In some embodiments, such as Figure 3 As shown, step S210 may also include steps S211, S212, S213 and S214.

[0039] Step S211: Determine the time interval according to the preset sampling frequency, and collect the vehicle's angular velocity, acceleration, wheel speed and steering wheel angle within the preset time period.

[0040] The preset sampling frequency refers to the time base parameter for synchronously acquiring data from multiple motion sensors, typically set to 50Hz, corresponding to a time interval of 20ms. This sampling frequency balances real-time performance with computational load, ensuring stable operation on automotive embedded platforms.

[0041] The steering wheel angle can be acquired by a high-precision rotary transformer or magnetic encoder at the steering column end, typically within the range of ±720°.

[0042] In some alternative embodiments, the sampling frequency can be adjusted within the range of 20Hz to 100Hz. For example, when focusing on low-speed fine-tuning, the sampling frequency can be increased to 80Hz to 100Hz to enhance transient response capture capability; when targeting high-speed cruise scenarios or constrained by processing unit resources, the sampling frequency can be reduced to 20Hz to 30Hz, and interpolation filtering and other methods can be used to compensate for the loss of timing information.

[0043] Step S212: Determine the component angular velocity of the rear wheel speed difference based on the wheel speed difference between the left and right wheels of the rear axle and the wheel track.

[0044] The wheel speed difference between the left and right rear wheels refers to the speed of the left rear wheel. With right rear wheel speed The difference, that is Wheelbase refers to the vertical distance W between the center points of the left and right wheels on the rear axle of a vehicle. It is usually between 1.5m and 1.65m. Wheelbase parameters can be fixed in the overall vehicle design drawings or measured and recorded in the non-volatile memory of the vehicle's electronic control unit during the production line calibration stage.

[0045] The differential angular velocity of the rear wheel speed can be calculated using the following formula: .in This refers to the differential angular velocity of the rear wheels, measured in rad / s. Its value directly reflects the vehicle's actual yaw dynamics caused by rear wheel lateral slippage or uneven road surfaces.

[0046] In some alternative embodiments, the wheel track W can be dynamically corrected using a temperature compensation model. For example, rear suspension temperature sensor readings can be incorporated, and a relationship can be established based on the thermal expansion and contraction characteristics of the aluminum alloy control arm, thereby improving performance under low or high temperature conditions. The accuracy.

[0047] Step S213: Determine the wheel speed and angular velocity of the vehicle based on the steering wheel angle, the vehicle's wheelbase, and the wheel speeds of the left and right rear wheels.

[0048] Wheelbase refers to the longitudinal distance between the center of the front axle and the center of the rear axle of a vehicle, which is usually 2.6m to 2.9m. It is also an inherent parameter of the vehicle and is pre-calibrated and stored.

[0049] Wheel speed and angular velocity can be calculated using the following formula: .in, Where L is the steering wheel angle and L is the wheelbase. Wheel speed and steering angle angular velocity, measured in rad / s. Wheel speed and steering angle angular velocity characterize the steering response trend of a vehicle under conditions of no sideslip and ideal adhesion.

[0050] Step S214: Construct a motion sensor dataset based on the vehicle's angular velocity, acceleration, wheel speed, steering wheel angle, rear wheel speed differential angular velocity, and wheel speed turning angle.

[0051] The motion sensor dataset is a structured temporal tensor with dimensions of . , This represents the total number of sampling points within a preset time period (e.g., 125 points for a 2.5s time window). The total dimension of the features is specifically composed as follows: 3D angular velocity ( ), 3D acceleration ( ), 4D wheel speed ( ), 1D steering wheel angle ( ), 1D rear wheel speed differential angular velocity ( ), 1D wheel speed, rotation angle, angular velocity ( The system comprises three features: a 1D wheel speed scaling factor and a 1D timestamp difference. The wheel speed scaling factor, provided by the fusion positioning module, corrects for wheel speed sensor bias and scale errors. The timestamp difference is the time interval between adjacent sampling points, used to normalize the temporal scale. This 15-dimensional feature set retains the original sensor physical quantities while injecting two derived variables with clear kinematic significance (wheel speed differential angular velocity and wheel speed turning angle angular velocity), forming a complete characterization of the vehicle's lateral motion.

[0052] Understandably, during the model training phase, the historical motion sensor dataset should include as much dynamic information about the vehicle as possible, such as straight driving, lane changes, turns, U-turns, highway ramps, and low-speed traffic jams. Additionally, a separate dataset can be collected for slippery conditions like rain or snow, covering various scenarios of daily driving for autonomous vehicles to improve the generalization ability of the vehicle's lateral speed prediction model.

[0053] Through steps S211 to S214, multi-source, multi-scale, and multi-physical mechanism fusion perception of the vehicle's lateral motion state is achieved. By introducing two derived features with clear kinematic interpretations—rear wheel speed differential angular velocity and wheel speed turning angle angular velocity—the data input to the lateral velocity prediction model not only includes raw sensor readings but also contains the relationship between vehicle structural parameters and steering dynamics. This significantly enhances the model's ability to identify non-zero lateral velocities during maneuvers such as turning and lane changes. Furthermore, all parameters can be obtained in real-time on mass-produced vehicles via standard CAN bus or in-vehicle Ethernet, eliminating the need to rely on physical parameters such as tire lateral stiffness and road adhesion coefficient, which are difficult to calibrate online. This effectively avoids the failure risk of traditional mechanical models under wet, slippery, and low-adhesion road conditions.

[0054] In some embodiments, such as Figure 4 As shown, step S220 also includes steps S221 and S222.

[0055] Step S221: Using a pre-trained lateral velocity prediction model, local feature extraction is performed on the motion sensor dataset to obtain a local feature sequence.

[0056] Local feature extraction can be achieved using a one-dimensional CNN (Convolutional Neural Network), with a specific configuration of 128 kernels, a kernel size of 3, a ReLU activation function, and a "same" padding method. Its function is to slide and scan the original motion sensor data sequence along the time dimension to identify short-term dynamic patterns, such as sudden acceleration edges, rising angular velocity edges, and step responses of wheel speed differential angular velocity, among other local transient features.

[0057] In some alternative embodiments, modules that can achieve local perception capabilities, such as separable convolutional layers, wavelet convolutional layers, or temporal convolutional units with gating mechanisms, can also be used for local feature extraction.

[0058] Step S222: Establish temporal dependencies for the local feature sequence, and determine the lateral speed of the vehicle at the current moment based on the temporal dependencies.

[0059] Establishing temporal dependencies can be accomplished using a three-layer stacked LSTM (Long Short-Term Memory) network. For example, each LSTM layer has 128 hidden units and adopts a bidirectional or unidirectional structure. The first LSTM layer receives the local feature sequence and outputs the hidden state sequence. The second and third layers are stacked sequentially, gradually enhancing the memory capacity and abstraction ability of historical states.

[0060] In some alternative embodiments, modules such as gated loop units, Transformer encoders, or neural differential equations can also be used to establish temporal dependencies under the same data conditions.

[0061] Determining the lateral speed of a vehicle at the current moment based on temporal dependencies can specifically include: taking the hidden state vector of the last time step of the third-layer LSTM as the global temporal representation of the entire sequence; after the hidden state vector is reduced in dimensionality by global average pooling, it is connected to a fully connected layer (8 neurons, ReLU function activated); and finally, the lateral speed value (unit: m / s) at the current moment is obtained by regression through a single neuron linear output layer.

[0062] By extracting short-term local features using CNN, the interference of IMU zero-bias drift, wheel speed sensor quantization error, and road excitation noise on the model input is significantly suppressed. By establishing temporal dependencies using LSTM, the problem of simple models failing to capture key dynamic patterns in long-term sequences is solved, ensuring that lateral velocity prediction fully considers the influence of historical motion states. Especially under conditions of high maneuverability and significant sideslip, the prediction results have excellent spatiotemporal consistency and physical rationality.

[0063] In some embodiments, the vehicle lateral speed prediction method further includes the following step: assigning corresponding weights to local features corresponding to different time steps in the local feature sequence.

[0064] The weights are used to measure the importance of the local features at each time step in predicting the lateral speed of the vehicle at the current moment.

[0065] The local features corresponding to different time steps refer to the feature vectors of each frame in the local feature sequence with indices t=1, 2, ..., t. Each vector corresponds to an abstract expression of the dynamic state of the vehicle at a certain sampling time, covering transient response features such as sudden changes in angular velocity, leaps in wheel speed difference, and acceleration of steering wheel angle.

[0066] Specifically, a scalar weight α can be calculated for each time step after the LSTM output using a self-attention mechanism. t ∈[0,1], satisfying ∑α t =1, this weight is obtained through backpropagation optimization during model training, and is used to reflect the contribution of features at each time step to the lateral velocity prediction task at the current time step.

[0067] In some alternative embodiments, weight allocation may also be implemented using scaled dot product attention, additive attention, or gated attention structures.

[0068] By embedding a dynamic weight allocation mechanism, the model can adaptively identify and reinforce key moments that have a dominant impact on lateral velocity, thereby improving the accuracy of lateral velocity prediction.

[0069] In some embodiments, the vehicle lateral velocity prediction method further includes the following step: determining the vehicle pose at the current moment based on the vehicle's lateral velocity at the current moment and its corresponding observation noise variance.

[0070] Among them, the observation noise variance is used to characterize the confidence level of the vehicle's lateral speed at the current moment.

[0071] The vehicle's pose at the current moment refers to its position (x, y) in the navigation coordinate system (i.e., the East-North-Sky coordinate system). n y n , z n ) and attitude (heading angle ψ) nPitch angle θ n , roll angle γ n The vehicle's pose can be updated within the error state Kalman filter framework, using lateral velocity as the observation. This is composed of a 6-DOF state vector.

[0072] Confidence level is a statistical indicator that quantifies the reliability of lateral velocity prediction results. Mathematically, it represents the diagonal element of the lateral velocity channel in the observation noise variance matrix. This confidence level directly affects the calculation of the Kalman gain. As the confidence level increases, the weights corresponding to lateral velocity corrections in the Kalman gain automatically decrease, thereby suppressing the contamination of pose estimation by low-quality lateral velocity prediction results. The observation noise variance can be determined based on the lateral velocity error learned in the lateral velocity prediction model.

[0073] By integrating the lateral velocity predicted by deep learning into the high-precision inertial navigation correction system, the predicted lateral velocity value truly reflects the non-zero sideslip characteristics of vehicle dynamics, and its accompanying observation noise variance can dynamically characterize the prediction reliability under different operating conditions. Therefore, in the error state Kalman filter, the lateral velocity observation value can strongly suppress lateral velocity drift at high confidence and automatically yield and avoid misguided guidance at low confidence, thus solving the model mismatch problem caused by the forced application of zero-speed constraints in traditional technologies.

[0074] For example, the lateral velocity is obtained through a lateral velocity prediction model. This lateral velocity is then used as input to a fusion positioning filtering algorithm (error state Kalman filter) to obtain the vehicle's optimal estimated pose, which includes lateral position estimation. The lateral position error in the vehicle coordinate system is obtained by subtracting this from the true lateral position value. The lateral position error can be used to determine the lateral velocity output by the lateral velocity prediction model and its effect on improving fusion positioning accuracy, as shown in Table 1. Metric Baseline Comparison Diff Diff% max 0.590 (unit: m) 0.365 (unit: m) -0.225 (unit: m) -38.15% The metrics include the maximum lateral position error (max), the Baseline column represents the lateral velocity improvement in fusion positioning accuracy without using the lateral velocity prediction model output in this application, the Comparison column represents the lateral velocity improvement in fusion positioning accuracy using the lateral velocity prediction model output in this application, the Diff column is the difference between the maximum lateral position error under the Comparison condition and the maximum lateral position error under the Baseline condition, representing the improvement of the maximum lateral position error under the two conditions, and the Diff% column represents the percentage improvement of the maximum lateral position error under the two conditions.

[0075] It can be seen that under the Comparison condition, the maximum lateral position error is 0.365m, which is 0.225m lower than the maximum lateral position error of 0.590m under the Baseline condition. The maximum position error is reduced by 38.15%, which is a significant effect and has engineering application value.

[0076] In some embodiments, such as Figure 5 As shown, the vehicle lateral speed prediction method also includes steps S510 and S520.

[0077] Step S510: Input the historical motion sensor dataset of the vehicle under different driving conditions into the initial lateral velocity prediction model to obtain the predicted lateral velocity.

[0078] The historical motion sensor dataset for vehicles under different driving conditions refers to a time-series multi-source sensor data collection covering the entire real-world road driving scenario. Its sampling resolution can also be 50Hz, and each time step includes 15 features: angular velocity (X / Y / Z axes), acceleration (X / Y / Z axes), left front / right front / left rear / right rear wheel speeds, steering wheel angle, wheel speed scaling factor, rear wheel speed differential angular velocity, wheel speed turning angle angular velocity, and timestamp difference. This dataset covers typical dynamic conditions such as straight-line constant speed driving, rapid acceleration, braking, lane changing, turning, U-turns, highway ramp entrances and exits, low-speed congested following, and sideslip on wet and snowy roads. Furthermore, each frame of data is strictly synchronized and aligned with the standard lateral velocity ground truth value generated by the multi-source fusion positioning system.

[0079] Different driving conditions not only reflect kinematic differences, but also include the dimension of environmental disturbance. For example, in the data collected on wet and slippery roads, the differential wheel speed angular velocity corresponding to the same steering wheel angle has a significant statistical offset from that on dry roads. This difference is fully preserved in the dataset and serves as a key basis for the model to learn the implicit representation of road surface adhesion state.

[0080] Among them, multi-source fusion positioning systems can include RTK (Real-time kinematic), LiDAR (Light Detection and Ranging), HDMap (High-Definition Map), etc.

[0081] Step S520: Based on the error between the predicted lateral velocity and the standard lateral velocity, the model parameters of the initial lateral velocity prediction model are corrected until the error converges to a preset error threshold. Based on the corrected model parameters, the pre-trained lateral velocity prediction model is determined.

[0082] The initial lateral velocity prediction model is a deep neural network architecture with end-to-end regression capability. Its basic structure can be a cascade of a one-dimensional convolutional neural network (Conv1D) and a long short-term memory network (LSTM), with a self-attention mechanism embedded. Specifically, the input layer receives a temporal tensor, the first layer is a Conv1D layer, followed by a MaxPooling1D layer to reduce dimensionality and enhance translation invariance, then three LSTM layers are stacked to model temporal dependencies, the output of the third LSTM layer is used as the query and value input to the self-attention layer to achieve dynamic weighted focusing on key time steps, and finally the lateral velocity scalar is regressed through a fully connected layer.

[0083] This model can be replaced by other architectures with temporal modeling capabilities, such as: TCN (Temporal Convolutional Network) to replace LSTM, Transformer encoder to replace LSTM+Attention combination, etc.

[0084] The error between the predicted lateral velocity and the standard lateral velocity can be calculated using the Huber loss function, and the model parameters can be corrected using the backpropagation algorithm.

[0085] The preset error threshold can be determined according to the actual measurement requirements. For example, the upper limit of the accuracy tolerance of the inertial navigation system for the lateral velocity observation is 0.025m / s, and the error threshold can be set to 0.022m / s to ensure that the model has sufficient robustness margin after going online.

[0086] The pre-trained lateral velocity prediction model determined based on the corrected model parameters can be defined as packaging all network weights and structural configurations saved in the final convergence state into a deployable model file and storing it in the controller memory.

[0087] By constructing a lateral velocity prediction model with a supervised learning paradigm at its core and using a historical sensor dataset covering all working conditions as training input, the model does not need to rely on physical parameters that are difficult to calibrate online, such as tire lateral stiffness, vertical load, and road friction coefficient. This fundamentally avoids the mismatch risk of traditional mechanical models in complex environments and can replace the zero-velocity assumption in traditional non-holonomic constraints, thereby significantly suppressing the divergence of inertial navigation errors and solving the problem of model mismatch caused by forced zero-velocity constraints during maneuvering, which in turn reduces positioning accuracy.

[0088] In some embodiments, such as Figure 6 As shown, the vehicle lateral velocity prediction method also includes the following steps: Step S610: Based on the vehicle's speed information in the navigation coordinate system and the pre-determined installation angle parameters of the inertial measurement unit relative to the vehicle, calculate the vehicle's speed information in the vehicle body coordinate system through coordinate transformation.

[0089] The installation angle parameter is used to correct the vehicle's speed information from the navigation coordinate system to the vehicle body coordinate system during coordinate transformation.

[0090] The navigation coordinate system refers to the ENU (East-North-Up) coordinate system. The velocity information in this coordinate system is output in real time by a high-precision multi-source fusion positioning system, which includes at least one or more of RTK, LiDAR, visual odometry and HDMap matching modules.

[0091] Inertial measurement units are typically mounted at a designated location on the vehicle chassis to collect raw data on angular velocity and acceleration.

[0092] Mounting angle parameters refer to the static attitude deviation of the IMU coordinate system relative to the vehicle coordinate system, which can be expressed in Euler angles as a set of pitch, roll, and yaw angles.

[0093] Among them, the pitch and roll components in the installation angle parameters can be obtained through offline calibration. For example, when the vehicle is stationary on a horizontal rigid platform, the IMU zero bias and the direction of the gravity vector are collected. Combined with the known platform normal, the rotational relationship between the IMU sensitive axis and the front-rear, left-right, and up-down axes of the vehicle body is calculated, thereby determining the pitch and roll components related to the direction of gravity.

[0094] However, when the vehicle is stationary and only gravity information is available, rotation around the direction of gravity is physically unobservable, making it impossible to determine the heading angle component. Therefore, the heading angle can be introduced as a parameter to be estimated into the state vector of an error-state Kalman filter. During vehicle movement, combined with observational information related to the vehicle's direction of motion provided by RTK or LiDAR, the heading angle can be estimated and dynamically corrected online, thus achieving online calibration of the heading angle component.

[0095] Coordinate transformation refers to mapping velocity information from the navigation coordinate system to the vehicle coordinate system through two stages of rotation. The first stage is from the navigation coordinate system to the IMU coordinate system, which depends on the direction cosine matrix from the navigation coordinate system to the IMU coordinate system. The second stage, from the IMU coordinate system to the vehicle coordinate system, depends on the direction cosine matrix from the IMU coordinate system to the vehicle coordinate system. .

[0096] The overall transformation relationship is: .

[0097] in, This is the vehicle's speed information in the vehicle's coordinate system. It is the velocity information in the navigation coordinate system.

[0098] In some optional embodiments, the coordinate transformation from the navigation coordinate system to the vehicle coordinate system can also be accomplished using quaternion operations. For example, the attitude relationship between the navigation coordinate system and the IMU coordinate system, and the attitude relationship between the IMU coordinate system and the vehicle coordinate system can be represented in quaternion form, and rotation operations can be performed using quaternions.

[0099] Step S620: Take the lateral velocity component of the vehicle's velocity information in the vehicle body coordinate system as the standard lateral velocity.

[0100] Lateral velocity component refers to The component of the vector along the Y-axis of the vehicle body refers to the instantaneous velocity of the vehicle's center of mass relative to the ground in the direction perpendicular to forward movement. This component directly corresponds to the lateral slip effect generated by tire lateral movement, exhibiting significant non-zero characteristics during maneuvers such as lane changes, cornering, and emergency avoidance.

[0101] By utilizing a two-stage coordinate transformation mechanism driven by high-precision ground truth velocity values ​​in the navigation coordinate system and IMU mounting angle parameters, the external absolute observed velocity is accurately mapped to the vehicle coordinate system, and the lateral component is precisely extracted as the standard lateral velocity. The navigation coordinate system velocity provides physical authenticity, while the mounting angle parameters provide systematic error correction capabilities. Together, they ensure that the obtained standard lateral velocity reflects the vehicle's true sideslip motion while eliminating the systematic mismatch introduced by IMU mounting deviations. This effectively solves the problem of distorted standard lateral velocity caused by IMU misalignment, leading to unreliable training labels for the lateral velocity prediction model. It lays a data foundation for building a high-confidence, high-generalization deep learning prediction model, and ultimately supports more accurate inertial navigation error suppression by the error state Kalman filter in the absence of absolute observations.

[0102] It should be understood that, although Figures 2 to 6 The steps in the flowchart are shown sequentially according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Figures 2 to 6Unless otherwise expressly stated herein, the steps illustrated and other steps involved in the embodiments are not subject to strict order restrictions and may be performed in other orders. Furthermore, at least some steps in the foregoing embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but may be performed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0103] In a second aspect, embodiments of this application provide a vehicle lateral speed prediction system, such as Figure 7 As shown, the vehicle lateral speed prediction system 700 includes: a sensor group 710 and a controller 720; the sensor group 710 is used to acquire a motion sensor dataset of the vehicle in response to a data acquisition command from the controller 720; the controller 720 is used to execute the vehicle lateral speed prediction method as provided in any embodiment of the first aspect of this application. Wherein, the controller 720 and... Figure 1 The controller 110 shown can be the same controller.

[0104] In a third aspect, embodiments of this application provide a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the vehicle lateral speed prediction method provided in any embodiment of the first aspect of this application.

[0105] In some embodiments, the computer device may be a controller, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data related to vehicle lateral speed prediction. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements the vehicle lateral speed prediction method in any embodiment of this document.

[0106] Those skilled in the art will understand that Figure 8The structures shown are merely block diagrams of some structures related to the embodiments of this application and do not constitute a limitation on the computer devices on which the embodiments of this application are applied. Specific computer devices may include more or fewer components than those shown in the figures, or combine certain components, or have different component arrangements.

[0107] In a fourth aspect, embodiments of this application provide a vehicle that includes a vehicle lateral speed prediction system as provided in embodiments of the second aspect of this application.

[0108] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The aforementioned computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments of this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0109] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0110] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A method for predicting the lateral speed of a vehicle, characterized in that, The method includes: Collect motion sensor data of the vehicle within a preset time period; the motion sensor data set includes at least the vehicle's angular velocity, acceleration, and wheel speed; The motion sensor dataset is input into a pre-trained lateral velocity prediction model to obtain the lateral velocity of the vehicle at the current moment; wherein, the lateral velocity prediction model is used to characterize the mapping relationship between the historical motion sensor dataset and the historical lateral velocity.

2. The method according to claim 1, characterized in that, The data set of motion sensors collected from the vehicle within a preset time period includes: The angular velocity, acceleration, wheel speed, and steering wheel angle of the vehicle are collected within the preset time period according to a preset sampling frequency. The component angular velocity of the rear wheel speed difference is determined based on the wheel speed difference between the left and right wheels of the rear axle and the wheel track. The wheel speed, turning angle, and angular velocity of the vehicle are determined based on the steering wheel angle, the wheelbase of the vehicle, and the wheel speeds of the left and right rear wheels. The motion sensor dataset is constructed based on the vehicle's angular velocity, acceleration, wheel speed, steering wheel angle, rear wheel speed differential angular velocity, and wheel speed rotation angle.

3. The method according to claim 1 or 2, characterized in that, The step of inputting the motion sensor dataset into a pre-trained lateral velocity prediction model to obtain the lateral velocity of the vehicle at the current moment includes: Using the pre-trained lateral velocity prediction model, local feature extraction is performed on the motion sensor dataset to obtain a local feature sequence; A temporal dependency relationship is established for the local feature sequence, and the lateral speed of the vehicle at the current moment is determined based on the temporal dependency relationship.

4. The method according to claim 3, characterized in that, The method further includes: A corresponding weight is assigned to the local features at different time steps in the local feature sequence; the weight is used to measure the importance of the local features at each time step in predicting the lateral speed of the vehicle at the current time.

5. The method according to claim 1, characterized in that, The method further includes: The pose of the vehicle at the current moment is determined based on its lateral velocity and the corresponding observation noise variance; wherein the observation noise variance is used to characterize the confidence level of the lateral velocity of the vehicle at the current moment.

6. The method according to claim 1, characterized in that, The method further includes: The historical motion sensor dataset of the vehicle under different driving conditions is input into the initial lateral velocity prediction model to obtain the predicted lateral velocity. Based on the error between the predicted lateral velocity and the standard lateral velocity, the model parameters of the initial lateral velocity prediction model are corrected until the error converges to a preset error threshold. Based on the corrected model parameters, the pre-trained lateral velocity prediction model is determined.

7. The method according to claim 6, characterized in that, The method further includes: Based on the vehicle's speed information in the navigation coordinate system and the predetermined installation angle parameter of the inertial measurement unit relative to the vehicle, the vehicle's speed information in the vehicle body coordinate system is calculated through coordinate transformation; wherein, the installation angle parameter is used to correct the vehicle's speed information in the navigation coordinate system to the vehicle body coordinate system in the coordinate transformation; The lateral velocity component of the vehicle's speed information in the vehicle body coordinate system is used as the standard lateral velocity.

8. A vehicle lateral speed prediction system, characterized in that, The system includes: a sensor array and a controller; The sensor group is used to collect motion sensor data sets of the vehicle in response to the data acquisition command of the controller; The controller is used to perform the vehicle lateral speed prediction method as described in any one of claims 1 to 7.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A vehicle, characterized in that, The vehicle includes the vehicle lateral speed prediction system as described in claim 8.

Images