A dynamic prediction method for safe vehicle speed on ice and snow road surface based on load transfer correction

Abstract

The application discloses a kind of ice and snow road safety vehicle speed dynamic prediction method based on load transfer correction, it is related to intelligent transportation, vehicle-road cooperation and vehicle active safety technical field.The method is first obtained by vehicle-mounted or roadsides visual perception equipment continuous ice and snow road image sequence, and pre-processing and time sequence sample construction are carried out;Then build by spatial feature extraction unit, ice and snow texture attention enhancement unit, time sequence feature modeling unit, time sequence attention unit and friction coefficient regression unit sequentially connected form double attention enhancement friction coefficient prediction model, and determine the final friction coefficient prediction model by training set training, verification set verification, determine model prediction error parameter by test set test.Real-time prediction, input continuous ice and snow road image sequence final friction coefficient prediction model, obtain current road friction coefficient prediction value, and conservative friction coefficient is obtained after filtering smoothing post-processing and safety side correction;Further combine vehicle barycenter height, wheelbase, track, weather information, available stopping distance, road curvature and driving condition, respectively establish load transfer correction safety vehicle speed calculation model under straight braking condition and turning condition, and dynamically output suggested safety vehicle speed.The application can improve the stability of ice and snow road friction coefficient prediction and the vehicle adaptability of safety vehicle speed calculation, and can provide real-time safety vehicle speed prediction for driver, intelligent driving system or vehicle-road cooperation platform.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation, vehicle-road cooperation and active vehicle safety technology, specifically to a method for predicting safe vehicle speed on icy and snowy roads based on road image temporal perception, dual-attention friction coefficient prediction and load transfer correction. Background Technology

[0002] Icy and snowy roads significantly reduce the coefficient of friction between tires and the road surface, making vehicles prone to problems such as increased braking distance, skidding, fishtailing, and even loss of control during braking, turning, and obstacle avoidance. Therefore, accurately sensing the adhesion status of icy and snowy roads and providing reasonable safe speed recommendations for vehicles is an urgent problem to be solved in the fields of intelligent transportation and active vehicle safety.

[0003] In existing technologies, some methods identify road surface conditions using cameras and deep learning models, classifying roads into categories such as dry, slippery, snow-covered, or icy. These methods can assess road risk conditions, but the output is mostly discrete categories, making it difficult to directly provide a continuous friction coefficient for safe speed calculations. Other methods, while capable of predicting the friction coefficient, often ignore the temporal variations in ice and snow textures and are disconnected from safe speed calculations, failing to form a complete closed loop.

[0004] Furthermore, existing methods for calculating safe vehicle speeds often employ general dynamic models, neglecting structural differences between vehicles (such as center of gravity height, wheelbase, and track width). Under conditions of low adhesion on ice and snow, vehicle braking or steering will result in significant load transfer. If a uniform safe vehicle speed model is used, it is highly likely to lead to inaccurate predictions of the safety boundaries for vehicles with a high center of gravity.

[0005] Current technologies lack a friction coefficient prediction model that combines spatial and temporal attention enhancement, while also considering load transfer and vehicle parameters to calculate safe vehicle speeds on icy and snowy roads. Therefore, a method for predicting safe vehicle speeds on icy and snowy roads that integrates temporal perception of road images, dual-attention friction coefficient prediction, vehicle parameter matching, and load transfer correction is particularly important. Summary of the Invention

[0006] This invention addresses the problems in existing technologies, such as the separation of icy and snowy road surface condition recognition and safe vehicle speed calculation, insufficient utilization of key textures and temporal changes in icy and snowy road surfaces in friction coefficient prediction, lack of vehicle parameter correction in traditional safe vehicle speed models, and insufficient consideration of load transfer effects on low-adhesion road surfaces. It provides a method for predicting safe vehicle speed on icy and snowy roads, providing a theoretical basis for vehicle-mounted real-time early warning systems.

[0007] This invention constructs a road surface friction coefficient prediction model with dual attention enhancement to extract spatial texture features and model temporal states from continuous icy and snowy road surface images. Model parameters are learned using a training set, the model is validated using a validation set to determine the final friction coefficient prediction model, and the model prediction error parameters are determined using a test set. In the real-time prediction stage, the current road surface friction coefficient prediction value output by the final friction coefficient prediction model is filtered, smoothed, and corrected for safety factors to obtain a conservative friction coefficient. Then, based on vehicle parameter databases, weather information, road parameters, and driving conditions, safe vehicle speeds are calculated for straight-line braking and turning conditions, respectively, thereby achieving safe vehicle speed outputs for different vehicle types, weather conditions, and road conditions.

[0008] The technical solution of this invention is: A dynamic prediction method for safe vehicle speed on icy and snowy roads based on load transfer correction includes the following steps: Step 1: Collect images of icy and snowy road surfaces as sample images, preprocess the sample images to obtain an icy and snowy road surface image dataset; construct a continuous sequence of icy and snowy road surface images in chronological order, and match the continuous sequence of icy and snowy road surface images with the measured road surface friction coefficient at the corresponding time to form a sample dataset with friction coefficient labels; divide the icy and snowy road surface image dataset into a training set, a validation set, and a test set.

[0009] Step 1.1: Preprocess the acquired road surface image, remove abnormal images, adjust the valid image to the preset size, and normalize the pixel values ​​to the preset range.

[0010] Step 1.2: Divide the continuous images into sliding windows according to the time sequence to construct an image sequence sample containing multiple continuous time steps. The image sequence sample is used to characterize the process of the change of the state of the icy and snowy road surface over time.

[0011] Step 2: Construct a dual-attention enhanced friction coefficient prediction model. This model starts from the input layer and sequentially connects to a spatial feature extraction unit, an ice and snow texture attention enhancement unit, a temporal feature modeling unit, a temporal attention unit, and a friction coefficient regression unit. The spatial feature extraction unit extracts spatial feature maps of the ice and snow road surface image frame by frame. The ice and snow texture attention enhancement unit enhances ice and snow texture regions related to the road friction coefficient. The temporal feature modeling unit extracts dynamic features of the ice and snow road surface state changing over time in a continuous image sequence. The temporal attention unit assigns time weights based on the contribution of different time steps to the current friction coefficient prediction result. The friction coefficient regression unit outputs the predicted friction coefficient values ​​for the continuous road surface. These predicted friction coefficient values ​​are then input to a filtering and smoothing post-processing unit for filtering and smoothing.

[0012] Step 2.1, Spatial Feature Extraction. MobileNetV2 is used as a lightweight convolutional neural network deep learning method to extract spatial texture features frame-by-frame from the input sequence of continuous icy and snowy road surface images, and pre-trained weights are loaded to enhance generalization ability. The initial convolutional layers are frozen, and only a few subsequent convolutional layers are fine-tuned to adapt to the features of the icy and snowy road surface.

[0013] Step 2.2, Snow and Ice Texture Attention Enhancement Unit. A CBAM attention structure is adopted, including channel attention branches and spatial attention branches, to enhance key texture features of snow and ice surfaces (such as snow accumulation, compacted snow, ice surface reflection, snow melt water film, and tire tracks), while suppressing background areas and irrelevant noise.

[0014] Step 2.3, Temporal Feature Construction. Global pooling is performed on the feature maps enhanced with attention to the ice and snow texture to obtain spatial feature vectors corresponding to each time step. These vectors are then combined into a temporal feature matrix according to the order of image acquisition time.

[0015] Step 2.4, Temporal Feature Modeling. The temporal feature matrix is ​​input into a Long Short-Term Memory (LSTM) network. A gating mechanism is used to model the change in the state of the icy and snowy road surface over time, obtaining the temporal hidden states corresponding to multiple time steps.

[0016] Step 2.5, Temporal Attention Unit. A Temporal Attention structure is employed at the output of the Long Short-Term Memory network. Temporal weights are assigned based on the contribution of the hidden states at different time steps to the current friction coefficient prediction, and the hidden states at each time step are weighted and fused to obtain temporal fusion features. This temporal attention unit reduces the impact of sudden changes in illumination in a single frame, snow reflections, local occlusion, or acquisition noise on the prediction results.

[0017] Step 2.6, Friction Coefficient Regression Unit. The time-series fusion features are input into the fully connected regression layer, which outputs continuous predicted values ​​of the road surface friction coefficient.

[0018] Step 2.7, Post-processing unit for filtering and smoothing. The predicted current road surface friction coefficient output by the final friction coefficient prediction model is smoothed using Kalman filtering to obtain the filtered friction coefficient.

[0019] Step 3: Train the dual-attention enhanced friction coefficient prediction model using the training set to obtain the trained dual-attention enhanced friction coefficient prediction model; validate the trained dual-attention enhanced friction coefficient prediction model using the validation set to obtain the final friction coefficient prediction model.

[0020] Step 4: Input the test set into the final friction coefficient prediction model, output the predicted value of the continuous road friction coefficient corresponding to the test sample, and compare the predicted value of the continuous road friction coefficient with the measured road friction coefficient to determine the model prediction error parameter.

[0021] Step 5: Input the real-time acquired continuous icy and snowy road surface image sequence into the final friction coefficient prediction model to obtain the current road surface friction coefficient prediction value; input the current road surface friction coefficient prediction value into the filtering and smoothing post-processing unit for filtering and smoothing to obtain the filtered friction coefficient; based on the model prediction error parameters obtained during model training, verification, or testing, perform a safety-side correction on the filtered friction coefficient to obtain a conservative friction coefficient. The conservative friction coefficient is expressed as:

[0022] in, For a conservative coefficient of friction, The friction coefficient after filtering. These are the model prediction error parameters. The model prediction error parameters can be the mean absolute error (MAE), root mean square error (RMSE), prediction uncertainty, or a preset safety correction.

[0023] Step 6: Obtain vehicle parameters, weather information, road parameters, and vehicle driving status information. The vehicle parameters include at least the vehicle center of gravity height, wheelbase, and track width. The track width parameters include the front track width, rear track width, and / or the average track width. The weather information is used to determine the maximum usable stopping distance under the current weather conditions. The road parameters include at least the road curvature or road turning radius.

[0024] Step 6-1: Obtain vehicle parameters. Based on the vehicle model entered by the user or the vehicle database matching results, obtain the vehicle's center of gravity height, wheelbase, front track, rear track, and average track width.

[0025] Step 6-2: Obtain weather information and determine the visible sensing distance. Weather information is obtained through user input, vehicle-mounted sensors, roadside weather stations, internet weather interfaces, or vehicle-road cooperative platforms. The maximum visible distance under the current weather conditions is mapped to the maximum available stopping distance. The worse the weather conditions, the smaller the value of the available stopping distance d, which serves as a safety constraint to further reduce the recommended safe speed.

[0026] Step 6-3: Obtain road parameters. Obtain road curvature, turning radius, slope, or speed limit information through high-precision maps, in-vehicle navigation systems, roadside sensors, visual recognition modules, or user input.

[0027] Step 7: Determine whether the vehicle is currently in a straight-ahead or turning condition based on the vehicle driving status information and the road parameters; when the road curvature is less than a preset curvature threshold, the steering wheel angle is less than a preset angle threshold, and the yaw rate is less than a preset yaw threshold, the vehicle is determined to be in a straight-ahead condition; when the road curvature is not less than the preset curvature threshold, the steering wheel angle is not less than the preset angle threshold, or the yaw rate is not less than the preset yaw threshold, the vehicle is determined to be in a turning condition.

[0028] Step 8: Based on the conservative friction coefficient, vehicle parameters, weather information, road parameters, and the vehicle's current driving conditions, call the load transfer correction safe speed calculation model to calculate the final recommended safe speed; when the vehicle is in a straight-line driving condition, call the straight-line safe speed calculation model that considers the load transfer between the brake axles; when the vehicle is in a turning condition, call the turning safe speed calculation model that considers the lateral load transfer.

[0029] Step 8-1, Calculation of safe speed for straight-ahead driving. When the vehicle is in straight-ahead driving condition, based on the conservative friction coefficient, gravitational acceleration g, vehicle center of gravity height h, wheelbase L, maximum available stopping distance d, driver reaction time, and safety margin, a safe speed model for straight-ahead driving considering inter-axle load transfer is constructed, and the recommended safe speed for straight-ahead driving is dynamically calculated using an iterative method.

[0030] Step 8-2, Calculation of safe vehicle speed under turning conditions. When the vehicle is turning, the system constructs a safe turning speed model that considers lateral load transfer based on the conservative friction coefficient, gravitational acceleration g, turning radius R, vehicle center of gravity height h, and wheelbase T, and dynamically calculates the recommended safe turning speed.

[0031] Step 9: Output the final recommended safe speed, and output a deceleration prompt or safety warning when the actual vehicle speed is greater than the final recommended safe speed.

[0032] The working principle of this invention is as follows: Continuous images of icy and snowy road surfaces are acquired through a visual perception device, and then preprocessed to form a temporal image sequence. A spatial feature extraction unit extracts road surface texture features, and an ice and snow texture attention enhancement unit strengthens key texture regions related to the friction coefficient. Subsequently, a temporal feature modeling unit and a temporal attention unit extract the dynamic features of the icy and snowy road surface state changing over time, and output a predicted friction coefficient value. After filtering, smoothing, and safety-side correction, the predicted friction coefficient value forms a conservative friction coefficient. Further, combining vehicle parameters, weather information, road parameters, and driving conditions, load transfer correction safety speed models for straight-ahead or turning conditions are invoked to calculate the recommended safe speed for the corresponding vehicle under the current icy and snowy road surface conditions, and this is displayed or warned through visual output.

[0033] The advantages of this invention are: by using a dual-attention enhanced road friction coefficient prediction model, it not only strengthens the key texture areas of icy and snowy roads, but also adaptively allocates contribution weights for different time steps, thereby improving the robustness of friction coefficient prediction to changes in illumination, snow surface reflection, and local occlusion; through a friction coefficient safety-side correction mechanism, it reduces the risk of the model overestimating road adhesion capacity; by introducing parameters such as vehicle center of gravity height, wheelbase, and track width, load transfer correction safety speed models are established for straight-line and turning conditions, respectively, to achieve safe speed output for different vehicle types on icy and snowy roads; by integrating vehicle parameters, weather information, road parameters, and driving conditions, it can provide intuitive and reliable safe speed suggestions for drivers, intelligent driving systems, or vehicle-road cooperative platforms. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall architecture of the safe vehicle speed prediction system for icy and snowy roads according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the dual-attention friction coefficient prediction model structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the ice and snow texture attention enhancement unit structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the temporal attention unit structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the safe vehicle speed calculation process for straight-going and turning conditions according to an embodiment of the present invention.

[0035] Figure 6 This is a flowchart of the method of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] Example 1

[0038] This embodiment provides a dynamic prediction method for safe vehicle speed on icy and snowy roads based on load transfer correction. This method achieves safe vehicle speed output for different vehicles under icy and snowy road conditions through continuous road image acquisition, dual-attention friction coefficient prediction, friction coefficient safety-side correction, vehicle parameter matching, weather and road parameter acquisition, driving condition judgment, and safe vehicle speed calculation. The overall system architecture is as follows: Figure 1 and Figure 6 As shown.

[0039] Step 1: Acquisition and preprocessing of image sequences of icy and snowy road surfaces.

[0040] In this embodiment, to construct the training and validation dataset for the dual-attention friction coefficient prediction model, high-definition roadside cameras are deployed in a winter icy and snowy road environment. These cameras, along with synchronous friction coefficient measurement equipment, acquire images of the road surface and the corresponding real values ​​of the road surface friction coefficient at different times. The cameras are positioned facing the road surface area to continuously acquire images of icy and snowy road surfaces on fixed sections. The acquired images are used to characterize visual features of icy and snowy road surfaces such as snow accumulation, compacted snow, thin ice, meltwater film, ice surface reflection, and tire tracks.

[0041] During the data acquisition process, the camera device acquires images of the road surface at preset time intervals. In this embodiment, keyframes are extracted from the continuous video stream at 60-second intervals, and the images are sorted according to the acquisition timestamp. Each frame is time-aligned with the road surface friction coefficient measured at the same or adjacent times, thereby forming an image sample of icy and snowy road surfaces labeled with the friction coefficient.

[0042] In the image preprocessing stage, firstly, abnormal images that are blurry, severely occluded, or whose road surfaces cannot be identified are removed; then, the retained images are uniformly adjusted to 128×128 pixels, and the pixel values ​​are normalized to the range of [0,1] to reduce the impact of different acquisition conditions on the model input.

[0043] After preprocessing, the image samples are arranged according to the image acquisition time sequence, and a time-series image sequence sample is constructed using a sliding window method. In this embodiment, five consecutive frames constitute a time-series input sample. The five frames are arranged in the acquisition time sequence, and the friction coefficient corresponding to the last frame of the sequence is used as the supervision label for the sequence sample. The image sequence constructed in this way contains both single-frame road surface spatial texture information and time-series information on the change of icy and snowy road surface conditions over time, which can be used as input for the subsequent dual-attention friction coefficient prediction model.

[0044] Furthermore, the dataset of icy and snowy road surface image samples labeled with friction coefficients is divided into training, validation, and test sets according to a preset ratio. The training set is used to learn the network parameters of the dual-attention enhanced friction coefficient prediction model; the validation set is used to evaluate the model's generalization performance, adjust hyperparameters, and determine the final friction coefficient prediction model during training; the test set does not participate in model parameter training and hyperparameter adjustment, but is used to independently test the final friction coefficient prediction model, and the model prediction error parameters are determined based on the deviation between the predicted friction coefficient and the measured friction coefficient of the test samples.

[0045] Step 2: Construction of the dual-attention friction coefficient prediction model.

[0046] like Figure 2As shown, the dual-attention friction coefficient prediction model in this embodiment includes a spatial feature extraction unit, an ice and snow texture attention enhancement unit, a temporal feature modeling unit, a temporal attention unit, and a friction coefficient regression unit connected sequentially along the feature processing direction. The model takes a continuous sequence of ice and snow road surface images as input and outputs a predicted value of the current road surface friction coefficient. This predicted value is then used as the model output and input into a filtering and smoothing post-processing unit to obtain the filtered friction coefficient.

[0047] Let the input image sequence be: (1) in, This represents a sequence of images of icy and snowy road surfaces arranged in chronological order. Let represent the road surface image at time step t, and n represent the length of the image sequence. In this embodiment, n is 5, meaning that 5 consecutive frames constitute an input sequence.

[0048] Step 2.1, Spatial feature extraction.

[0049] For each frame of the input sequence Spatial feature extraction was performed using the MobileNetV2 network to obtain spatial feature maps for the corresponding time steps. : (2) in, This represents the MobileNetV2 spatial feature extraction network. This represents the trainable parameters of the network. This represents the spatial feature map corresponding to the image in frame t.

[0050] In this embodiment, MobileNetV2 loads pre-trained weights and freezes several convolutional layers at the beginning, while only fine-tuning several layers at the end to retain the ability to extract general image texture features, while adapting to visual features such as snow accumulation, thin ice, compacted snow, meltwater film, ice surface reflection, and tire tracks on icy and snowy roads.

[0051] Step 2.2, Ice and Snow Texture Attention Enhancement.

[0052] Spatial feature map Input ice and snow texture attention enhancement unit. The ice and snow texture attention enhancement unit adopts a CBAM structure, including channel attention branch and spatial attention branch, to enhance key ice and snow texture information related to the road surface friction coefficient from the channel dimension and spatial dimension.

[0053] First, the channel attention branch is applied to the input feature map. Perform global average pooling and global max pooling respectively to obtain the channel description vector: (3) (4) Subsequently, the two channel description vectors are input into a shared multilayer perceptron, and channel attention weights are generated using the Sigmoid function. : (5) Where AvgPool represents global average pooling, MaxPool represents global max pooling, and MLP represents shared multilayer perceptron. This represents the Sigmoid activation function. This represents the channel attention weight.

[0054] Applying channel attention weights to the original feature map yields a channel-enhanced feature map. : (6) in, This indicates element-wise multiplication.

[0055] Then, the spatial attention branch enhances the channel feature map. Average pooling and max pooling are performed along the channel dimension, followed by feature concatenation: (7) (8) (9) in, This indicates that average pooling is performed along the channel dimension. This indicates that max pooling is performed along the channel dimension. ] indicates feature concatenation. This represents a convolution operation with a kernel size of k×k. This represents the spatial attention weights.

[0056] Finally, spatial attention weights are applied to the channel-enhanced feature map to obtain the feature map after dual attention enhancement. : (10) Through the aforementioned channel attention and spatial attention processing, the model can enhance key features related to the road surface friction coefficient, such as snow particles, compacted snow, ice surface reflection, snow melt water film, and tire tracks, while suppressing the influence of irrelevant features such as background areas, shadows, and road edge debris on the prediction results.

[0057] Step 2.3, Time-series feature construction.

[0058] Feature maps enhanced by dual attention Perform global average pooling to obtain the spatial feature vector at time step t. : (11) Here, GAP represents the Global Average Pooling operation. Let represent the spatial feature vector corresponding to the image in frame t.

[0059] Arrange the spatial eigenvectors of n consecutive time steps in chronological order to obtain the temporal feature matrix Z: (12) Step 2.4, Temporal Feature Modeling.

[0060] The temporal feature matrix Z is input into a Long Short-Term Memory (LSTM) network to extract the dynamic features of the icy and snowy road surface state over time. The state update process of the LSTM at time step t is represented as: (13) (14) (15) (16) (17) (18) in, Indicates the input gate. Represents the Gate of Oblivion Indicates the output gate. Indicates candidate memory states, Indicates the current memory state. Indicates the current hidden state; , , , and , , , These represent the weight matrices respectively; , , , Indicates the bias term; Let denot sigmoid and tanh denote the hyperbolic tangent function. This indicates element-wise multiplication.

[0061] Through the above LSTM temporal modeling, the model can utilize the temporal correlation between consecutive multiple frames of images to capture the changes in the adhesion state of icy and snowy roads caused by snow compaction, local melting, thin ice formation, or changes in snow surface reflection.

[0062] Step 2.5, Temporal attention enhancement.

[0063] The hidden states from multiple time steps output by the LSTM are input into a temporal attention unit. This temporal attention unit generates temporal weights based on the contribution of different time-step hidden states to the current friction coefficient prediction, and then performs weighted fusion of the hidden states from each time step. Let the hidden state sequence output by the LSTM be: (19) The temporal attention unit first calculates the importance score of the hidden state at each time step. : (20) in, , and These are the trainable parameters for the temporal attention unit. Let represent the attention score at time step t.

[0064] Then, the scores at each time step are normalized using the Softmax function to obtain the temporal attention weights. : (twenty one) Finally, the hidden states at each time step are weighted and fused to obtain the temporal fusion features. : (twenty two) in, This represents the contribution weight of the t-th time step to the current friction coefficient prediction. This represents the temporal characteristics after fusion.

[0065] Through this temporal attention unit, the model can adaptively determine the importance of different time frames. When a frame is affected by sudden changes in illumination, snow reflection, local occlusion, or acquisition noise, the model can reduce the weight of that time step, thereby improving the stability and robustness of friction coefficient prediction.

[0066] Step 2.6, Friction coefficient regression output.

[0067] Temporal fusion features Inputting a fully connected regression layer yields the predicted values ​​of the road surface friction coefficient. : (twenty three) in, This is the weight matrix of the regression layer. For the regression layer bias term, This is the predicted value of the friction coefficient of the continuous road surface output by the model.

[0068] During the model training phase, the measured friction coefficients corresponding to the icy and snowy road surface image sequences in the training set are used as supervision labels, and the mean squared error is used as the loss function. (twenty four) Where N represents the number of training samples. This represents the measured friction coefficient of the i-th sample. Let represent the predicted friction coefficient of the i-th sample.

[0069] Step 3, Model Training and Validation. After training, input the validation set into the trained dual-attention enhanced friction coefficient prediction model. Evaluate the model's generalization performance based on the loss value, mean absolute error, or root mean square error on the validation set. Adjust the learning rate, training epochs, batch size, number of LSTM hidden units, or attention layer parameters based on the evaluation results to determine the final friction coefficient prediction model.

[0070] Step 4: Model Testing and Error Parameter Determination. The test set is input into the final friction coefficient prediction model, which outputs the predicted friction coefficient of the continuous road surface corresponding to the test sample. This predicted value is then compared with the measured friction coefficient of the road surface corresponding to the test sample. The mean absolute error, root mean square error, prediction uncertainty, or a combination thereof are calculated and used as the model prediction error parameters for subsequent safety-side corrections.

[0071] Step 5.1, post-processing after filtering and smoothing.

[0072] To reduce instantaneous fluctuations in the friction coefficient prediction sequence, a filtering and smoothing post-processing unit is used to recursively smooth the current road surface friction coefficient prediction value output by the final friction coefficient prediction model. In this embodiment, the filtering and smoothing post-processing unit employs a Kalman filter.

[0073] The state prediction process of Kalman filtering is represented as follows: (25) (26) in, This represents the prior estimate of the friction coefficient at time k. Let A represent the posterior friction coefficient estimate from the previous time step, and let A represent the state transition matrix. Let Q represent the prior estimate covariance, and let Q represent the process noise covariance. This indicates process noise.

[0074] The observation update process is represented as follows: (27) (28) (29) in, H represents the Kalman gain, H represents the observation matrix, and R represents the observation noise covariance. This represents the observed friction coefficient output by the model at time k, i.e. , This represents the filtered estimate of the friction coefficient. This represents the posterior estimate of the covariance.

[0075] In this embodiment, both A and H are set as identity matrices to characterize the continuous variation of the friction coefficient of icy and snowy road surfaces at adjacent time points. Through the aforementioned Kalman filtering process, high-frequency fluctuations caused by image acquisition noise, lighting changes, snow fog, or snow surface reflection can be reduced, resulting in a more stable friction coefficient output.

[0076] Step 5.2, Friction coefficient safety side correction

[0077] Based on the model prediction error parameters obtained during the model training, validation, or testing phases, the filtered friction coefficient is conservatively corrected to obtain the conservative friction coefficient.

[0078] In one embodiment, the conservative coefficient of friction is calculated as follows: (30) in, For a conservative coefficient of friction, The friction coefficient after filtering. These are the model prediction error parameters. The mean absolute error (MAE), root mean square error (RMSE), prediction uncertainty, or preset safety correction can be used, obtained during the training, validation, or testing phases.

[0079] This safety-side correction mechanism can prevent the recommended safe speed from being too high due to an overestimation of the friction coefficient, even if the model has a certain prediction error in the road friction coefficient.

[0080] Step 6: Obtain vehicle parameters, weather information, road parameters, and vehicle driving status information.

[0081] 6.1 Vehicle Parameter Acquisition. The vehicle parameter matching module acquires the vehicle's center of gravity height h, wheelbase L, front track width, rear track width, and average track width T from the established vehicle parameter database based on the user-input vehicle model or the matching results from the vehicle database. The vehicle parameters include at least the vehicle's center of gravity height, wheelbase, and track width parameters. The track width parameters include the front track width, rear track width, and / or the average track width. The vehicle parameter database may include publicly available parameters from manufacturers, vehicle registration parameters, or information collected by onboard sensors, and is regularly updated to cover newly added vehicle models.

[0082] 6.2 Weather Information Acquisition. The weather information is collected through a weather information acquisition module and can be provided by user input, onboard sensors, roadside weather stations, internet weather interfaces, or vehicle-to-infrastructure (V2I) platforms. The weather information acquisition module collects current weather conditions and, based on the defensive driving safety principle that 'the vehicle's safe braking distance must not exceed the current visible perception boundary,' maps the maximum visible distance under the current weather conditions to the current maximum usable stopping distance d. For example, when severe weather severely limits visibility, the value of d will decrease accordingly, thus ensuring that the vehicle can come to a complete stop before exceeding the unknown perception boundary. This parameter will be used as a safety constraint for subsequent safe speed calculations.

[0083] 6.3 Road Parameter Acquisition. The road parameter acquisition module acquires information such as road curvature, turning radius R, road slope, and speed limit. In one implementation, the turning radius R can be provided by a high-precision map, an in-vehicle navigation system, roadside sensors, a visual recognition module, or user input, and is updated in real time to reflect the actual curvature of the road segment currently being traveled by the vehicle.

[0084] Step 7, Determine Driving Condition

[0085] The driving condition determination module determines whether the vehicle is currently in a straight-ahead or turning condition. This driving condition can be determined based on steering wheel angle, yaw rate, vehicle trajectory curvature, road curvature on a high-precision map, visual recognition results, or user input. Specifically, when the road curvature is less than a preset curvature threshold, the steering wheel angle is less than a preset angle threshold, and the yaw rate is less than a preset yaw threshold, the vehicle is determined to be in a straight-ahead condition. When the road curvature is not less than the preset curvature threshold, the steering wheel angle is not less than the preset angle threshold, or the yaw rate is not less than the preset yaw threshold, the vehicle is determined to be in a turning condition. The turning radius R can be provided by a high-precision map, an in-vehicle navigation system, roadside equipment, a visual recognition module, or user input.

[0086] Step 8, Calculate safe speed

[0087] The safe speed calculation module calculates a recommended safe speed based on the conservative friction coefficient obtained in step 5, the vehicle parameters obtained in step 6, the weather information and road parameters, and the driving conditions obtained in step 7.

[0088] In this embodiment, if the vehicle is traveling straight, the safe speed calculation model for straight-traffic conditions is invoked; if the vehicle is turning, the safe speed calculation model for turning conditions is invoked. Straight-traffic conditions primarily consider the braking safety of the vehicle on low-friction surfaces, while turning conditions primarily consider the lateral stability of the vehicle on low-friction surfaces.

[0089] Step 8.1, Calculation of safe speed for straight-going conditions

[0090] When the vehicle is traveling straight, the safe speed calculation model is based on a conservative friction coefficient. Gravitational acceleration g, vehicle center of gravity height h, wheelbase L, usable stopping distance d, driver reaction time, and safety margin Calculate the recommended safe speed for going straight. In this embodiment, the speed unit in the safe vehicle speed calculation formula is m / s.

[0091] During straight-line braking, the vehicle experiences inter-axle load transfer. Vehicles with a high center of gravity experience a more significant decrease in the normal load on the rear axle during braking, thus affecting braking stability on icy and snowy roads. Therefore, this embodiment incorporates the vehicle's center of gravity height h and wheelbase L into the calculation of the safe straight-line speed to correct the traditional safe braking speed model.

[0092] Since the driver's reaction distance is related to the vehicle's speed, the reaction distance is calculated based on the current iterative vehicle speed, and the reaction distance and safety margin are subtracted from the available stopping distance to obtain the actual distance that can be used for braking. : (31) in, This represents the actual distance that can be used for braking, and d represents the maximum usable stopping distance determined by current weather visibility, used to characterize the current visually perceived safety boundary. This indicates the recommended safe speed for driving straight. Indicates the driver's reaction time. This represents the safety margin. The formula shows that if severe weather significantly reduces the usable stopping distance *d*, then after deducting the driver's reaction distance and the safety margin, the actual distance available for braking is... This will significantly reduce the speed, and then through formula (32), the recommended safe speed of the final output of the system will be forced to decrease, thereby achieving adaptive speed adjustment under the condition of limited visibility.

[0093] After obtaining the actual distance that can be used for braking, the straight-line safe speed, considering the inter-axle load transfer correction, can be expressed as: (32) Where is the conservative friction coefficient obtained in step 5, g is the gravitational acceleration, h is the height of the vehicle's center of gravity, and L is the vehicle's wheelbase.

[0094] In actual calculations, since the reaction distance in equation (31) is related to the vehicle speed, the system uses an iterative approach to solve for the recommended safe speed for straight-line driving. Specifically, firstly, an initial vehicle speed estimate is given, the driver's reaction distance is calculated based on the current iterative vehicle speed, and the driver's reaction distance and safety margin are subtracted from the maximum available stopping distance to obtain the actual distance that can be used for braking; then, the recommended safe speed for straight-line driving is updated based on the actual distance that can be used for braking. When the difference between the recommended safe speeds for straight-line driving obtained from two adjacent iterations is less than a preset convergence threshold, the converged recommended safe speed for straight-line driving is output; when the actual distance that can be used for braking is less than or equal to zero, it is determined that the visible perception boundary under the current weather conditions is insufficient to support the vehicle to complete safe braking, and a deceleration prompt, safety warning information, or parking risk prompt is output.

[0095] Step 8.2, Calculation of safe speed under turning conditions

[0096] When the vehicle is turning, according to the conservative coefficient of friction Calculate the recommended safe turning speed based on the following factors: gravitational acceleration g, turning radius R, vehicle center of gravity height h, and wheelbase T. .

[0097] During cornering, vehicles experience lateral acceleration, resulting in lateral load transfer. For high-center-of-gravity vehicles such as SUVs, MPVs, vans, or trucks, this lateral load transfer significantly affects vehicle stability on icy or snowy curves. Therefore, this embodiment incorporates the vehicle's center-of-gravity height h and wheelbase T into the calculation of safe cornering speed, thus correcting the traditional critical cornering speed model.

[0098] The safe turning speed considering lateral load transfer correction can be expressed as: (33) Where R represents the road turning radius, T represents the vehicle track width or the average front and rear track width, and the meanings of other parameters are the same as in step 8.1.

[0099] Step 8.3: Final recommendation of safe speed.

[0100] Based on the driving conditions obtained in step 7, the system selects between the recommended safe speed for straight driving and the recommended safe speed for turning, and obtains the final recommended safe speed: (34) Finally, the visualization output module displays the final recommended safe speed to the driver, or uses it as a recommended safe speed for intelligent driving systems and vehicle-to-infrastructure (V2I) platforms. If the current actual vehicle speed exceeds the final recommended safe speed, the system can issue a deceleration prompt or safety warning; if the actual available braking distance is less than or equal to zero, the system can further output a parking risk warning.

[0101] Table 1. Parameter Table for Dynamic Prediction Method of Safe Vehicle Speed ​​on Icy and Snowy Roads

[0102] Step 9: Safe speed output and risk warning.

[0103] The system outputs the final recommended safe speed, conservative coefficient of friction, vehicle model, weather conditions, road parameters, and current driving conditions to the in-vehicle human-machine interface, mobile terminal, roadside display device, intelligent driving system, or vehicle-road cooperative platform, and outputs deceleration prompts or safety warnings when the actual vehicle speed exceeds the final recommended safe speed.

[0104] Using the above method, this embodiment can realize the complete process from image perception of icy and snowy road surfaces to output of safe vehicle speed.

[0105] Example 2

[0106] This embodiment provides a dynamic prediction device for safe vehicle speed on icy and snowy roads based on load transfer correction. It includes an image sequence construction module, a friction coefficient prediction model construction module, a model training and verification testing module, a friction coefficient prediction module, a friction coefficient post-processing module, a safety-side correction module, a parameter acquisition module, a working condition judgment module, a safe vehicle speed calculation module, and an output warning module. Each module is used to execute the corresponding method steps in Embodiment 1. Related technical details can be found in the description of Embodiment 1 and will not be repeated here.

[0107] Example 3

[0108] This embodiment provides an electronic device, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the dynamic prediction method for safe vehicle speed on icy and snowy roads described in Embodiment 1.

[0109] The electronic device may be an in-vehicle computing terminal, a roadside computing device, an edge computing device, a mobile terminal, a server, or an intelligent driving control unit.

[0110] Example 4

[0111] This embodiment provides a computer-readable storage medium on which a program is stored. When executed by a processor, the program implements the dynamic prediction method for safe vehicle speed on icy and snowy roads described in Embodiment 1.

[0112] The various embodiments in this invention are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. For the system embodiments, electronic device embodiments, and computer-readable storage medium embodiments, their basic principles correspond to those of the method embodiments, and relevant details can be found in the description of the method embodiments.

[0113] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications, equivalent substitutions, and improvements can be made to the present invention without departing from its spirit and principles. All modifications, equivalent substitutions, and improvements made within the scope of the present invention's technical concept should be included within the protection scope of the present invention.

Claims

1. A method for dynamic prediction of safe vehicle speed on icy and snowy roads based on load transfer correction, characterized in that, Includes the following steps: Step 1: Collect images of icy and snowy road surfaces as sample images, preprocess the sample images to obtain an icy and snowy road surface image dataset; construct a continuous sequence of icy and snowy road surface images in chronological order, and match the continuous sequence of icy and snowy road surface images with the measured road surface friction coefficient at the corresponding time to form a sample dataset with friction coefficient labels; divide the icy and snowy road surface image dataset into a training set, a validation set, and a test set. Step 2: Build a dual-attention enhanced friction coefficient prediction model; the structure of the dual-attention enhanced friction coefficient prediction model is as follows: starting from the input layer, it is connected in sequence to the spatial feature extraction unit, the ice and snow texture attention enhancement unit, the temporal feature modeling unit, the temporal attention unit, and the friction coefficient regression unit; The spatial feature extraction unit extracts spatial feature maps frame by frame from each time step image in the icy and snowy road surface image sequence sample; the icy and snowy texture attention enhancement unit enhances icy and snowy texture features related to the road surface friction coefficient; the temporal feature modeling unit extracts dynamic features of the icy and snowy road surface state changing over time in a continuous image sequence; the temporal attention unit assigns temporal weights based on the contribution of different time step images to the current friction coefficient prediction result; and the friction coefficient regression unit outputs the predicted value of the continuous road surface friction coefficient. Step 3: Train the dual-attention enhanced friction coefficient prediction model using the training set to obtain the trained dual-attention enhanced friction coefficient prediction model; validate the trained dual-attention enhanced friction coefficient prediction model using the validation set to obtain the final friction coefficient prediction model. Step 4. Input the test set into the final friction coefficient prediction model, output the predicted value of the continuous road friction coefficient corresponding to the test sample, and compare the predicted value of the continuous road friction coefficient with the measured road friction coefficient to determine the model prediction error parameter. Step 5: Input the real-time acquired continuous icy and snowy road surface image sequence into the final friction coefficient prediction model to obtain the current road surface friction coefficient prediction value; input the current road surface friction coefficient prediction value into the filtering and smoothing post-processing unit for filtering and smoothing to obtain the filtered friction coefficient; based on the model prediction error parameters obtained during the model training, verification, or testing phases, perform a safety-side correction on the filtered friction coefficient to obtain a conservative friction coefficient. Step 6: Obtain vehicle parameters, weather information, road parameters, and vehicle driving status information. The vehicle parameters include at least the vehicle center of gravity height, wheelbase, and track width. The track width parameters include the front track width, rear track width, and / or the average track width. The weather information is used to determine the maximum usable stopping distance under the current weather conditions. The road parameters include at least the road curvature or road turning radius. Step 7: Determine whether the vehicle is currently in a straight-ahead or turning condition based on the vehicle driving status information and the road parameters; Step 8: Based on the conservative friction coefficient, vehicle parameters, weather information, road parameters, and the vehicle's current driving conditions, call the load transfer correction safe speed calculation model to calculate the final recommended safe speed; when the vehicle is in a straight-line driving condition, call the straight-line safe speed calculation model that considers the load transfer between brake axles; when the vehicle is in a turning condition, call the turning safe speed calculation model that considers the lateral load transfer. Step 9: Output the final recommended safe speed, and output a deceleration prompt or safety warning when the actual vehicle speed is greater than the final recommended safe speed.

2. The method for dynamic prediction of safe vehicle speed on icy and snowy roads according to claim 1, characterized in that, Step 1, in constructing the image sequence samples of the icy and snowy road surface, specifically includes: Step 1-1: Collect road surface images at preset time intervals using vehicle-mounted visual perception devices or roadside visual perception devices; Steps 1-2: Remove blurry, severely occluded, abnormally exposed, or unidentifiable road surface images; Steps 1-3: Adjust the retained image to the preset size and normalize the image pixel values ​​to the preset range; Steps 1-4: Sort the retained images according to the image acquisition timestamp, and use a sliding window method to extract a sequence of icy and snowy road surface images consisting of n frames from the continuous images.

3. The method for dynamic prediction of safe vehicle speed on icy and snowy roads based on load transfer correction according to claim 1, characterized in that, The spatial feature extraction unit in step 2 uses a lightweight convolutional neural network; the lightweight convolutional neural network is MobileNetV2, and the MobileNetV2 is loaded with pre-trained weights, the front convolutional layer is frozen and the rear convolutional layer is fine-tuned, in order to extract spatial texture features corresponding to snow, compacted snow, thin ice, snow melt water film, ice surface reflection and tire rolling marks. The ice and snow texture attention enhancement unit adopts a CBAM attention structure, which includes a channel attention branch and a spatial attention branch, and the channel attention branch and the spatial attention branch are connected serially along the feature processing direction; The channel attention branch is located before the spatial attention branch. It is used to perform global average pooling and global max pooling on the input feature map respectively, and generate channel attention weights through a shared multilayer perceptron and a sigmoid function. The channel attention weights are then applied to the input feature map to obtain a channel-enhanced feature map. The spatial attention branch is used to receive the channel enhancement feature map, and to perform average pooling and max pooling on the channel enhancement feature map in the channel dimension respectively. After feature concatenation and convolution operation, spatial attention weights are generated, and the spatial attention weights are applied to the channel enhancement feature map to obtain the ice and snow texture attention enhancement feature map. The snow and ice texture attention enhancement unit enhances key texture areas of snow and ice road surfaces corresponding to snow accumulation, compacted snow, thin ice, snow melt water film, ice surface reflection, and tire rolling marks through a serial attention enhancement method that first enhances channels and then spaces, while suppressing background areas, shadows, road edge debris, and acquisition noise. The temporal feature modeling unit and the temporal attention unit are sequentially connected between the snow texture attention enhancement unit and the friction coefficient regression unit along the feature processing direction; The temporal feature modeling unit is used to receive the ice and snow texture attention enhancement feature maps corresponding to multiple time steps output by the ice and snow texture attention enhancement unit, and process them in the following order: Step 2-1: Perform global average pooling on the ice and snow texture attention enhancement feature map corresponding to each time step to obtain the spatial feature vector corresponding to each time step; Step 2-2: Arrange and combine the spatial feature vectors corresponding to multiple consecutive time steps according to the image acquisition time order to obtain the temporal feature matrix; Steps 2-3: Input the temporal feature matrix into the Long Short-Term Memory (LSTM) network, and use the LSTM network to model the changes in the state of the icy and snowy road surface between consecutive time steps to obtain the hidden state corresponding to each time step. The temporal attention unit is used to receive the hidden states at each time step output by the Long Short-Term Memory (LSTM) network, and processes them in the following order: Steps 2-4: Calculate the attention score for each time step based on the contribution of the hidden state at each time step to the current friction coefficient prediction result; Steps 2-5: Perform Softmax normalization on the attention scores to obtain the temporal attention weights corresponding to each time step; Steps 2-6: Use the aforementioned temporal attention weights to perform weighted fusion of the hidden states at each time step to obtain temporal fusion features; Steps 2-7: Input the time-series fusion features into the friction coefficient regression unit to output the predicted value of the friction coefficient of the continuous road surface; The predicted value of the continuous road surface friction coefficient is used as the output of the dual attention enhanced friction coefficient prediction model and is input into the filtering and smoothing post-processing unit in step 5 for filtering and smoothing processing. When the temporal attention unit is affected by sudden changes in illumination, snow reflection, partial occlusion, snow fog, or acquisition noise at a certain time step, it reduces the temporal attention weight corresponding to the hidden state at that time step, so as to improve the stability of the predicted value of the friction coefficient of the continuous road surface. The friction coefficient regression unit includes a fully connected regression layer, which is used to map the temporal fusion features into continuous road surface friction coefficient prediction values. During the model training phase, the measured road surface friction coefficient is used as the supervision label, and the mean square error loss function is used to train the dual attention enhanced friction coefficient prediction model. The trained model is validated using a validation set, and the model prediction error parameters are determined using a test set. The filtering and smoothing post-processing unit uses the Kalman filtering method to filter and smooth the current road surface friction coefficient prediction value. The Kalman filtering method includes a state prediction process and an observation update process, and the state transition matrix and observation matrix are set as identity matrices to characterize the continuous change characteristics of the ice and snow road surface friction coefficient at adjacent time points.

4. The method for dynamic prediction of safe vehicle speed on icy and snowy roads based on load transfer correction according to claim 1, characterized in that, In step 1, during the model training phase, each frame of the image is time-aligned with the measured road friction coefficient at the same or adjacent time points, and the measured road friction coefficient corresponding to the last frame of the icy and snowy road image sequence is used as the supervision label for that icy and snowy road image sequence.

5. The method for dynamic prediction of safe vehicle speed on icy and snowy roads based on load transfer correction according to claim 1, characterized in that, The safety correction in step 5 is as follows: determine the model prediction error parameter based on one or more of the following: mean absolute error, root mean square error, prediction uncertainty, or preset safety correction amount; and subtract the model prediction error parameter or the preset safety correction amount determined by the model prediction error parameter from the filtered and smoothed friction coefficient to obtain the conservative friction coefficient.

6. The method for dynamic prediction of safe vehicle speed on icy and snowy roads according to claim 1, characterized in that, Step 6, which involves obtaining vehicle parameters, weather information, and road parameters, specifically includes: Step 6-1: Based on the vehicle model entered by the user or the vehicle database matching results, obtain the vehicle's center of gravity height, wheelbase, front track, rear track, and average track width. Step 6-2: Obtain weather information through user input, vehicle-mounted sensors, roadside weather stations, Internet weather interfaces, or vehicle-road cooperative platforms, and map the maximum visibility distance under the current weather conditions to the maximum available stopping distance; Step 6-3: Obtain information on road curvature, road turning radius, road slope, or road speed limit through high-precision maps, in-vehicle navigation systems, roadside sensors, visual recognition modules, or user input.

7. The method for dynamic prediction of safe vehicle speed on icy and snowy roads according to claim 1, characterized in that, Step 7, determining the vehicle's current driving condition, specifically includes: when the road curvature is less than a preset curvature threshold, the steering wheel angle is less than a preset angle threshold, and the yaw rate is less than a preset yaw threshold, the vehicle is determined to be in a straight-ahead driving condition; when the road curvature is not less than the preset curvature threshold, the steering wheel angle is not less than the preset angle threshold, or the yaw rate is not less than the preset yaw threshold, the vehicle is determined to be in a turning condition.

8. The method for dynamic prediction of safe vehicle speed on icy and snowy roads based on load transfer correction according to claim 1, characterized in that, The straight-ahead safe speed calculation model described in step 8 calculates the recommended safe speed for straight-ahead driving based on the conservative friction coefficient, gravitational acceleration, vehicle center of gravity height, vehicle wheelbase, maximum available stopping distance, driver reaction time, and safety margin. Specifically, the straight-ahead safe speed calculation model calculates the driver's reaction distance based on the current iterative speed, subtracts the driver's reaction distance and safety margin from the maximum available stopping distance to obtain the actual distance available for braking, and corrects the braking safe speed using an inter-axle load transfer correction term determined by the vehicle's center of gravity height and wheelbase. When the difference between the straight-going suggested safe speeds obtained from two adjacent iterations is less than the preset convergence threshold, the converged straight-going suggested safe speed is output; when the actual distance available for braking is less than or equal to zero, it is determined that the visible perception boundary under the current weather conditions is insufficient to support the vehicle to complete safe braking, and a deceleration prompt, safety warning information or parking risk prompt is output. The safe turning speed calculation model calculates the recommended safe turning speed based on the conservative friction coefficient, gravitational acceleration, road turning radius, vehicle center of gravity height, and vehicle track width or average track width. The safe turning speed calculation model uses a lateral load transfer correction term determined by the vehicle center of gravity height and vehicle track width or average track width to correct the critical turning speed, so as to reduce the risk of sideslip or instability of high center of gravity vehicles when driving on curves on icy and snowy roads.

9. The method for dynamic prediction of safe vehicle speed on icy and snowy roads based on load transfer correction according to claim 1, characterized in that, Step 9 includes: outputting the final recommended safe speed, conservative coefficient of friction, vehicle model, weather conditions, road parameters and current driving conditions to the vehicle human-machine interface, mobile terminal, roadside display device, intelligent driving system or vehicle-road cooperative platform.

10. A dynamic prediction device for safe vehicle speed on icy and snowy roads based on load transfer correction, characterized in that, include: The image sequence construction module is used to acquire images of icy and snowy road surfaces, preprocess the images, and construct a continuous image sequence of icy and snowy road surfaces. The friction coefficient prediction model construction module is used to build a dual-attention enhanced friction coefficient prediction model. The dual-attention enhanced friction coefficient prediction model starts from the input layer and is sequentially connected to the spatial feature extraction unit, the ice and snow texture attention enhancement unit, the temporal feature modeling unit, the temporal attention unit, and the friction coefficient regression unit. The model training, validation, and testing module is used to divide the icy and snowy road surface image dataset into a training set, a validation set, and a test set. The training set is used to train the dual-attention enhanced friction coefficient prediction model, and the validation set is used to validate the trained dual-attention enhanced friction coefficient prediction model to obtain the final friction coefficient prediction model. The test set is used to determine the model prediction error parameters. The friction coefficient prediction module is used to input the real-time acquired continuous snow and ice road surface image sequence into the final friction coefficient prediction model to obtain the current road surface friction coefficient prediction value; The friction coefficient post-processing module is used to filter and smooth the current road surface friction coefficient prediction value to obtain the filtered friction coefficient. The safety-side correction module is used to perform safety-side correction on the filtered friction coefficient based on the model prediction error parameters to obtain a conservative friction coefficient. The parameter acquisition module is used to acquire vehicle parameters, weather information, road parameters, and vehicle driving status information. The operating condition judgment module is used to determine whether the vehicle is currently in a straight-going or turning condition; The safe speed calculation module is used to call the straight-going safe speed calculation model that takes into account the load transfer between the brake axles or the turning safe speed calculation model that takes into account the lateral load transfer, and calculate the final recommended safe speed. The output warning module is used to output the final recommended safe speed and output a deceleration prompt or safety warning when the actual vehicle speed is greater than the final recommended safe speed.

Images