A mountain road working condition adaptive prediction method for vehicle chassis pre-control

Abstract

The application relates to the field of vehicle control systems, in particular to a mountain road working condition adaptive prediction method and device for vehicle chassis pre-control, which comprises the following steps: acquiring a continuous video stream of a road ahead of a vehicle, real-time dynamic data of the vehicle, inherent model parameters of the vehicle and external road data; performing time sequence processing on the continuous video stream, extracting road geometric features of the road ahead and predicting a first road surface type probability; performing time synchronization processing on the real-time dynamic data of the vehicle, the inherent model parameters of the vehicle and the external road data, estimating dynamics parameters related to the state of the road ahead, and then mapping a second road surface type probability obtained; adopting a Kalman filter to obtain a final road surface type probability; determining the state of the road ahead; calculating safety constraint parameters of the vehicle; generating a pre-control instruction to adjust a chassis system; and dynamically correcting parameters of the model based on actual driving responses of the vehicle. The method can reduce the risk of losing control of the vehicle during driving.

Description

Technical Field

[0001] This application relates to the field of vehicle control systems, specifically to an adaptive prediction method and device for mountain road conditions for vehicle chassis pre-control. Background Technology

[0002] With the development of the automotive industry, users' requirements for vehicle performance have expanded from power and comfort to a higher demand for driving stability. Traditional chassis systems are gradually proving inadequate in dealing with complex or extreme conditions such as emergency braking and slippery roads, and are unable to actively adjust braking force distribution or steering characteristics based on real-time tire adhesion conditions, thus limiting their effectiveness in active safety.

[0003] On the other hand, advancements in automotive electronics technology have supported the development of intelligent chassis. Various sensors (such as vehicle speed, wheel speed, and yaw rate sensors) can accurately acquire real-time vehicle operating parameters, providing a data foundation for stability control. Therefore, there is an urgent need to utilize intelligent chassis to achieve vehicle driving stability. Summary of the Invention

[0004] To reduce the risk of loss of control of vehicles during driving, this application provides an adaptive prediction method and device for mountain road conditions for vehicle chassis pre-control.

[0005] In a first aspect, embodiments of this application provide an adaptive prediction method for mountain road conditions for vehicle chassis pre-control, including:

[0006] Acquire continuous video streams of the road ahead of the vehicle, real-time dynamic data of the vehicle, inherent model parameters of the vehicle, and external road data;

[0007] The continuous video stream is processed in a time sequence. A deep learning model containing a feature extraction network and a recurrent neural network is used to extract the road geometry features of the road ahead and predict the probability of the first road surface type.

[0008] The system performs time synchronization processing on real-time vehicle dynamic data, vehicle inherent model parameters, and external road data. Based on the vehicle dynamics model, it estimates the dynamic parameters related to the road conditions ahead and then maps the probability of the second road surface type.

[0009] A Kalman filter is used to fuse the first road surface type probability and the second road surface type probability to obtain the final road surface type probability; the road condition ahead is determined based on the final road surface type probability.

[0010] Based on the road conditions ahead and the predicted road geometry, and combined with the current vehicle status, calculate the safety constraint parameters for the vehicle to pass through the road ahead; and generate pre-control commands for adjusting the vehicle chassis system based on the safety constraint parameters.

[0011] Based on the actual driving response of the vehicle under the pre-control command, an estimate of the actual state of the road that has been traversed is generated; the estimate of the actual state is compared with the previously corresponding road state prediction result to generate an error signal, and the parameters of the deep learning model and / or vehicle dynamics model are dynamically corrected using the error signal.

[0012] In one possible implementation, the temporal processing of the continuous video stream, using a deep learning model comprising a feature extraction network and a recurrent neural network, extracts road geometry features of the road ahead and predicts the probability of a first road surface type, including:

[0013] The video frames of the continuous video stream are input into a depth feature extraction network, and visual features are extracted through a ResNet backbone network. Based on the visual features, road geometric parameters, road surface texture type, and depth maps corresponding to each video frame are output; wherein, the depth map includes depth information of lane edges;

[0014] Obtain a sequence of road geometry parameters, road surface texture type, and depth map output from each video frame in chronological order, input the sequence into a long short-term memory network, and predict the road geometry parameters, road surface texture type, and depth information of the road ahead.

[0015] The real-time dynamic data of the vehicle is fused with the road geometry parameters, road surface texture type and depth information predicted by the long short-term memory network. The fused time-series data is then analyzed by an online learning network to output the probability of the first road surface type.

[0016] In one possible implementation, the step of time-synchronizing the vehicle's real-time dynamic data, the vehicle's inherent model parameters, and external road data, and estimating the dynamic parameters related to the road conditions ahead based on the vehicle dynamics model, and then mapping the resulting second road surface type probability, includes:

[0017] Using the acquisition time of each video frame in the continuous video stream as a time reference, the real-time dynamic data of the vehicle, the inherent model parameters of the vehicle, and the external road data are aligned to the same point in time through data interpolation methods to form a time-synchronized snapshot of vehicle status data.

[0018] Based on the vehicle state data snapshot, dynamic parameters related to the road conditions ahead are calculated using a vehicle dynamics model. These dynamic parameters include: vehicle slip ratio S, turning radius R, and road adhesion coefficient. m and slope i ;

[0019] in, , , , ; oh The yaw rate is angular velocity. V For vehicle speed, L This is the length of the ramp. d Indicates the steering angle of the vehicle's front wheels. T The driving or braking torque acting on the wheels, F z The vertical normal load on the tire. h This represents the vertical height difference of the ramp;

[0020] Slip ratio S, turning radius R, and road adhesion coefficient m ,slope i Multimodal fusion is performed to form a feature vector containing temporal information;

[0021] The feature vector is input into a mapping network and processed to obtain the probability of the second road surface type.

[0022] In one possible implementation, the step of employing a Kalman filter to fuse the first road surface type probability and the second road surface type probability to obtain a final road surface type probability; determining the road condition ahead based on the final road surface type probability includes:

[0023] A Kalman filter is used to fuse the probabilities of the first road surface type and the second road surface type to obtain the final road surface type probability value. P M ;

[0024] When the final road surface type probability value P M If the probability exceeds a preset threshold, the road ahead is determined to be a mountain road.

[0025] In one possible implementation, calculating the safety constraint parameters for the vehicle to pass through the road ahead based on the road condition ahead and the predicted road geometry, combined with the current vehicle state, includes:

[0026] Based on the road surface adhesion coefficient in the road condition ahead, the road curvature in the predicted road geometry, and combined with the current vehicle speed and yaw rate, the vehicle dynamics model calculates the maximum speed and / or upper limit of driving force for the vehicle to pass through the road ahead, which serve as safety constraint parameters.

[0027] In one possible implementation, generating pre-control commands for adjusting the vehicle chassis system based on the safety constraint parameters includes:

[0028] Based on the maximum speed and / or maximum driving force values ​​in the safety constraint parameters, instructions are generated to make forward adjustments to at least one of the vehicle's drive system, braking system, steering system, or suspension system to adjust its motion state before the vehicle enters the corresponding road.

[0029] In one possible implementation, the step of calculating the vehicle's maximum speed and / or upper limit of driving force for traversing the road ahead based on the road surface adhesion coefficient in the road condition ahead, the road curvature in the predicted road geometry, and in combination with the current vehicle speed and yaw rate, through a vehicle dynamics model includes:

[0030] If the road curvature in the predicted road geometry is not 0, the upper limit of the driving force for the vehicle to pass through the road ahead is calculated using the friction circle model based on the road surface adhesion coefficient and total tire grip in the road ahead condition; if the road curvature in the predicted road geometry is 0, the upper limit of the driving force for the vehicle to pass through the road ahead is the total tire grip.

[0031] Based on the current vehicle speed, yaw rate, distance from the front and rear axles to the vehicle's center of gravity, centripetal force, and lateral force of the front and rear wheels, the first limit speed of the vehicle passing through the road ahead is calculated using a two-degree-of-freedom vehicle model.

[0032] Based on the current vehicle speed and yaw rate, the vehicle sideslip angle is calculated. Based on the current vehicle speed, yaw rate, vehicle sideslip angle, and vehicle stability region, the second limit speed of the vehicle passing through the road ahead is calculated using a phase diagram model.

[0033] The smaller of the first and second maximum speed limits shall be taken as the maximum speed limit for the vehicle to pass through the road ahead.

[0034] Secondly, embodiments of this application provide an adaptive prediction device for mountain road conditions for vehicle chassis pre-control, comprising:

[0035] The acquisition module is used to acquire continuous video streams of the road ahead of the vehicle, real-time dynamic data of the vehicle, inherent model parameters of the vehicle, and external road data.

[0036] The first road surface type probability prediction module is used to perform time-series processing on the continuous video stream, and extract the road geometric features of the road ahead and predict the probability of the first road surface type through a deep learning model that includes a feature extraction network and a recurrent neural network.

[0037] The second road surface type probability prediction module is used to perform time synchronization processing on real-time vehicle dynamic data, vehicle inherent model parameters and external road data, and based on the vehicle dynamics model, estimate the dynamic parameters related to the road conditions ahead, and then map the second road surface type probability.

[0038] The final road surface type probability prediction module is used to fuse the first road surface type probability and the second road surface type probability using a Kalman filter to obtain the final road surface type probability; and to determine the road condition ahead based on the final road surface type probability.

[0039] The generation module is used to calculate the safety constraint parameters for the vehicle to pass through the road ahead based on the road conditions ahead and the predicted road geometry features, and in combination with the current vehicle state; and to generate pre-control commands for adjusting the vehicle chassis system based on the safety constraint parameters.

[0040] The correction module is used to generate an estimate of the actual state of the road that has been traversed based on the actual driving response of the vehicle under the pre-control command; compare the estimate of the actual state with the previously corresponding road state prediction result to generate an error signal; and use the error signal to dynamically correct the parameters of the deep learning model and / or the vehicle dynamics model.

[0041] In one possible implementation, the first road surface type probability prediction module is specifically used for:

[0042] The video frames of the continuous video stream are input into a depth feature extraction network, and visual features are extracted through a ResNet backbone network. Based on the visual features, road geometric parameters, road surface texture type, and depth maps corresponding to each video frame are output; wherein, the depth map includes depth information of lane edges;

[0043] Obtain a sequence of road geometry parameters, road surface texture type, and depth map output from each video frame in chronological order, input the sequence into a long short-term memory network, and predict the road geometry parameters, road surface texture type, and depth information of the road ahead.

[0044] The real-time dynamic data of the vehicle is fused with the road geometry parameters, road surface texture type and depth information predicted by the long short-term memory network. The fused time-series data is then analyzed by an online learning network to output the probability of the first road surface type.

[0045] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned adaptive prediction methods for mountain road conditions for vehicle chassis pre-control.

[0046] This application provides an adaptive prediction method for mountain road conditions for vehicle chassis pre-control, comprising: acquiring a continuous video stream of the road ahead, real-time vehicle dynamic data, vehicle inherent model parameters, and external road data; performing time-series processing on the continuous video stream, extracting road geometric features of the road ahead and predicting a first road surface type probability using a deep learning model including a feature extraction network and a recurrent neural network; performing time synchronization processing on the real-time vehicle dynamic data, vehicle inherent model parameters, and external road data, and estimating dynamic parameters related to the road ahead state based on the vehicle dynamics model, thereby mapping a second road surface type probability; and using a Kalman filter to determine the probability of the first road surface type. The probability is fused with the second road surface type probability to obtain the final road surface type probability; the road condition ahead is determined based on the final road surface type probability; based on the road condition ahead and the predicted road geometry features, and combined with the current vehicle state, safety constraint parameters for the vehicle to pass through the road ahead are calculated; based on the safety constraint parameters, pre-control commands for adjusting the vehicle chassis system are generated; based on the actual driving response of the vehicle under the pre-control commands, an estimate of the actual state of the already passed road is generated; the actual state estimate is compared with the previously corresponding road condition prediction result to generate an error signal, and the parameters of the deep learning model and / or vehicle dynamics model are dynamically corrected using the error signal. The method of this application can reduce the risk of loss of control of the vehicle during driving. Attached Figure Description

[0047] Figure 1 This is a schematic flowchart of an adaptive prediction method for mountain road conditions for vehicle chassis pre-control, provided in an embodiment of this application.

[0048] Figure 2 An overall architecture diagram for implementing an adaptive prediction method for mountain road conditions for vehicle chassis pre-control, provided in an embodiment of this application;

[0049] Figure 3 A schematic flowchart of another adaptive prediction method for mountain road conditions for vehicle chassis pre-control provided in this application embodiment;

[0050] Figure 4 This is a schematic diagram of the slope prediction method provided in the embodiments of this application;

[0051] Figure 5 This is a schematic diagram of the road surface adhesion coefficient prediction method provided in the embodiments of this application;

[0052] Figure 6 This is a schematic diagram illustrating the calculation of safety constraint parameters provided in an embodiment of this application. Detailed Implementation

[0053] The present invention will be described in detail below through embodiments.

[0054] With the development of the automotive industry, users' requirements for vehicle performance have expanded from power and comfort to a higher demand for driving stability. Traditional chassis systems are gradually proving inadequate in dealing with complex or extreme conditions such as emergency braking and slippery roads, and are unable to actively adjust braking force distribution or steering characteristics based on real-time tire adhesion conditions, thus limiting their effectiveness in active safety.

[0055] On the other hand, advancements in automotive electronics technology have provided support for the development of intelligent chassis. Various sensors (such as vehicle speed, wheel speed, and yaw rate sensors) can accurately acquire real-time vehicle operating parameters, providing a data foundation for achieving stability control. Therefore, there is an urgent need to utilize intelligent chassis to achieve vehicle driving stability.

[0056] Firstly, see [the following] Figure 1 This application provides an adaptive prediction method for mountain road conditions for vehicle chassis pre-control, including:

[0057] S101 acquires continuous video streams of the road ahead of the vehicle, real-time dynamic data of the vehicle, inherent model parameters of the vehicle, and external road data.

[0058] Continuous video streams of the road ahead can be acquired via the vehicle's front-facing camera; real-time vehicle dynamic data is obtained via the CAN (Controller Area Network) bus, including vehicle speed, wheel speed, steering wheel angle, yaw rate, and acceleration; inherent vehicle model parameters are read from the ECU (Electronic Control Unit) calibration file, including vehicle weight, wheelbase, track width, center of gravity height, tire stiffness characteristics, suspension parameters, vehicle length, width, height, engine power and torque characteristics, transmission ratios, and tire size; external road data is obtained through vehicle-to-everything (V2X) networks or high-precision map APIs, including road topology data. All this data collectively forms the foundation for a comprehensive perception of the vehicle's internal and external environment.

[0059] S102, perform time-series processing on the continuous video stream, and extract the road geometry features of the road ahead and predict the probability of the first road surface type through a deep learning model that includes a feature extraction network and a recurrent neural network.

[0060] The video frames of the continuous video stream are input into a depth feature extraction network, and visual features are extracted through a ResNet backbone network. Based on the visual features, road geometric parameters, road surface texture type, and depth maps corresponding to each video frame are output; wherein, the depth map includes depth information of lane edges;

[0061] Obtain a sequence of road geometry parameters, road surface texture type, and depth map output from each video frame in chronological order, input the sequence into a long short-term memory network, and predict the road geometry parameters, road surface texture type, and depth information of the road ahead.

[0062] The real-time dynamic data of the vehicle is fused with the road geometry parameters, road surface texture type and depth information predicted by the long short-term memory network. The fused time-series data is then analyzed by an online learning network to output the probability of the first road surface type.

[0063] The deep feature extraction network is a multi-task network consisting of a ResNet backbone and multiple parallel heads, trained to perform multiple tasks simultaneously. First, the ResNet backbone processes single-frame video images, extracting core visual features. Then, these features are simultaneously fed into different heads for decoding. One head outputs the road's geometric parameters (curvature, slope), another outputs the road surface texture type (asphalt, dirt, etc.), and a third outputs a specific depth map—a depth map the same size as the original image. Each pixel value in this map directly represents the physical distance between that point and the camera, revealing information such as lane boundaries and guardrails. To process consecutive video frames, the processing results of each frame are concatenated into a time series, which is then fed into a separate Long Short-Term Memory (LSTM) network. This network analyzes the data's changes over time to predict future road conditions.

[0064] In one example, the graph-format external road data acquired by the vehicle-to-everything (V2X) network can be input into a deep feature extraction network to obtain the road geometry parameters, road surface texture type, and depth information of the road ahead. This serves as a verification and supplement to the method of extracting the road geometry parameters, road surface texture type, and depth information of the road ahead from continuous video stream processing.

[0065] The probability of the first road surface type is the probability that the road surface ahead is of a certain type.

[0066] S103 performs time synchronization processing on real-time vehicle dynamic data, vehicle inherent model parameters, and external road data, and estimates dynamic parameters related to the road conditions ahead based on the vehicle dynamics model, thereby mapping the probability of the second road surface type.

[0067] Using the acquisition time of each video frame in the continuous video stream as a time reference, the real-time dynamic data of the vehicle, the inherent model parameters of the vehicle, and the external road data are aligned to the same point in time through data interpolation methods to form a time-synchronized snapshot of vehicle status data.

[0068] Based on the vehicle state data snapshot, dynamic parameters related to the road conditions ahead are calculated using a vehicle dynamics model. These dynamic parameters include: vehicle slip ratio S, turning radius R, and road adhesion coefficient. m and slope i ;

[0069] in, , , , ; oh The yaw rate is angular velocity. V For vehicle speed, L This is the length of the ramp. d Indicates the steering angle of the vehicle's front wheels. T The driving or braking torque acting on the wheels, F z The vertical normal load on the tire. h This represents the vertical height difference of the ramp;

[0070] Slip ratio S, turning radius R, and road adhesion coefficient m ,slope i Multimodal fusion is performed to form a feature vector containing temporal information;

[0071] The feature vector is input into a mapping network and processed to obtain the probability of the second road surface type.

[0072] S104, A Kalman filter is used to fuse the first road surface type probability and the second road surface type probability to obtain the final road surface type probability; the road condition ahead is determined based on the final road surface type probability.

[0073] A Kalman filter is used to fuse the probabilities of the first road surface type and the second road surface type to obtain the final road surface type probability value. P M ;

[0074] When the final road surface type probability value P M If the probability exceeds a preset threshold, the road ahead is determined to be a mountain road.

[0075] The preset probability threshold can be 0.6.

[0076] S105, based on the road condition ahead and the predicted road geometry, and combined with the current vehicle condition, calculate the safety constraint parameters for the vehicle to pass through the road ahead; and generate pre-control commands for adjusting the vehicle chassis system according to the safety constraint parameters.

[0077] Based on the road surface adhesion coefficient in the road condition ahead, the road curvature in the predicted road geometry, and combined with the current vehicle speed and yaw rate, the vehicle dynamics model calculates the maximum speed and / or upper limit of driving force for the vehicle to pass through the road ahead, which serve as safety constraint parameters.

[0078] Specifically, if the road curvature in the predicted road geometry is not 0, the upper limit of the driving force for the vehicle to pass through the road ahead is calculated using the friction circle model based on the road surface adhesion coefficient and total tire grip in the road ahead condition; if the road curvature in the predicted road geometry is 0, the upper limit of the driving force for the vehicle to pass through the road ahead is the total tire grip.

[0079] Based on the current vehicle speed, yaw rate, distance from the front and rear axles to the vehicle's center of gravity, centripetal force, and lateral force of the front and rear wheels, the first limit speed of the vehicle passing through the road ahead is calculated using a two-degree-of-freedom vehicle model.

[0080] Based on the current vehicle speed and yaw rate, the vehicle sideslip angle is calculated. Based on the current vehicle speed, yaw rate, vehicle sideslip angle, and vehicle stability region, the second limit speed of the vehicle passing through the road ahead is calculated using a phase diagram model.

[0081] The smaller of the first and second maximum speed limits shall be taken as the maximum speed limit for the vehicle to pass through the road ahead.

[0082] Road curvature describes the bending of a road; a curvature of 0 indicates a straight road segment. Total tire grip is constant; in the friction circle model, the radius represents the total tire grip. When a vehicle is turning, the total tire grip needs to be simultaneously distributed to the lateral force used for steering. F y and longitudinal force used for acceleration and deceleration F x When a vehicle is turning (with lateral forces), the longitudinal force available for acceleration and deceleration decreases. Based on the lateral force and the road surface adhesion coefficient, it can be calculated how much longitudinal force (driving force or braking force) the tires can provide without slipping in this situation.

[0083] The two-degree-of-freedom model is based on the current vehicle speed, yaw rate, and vehicle geometry parameters (such as the distance from the front and rear axles to the center of gravity). l f and l r Using Newton's laws, analyze the centripetal force when a vehicle is turning. F cf Lateral forces of the front and rear wheels F f , F rThe balance between these factors is calculated to determine the maximum safe speed for a vehicle to maintain a stable tracking posture (without sideslip) under the current turning conditions.

[0084] Because the rear wheels tend to steer excessively, considering only them would yield a maximum speed. First, the vehicle's sideslip angle is calculated based on the current vehicle speed and yaw rate. Then, the current vehicle speed, yaw rate, and sideslip angle are fed together into a phase diagram model for analysis. The phase diagram describes the vehicle's dynamic stability, defining a closed curved region representing the vehicle's stability domain. As long as the vehicle's state remains within this region, it is stable; if it exceeds the boundary, it will lose control (e.g., fishtailing). Slippery road surfaces or excessively high speeds will cause this stability domain to contract, making the vehicle more prone to loss of control.

[0085] Based on the maximum speed and / or maximum driving force values ​​in the safety constraint parameters, instructions are generated to make forward adjustments to at least one of the vehicle's drive system, braking system, steering system, or suspension system to adjust its motion state before the vehicle enters the corresponding road.

[0086] S106, Based on the actual driving response of the vehicle under the pre-control command, generate an estimate of the actual state of the road that has been traversed; compare the estimate of the actual state with the previously corresponding road state prediction result to generate an error signal, and use the error signal to dynamically correct the parameters of the deep learning model and / or the vehicle dynamics model.

[0087] This application obtains a more realistic reference value by utilizing the vehicle's actual physical response, quantifies the gap between the reference value and the road state prediction results output by the model, and uses this gap as a driving force to make a small, directional adjustment to the weight parameters within the deep learning model and / or vehicle dynamics model through a gradient descent algorithm. This cycle is repeated continuously at an extremely high frequency (e.g., tens of times per second), the effect of which is that the deep learning model and / or vehicle dynamics model are fine-tuning and self-learning every kilometer the vehicle travels. It becomes increasingly adapted to the specific characteristics of the vehicle (such as tire wear, load changes) and the road conditions it frequently travels on, thus achieving true adaptive evolution.

[0088] This application provides an adaptive prediction method for mountain road conditions for vehicle chassis pre-control, comprising: acquiring a continuous video stream of the road ahead, real-time vehicle dynamic data, vehicle inherent model parameters, and external road data; performing time-series processing on the continuous video stream, extracting road geometric features of the road ahead and predicting a first road surface type probability using a deep learning model including a feature extraction network and a recurrent neural network; performing time synchronization processing on the real-time vehicle dynamic data, vehicle inherent model parameters, and external road data, and estimating dynamic parameters related to the road ahead state based on the vehicle dynamics model, thereby mapping a second road surface type probability; and using a Kalman filter to determine the probability of the first road surface type. The probability is fused with the second road surface type probability to obtain the final road surface type probability; the road condition ahead is determined based on the final road surface type probability; based on the road condition ahead and the predicted road geometry features, and combined with the current vehicle state, safety constraint parameters for the vehicle to pass through the road ahead are calculated; based on the safety constraint parameters, pre-control commands for adjusting the vehicle chassis system are generated; based on the actual driving response of the vehicle under the pre-control commands, an estimate of the actual state of the already passed road is generated; the actual state estimate is compared with the previously corresponding road condition prediction result to generate an error signal, and the parameters of the deep learning model and / or vehicle dynamics model are dynamically corrected using the error signal. The method of this application can reduce the risk of loss of control of the vehicle during driving.

[0089] See Figure 2 This is an overall architecture diagram of this application. The left side includes three branches: forward road information detection, vehicle dynamic state extraction, and vehicle dynamic inspection. Forward road information detection is the foundation for vehicle dynamic stability pre-control. After road time limit prediction and vehicle dynamic stability prediction, forward road information detection enters vehicle dynamic stability pre-control. Vehicle dynamic state extraction, after road time limit estimation, enters vehicle dynamic stability forward control. Based on vehicle dynamic state extraction and vehicle dynamic inspection, vehicle dynamic stability feedback control is achieved. Arbitration / coordination is performed based on these three branches, outputting vehicle dynamic control commands (including brake distribution and steering correction parameters), road risk level assessment results (divided into high, medium, and low levels), and real-time stability control strategy suggestions (such as ESP intervention threshold adjustment).

[0090] Unlike traditional vehicle stability systems that passively intervene when the vehicle is about to lose control, this invention utilizes predicted road conditions (such as curvature, gradient, and coefficient of friction) to achieve pre-control. It is no longer a passive response but a proactive measure: before entering a dangerous situation (such as a slippery, sharp bend), it pre-emptively and smoothly adjusts the braking, power, steering, and suspension to bring the vehicle to its safest and most stable state. For example, it gently decelerates in advance, actively assists steering, pre-manages torque output, and pre-sets the optimal anti-roll state for the suspension. This pre-control approach nipps the risk of loss of control in the bud, greatly improving the vehicle's safety, comfort, and driving smoothness in complex road conditions, achieving a fundamental upgrade from reactive rescue to predictive avoidance.

[0091] See Figure 3 This is a flowchart illustrating the road state adaptive prediction method for vehicle chassis pre-control provided in this application. The visual perception module includes three parts: feature extraction, parallel decoding, and time-series analysis. It uses video input, real-time dynamic data of the vehicle collected by the IMU (Inertial Measurement Unit), and external road data. V wheel As input, the visual perception module outputs road geometry detection (slope, curvature), road texture classification (asphalt, soil, etc.), lane boundaries, and guardrail information, all containing temporal information. Then, the output of the visual perception module is fused with real-time vehicle dynamic data. This fused temporal data is analyzed using a more advanced network (LSTM or RNN) to output the final prediction result, such as the probability of mountain road surface type. P M1 This can also be based on external road data. V wheel The system acquires information on the road surface geometry (slope, curvature), road surface texture classification (asphalt, soil, etc.), lane boundaries, and guardrails ahead, which serves as a supplement and verification to the video detection results. If the data is video, it uses a visual perception module to extract features, perform parallel decoding, and analyze time series data to predict the road surface geometry (slope, curvature), road surface texture classification (asphalt, soil, etc.), lane boundaries, and guardrails ahead.

[0092] Vehicle information parameters include real-time vehicle dynamic data, inherent vehicle model parameters, and external road data. These parameters, related to the road conditions ahead, are calculated by the vehicle dynamics estimation module, including: vehicle slip ratio S, turning radius R, and road adhesion coefficient. m and slope i Then, these parameters are fused using multimodal methods to form a feature vector from the time series. Mapped from feature vectors PM2 Kalman fusion is performed based on these two road surface type probabilities to obtain the final road surface type probability. If this probability is greater than 0.6, the road condition is determined to be a mountain road.

[0093] In this application, the parameters of the online network will also be updated through backpropagation based on the predicted road surface type probability.

[0094] This application introduces autonomous and online learning mechanisms to enable the vehicle perception system to predict road conditions in real time and adjust the chassis system accordingly. By dynamically extracting vehicle operating parameters and identifying and analyzing vehicle-road features, the system can predict future driving routes in advance. Simultaneously, the system continuously updates model parameters based on a feedback learning mechanism, thereby improving the accuracy of road surface type recognition. Through these methods, the intelligent chassis's predictive control can anticipate road types, enhancing the vehicle's stability control during driving and providing the driver with a safer and smoother driving experience.

[0095] See Figure 4 This diagram illustrates a slope prediction process. Firstly, it takes video input as input and performs feature extraction, including edge pixel depth extraction and prediction via a depth estimation network. The edge pixel depth and prediction are then fused across multiple modules to output the slope projection length *a*. Next, edge extraction is performed to generate a sharpened image, and the slope length *L* is output via coordinate transformation. Based on *a* and *L*, the slope angle is predicted. i 1. On the other hand, output through vehicle dynamics model i 2. Kalman calculus is based on the fusion of two slope angles to obtain the final result. i In this application, the parameters of the online network will also be updated through backpropagation based on the predicted slope angle.

[0096] See Figure 5 This is a schematic diagram of the road surface adhesion coefficient prediction process provided in this application. Firstly, it uses video input to extract road surface conditions and fuse road surface feature data to obtain a road surface feature space. Based on the road surface feature space at different times, different road surface information networks are obtained. Then, road surface prediction is performed based on these different road surface information networks, the road surface feature space, and the road surface conditions. Finally, the road surface prediction and the different road surface information networks are passed through a probe network to obtain... On the other hand, the vehicle motion model predicts outputs based on vehicle parameters. Kalman derives the final result based on the fusion of these two road surface adhesion coefficients. m In this application, the parameters of the online network will also be updated through backpropagation based on the predicted road surface adhesion coefficient.

[0097] See Figure 6This application provides a vehicle motion pre-control method, lateral force F y and predicted m futr Input to total tire grip F lim In the friction circle model, the upper limit of the driving force is obtained. Current vehicle speed V cuur and yaw rate oh It will be fed into a two-degree-of-freedom vehicle model, which is based on the vehicle's geometric parameters (such as the distance from the front and rear axles to the center of gravity). l f and l r Using Newton's laws, analyze the centripetal force when a vehicle is turning. F cf Lateral forces of the front and rear wheels F f , F r The balance between these factors is calculated to determine the maximum safe speed for the vehicle to maintain a stable tracking posture (without sideslip) under the current turning condition. Since the rear wheels tend to steer excessively, considering the rear wheels alone also yields a speed limit. First, the current speed... V cuur and yaw rate oh Sent to a place called β calc The module is used to calculate the predicted vehicle center of gravity sideslip angle. β pred ,Then, m futr 、V cuur ,ω,b pred They are fed into a phase diagram model for analysis, which describes the vehicle's dynamic stability. The closed curve region in the diagram represents the vehicle's stability domain. As long as the vehicle state (by...) β pred and oh As long as the defined point (located within this area) is within the vehicle's range, the vehicle will be stable. If it crosses the boundary, it will lose control (e.g., fishtail). The text below the diagram explains: m A decrease in velocity (v) indicates decreased stability; a decrease in velocity (v) indicates increased stability. This means that slippery road surfaces or excessive speeds will cause this stability region to shrink, making the vehicle more prone to loss of control. By analyzing the vehicle's current position in the phase diagram, the maximum safe speed required to prevent the vehicle from entering an unstable state (especially excessive steering, where the lateral force of the rear wheels cannot meet the centrifugal force, resulting in sideslip) can be calculated.

[0098] Then calculate the upper limit of the driving force. and two extreme speeds V lim Arbitration / coordination will be conducted to obtain a final result.

[0099] Secondly, embodiments of this application provide an adaptive prediction device for mountain road conditions for vehicle chassis pre-control, comprising:

[0100] The acquisition module is used to acquire continuous video streams of the road ahead of the vehicle, real-time dynamic data of the vehicle, inherent model parameters of the vehicle, and external road data.

[0101] The first road surface type probability prediction module is used to perform time-series processing on the continuous video stream, and extract the road geometric features of the road ahead and predict the probability of the first road surface type through a deep learning model that includes a feature extraction network and a recurrent neural network.

[0102] The second road surface type probability prediction module is used to perform time synchronization processing on real-time vehicle dynamic data, vehicle inherent model parameters and external road data, and based on the vehicle dynamics model, estimate the dynamic parameters related to the road conditions ahead, and then map the second road surface type probability.

[0103] The final road surface type probability prediction module is used to fuse the first road surface type probability and the second road surface type probability using a Kalman filter to obtain the final road surface type probability; and to determine the road condition ahead based on the final road surface type probability.

[0104] The generation module is used to calculate the safety constraint parameters for the vehicle to pass through the road ahead based on the road conditions ahead and the predicted road geometry features, and in combination with the current vehicle state; and to generate pre-control commands for adjusting the vehicle chassis system based on the safety constraint parameters.

[0105] The correction module is used to generate an estimate of the actual state of the road that has been traversed based on the actual driving response of the vehicle under the pre-control command; compare the estimate of the actual state with the previously corresponding road state prediction result to generate an error signal; and use the error signal to dynamically correct the parameters of the deep learning model and / or the vehicle dynamics model.

[0106] In one possible implementation, the first road surface type probability prediction module is specifically used for:

[0107] The video frames of the continuous video stream are input into a depth feature extraction network, and visual features are extracted through a ResNet backbone network. Based on the visual features, road geometric parameters, road surface texture type, and depth maps corresponding to each video frame are output; wherein, the depth map includes depth information of lane edges;

[0108] Obtain a sequence of road geometry parameters, road surface texture type, and depth map output from each video frame in chronological order, input the sequence into a long short-term memory network, and predict the road geometry parameters, road surface texture type, and depth information of the road ahead.

[0109] The real-time dynamic data of the vehicle is fused with the road geometry parameters, road surface texture type and depth information predicted by the long short-term memory network. The fused time-series data is then analyzed by an online learning network to output the probability of the first road surface type.

[0110] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned adaptive prediction methods for mountain road conditions for vehicle chassis pre-control.

[0111] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0112] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0113] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are described simply because their systems are similar to the method embodiments; relevant parts can be referred to the descriptions of the method embodiments.

[0114] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A mountain road working condition adaptive prediction method for vehicle chassis pre-control, characterized in that, include: Acquire continuous video streams of the road ahead of the vehicle, real-time dynamic data of the vehicle, inherent model parameters of the vehicle, and external road data; The continuous video stream is processed in a time sequence. A deep learning model containing a feature extraction network and a recurrent neural network is used to extract the road geometry features of the road ahead and predict the probability of the first road surface type. The system performs time synchronization processing on real-time vehicle dynamic data, vehicle inherent model parameters, and external road data. Based on the vehicle dynamics model, it estimates the dynamic parameters related to the road conditions ahead and then maps the probability of the second road surface type. A Kalman filter is used to fuse the first road surface type probability and the second road surface type probability to obtain the final road surface type probability; the road condition ahead is determined based on the final road surface type probability. Based on the road conditions ahead and the predicted road geometry, and combined with the current vehicle status, calculate the safety constraint parameters for the vehicle to pass through the road ahead; and generate pre-control commands for adjusting the vehicle chassis system based on the safety constraint parameters. Based on the actual driving response of the vehicle under the pre-control command, an estimate of the actual state of the road that has been traversed is generated. The actual state estimate is compared with the previously corresponding road state prediction result to generate an error signal, and the parameters of the deep learning model and / or vehicle dynamics model are dynamically corrected using the error signal. The step of performing time-series processing on the continuous video stream, using a deep learning model including a feature extraction network and a recurrent neural network, to extract road geometry features of the road ahead and predict the probability of a first road surface type, includes: The video frames of the continuous video stream are input into a depth feature extraction network, and visual features are extracted through a ResNet backbone network. Based on the visual features, road geometric parameters, road surface texture type, and depth maps corresponding to each video frame are output; wherein, the depth map includes depth information of lane edges; Obtain a sequence of road geometry parameters, road surface texture type, and depth map output from each video frame in chronological order, input the sequence into a long short-term memory network, and predict the road geometry parameters, road surface texture type, and depth information of the road ahead. The real-time dynamic data of the vehicle is fused with the road geometry parameters, road surface texture type and depth information predicted by the long short-term memory network. The fused time series data is then analyzed by an online learning network to output the probability of the first road surface type. The process of time-synchronizing vehicle real-time dynamic data, vehicle inherent model parameters, and external road data, and estimating dynamic parameters related to the road condition ahead based on the vehicle dynamics model, and then mapping the resulting second road surface type probability, includes: Using the acquisition time of each video frame in the continuous video stream as a time reference, the real-time dynamic data of the vehicle, the inherent model parameters of the vehicle, and the external road data are aligned to the same point in time through data interpolation methods to form a time-synchronized snapshot of vehicle status data. Based on the vehicle state data snapshot, a dynamics parameter related to the front road state is calculated through a vehicle dynamics model, the dynamics parameter including: a slip ratio S of the vehicle, a steering radius R, a road adhesion coefficient μ and a slope θ ​ in, , , , ; ω The yaw rate is angular velocity. V For vehicle speed, L This is the length of the ramp. δ Indicates the steering angle of the vehicle's front wheels. T The driving or braking torque acting on the wheels, F z The vertical normal load on the tire. h This represents the vertical height difference of the ramp; Slip ratio S, turning radius R, and road adhesion coefficient μ ,slope θ Multimodal fusion is performed to form a feature vector containing temporal information; The feature vector is input into a mapping network and processed to obtain the probability of the second road surface type.

2. The method according to claim 1, characterized in that, The Kalman filter is used to fuse the first road surface type probability and the second road surface type probability to obtain the final road surface type probability. Determining the road condition ahead based on the final road surface type probability includes: A Kalman filter is used to fuse the probabilities of the first road surface type and the second road surface type to obtain the final road surface type probability value. P M ; When the final road surface type probability value P M If the probability exceeds a preset threshold, the road ahead is determined to be a mountain road.

3. The method according to claim 1, characterized in that, The calculation of safety constraint parameters for the vehicle to pass through the road ahead, based on the road condition ahead and predicted road geometry, and in conjunction with the current vehicle state, includes: Based on the road surface adhesion coefficient in the road condition ahead, the road curvature in the predicted road geometry, and combined with the current vehicle speed and yaw rate, the vehicle dynamics model calculates the maximum speed and / or upper limit of driving force for the vehicle to pass through the road ahead, which serve as safety constraint parameters.

4. The method according to claim 3, characterized in that, The step of generating pre-control commands for adjusting the vehicle chassis system based on the safety constraint parameters includes: Based on the maximum speed and / or maximum driving force values ​​in the safety constraint parameters, instructions are generated to make forward adjustments to at least one of the vehicle's drive system, braking system, steering system, or suspension system to adjust its motion state before the vehicle enters the corresponding road.

5. The method according to claim 3, characterized in that, The method of calculating the vehicle's maximum speed and / or upper limit of driving force for traversing the road ahead based on the road surface adhesion coefficient in the road condition ahead, the road curvature in the predicted road geometry, and in combination with the current vehicle speed and yaw rate, through a vehicle dynamics model, includes: If the road curvature in the predicted road geometry is not 0, the upper limit of the driving force for the vehicle to pass through the road ahead is calculated using the friction circle model based on the road surface adhesion coefficient and total tire grip in the road ahead condition; if the road curvature in the predicted road geometry is 0, the upper limit of the driving force for the vehicle to pass through the road ahead is the total tire grip. Based on the current vehicle speed, yaw rate, distance from the front and rear axles to the vehicle's center of gravity, centripetal force, and lateral force of the front and rear wheels, the first limit speed of the vehicle passing through the road ahead is calculated using a two-degree-of-freedom vehicle model. Based on the current vehicle speed and yaw rate, the vehicle sideslip angle is calculated. Based on the current vehicle speed, yaw rate, vehicle sideslip angle, and vehicle stability region, the second limit speed of the vehicle passing through the road ahead is calculated using a phase diagram model. The smaller of the first and second maximum speed limits shall be taken as the maximum speed limit for the vehicle to pass through the road ahead.

6. An adaptive prediction device for mountain road conditions for vehicle chassis pre-control, characterized in that, include: The acquisition module is used to acquire continuous video streams of the road ahead of the vehicle, real-time dynamic data of the vehicle, inherent model parameters of the vehicle, and external road data. The first road surface type probability prediction module is used to perform time-series processing on the continuous video stream, and extract the road geometric features of the road ahead and predict the probability of the first road surface type through a deep learning model that includes a feature extraction network and a recurrent neural network. The second road surface type probability prediction module is used to perform time synchronization processing on real-time vehicle dynamic data, vehicle inherent model parameters and external road data, and based on the vehicle dynamics model, estimate the dynamic parameters related to the road conditions ahead, and then map the second road surface type probability. The final road surface type probability prediction module is used to fuse the first road surface type probability and the second road surface type probability using a Kalman filter to obtain the final road surface type probability; and to determine the road condition ahead based on the final road surface type probability. The generation module is used to calculate the safety constraint parameters for the vehicle to pass through the road ahead based on the road conditions ahead and the predicted road geometry features, and in combination with the current vehicle status. Based on the safety constraint parameters, generate pre-control commands for adjusting the vehicle chassis system; The correction module is used to generate an estimate of the actual state of the road that has been traversed based on the actual driving response of the vehicle under the pre-control command. The actual state estimate is compared with the previously corresponding road state prediction result to generate an error signal, and the parameters of the deep learning model and / or vehicle dynamics model are dynamically corrected using the error signal. The first road surface type probability prediction module is specifically used for: The video frames of the continuous video stream are input into a depth feature extraction network, and visual features are extracted through a ResNet backbone network. Based on the visual features, road geometric parameters, road surface texture type, and depth maps corresponding to each video frame are output; wherein, the depth map includes depth information of lane edges; Obtain a sequence of road geometry parameters, road surface texture type, and depth map output from each video frame in chronological order, input the sequence into a long short-term memory network, and predict the road geometry parameters, road surface texture type, and depth information of the road ahead. The real-time dynamic data of the vehicle is fused with the road geometry parameters, road surface texture type and depth information predicted by the long short-term memory network. The fused time series data is then analyzed by an online learning network to output the probability of the first road surface type. The second road surface type probability prediction module is specifically used for: Using the acquisition time of each video frame in the continuous video stream as a time reference, the real-time dynamic data of the vehicle, the inherent model parameters of the vehicle, and the external road data are aligned to the same point in time through data interpolation methods to form a time-synchronized snapshot of vehicle status data. Based on the vehicle state data snapshot, dynamic parameters related to the road conditions ahead are calculated using a vehicle dynamics model. These dynamic parameters include: vehicle slip ratio S, turning radius R, and road adhesion coefficient. μ and slope θ ; in, , , , ; ω The yaw rate is angular velocity. V For vehicle speed, L This is the length of the ramp. δ Indicates the steering angle of the vehicle's front wheels. T The driving or braking torque acting on the wheels, F z The vertical normal load on the tire. h This represents the vertical height difference of the ramp; Slip ratio S, turning radius R, and road adhesion coefficient μ ,slope θ Multimodal fusion is performed to form a feature vector containing temporal information; The feature vector is input into a mapping network and processed to obtain the probability of the second road surface type.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-5.

Images