Vehicle chassis state estimation model training method, estimation method and device

Abstract

The invention relates to the technical field of automatic driving, and provides a vehicle chassis state estimation model training method, estimation method and device, and the estimation method comprises the steps: obtaining the real-time motion data of a cab and the real-time rotation angle of a steering wheel corresponding to the real-time motion data of the cab; inputting the real-time motion data of the cab and the real-time rotation angle of the steering wheel into a vehicle chassis state estimation model to obtain the real-time offset of the pose data of the vehicle chassis output by the vehicle chassis state estimation model and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab; and determining real-time motion data of the chassis based on the real-time difference value, and determining real-time pose data of the chassis based on the pose data real-time offset so as to obtain a real-time state of the chassis. The real-time motion state of the vehicle chassis can be accurately estimated according to the real-time motion data of the cab and the real-time rotation angle of the steering wheel, and the stability and accuracy of an automatic driving system on vehicle control are improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a vehicle chassis state estimation model training method, estimation method and device. Background Art

[0002] In the heavy truck automatic driving system, the control of the vehicle chassis is the key link to ensure the safe and stable driving of the vehicle. In the traditional control method, the attitude sensors installed on the cab, such as gyroscopes and accelerometers, are usually used to measure the vehicle's motion data (linear velocity, linear acceleration and angular velocity) in real time, and the posture data (position and attitude angle, attitude angle includes roll angle, yaw angle and pitch angle) of the vehicle chassis are estimated based on the motion data combined with the kinematic or dynamic model, and the motion data and posture data are used to characterize the motion state of the vehicle chassis. However, due to the non-rigid connection structure characteristics between the cab and the vehicle chassis of the heavy truck, the motion data obtained by the sensors installed in the cab and the estimated posture data cannot accurately characterize the motion state of the chassis. Therefore, the automatic driving with this motion state as a reference makes the stability and accuracy of the overall automatic driving vehicle control poor. Summary of the invention

[0003] The present invention provides a vehicle chassis state estimation model training method, an estimation method and a device, which are used to solve the problem in the prior art that the motion state of the vehicle chassis cannot be accurately estimated, resulting in poor stability and accuracy of vehicle automatic driving control.

[0004] The present invention provides a vehicle chassis state estimation model training method, comprising the following steps.

[0005] Acquire sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true posture data of the vehicle chassis corresponding to the sample motion data.

[0006] The sample motion data and sample turning angle are input into a vehicle chassis state estimation model, and the offset of the true value pose data is used as the first label, and the difference between the true value motion data and the corresponding sample motion data is used as the second label. The vehicle chassis state estimation model is trained, and the offset of the true value pose data is the offset of the true value pose data collected twice adjacently.

[0007] Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

[0008] According to a vehicle chassis state estimation model training method provided by the present invention, the vehicle chassis state estimation model includes: a motion feature extraction module, a corner feature extraction module, a cross-attention module, a first multi-layer perceptron and a second multi-layer perceptron, and the cross-attention module connects the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron and the second multi-layer perceptron.

[0009] The motion feature extraction module is used to extract the motion feature vector of the sample motion data.

[0010] The corner feature extraction module is used to extract the corner feature vector of the sample corner.

[0011] The cross-attention module is used to make the motion feature vector respond to the correlation of the corner feature vector.

[0012] The first multi-layer perceptron is used to regress the pose data according to the motion feature vector and the rotation angle feature vector after the response correlation and predict the offset and output it.

[0013] The second multi-layer perceptron is used to regress the motion data prediction difference according to the motion feature vector and the corner feature vector after the response correlation and output it.

[0014] According to a vehicle chassis state estimation model training method provided by the present invention, the sample motion data, true value motion data and real-time motion data respectively include their corresponding three-axis speeds, three-axis accelerations and three-axis angular velocities.

[0015] The motion feature extraction module includes: a splicing module and a first convolution module, a second convolution module and a third convolution module connected to the splicing module, and the splicing module is connected to the cross attention module.

[0016] The first convolution module is used to convolve the sample three-axis velocity corresponding to the sample motion data to obtain a three-axis velocity feature vector.

[0017] The second convolution module is used to convolve the sample three-axis acceleration corresponding to the sample motion data to obtain a three-axis acceleration feature vector.

[0018] The third convolution module is used to convolve the sample three-axis angular velocity corresponding to the sample motion data to obtain a three-axis angular velocity feature vector.

[0019] The splicing module is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain the motion feature vector.

[0020] According to a vehicle chassis state estimation model training method provided by the present invention, the sample motion data and the sample turning angle are input into the vehicle chassis state estimation model, the offset of the true value posture data is used as the first label, and the difference between the true value motion data and the corresponding sample motion data is used as the second label, and the vehicle chassis state estimation model is trained, including the following steps.

[0021] The sample motion data and the sample rotation angle are respectively input into the motion feature extraction module and the rotation angle feature extraction module to obtain a motion feature vector and a rotation angle feature vector.

[0022] The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector.

[0023] The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the predicted offset of the posture data.

[0024] The second multi-layer perceptron regresses the motion feature vector and the corner feature vector after the response correlation and outputs the motion data prediction difference.

[0025] The predicted offset of the pose data together with the first label corresponding to the sample motion data are substituted into the first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data are substituted into the second loss function. When both the first loss function and the second loss function converge, the training of the vehicle chassis state estimation model is completed.

[0026] According to a vehicle chassis state estimation model training method provided by the present invention, the sample motion data and the sample turning angle are input into the vehicle chassis state estimation model, the offset of the true value posture data is used as the first label, and the difference between the true value motion data and the corresponding sample motion data is used as the second label. The vehicle chassis state estimation model is trained, including the following two training stages.

[0027] The first stage of training: the sample motion data and the sample rotation angle are input into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector.

[0028] The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector.

[0029] The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the first position data prediction offset.

[0030] The predicted offset of the first pose data and the first label corresponding to the sample motion data are brought into a first loss function. When the first loss function converges, the first stage of training is completed.

[0031] Second stage training: inputting the sample motion data and the sample rotation angle into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector.

[0032] The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector.

[0033] The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the second posture data prediction offset.

[0034] The second multi-layer perceptron regresses the motion feature vector and the corner feature vector after the response correlation and outputs the motion data prediction difference.

[0035] The predicted offset of the second pose data together with the first label corresponding to the sample motion data are substituted into the first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data are substituted into the second loss function. When both the first loss function and the second loss function converge, the second stage of training is completed to obtain the vehicle chassis state estimation model.

[0036] According to a vehicle chassis state estimation model training method provided by the present invention, the first loss function L 1 for.

[0037] L 1= λ t L t + λ R L R .

[0038] λ t + λ R =1.

[0039] L t = pred t - gt t .

[0040] L R =||( pred x - gt x )+( pred y - gt y )+(pred z - gt z )||.

[0041] The second loss function L 2 for.

[0042] L 2= λ 1||( pred 1 x - gt 1 x )+( pred 1 y - gt 1 y )|| 2 + λ 2||( pred 2 x - gt 2 x )+( pred 2 y - gt 2 y )|| 2 + λ 3||( pred 3 x - gt 3 x )+( pred 3 y - gt 3 y )|| 2 .

[0043] λ 1+ λ 2+ λ 3=1.

[0044] in, L t represents the position loss function, L R represents the attitude angle loss function, λ t and λ R Respectively L t and L R Weighting coefficient, pred t Represents the predicted offset of the position in the predicted offset of the pose data, gt t Indicates the offset of the position in the first label, pred xIndicates the predicted offset of the roll angle in the predicted offset of the attitude data. pred y Indicates the predicted offset of the yaw angle in the predicted offset of the attitude data. pred z Indicates the predicted offset of the pitch angle in the predicted offset of the attitude data. gt x Indicates the offset of the roll angle in the first label, gt y Indicates the offset of the yaw angle in the first tag, gt z represents the offset of the pitch angle in the first label, and || || represents the second norm.

[0045] pred 1 x Indicates the linear velocity prediction difference in motion data x The weight of the axis, pred 1 y Indicates the linear velocity prediction difference in motion data y The weight of the axis, pred 2 x Indicates the predicted difference of linear acceleration in motion data x The weight of the axis, pred 2 y Indicates the predicted difference of linear acceleration in motion data y The weight of the axis, pred 3 x Indicates that the angular velocity prediction difference in the motion data is x The weight of the axis, pred 3 y Indicates that the angular velocity prediction difference in the motion data is y The weight of the axis, gt 1 x Indicates that the linear velocity difference in the second label is x The weight of the axis, gt 1 y Indicates that the linear velocity difference in the second label is y The weight of the axis, gt 2 x Indicates that the midline acceleration difference of the second label is x The weight of the axis, gt 2 y Indicates that the midline acceleration difference of the second label is y The weight of the axis, gt 3 x Indicates that the angular velocity difference in the second tag is x The weight of the axis, gt 3 y Indicates that the angular velocity difference in the second tag is y The weight of the axis, λ1. λ 2 and λ 3 are weighting coefficients respectively.

[0046] According to a vehicle chassis state estimation model training method provided by the present invention, the true value motion data and true value posture data of the vehicle chassis corresponding to the sample motion data are obtained, including the following steps.

[0047] The original motion data and original position and posture data of the vehicle chassis corresponding to the sample motion data are obtained.

[0048] The original motion data and the original posture data are low-pass filtered to obtain the true motion data and the true posture data.

[0049] The present invention also provides a vehicle chassis state estimation method, which includes the following steps.

[0050] Real-time motion data of the cab and a real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab are obtained.

[0051] The real-time motion data of the cab and the real-time angle of the steering wheel are input into the vehicle chassis state estimation model to obtain the real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, as well as the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab.

[0052] The real-time motion data of the chassis is determined based on the real-time difference, and the real-time posture data of the chassis is determined based on the real-time offset of the posture data, so as to obtain the real-time motion state of the chassis.

[0053] Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

[0054] According to a vehicle chassis state estimation method provided by the present invention, the vehicle chassis state estimation model includes: a motion feature extraction module, a corner feature extraction module, a cross-attention module, a first multi-layer perceptron and a second multi-layer perceptron, and the cross-attention module connects the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron and the second multi-layer perceptron.

[0055] The motion feature extraction module is used to extract the motion feature vector of the real-time motion data of the cab.

[0056] The corner feature extraction module is used to extract the corner feature vector of the real-time corner.

[0057] The cross-attention module is used to make the motion feature vector respond to the correlation of the corner feature vector.

[0058] The first multi-layer perceptron is used to regress and output the real-time posture data of the vehicle chassis according to the motion feature vector and the angle feature vector after response correlation.

[0059] The second multi-layer perceptron is used to regress and output the real-time motion data of the vehicle chassis according to the motion feature vector and the turning angle feature vector after response correlation.

[0060] According to a vehicle chassis state estimation method provided by the present invention, the sample motion data, true value motion data and real-time motion data respectively include the three-axis velocity, three-axis acceleration and three-axis angular velocity corresponding to each other.

[0061] The motion feature extraction module includes: a splicing module and a first convolution module, a second convolution module and a third convolution module connected to the splicing module, and the splicing module is connected to the cross attention module.

[0062] The first convolution module is used to convolve the real-time three-axis velocity corresponding to the real-time motion data of the cab to obtain a three-axis velocity feature vector.

[0063] The second convolution module is used to convolve the real-time three-axis acceleration corresponding to the real-time motion data of the cab to obtain a three-axis acceleration feature vector.

[0064] The third convolution module is used to convolve the real-time three-axis angular velocity corresponding to the real-time motion data of the cab to obtain a three-axis angular velocity feature vector.

[0065] The splicing module is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain the motion feature vector.

[0066] The present invention also provides a vehicle chassis state estimation model training device, comprising the following functional units.

[0067] The training data acquisition unit is used to acquire sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true posture data of the vehicle chassis corresponding to the sample motion data.

[0068] A model training unit is used to input the sample motion data and sample turning angle into a vehicle chassis state estimation model, use the offset of the true value pose data as a first label, and the difference between the true value motion data and the corresponding sample motion data as a second label to train the vehicle chassis state estimation model, wherein the offset of the true value pose data is the offset of the true value pose data collected twice adjacently.

[0069] Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

[0070] The present invention also provides a vehicle chassis state estimation device, comprising the following functional units.

[0071] The real-time data acquisition unit is used to acquire the real-time motion data of the cab and the real-time turning angle of the steering wheel corresponding to the real-time motion data of the cab.

[0072] The model execution unit is used to input the real-time motion data of the cab and the real-time angle of the steering wheel into the vehicle chassis state estimation model to obtain the real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, as well as the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab.

[0073] The chassis state determination unit is used to determine the real-time motion data of the chassis based on the real-time difference, and determine the real-time posture data of the chassis based on the real-time offset of the posture data, so as to obtain the real-time state of the chassis.

[0074] Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

[0075] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements a vehicle chassis state estimation model training method as described in any one of the above, or implements a vehicle chassis state estimation method as described in any one of the above.

[0076] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the vehicle chassis state estimation model training method described in any one of the above-mentioned methods is implemented, or the vehicle chassis state estimation method described in any one of the above-mentioned methods is implemented.

[0077] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements a vehicle chassis state estimation model training method as described in any one of the above, or a vehicle chassis state estimation method as described in any one of the above.

[0078] The vehicle chassis state estimation model training method, estimation method and device provided by the present invention use the collected sample motion data of the cab and the sample turning angle of the steering wheel as training samples, the offset of the true value posture data as the first label, and the difference between the true value motion data and the corresponding sample motion data as the second label to train the vehicle chassis state estimation model, so that the trained vehicle chassis state estimation model can accurately estimate the real-time offset of the posture data of the vehicle chassis, and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab through the real-time collected real-time motion data of the cab and the real-time turning angle of the steering wheel, and can accurately obtain the real-time motion data and real-time posture data of the vehicle chassis according to the real-time offset and real-time difference of the posture data, so as to accurately estimate the motion state of the vehicle chassis. Moreover, during the training process, the steering wheel angle is associated, so that the real-time offset and real-time difference of the posture data of the chassis output by the vehicle chassis state estimation model are both correlated with the operating state of the vehicle, thereby providing reliable reference data for the vehicle's automatic driving system, and improving the stability and accuracy of the automatic driving system's control of the vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0080] Figure 1 It is a flow chart of the vehicle chassis state estimation model training method provided by the present invention.

[0081] Figure 2 It is a schematic diagram of the structure of a vehicle chassis state estimation model in the vehicle chassis state estimation model training method provided by the present invention.

[0082] Figure 3It is a schematic diagram of the first stage training of the vehicle chassis state estimation model in the vehicle chassis state estimation model training method provided by the present invention.

[0083] Figure 4 It is a schematic diagram of the second stage training of the vehicle chassis state estimation model in the vehicle chassis state estimation model training method provided by the present invention.

[0084] Figure 5 It is a flow chart of the vehicle chassis state estimation method provided by the present invention.

[0085] Figure 6 It is a schematic diagram of the application of the vehicle chassis state estimation model in the vehicle chassis state estimation method provided by the present invention.

[0086] Figure 7 It is a structural schematic diagram of a vehicle chassis state estimation model training device provided by the present invention.

[0087] Figure 8 It is a structural schematic diagram of a vehicle chassis state estimation device provided by the present invention.

[0088] Figure 9 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0089] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0090] The vehicle chassis state estimation model training method of the embodiment of the present invention is as follows: Figure 1 As shown, the process includes the following steps S110 and S120.

[0091] Step S110: Obtain sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true posture data of the vehicle chassis corresponding to the sample motion data. The sample motion data of the cab is IMU (Inertial Measurement Unit) data output by the combined navigation sensor installed in the cab, including: three-axis velocity, three-axis acceleration and three-axis angular velocity. The sample angle of the steering wheel can be directly obtained from the vehicle control system (the vehicle usually has its own angle sensor, which can collect steering wheel angle data and provide it to the vehicle control system). The sample angle of the steering wheel corresponding to the sample motion data means that in each set of sample data, the sample angle and sample motion data are collected at the same time.

[0092] Specifically, a combined navigation sensor is installed in the cab, and the sample motion data output by the combined navigation sensor is collected at a preset collection frequency (for example: 50Hz), and the sample angle of the steering wheel is collected at the same collection frequency, thereby obtaining multiple groups of sample data, each group of sample data includes sample motion data and sample angles collected at the same time, and the sample data can be stored in a training database. Among them, the true value motion data of the chassis also includes: three-axis speed, three-axis acceleration and three-axis angular velocity, and the true value posture data includes: position (chassis position) and attitude angle. The true value motion data and true value posture data of the chassis are obtained by collecting the combined navigation sensor installed on the chassis. The true value motion data and true value posture data of the vehicle chassis corresponding to the sample motion data refer to: for each sample motion data, the corresponding true value motion data and true value posture data are collected at the same time as the sample motion data.

[0093] Specifically, a combined navigation sensor is installed on the chassis, and the original motion data and original posture data output by the combined navigation sensor are collected at the above-mentioned preset collection frequency (for example: 50 Hz). The original motion data and original posture data can be directly used as true motion data and true posture data, thereby obtaining multiple groups of label data. A group of label data and a group of sample data collected at the same time form a group of training data, and the training data can be pre-stored in a training database.

[0094] During training, sample motion data, sample rotation angles, true motion data, and true pose data are obtained from the training database.

[0095] Step S120: Input the sample motion data and sample turning angle into a vehicle chassis state estimation model, use the offset of the true value pose data as the first label, and the difference between the true value motion data and the corresponding sample motion data as the second label, and train the vehicle chassis state estimation model. The offset of the true value pose data is the offset of the true value pose data collected twice adjacently, that is, the offset of the two true value pose data collected at two adjacent data collection moments during the training data collection process. Specifically, at the 0th moment, sample motion data 0, sample rotation angle 0, true motion data 0 and true posture data 0 are collected. At this time, the first label can be set to 0, and the second label is: true motion data 0-sample motion data 0; at the 1st moment, sample motion data 1, sample rotation angle 1, true motion data 1 and true posture data 1 are collected. At this time, the first label is: true posture data 1-true posture data 0, and the second label is: true motion data 1-sample motion data 1, and so on. At the i-th moment, sample motion data i, sample rotation angle i, true motion data i and true posture data i are collected. At this time, the first label is: true posture data i-true posture data (i-1), and the second label is true motion data i-sample motion data i. Preferably, the first label and the second label can be pre-stored in the above-mentioned training database, and stored in corresponding rows with the corresponding sample motion data.

[0096] Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

[0097] In the vehicle chassis state estimation model training method of the present embodiment, the sample motion data of the cab and the sample turning angle of the steering wheel collected are used as training samples, the offset of the true value posture data is used as the first label, and the difference between the true value motion data and the corresponding sample motion data is used as the second label. The vehicle chassis state estimation model is trained so that the trained vehicle chassis state estimation model can accurately estimate the real-time offset of the posture data of the vehicle chassis and the real-time difference of the real-time motion data of the chassis and the real-time motion data of the cab through the real-time motion data of the cab collected in real time and the real-time turning angle of the steering wheel. The real-time motion data and real-time posture data of the vehicle chassis can be accurately obtained according to the real-time offset and real-time difference of the posture data, thereby accurately estimating the motion state of the vehicle chassis. Moreover, during the training process, the steering wheel angle is associated, so that the real-time offset and real-time difference of the posture data of the chassis output by the vehicle chassis state estimation model are both correlated with the operating state of the vehicle, thereby providing reliable reference data for the vehicle's automatic driving system, and improving the stability and accuracy of the automatic driving system's control of the vehicle.

[0098] It should be noted that: in order to reduce costs, a low-cost combined navigation sensor is installed in the cab, and a high-precision combined navigation sensor is installed in the chassis, so that the accuracy of the true motion data and the true pose data is high, that is, the accuracy of the first label and the second label is high. Therefore, the real-time offset and real-time difference of the pose data of the chassis estimated by the trained vehicle chassis state estimation model are of high accuracy. Moreover, in the model application stage, only the low-cost combined navigation sensor of the cab is retained on the vehicle, which greatly saves the vehicle cost. The vehicle chassis state estimation model trained by the method of this embodiment is particularly suitable for vehicles with non-rigid connection between the cab and the chassis, such as heavy trucks, and realizes accurate estimation of the motion state of the chassis that is non-rigidly connected to the cab by the motion data of the cab and the steering wheel angle.

[0099] In some embodiments, Figure 2 As shown, the vehicle chassis state estimation model includes: a motion feature extraction module, a corner feature extraction module, a cross-attention module 6, a first multi-layer perceptron 71 and a second multi-layer perceptron 72, and the cross-attention module 6 connects the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron 71 and the second multi-layer perceptron 72.

[0100] The motion feature extraction module is used to extract the motion feature vector of the sample motion data.

[0101] The corner feature extraction module is used to extract the corner feature vector of the sample corner.

[0102] The cross attention module 6 is used to make the motion feature vector respond to the correlation of the corner feature vector.

[0103] The first multi-layer perceptron 71 is used to regress the posture data according to the motion feature vector and the rotation angle feature vector after the response correlation and predict the offset and output it.

[0104] The second multi-layer perceptron 72 is used to regress the motion data prediction difference according to the motion feature vector and the rotation angle feature vector after the response correlation and output it, that is, to regress the difference with the sample motion data.

[0105] Specifically, the sample motion data, true value motion data and real-time motion data include the three-axis velocity, three-axis acceleration and three-axis angular velocity corresponding to each other.

[0106] Based on this, the motion feature extraction module includes: a splicing module 5 and a first convolution module 1, a second convolution module 2 and a third convolution module 3 connected to the splicing module 5, and the splicing module 5 is connected to the cross-attention module 6.

[0107] The first convolution module 1 is used to convolve the sample three-axis velocity corresponding to the sample motion data to obtain a three-axis velocity feature vector.

[0108] The second convolution module 2 is used to convolve the sample three-axis acceleration corresponding to the sample motion data to obtain a three-axis acceleration feature vector.

[0109] The third convolution module 3 is used to convolve the sample three-axis angular velocity corresponding to the sample motion data to obtain a three-axis angular velocity feature vector.

[0110] The splicing module 5 is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector (can be spliced ​​by a concat function) to obtain the motion feature vector.

[0111] The corner feature extraction module is the fourth convolution module 4, which obtains the corner feature vector by performing convolution on the sample corner.

[0112] In some embodiments, based on the above model structure, step S120 specifically includes the following steps.

[0113] The sample motion data and the sample rotation angle are respectively input into the motion feature extraction module and the rotation angle feature extraction module to obtain a motion feature vector and a rotation angle feature vector. Specifically, the sample three-axis velocity, the sample three-axis acceleration, the sample three-axis angular velocity and the sample rotation angle are respectively input into the first convolution module 1, the second convolution module 2, the third convolution module 3 and the fourth convolution module 4 to obtain a three-axis velocity feature vector, a three-axis acceleration feature vector, a three-axis angular velocity feature vector and a rotation angle feature vector. The splicing module 5 splices the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain a motion feature vector.

[0114] The cross attention module 6 makes the motion feature vector respond to the correlation of the corner feature vector.

[0115] The first multi-layer perceptron 71 regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the predicted offset of the posture data.

[0116] The second multi-layer perceptron 72 regresses the motion feature vector and the corner feature vector after the response correlation and outputs the motion data prediction difference.

[0117] The predicted offset of the pose data together with the first label corresponding to the sample motion data is brought into a preset first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data is brought into a preset second loss function. When both the first loss function and the second loss function converge, the training of the vehicle chassis state estimation model is completed. Specifically, the convergence of the first loss function and the second loss function can be the convergence of the first loss function and the second loss function respectively, or the total loss function of the weighted sum of the first loss function and the second loss function converges.

[0118] In this embodiment, the first loss function L 1 can be: L 1= λ t L t + λ R L R .

[0119] λ t + λ R =1.

[0120] L t = pred t - gt t .

[0121] L R =||( pred x - gt x )+( pred y - gt y )+( pred z - gt z )||.

[0122] The second loss function L 2 can be: L 2= λ 1||( pred 1 x - gt 1 x )+( pred 1 y - gt 1 y )|| 2 + λ2||( pred 2 x - gt 2 x )+( pred 2 y - gt 2 y )|| 2 + λ 3||( pred 3 x - gt 3 x )+( pred 3 y - gt 3 y )|| 2 .

[0123] λ 1+ λ 2+ λ 3=1.

[0124] in, L t represents the position loss function, L R represents the attitude angle loss function, λ t and λ R Respectively L t and L R Weighting coefficient, pred t Represents the predicted offset of the position in the predicted offset of the pose data, gt t Indicates the offset of the position in the first label, pred x Indicates the predicted offset of the roll angle in the predicted offset of the attitude data. pred y Indicates the predicted offset of the yaw angle in the predicted offset of the attitude data. pred z Indicates the predicted offset of the pitch angle in the predicted offset of the attitude data. gt x Indicates the offset of the roll angle in the first label, gt y Indicates the offset of the yaw angle in the first tag, gt z represents the offset of the pitch angle in the first label, and || || represents the second norm.

[0125] pred 1 x Indicates the linear velocity prediction difference in motion datax The weight of the axis, pred 1 y Indicates the linear velocity prediction difference in motion data y The weight of the axis, pred 2 x Indicates the predicted difference of linear acceleration in motion data x The weight of the axis, pred 2 y Indicates the predicted difference of linear acceleration in motion data y The weight of the axis, pred 3 x Indicates that the angular velocity prediction difference in the motion data is x The weight of the axis, pred 3 y Indicates that the angular velocity prediction difference in the motion data is y The weight of the axis, gt 1 x Indicates that the linear velocity difference in the second label is x The weight of the axis, gt 1 y Indicates that the linear velocity difference in the second label is y The weight of the axis, gt 2 x Indicates that the midline acceleration difference of the second label is x The weight of the axis, gt 2 y Indicates that the midline acceleration difference of the second label is y The weight of the axis, gt 3 x Indicates that the angular velocity difference in the second tag is x The weight of the axis, gt 3 y Indicates that the angular velocity difference in the second tag is y The weight of the axis, λ 1. λ 2 and λ 3 are weighting coefficients respectively. x axis, y Axis and z The coordinate system of the axis can be the world coordinate system.

[0126] In the above model training method, there are two loss functions. Although the training can be completed in the end, the model convergence time is long, resulting in a long model training time, high training difficulty and low efficiency. Therefore, in some embodiments, based on the above model structure, the model training in step S120 can be divided into two-stage training, and step S120 specifically includes two training stages.

[0127] The first stage of training, such as Figure 3 shown.

[0128] The sample motion data and the sample rotation angle are respectively input into the motion feature extraction module and the rotation angle feature extraction module to obtain a motion feature vector and a rotation angle feature vector. Specifically, the sample three-axis velocity, the sample three-axis acceleration, the sample three-axis angular velocity and the sample rotation angle are respectively input into the first convolution module 1, the second convolution module 2, the third convolution module 3 and the fourth convolution module 4 to obtain a three-axis velocity feature vector, a three-axis acceleration feature vector, a three-axis angular velocity feature vector and a rotation angle feature vector. The splicing module 5 splices the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain a motion feature vector.

[0129] The cross attention module 6 makes the motion feature vector respond to the correlation of the corner feature vector.

[0130] The first multi-layer perceptron 71 regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the first position data prediction offset.

[0131] The first posture data predicted offset and the first label corresponding to the sample motion data are brought into a preset first loss function. When the first loss function converges, the first stage of training is completed.

[0132] The second stage training is to continue the training based on the first stage training, that is, to use the parameters of the model obtained after the first stage training, to add the second multi-layer perceptron 72 to continue the training. The second stage training is as follows: Figure 4 shown.

[0133] The sample motion data and the sample rotation angle are input into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector. The sample motion data and the sample rotation angle used in the second stage of training may be the same as or different from the sample motion data and the sample rotation angle used in the first stage of training.

[0134] The cross attention module 6 makes the motion feature vector respond to the correlation of the corner feature vector.

[0135] The first multi-layer perceptron 71 regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the second posture data predicted offset.

[0136] The second multi-layer perceptron 72 regresses the motion feature vector and the corner feature vector after the response correlation and outputs the motion data prediction difference.

[0137] The predicted offset of the second pose data together with the first label corresponding to the sample motion data are brought into the preset first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data are brought into the preset second loss function. When both the first loss function and the second loss function converge, the second stage of training is completed to obtain the vehicle chassis state estimation model. Among them, the first loss function and the second loss function can adopt the above loss function L 1 and L 2.

[0138] In this embodiment, the difficulty of model training is reduced by a two-stage training method, the model converges faster during training, and the model training efficiency is improved.

[0139] In some embodiments, obtaining true motion data and true posture data of a vehicle chassis corresponding to the sample motion data includes: obtaining original motion data and original posture data of the vehicle chassis corresponding to the sample motion data, wherein the original motion data and original posture data are data directly output by a combined navigation sensor installed on the chassis, and include vibration components of the vehicle mechanical structure itself, for example, the high-frequency vibration of the engine is transmitted to the chassis, resulting in the original motion data and original posture data being superimposed with motion components and posture components generated by the vibration of the vehicle mechanical structure itself.

[0140] In order to obtain more accurate chassis motion data and posture data, the original motion data and original posture data are low-pass filtered to obtain the true motion data and true posture data, that is, the motion components and posture components generated by the vibration of the vehicle mechanical structure itself are removed, and more accurate vehicle motion data and posture data are obtained as labels. This makes the estimation result of the vehicle chassis state estimation model that is finally trained more accurate.

[0141] The present invention also provides a vehicle chassis state estimation method, such as Figure 5 As shown, the following steps are included.

[0142] Step S510: Acquire the real-time motion data of the cab and the real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab. Specifically, the real-time motion data of the cab is collected at a preset collection frequency by a combined navigation sensor arranged in the cab, and the real-time steering angle of the steering wheel is collected at the preset collection frequency by an angle sensor of the steering wheel.

[0143] Step S520: Input the real-time motion data of the cab and the real-time angle of the steering wheel into the vehicle chassis state estimation model to obtain the real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, as well as the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab.

[0144] Step S530: Determine the real-time motion data of the chassis based on the real-time difference, and determine the real-time posture data of the chassis based on the real-time offset of the posture data to obtain the real-time motion state of the chassis. Specifically, the real-time motion data of the chassis is the real-time difference plus the corresponding real-time motion data of the cab, and the linear velocity, linear acceleration and angular velocity are added to the corresponding real-time difference respectively; the real-time posture data of the current chassis is obtained by adding the real-time posture data of the chassis determined last time to the real-time offset of the currently output posture data, and the position, roll angle, yaw angle and pitch angle are added to the corresponding real-time offset respectively.

[0145] Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

[0146] The vehicle chassis state estimation method of this embodiment uses the collected sample motion data of the cab and the sample angle of the steering wheel as training samples, the offset of the true value posture data as the first label, and the difference between the true value motion data and the corresponding sample motion data as the second label, to train the vehicle chassis state estimation model, so that the trained vehicle chassis state estimation model can accurately estimate the real-time offset of the posture data of the vehicle chassis and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab through the real-time collected real-time motion data of the cab and the real-time angle of the steering wheel. The real-time motion data and real-time posture data of the vehicle chassis can be accurately obtained according to the real-time offset and real-time difference of the posture data, thereby accurately estimating the motion state of the vehicle chassis. In addition, during the training process, the steering wheel angle is associated, so that the real-time offset and real-time difference of the posture data of the chassis output by the vehicle chassis state estimation model are both correlated with the operating state of the vehicle, thereby providing reliable reference data for the vehicle's automatic driving system, and improving the stability and accuracy of the automatic driving system's control of the vehicle.

[0147] In some embodiments, Figure 2 and 6 As shown, the vehicle chassis state estimation model includes: a motion feature extraction module, a corner feature extraction module, a cross-attention module 6, a first multi-layer perceptron 71 and a second multi-layer perceptron 72, and the cross-attention module 6 connects the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron 71 and the second multi-layer perceptron 72.

[0148] The motion feature extraction module is used to extract the motion feature vector of the real-time motion data.

[0149] The corner feature extraction module is used to extract the corner feature vector of the real-time corner. Specifically, the corner feature extraction module is a fourth convolution module 4, which obtains the corner feature vector corresponding to the real-time corner by performing convolution on the real-time corner.

[0150] The cross attention module 6 is used to make the motion feature vector respond to the correlation of the angle feature vector.

[0151] The first multi-layer perceptron 71 is used to regress and output the real-time position and posture data of the vehicle chassis according to the motion feature vector and the angle feature vector after the response correlation.

[0152] The second multi-layer perceptron 72 is used to regress and output the real-time motion data of the vehicle chassis according to the motion feature vector and the turning angle feature vector after the response correlation.

[0153] Specifically, the sample motion data, true value motion data and real-time motion data respectively include the three-axis speed, three-axis acceleration and three-axis angular velocity corresponding to each other. The motion feature extraction module includes: a splicing module 5 and a first convolution module 1, a second convolution module 2 and a third convolution module 3 connected to the splicing module 5, and the splicing module 5 is connected to the cross attention module 6.

[0154] The first convolution module 1 is used to convolve the real-time three-axis velocity corresponding to the real-time motion data of the cab to obtain a three-axis velocity feature vector.

[0155] The second convolution module 2 is used to convolve the real-time three-axis acceleration corresponding to the real-time motion data of the cab to obtain a three-axis acceleration feature vector.

[0156] The third convolution module 3 is used to convolve the real-time three-axis angular velocity corresponding to the real-time motion data of the cab to obtain a three-axis angular velocity feature vector.

[0157] The splicing module 5 is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector (can be spliced ​​by a concat function) to obtain a motion feature vector corresponding to the real-time motion data.

[0158] The vehicle chassis state estimation model in this embodiment can be trained using the above-mentioned vehicle chassis state estimation model training method, which will not be repeated here.

[0159] The vehicle chassis state estimation model training device provided by the present invention is described below. The vehicle chassis state estimation model training device described below and the vehicle chassis state estimation model training method described above can be referenced to each other.

[0160] like Figure 7As shown, the vehicle chassis state estimation model training device of an embodiment of the present invention includes the following functional units.

[0161] The training data acquisition unit 710 is used to acquire sample motion data of the cab and sample steering angles corresponding to the sample motion data, as well as true motion data and true posture data of the vehicle chassis corresponding to the sample motion data.

[0162] The model training unit 720 is used to input the sample motion data and the sample turning angle into the vehicle chassis state estimation model, use the offset of the true value pose data as the first label, and the difference between the true value motion data and the corresponding sample motion data as the second label, to train the vehicle chassis state estimation model, and the offset of the true value pose data is the offset of the true value pose data collected twice adjacently.

[0163] Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

[0164] Optionally, the vehicle chassis state estimation model includes: a motion feature extraction module, a corner feature extraction module, a cross-attention module, a first multi-layer perceptron and a second multi-layer perceptron, and the cross-attention module connects the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron and the second multi-layer perceptron.

[0165] The motion feature extraction module is used to extract the motion feature vector of the sample motion data.

[0166] The corner feature extraction module is used to extract the corner feature vector of the sample corner.

[0167] The cross-attention module is used to make the motion feature vector respond to the correlation of the corner feature vector.

[0168] The first multi-layer perceptron is used to regress the pose data according to the motion feature vector and the rotation angle feature vector after the response correlation and predict the offset and output it.

[0169] The second multi-layer perceptron is used to regress the motion data prediction difference according to the motion feature vector and the corner feature vector after the response correlation and output it.

[0170] Optionally, the sample motion data, true value motion data and real-time motion data respectively include corresponding three-axis velocities, three-axis accelerations and three-axis angular velocities.

[0171] The motion feature extraction module includes: a splicing module and a first convolution module, a second convolution module and a third convolution module connected to the splicing module, and the splicing module is connected to the cross attention module.

[0172] The first convolution module is used to convolve the sample three-axis velocity corresponding to the sample motion data to obtain a three-axis velocity feature vector.

[0173] The second convolution module is used to convolve the sample three-axis acceleration corresponding to the sample motion data to obtain a three-axis acceleration feature vector.

[0174] The third convolution module is used to convolve the sample three-axis angular velocity corresponding to the sample motion data to obtain a three-axis angular velocity feature vector.

[0175] The splicing module is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain the motion feature vector.

[0176] Optionally, the model training unit 720 is specifically used to: input the sample motion data and the sample rotation angle into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector.

[0177] The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector.

[0178] The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the predicted offset of the posture data.

[0179] The second multi-layer perceptron regresses the motion feature vector and the corner feature vector after the response correlation and outputs the motion data prediction difference.

[0180] The predicted offset of the pose data together with the first label corresponding to the sample motion data are substituted into the first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data are substituted into the second loss function. When both the first loss function and the second loss function converge, the training of the vehicle chassis state estimation model is completed.

[0181] Optionally, the model training unit 720 specifically includes: a first training unit and a second training unit, which are used for first stage training and second stage training respectively.

[0182] The first stage of training includes: inputting the sample motion data and the sample rotation angle into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector.

[0183] The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector.

[0184] The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the first position data prediction offset.

[0185] The predicted offset of the first pose data and the first label corresponding to the sample motion data are brought into a first loss function. When the first loss function converges, the first stage of training is completed.

[0186] The second stage of training includes: inputting the sample motion data and the sample rotation angle into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector.

[0187] The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector.

[0188] The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the second posture data prediction offset.

[0189] The second multi-layer perceptron regresses the motion feature vector and the corner feature vector after the response correlation and outputs the motion data prediction difference.

[0190] The predicted offset of the second pose data together with the first label corresponding to the sample motion data are substituted into the first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data are substituted into the second loss function. When both the first loss function and the second loss function converge, the second stage of training is completed to obtain the vehicle chassis state estimation model.

[0191] Optionally, the first loss function L 1 for.

[0192] L 1= λ t L t + λ R L R .

[0193] λ t + λ R =1.

[0194] L t = pred t - gtt .

[0195] L R =||( pred x - gt x )+( pred y - gt y )+( pred z - gt z )||.

[0196] The second loss function L 2 for.

[0197] L 2= λ 1||( pred 1 x - gt 1 x )+( pred 1 y - gt 1 y )|| 2 + λ 2||( pred 2 x - gt 2 x )+( pred 2 y - gt 2 y )|| 2 + λ 3||( pred 3 x - gt 3 x )+( pred 3 y - gt 3 y )|| 2 .

[0198] λ 1+ λ 2+ λ 3=1.

[0199] in, L t represents the position loss function, L R represents the attitude angle loss function, λ t and λ R RespectivelyL t and L R Weighting coefficient, pred t Represents the predicted offset of the position in the predicted offset of the pose data, gt t Indicates the offset of the position in the first label, pred x Indicates the predicted offset of the roll angle in the predicted offset of the attitude data. pred y Indicates the predicted offset of the yaw angle in the predicted offset of the attitude data. pred z Indicates the predicted offset of the pitch angle in the predicted offset of the attitude data. gt x Indicates the offset of the roll angle in the first label, gt y Indicates the offset of the yaw angle in the first tag, gt z represents the offset of the pitch angle in the first label, and || || represents the second norm.

[0200] pred 1 x Indicates the linear velocity prediction difference in motion data x The weight of the axis, pred 1 y Indicates the linear velocity prediction difference in motion data y The weight of the axis, pred 2 x Indicates the predicted difference of linear acceleration in motion data x The weight of the axis, pred 2 y Indicates the predicted difference of linear acceleration in motion data y The weight of the axis, pred 3 x Indicates that the angular velocity prediction difference in the motion data is x The weight of the axis, pred 3 y Indicates that the angular velocity prediction difference in the motion data is y The weight of the axis, gt 1 x Indicates that the linear velocity difference in the second label is x The weight of the axis, gt 1 y Indicates that the linear velocity difference in the second label is y The weight of the axis, gt 2 x Indicates that the midline acceleration difference of the second label is x The weight of the axis, gt 2y Indicates that the midline acceleration difference of the second label is y The weight of the axis, gt 3 x Indicates that the angular velocity difference in the second tag is x The weight of the axis, gt 3 y Indicates that the angular velocity difference in the second tag is y The weight of the axis, λ 1. λ 2 and λ 3 are weighting coefficients respectively.

[0201] Optionally, the training data acquisition unit 710 is specifically used to obtain original motion data and original posture data of the vehicle chassis corresponding to the sample motion data; and perform low-pass filtering on the original motion data and original posture data to obtain the true motion data and true posture data.

[0202] The vehicle chassis state estimating device provided by the present invention is described below. The vehicle chassis state estimating device described below and the vehicle chassis state estimating method described above can be referred to each other.

[0203] like Figure 8 As shown, the vehicle chassis state estimation device according to an embodiment of the present invention includes the following functional units.

[0204] The real-time data acquisition unit 810 is used to acquire the real-time motion data of the cab and the real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab.

[0205] The model execution unit 820 is used to input the real-time motion data of the cab and the real-time angle of the steering wheel into the vehicle chassis state estimation model to obtain the real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, as well as the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab.

[0206] The chassis state determination unit 830 is used to determine the real-time motion data of the chassis based on the real-time difference, and determine the real-time posture data of the chassis based on the real-time offset of the posture data, so as to obtain the real-time state of the chassis.

[0207] Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

[0208] Optionally, the vehicle chassis state estimation model includes: a motion feature extraction module, a corner feature extraction module, a cross-attention module, a first multi-layer perceptron and a second multi-layer perceptron, and the cross-attention module connects the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron and the second multi-layer perceptron.

[0209] The motion feature extraction module is used to extract the motion feature vector of the real-time motion data of the cab.

[0210] The corner feature extraction module is used to extract the corner feature vector of the real-time corner.

[0211] The cross-attention module is used to make the motion feature vector respond to the correlation of the corner feature vector.

[0212] The first multi-layer perceptron is used to regress and output the real-time posture data of the vehicle chassis according to the motion feature vector and the angle feature vector after response correlation.

[0213] The second multi-layer perceptron is used to regress and output the real-time motion data of the vehicle chassis according to the motion feature vector and the turning angle feature vector after response correlation.

[0214] Optionally, the sample motion data, true value motion data and real-time motion data respectively include corresponding three-axis velocities, three-axis accelerations and three-axis angular velocities.

[0215] The motion feature extraction module includes: a splicing module and a first convolution module, a second convolution module and a third convolution module connected to the splicing module, and the splicing module is connected to the cross attention module.

[0216] The first convolution module is used to convolve the real-time three-axis velocity corresponding to the real-time motion data of the cab to obtain a three-axis velocity feature vector.

[0217] The second convolution module is used to convolve the real-time three-axis acceleration corresponding to the real-time motion data of the cab to obtain a three-axis acceleration feature vector.

[0218] The third convolution module is used to convolve the real-time three-axis angular velocity corresponding to the real-time motion data of the cab to obtain a three-axis angular velocity feature vector.

[0219] The splicing module is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain the motion feature vector.

[0220] Figure 9 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 9As shown, the electronic device may include: a processor (processor) 910, a communication interface (Communications Interface) 920, a memory (memory) 930 and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other through the communication bus 940. The processor 910 may call the logic instructions in the memory 930 to execute the vehicle chassis state estimation model training method, which includes the following steps.

[0221] Acquire sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true posture data of the vehicle chassis corresponding to the sample motion data.

[0222] The sample motion data and sample turning angle are input into a vehicle chassis state estimation model, and the offset of the true value pose data is used as the first label, and the difference between the true value motion data and the corresponding sample motion data is used as the second label. The vehicle chassis state estimation model is trained, and the offset of the true value pose data is the offset of the true value pose data collected twice adjacently.

[0223] Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

[0224] Alternatively, a vehicle chassis state estimation method is executed, the method comprising the following steps.

[0225] Real-time motion data of the cab and a real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab are obtained.

[0226] The real-time motion data of the cab and the real-time angle of the steering wheel are input into the vehicle chassis state estimation model to obtain the real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, as well as the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab.

[0227] The real-time motion data of the chassis is determined based on the real-time difference, and the real-time posture data of the chassis is determined based on the real-time offset of the posture data, so as to obtain the real-time motion state of the chassis.

[0228] Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

[0229] In addition, the logic instructions in the above-mentioned memory 930 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0230] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the vehicle chassis state estimation model training method provided by the above-mentioned methods, which includes the following steps.

[0231] Acquire sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true posture data of the vehicle chassis corresponding to the sample motion data.

[0232] The sample motion data and sample turning angle are input into a vehicle chassis state estimation model, and the offset of the true value pose data is used as the first label, and the difference between the true value motion data and the corresponding sample motion data is used as the second label. The vehicle chassis state estimation model is trained, and the offset of the true value pose data is the offset of the true value pose data collected twice adjacently.

[0233] Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

[0234] Alternatively, a vehicle chassis state estimation method is executed, the method comprising the following steps.

[0235] Real-time motion data of the cab and a real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab are obtained.

[0236] The real-time motion data of the cab and the real-time angle of the steering wheel are input into the vehicle chassis state estimation model to obtain the real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, as well as the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab.

[0237] The real-time motion data of the chassis is determined based on the real-time difference, and the real-time posture data of the chassis is determined based on the real-time offset of the posture data, so as to obtain the real-time motion state of the chassis.

[0238] Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

[0239] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the vehicle chassis state estimation model training method provided by the above-mentioned methods is implemented, and the method includes the following steps.

[0240] Acquire sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true posture data of the vehicle chassis corresponding to the sample motion data.

[0241] The sample motion data and sample turning angle are input into a vehicle chassis state estimation model, and the offset of the true value pose data is used as the first label, and the difference between the true value motion data and the corresponding sample motion data is used as the second label. The vehicle chassis state estimation model is trained, and the offset of the true value pose data is the offset of the true value pose data collected twice adjacently.

[0242] Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

[0243] Alternatively, a vehicle chassis state estimation method is executed, the method comprising the following steps.

[0244] Real-time motion data of the cab and a real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab are obtained.

[0245] The real-time motion data of the cab and the real-time angle of the steering wheel are input into the vehicle chassis state estimation model to obtain the real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, as well as the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab.

[0246] The real-time motion data of the chassis is determined based on the real-time difference, and the real-time posture data of the chassis is determined based on the real-time offset of the posture data, so as to obtain the real-time motion state of the chassis.

[0247] Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

[0248] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0249] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0250] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vehicle chassis state estimation model training method, characterized in that: include: Acquire sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true pose data of the vehicle chassis corresponding to the sample motion data; The sample motion data and the sample turning angle are input into a vehicle chassis state estimation model, the offset of the true value pose data is used as a first label, and the difference between the true value motion data and the corresponding sample motion data is used as a second label, and the vehicle chassis state estimation model is trained, wherein the offset of the true value pose data is the offset of the true value pose data collected twice adjacently; Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

2. The vehicle chassis state estimation model training method according to claim 1, characterized in that: The vehicle chassis state estimation model comprises: a motion feature extraction module, a corner feature extraction module, a cross attention module, a first multi-layer perceptron and a second multi-layer perceptron, wherein the cross attention module is connected to the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron and the second multi-layer perceptron; The motion feature extraction module is used to extract the motion feature vector of the sample motion data; The corner feature extraction module is used to extract the corner feature vector of the sample corner; The cross attention module is used to make the motion feature vector respond to the correlation of the corner feature vector; The first multi-layer perceptron is used to regress the position data according to the motion feature vector and the rotation angle feature vector after the response correlation and predict the offset and output it; The second multi-layer perceptron is used to regress the motion data prediction difference according to the motion feature vector and the corner feature vector after the response correlation and output it.

3. The vehicle chassis state estimation model training method according to claim 2, characterized in that: The sample motion data, true value motion data and real-time motion data respectively include the three-axis velocity, three-axis acceleration and three-axis angular velocity corresponding to each other; The motion feature extraction module includes: a splicing module and a first convolution module, a second convolution module and a third convolution module connected to the splicing module, and the splicing module is connected to the cross attention module; The first convolution module is used to convolve the sample three-axis velocity corresponding to the sample motion data to obtain a three-axis velocity feature vector; The second convolution module is used to convolve the sample three-axis acceleration corresponding to the sample motion data to obtain a three-axis acceleration feature vector; The third convolution module is used to convolve the sample three-axis angular velocity corresponding to the sample motion data to obtain a three-axis angular velocity feature vector; The splicing module is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain the motion feature vector.

4. The vehicle chassis state estimation model training method according to claim 2, characterized in that: The sample motion data and the sample turning angle are input into a vehicle chassis state estimation model, the offset of the true value pose data is used as a first label, and the difference between the true value motion data and the corresponding sample motion data is used as a second label, and the vehicle chassis state estimation model is trained, including: Inputting the sample motion data and the sample rotation angle into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector; The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector; The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the predicted offset of the posture data; The second multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the motion data prediction difference; The predicted offset of the pose data together with the first label corresponding to the sample motion data are substituted into the first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data are substituted into the second loss function. When both the first loss function and the second loss function converge, the training of the vehicle chassis state estimation model is completed.

5. The vehicle chassis state estimation model training method according to claim 2, characterized in that: The sample motion data and the sample turning angle are input into a vehicle chassis state estimation model, the offset of the true value pose data is used as a first label, and the difference between the true value motion data and the corresponding sample motion data is used as a second label, and the vehicle chassis state estimation model is trained, including: First stage training: Inputting the sample motion data and the sample rotation angle into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector; The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector; The first multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the first position data prediction offset; Bringing the predicted offset of the first pose data together with the first label corresponding to the sample motion data into a first loss function, and when the first loss function converges, the first stage of training is completed; Second stage training: Inputting the sample motion data and the sample rotation angle into the motion feature extraction module and the rotation angle feature extraction module respectively to obtain a motion feature vector and a rotation angle feature vector; The cross-attention module causes the motion feature vector to respond to the correlation of the corner feature vector; The first multi-layer perceptron regresses and outputs a second posture data prediction offset according to the motion feature vector and the rotation angle feature vector after the response correlation; The second multi-layer perceptron regresses the motion feature vector and the rotation angle feature vector after the response correlation and outputs the motion data prediction difference; The predicted offset of the second pose data together with the first label corresponding to the sample motion data are substituted into the first loss function, and the predicted difference of the motion data together with the second label corresponding to the sample motion data are substituted into the second loss function. When both the first loss function and the second loss function converge, the second stage of training is completed to obtain the vehicle chassis state estimation model.

6. The vehicle chassis state estimation model training method according to claim 4 or 5, characterized in that: The first loss function L 1 is: L 1= λ t L t + λ R L R ； λ t + λ R =1； L t = pred t - gt t ； L R =||( pred x - gt x )+( pred y - gt y )+( pred z - gt z )||； The second loss function L 2 is: L 2= λ 1||( pred 1 x - gt 1 x )+( pred 1 y - gt 1 y )|| 2 + λ 2||( pred 2 x - gt 2 x )+( pred 2 y - gt 2 y )|| 2 + λ 3||( pred 3 x - gt 3 x )+( pred 3 y - gt 3 y )|| 2 ； λ 1+ λ 2+ λ 3=1； in, L t represents the position loss function, L R represents the attitude angle loss function, λ t and λ R Respectively L t and L R Weighting coefficient, pred t Represents the predicted offset of the position in the predicted offset of the pose data, gt t Indicates the offset of the position in the first label, pred x Indicates the predicted offset of the roll angle in the predicted offset of the attitude data. pred y Indicates the predicted offset of the yaw angle in the predicted offset of the attitude data. pred z Indicates the predicted offset of the pitch angle in the predicted offset of the attitude data. gt x Indicates the offset of the roll angle in the first label, gt y Indicates the offset of the yaw angle in the first tag, gt z represents the offset of the pitch angle in the first label, || || represents the second norm; pred 1 x Indicates the linear velocity prediction difference in motion data x The weight of the axis, pred 1 y Indicates the linear velocity prediction difference in motion data y The weight of the axis, pred 2 x Indicates the predicted difference of linear acceleration in motion data x The weight of the axis, pred 2 y Indicates the predicted difference of linear acceleration in motion data y The weight of the axis, pred 3 x Indicates that the angular velocity prediction difference in the motion data is x The weight of the axis, pred 3 y Indicates that the angular velocity prediction difference in the motion data is y The weight of the axis, gt 1 x Indicates that the linear velocity difference in the second label is x The weight of the axis, gt 1 y Indicates that the linear velocity difference in the second label is y The weight of the axis, gt 2 x Indicates that the midline acceleration difference of the second label is x The weight of the axis, gt 2 y Indicates that the midline acceleration difference of the second label is y The weight of the axis, gt 3 x Indicates that the angular velocity difference in the second tag is x The weight of the axis, gt 3 y Indicates that the angular velocity difference in the second tag is y The weight of the axis, λ 1. λ 2 and λ 3 are weighting coefficients respectively.

7. The vehicle chassis state estimation model training method according to claim 1, characterized in that: Acquiring true value motion data and true value posture data of the vehicle chassis corresponding to the sample motion data, including: Acquire original motion data and original position and posture data of the vehicle chassis corresponding to the sample motion data; The original motion data and the original posture data are low-pass filtered to obtain the true motion data and the true posture data.

8. A vehicle chassis state estimation method, characterized in that: include: Acquiring real-time motion data of the cab and a real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab; Inputting the real-time motion data of the cab and the real-time angle of the steering wheel into a vehicle chassis state estimation model to obtain a real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, and a real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab; Determine the real-time motion data of the chassis based on the real-time difference, and determine the real-time posture data of the chassis based on the real-time offset of the posture data to obtain the real-time motion state of the chassis; Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

9. The vehicle chassis state estimation method according to claim 8, characterized in that: The vehicle chassis state estimation model comprises: a motion feature extraction module, a corner feature extraction module, a cross attention module, a first multi-layer perceptron and a second multi-layer perceptron, wherein the cross attention module is connected to the motion feature extraction module, the corner feature extraction module, the first multi-layer perceptron and the second multi-layer perceptron; The motion feature extraction module is used to extract the motion feature vector of the real-time motion data of the cab; The corner feature extraction module is used to extract the corner feature vector of the real-time corner; The cross attention module is used to make the motion feature vector respond to the correlation of the corner feature vector; The first multi-layer perceptron is used to regress and output the real-time posture data of the vehicle chassis according to the motion feature vector and the rotation angle feature vector after the response correlation; The second multi-layer perceptron is used to regress and output the real-time motion data of the vehicle chassis according to the motion feature vector and the turning angle feature vector after response correlation.

10. The vehicle chassis state estimation method according to claim 9, characterized in that: The sample motion data, true value motion data and real-time motion data respectively include the three-axis velocity, three-axis acceleration and three-axis angular velocity corresponding to each other; The motion feature extraction module includes: a splicing module and a first convolution module, a second convolution module and a third convolution module connected to the splicing module, and the splicing module is connected to the cross attention module; The first convolution module is used to convolve the real-time three-axis velocity corresponding to the real-time motion data of the cab to obtain a three-axis velocity feature vector; The second convolution module is used to convolve the real-time three-axis acceleration corresponding to the real-time motion data of the cab to obtain a three-axis acceleration feature vector; The third convolution module is used to convolve the real-time three-axis angular velocity corresponding to the real-time motion data of the cab to obtain a three-axis angular velocity feature vector; The splicing module is used to splice the three-axis velocity feature vector, the three-axis acceleration feature vector and the three-axis angular velocity feature vector to obtain the motion feature vector.

11. A vehicle chassis state estimation model training device, characterized in that: include: A training data acquisition unit, used to acquire sample motion data of the cab and sample angles of the steering wheel corresponding to the sample motion data, as well as true motion data and true pose data of the vehicle chassis corresponding to the sample motion data; A model training unit, used for inputting the sample motion data and the sample turning angle into a vehicle chassis state estimation model, using the offset of the true value pose data as a first label, and the difference between the true value motion data and the corresponding sample motion data as a second label, to train the vehicle chassis state estimation model, wherein the offset of the true value pose data is the offset of the true value pose data collected twice adjacently; Among them, the vehicle chassis state estimation model is used to estimate the real-time offset of the vehicle chassis posture data and the real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab based on the real-time motion data of the cab and the real-time angle of the steering wheel corresponding to the real-time motion data.

12. A vehicle chassis state estimation device, characterized in that: include: A real-time data acquisition unit, used to acquire real-time motion data of the cab and a real-time steering angle of the steering wheel corresponding to the real-time motion data of the cab; A model execution unit, used for inputting the real-time motion data of the cab and the real-time angle of the steering wheel into a vehicle chassis state estimation model, so as to obtain a real-time offset of the vehicle chassis posture data output by the vehicle chassis state estimation model, and a real-time difference between the real-time motion data of the chassis and the real-time motion data of the cab; A chassis state determination unit, used to determine the real-time motion data of the chassis based on the real-time difference, and determine the real-time posture data of the chassis based on the real-time offset of the posture data, so as to obtain the real-time state of the chassis; Among them, the vehicle chassis state estimation model takes the sample motion data of the cab and the sample angle of the steering wheel corresponding to the sample motion data as input, the offset of the true value posture data of the chassis as the first label, and the difference between the true value motion data of the chassis and the corresponding sample motion data as the second label training, and the offset of the true value posture data is the offset of the true value posture data collected twice adjacently.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the vehicle chassis state estimation model training method as described in any one of claims 1 to 7, or the vehicle chassis state estimation method as described in any one of claims 8 to 10 is implemented.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the vehicle chassis state estimation model training method as described in any one of claims 1 to 7 or the vehicle chassis state estimation method as described in any one of claims 8 to 10 is implemented.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the vehicle chassis state estimation model training method as described in any one of claims 1 to 7 or the vehicle chassis state estimation method as described in any one of claims 8 to 10 is implemented.

Images