A road surface adhesion coefficient estimation method and system based on visual dynamics adaptive fusion

Abstract

This invention belongs to the field of autonomous driving control technology, specifically relating to a method and system for estimating road surface adhesion coefficient based on visual-dynamic adaptive fusion. The method includes acquiring a visual image of the road surface ahead of the vehicle; determining prior information on the visual friction coefficient based on road surface category information obtained from the visual image; determining the braking pressure level; applying braking pulses to the wheels to excite the tire-road system; acquiring the vehicle's dynamic response information during the application of the braking pulses; obtaining an estimated value of the dynamic friction coefficient; determining fusion weights based on the prior information on the visual friction coefficient and the estimated value of the dynamic friction coefficient; and obtaining the final estimated value of the road surface adhesion coefficient based on the fusion weights. This invention guides graded braking through visual classification results, actively activates dynamic estimation under low slip conditions, and dynamically adjusts the fusion weights based on the visual interval width, dynamic covariance, and the deviation between the two, achieving adaptive integration of observable and multi-source information under low-excitation conditions.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving control technology, specifically relating to a method and system for estimating road surface adhesion coefficient based on visual dynamics adaptive fusion. Background Technology

[0002] The tire-road friction coefficient (TRFC) is a quantitative indicator characterizing the maximum available grip between the tire and the road surface. It directly determines the physical limits of a vehicle's braking, steering, and acceleration, and is a core input parameter for anti-lock braking systems (ABS), electronic stability control (ESC), and trajectory planning and motion control in autonomous vehicles. Therefore, achieving accurate, robust, and real-time estimation of the tire-road friction coefficient is of great significance for improving the active safety performance and driving stability of intelligent vehicles under complex and varied road conditions.

[0003] Currently, existing road adhesion coefficient estimation techniques are mainly divided into three categories. The first category is estimation methods based on vehicle dynamics. These methods utilize the vehicle's onboard wheel speed sensors, accelerometers, etc., combined with vehicle dynamics models and state observers (such as Kalman filters) to calculate the adhesion coefficient. However, dynamic methods are highly dependent on the degree of tire "excitation," i.e., the magnitude of the longitudinal slip ratio. Under low-excitation conditions such as straight-line cruising or slow deceleration, the tire slip ratio is extremely small, the nonlinear relationship between tire force and adhesion coefficient is not obvious, and the system is in an "unobservable" state. At this time, the estimation algorithm often diverges or fails, and cannot accurately output the adhesion coefficient. Although some studies have proposed to solve this problem by actively applying braking to generate slip ratio, blindly applying continuous braking can interfere with normal vehicle driving and even cause danger. How to achieve sufficient excitation while ensuring driving safety is a technical problem that urgently needs to be solved in this field.

[0004] The second category is estimation methods based on visual perception. These methods utilize onboard cameras to capture road surface images and extract road texture features using deep learning models such as convolutional neural networks (CNNs) or Transformers to classify road surface types and thus predict the range of adhesion coefficients. Visual methods have the advantage of being forward-looking, independent of vehicle motion, and can provide friction coefficient predictions before the vehicle enters the road surface. However, they are significantly affected by ambient lighting, shadows, and water reflections, resulting in insufficient robustness. More importantly, visual methods typically only output discrete semantic categories or coarse numerical ranges, making it difficult to provide the precise, continuous values ​​reflecting the current physical contact state required by the vehicle chassis control system.

[0005] The third category is the fusion estimation method combining vision and dynamics. Existing research has attempted to combine visual perception with dynamics estimation to leverage their complementary advantages. However, most existing fusion strategies employ simple weighted averaging or rule switching, failing to adequately characterize the uncertainties of each information source and effectively resolving the contradiction between "insufficient stimulus" and "driving stability." Specifically, existing fusion methods still assign high weights to unreliable dynamics estimations or lack effective correction mechanisms when visual perception is interfered with, resulting in insufficient fusion accuracy and robustness. Furthermore, when vehicles encounter sudden changes in road adhesion coefficients (such as suddenly transitioning from dry asphalt to icy or snowy surfaces), the response speed of existing fusion methods is often lagging, making it difficult to track the actual changes in the road surface in a timely manner.

[0006] In summary, how to achieve dynamic observability under low-excitation conditions without excessive braking interference, how to effectively transform visual semantic information into prior knowledge with physical scale, and how to achieve adaptive and robust fusion when there is conflict among multiple sources of information are all technical problems that urgently need to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a road surface adhesion coefficient estimation method and system based on visual-dynamic adaptive fusion, so as to solve the problems in existing road surface adhesion coefficient estimation methods, such as the unobservable dynamic estimation under low excitation conditions, the lack of physical numerical accuracy in visual estimation, and the inability to balance excitation requirements and driving safety when the two are fused.

[0008] The present invention achieves the above objectives through the following technical solutions: This invention proposes a method for estimating road surface adhesion coefficient based on visual dynamics adaptive fusion, the method comprising: A visual image of the road surface in front of the vehicle is acquired, and road surface category information is obtained based on the visual image, wherein the road surface category information is used to indicate the semantic category of the road surface on which the vehicle is currently traveling; Based on the road surface category information, the prior information of the visual friction coefficient of the road surface is determined, wherein the prior information of the visual friction coefficient includes the friction coefficient range and / or the center value of the friction coefficient. Based on the prior information of the visual friction coefficient, the braking pressure level is determined, and braking pulses are applied to the wheels of the vehicle according to the braking pressure level to excite the tire-road system. The vehicle's dynamic response information during the application of the braking pulse is obtained, and an estimated value of the dynamic friction coefficient is obtained based on the dynamic response information, wherein the dynamic response information includes wheel speed, vehicle speed and / or longitudinal acceleration; Based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient, the fusion weight is determined, and based on the fusion weight, the final estimated value of the adhesion coefficient of the road surface is obtained.

[0009] As a further optimization of the present invention, obtaining road surface category information based on the visual image includes: The visual image is input into an interactive road surface perception network to obtain the road surface category information. The interactive road surface perception network includes a four-level convolutional encoder, and each stage of the four-level convolutional encoder is connected to a cross-scale enhancement module.

[0010] As a further optimization of the present invention, the cross-scale enhancement module processes the input feature map in the following manner: The input feature map is captured by a first depthwise separable convolutional branch to obtain the first branch features. The input feature map is captured by a second depthwise separable convolutional branch to obtain the second branch features, wherein the second depthwise separable convolutional branch includes cascaded 5×5 dilated convolutions; Channel splicing and point-by-point convolution are performed on the first branch features and the second branch features to obtain multi-scale fused features; The multi-scale fusion features are processed by spatial attention branch and channel attention branch to obtain an attention map; The input feature map is multiplied element-wise with the attention map to obtain the enhanced output feature map.

[0011] As a further optimization of the present invention, obtaining the estimated value of the dynamic friction coefficient based on the dynamic response information includes: Based on the longitudinal acceleration of the vehicle, the vertical load on the front axle of the vehicle is dynamically corrected to obtain the corrected vertical load. The tire longitudinal slip ratio is determined based on the corrected vertical load, the wheel speed, and the vehicle speed. Based on the tire longitudinal slip ratio and the corrected vertical load, the friction coefficient of the vehicle is estimated using a constrained unscented Kalman filter to obtain the estimated value of the dynamic friction coefficient. Specifically, when the longitudinal slip ratio of the tire is lower than a preset slip ratio threshold, the constrained unscented Kalman filter skips the measurement update step and keeps the state frozen.

[0012] As a further optimization of the present invention, applying braking pulses to the wheels of the vehicle according to the braking pressure level includes: Braking pulses of the braking pressure level are applied to the wheels of the vehicle in the form of trapezoidal braking pulses; During the application of the braking pulse, the tire slip ratio of the vehicle is monitored in real time; If the tire slip ratio exceeds a preset safety threshold, the application of the braking pulse will be terminated immediately.

[0013] As a further optimization of the present invention, the method further includes: After the braking pulse is applied or terminated, convergence state information is determined based on the comparison results of the tire slip ratio and multiple slip ratio thresholds, wherein the convergence state information is used to indicate the friction range in which the estimation result of the constrained unscented Kalman filter is located. Based on the convergence state information, the dynamic prior interval and prior confidence level are determined.

[0014] As a further optimization of the present invention, determining the fusion weight based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient includes: The visual interval width is determined based on the prior information of the visual friction coefficient. Obtain the posterior covariance of the estimated dynamic friction coefficient; Determine the deviation between the estimated value of the dynamic friction coefficient and the target interval, wherein the target interval is determined based on the prior information of the visual friction coefficient and the prior dynamic interval; The visual interval width, the posterior covariance, and the bias are input into the fuzzy inference system to obtain the basic fusion weights. The basic fusion weights are corrected based on the prior confidence level to obtain the fusion weights.

[0015] As a further optimization of the present invention, obtaining the final adhesion coefficient estimate of the road surface based on the fusion weight includes: The target center value is determined based on the prior information of the visual friction coefficient and the prior dynamic interval. The final estimated adhesion coefficient is calculated using the following formula: in, This is the estimated final adhesion coefficient. The fusion weights are... This is the estimated value of the dynamic friction coefficient. The target center value; The final adhesion coefficient estimate is constrained to a preset physical boundary range.

[0016] As a further optimization of the present invention, the prior information of the visual friction coefficient includes a friction coefficient range, and determining the prior information of the visual friction coefficient of the road surface based on the road surface category information includes: According to the preset mapping relationship, the road surface category information is mapped to the corresponding friction coefficient range; The range of friction coefficients is determined as the prior information of the visual friction coefficient; The method further includes: When a sudden change in the road surface category information is detected, the process noise covariance and state error covariance of the constrained unscented Kalman filter are adaptively reset.

[0017] This invention also proposes a road surface adhesion coefficient estimation system based on visual dynamics adaptive fusion, used to implement the road surface adhesion coefficient estimation method described above. The system includes: A visual perception module is used to acquire a visual image of the road surface in front of the vehicle, and obtain road surface category information based on the visual image, wherein the road surface category information is used to indicate the semantic category of the road surface on which the vehicle is currently traveling; and determine the visual friction coefficient prior information of the road surface based on the road surface category information, wherein the visual friction coefficient prior information includes friction coefficient range and / or friction coefficient center value. The dynamic estimation module is used to determine the braking pressure level based on the prior information of the visual friction coefficient, and apply a braking pulse to the wheels of the vehicle according to the braking pressure level to excite the tire-road system; and to acquire the dynamic response information of the vehicle during the application of the braking pulse, and obtain an estimated value of the dynamic friction coefficient based on the dynamic response information, wherein the dynamic response information includes wheel speed, vehicle speed and / or longitudinal acceleration; The fusion module is used to determine the fusion weight based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient, and to obtain the final estimated value of the adhesion coefficient of the road surface based on the fusion weight.

[0018] The beneficial effects of this invention are as follows: 1. The estimation method proposed in this invention adaptively determines the braking pressure level and applies braking pulses based on the visual road surface classification results under low slip rate conditions such as straight-line cruising and slow deceleration. This induces a controllable slip rate in the tires, activates the observability of the constrained unscented Kalman filter, and enables the dynamic estimator to obtain a convergent friction coefficient estimate without the need for continuous large-amplitude braking. This avoids decreased driving stability and occupant discomfort caused by blind braking. Furthermore, addressing the problem that pure visual methods can only output discrete semantic categories and cannot provide accurate continuous numerical values, this invention transforms the visual classification results into friction coefficient intervals and center values ​​through a pre-defined mapping relationship, providing a priori physical benchmark for dynamic estimation.

[0019] 2. This invention obtains uncertainty representations from both the visual and dynamic modules, including the visual interval width, the dynamic posterior covariance, and the deviation between the two. Based on these three variables, fuzzy inference dynamically determines the fusion weights. When visual recognition is clear and dynamics have converged, dynamic estimation takes precedence; when dynamics diverge due to insufficient excitation, visual prior estimation takes precedence; when the two information sources conflict, selective retention is based on confidence levels. The final output adhesion coefficient estimate remains continuous and stable during steady-state driving and rapidly tracks changes in the true value after sudden road surface changes. Attached Figure Description

[0020] Figure 1 This is an overall framework diagram of the road surface adhesion coefficient estimation method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the structure of the Interactive Road Surface Perception Network (IRTP-Net) provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the cross-scale enhancement module (CSE) provided in an embodiment of the present invention; Figure 4 This is one of the schematic diagrams of the braking pressure pulse design provided in the embodiments of the present invention; Figure 5 This is the second schematic diagram of the braking pressure pulse design provided in the embodiment of the present invention; Figure 6 This is a vehicle dynamics schematic diagram provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the fuzzy rule matrix for inferring fusion weights in the adaptive fuzzy fusion framework provided in this embodiment of the invention; Figure 8 This is a structural block diagram of the road adhesion coefficient estimation system provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the road surface friction coefficient setting value in a high-low-high friction transition scenario in the experimental verification of this invention; Figure 10This is a graph showing the change of braking pressure over time in a high-low-high friction transition scenario during the experimental verification of this invention. Figure 11 This is a graph showing the change of vehicle speed over time in a high-low-high friction transition scenario during the experimental verification of this invention. Figure 12 This is a comparison chart of the estimated and actual longitudinal force values ​​in a high-low-high friction transition scenario in the first test verification of this invention. Figure 13 This is a comparison chart of the estimated and actual vertical load values ​​in a high-low-high friction transition scenario in the first test verification of this invention. Figure 14 This is a curve showing the change in tire slip ratio during a high-low-high friction transition scenario in the experimental verification of this invention. Figure 15 This is a comparison chart of the adhesion coefficient estimation results in a high-low-high friction transition scenario in the first test verification of this invention; Figure 16 This is a schematic diagram of the road surface friction coefficient setting value in the medium-high-low friction transition scenario in the second test verification of this invention; Figure 17 This is a graph showing the change of braking pressure over time in the medium-high-low friction transition scenario of the second test verification of this invention; Figure 18 This is a graph showing the change of vehicle speed over time in the medium-high-low friction transition scenario in Experimental Verification II of this invention. Figure 19 This is a comparison chart of the estimated and actual longitudinal force values ​​in the medium-high-low friction transition scenario in Experimental Verification II of this invention; Figure 20 This is a comparison chart of the estimated and actual vertical load values ​​in the medium-high-low friction transition scenario of the second test verification of this invention; Figure 21 This is a curve showing the change in tire slip ratio during the medium-high-low friction transition scenario in Experimental Verification II of this invention. Figure 22 This is a comparison chart of the adhesion coefficient estimation results in the medium-high-low friction transition scenario in the second test verification of this invention. Detailed Implementation

[0021] The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to make and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0022] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. When used in this specification, the terms “comprising,” “including,” and / or “containing” mean that the associated integers, steps, operations, elements, and / or components are present, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or that other features, integers, steps, operations, elements, components, and / or groups may be added to the system / method.

[0023] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0024] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0025] The technical solutions in this application will now be described with reference to the accompanying drawings.

[0026] First Embodiment Please see Figure 1 and Figure 2 This embodiment provides a road adhesion coefficient estimation method based on visual-dynamic adaptive fusion. This method can be used in both autonomous and non-autonomous vehicles to estimate the adhesion coefficient between the tires and the road surface in real time, providing key input parameters for the vehicle's active safety control system (such as ABS and ESC) and trajectory planning module. The overall framework of this embodiment includes three functional modules: a visual perception module, a dynamic estimation module, and an adaptive fuzzy fusion module. The visual perception module first provides a priori intervals for the friction coefficient through road image classification, achieving look-ahead perception; the dynamic estimation module refines the estimation using vehicle dynamic response under visually guided graded braking pulse excitation; the fusion module integrates the two information sources through a multi-dimensional fuzzy inference system and outputs the final adhesion coefficient estimate.

[0027] Please see Figure 1 The road surface adhesion coefficient estimation method provided in this embodiment includes the following steps: Step S1: Obtain a visual image of the road surface in front of the vehicle, and obtain road surface category information based on the visual image.

[0028] The road surface category information is used to indicate the semantic category of the road surface the vehicle is currently traveling on, such as dry asphalt road surface, wet asphalt road surface, icy and snowy road surface, etc. Specifically, the vehicle is equipped with an onboard camera, which can be a forward-facing camera installed on the windshield of the vehicle, used to capture images of the road surface in front of the vehicle's direction of travel. The camera can be one or more of a monocular camera, a binocular camera, or a surround-view camera; this embodiment does not limit this. In one implementation, the camera continuously captures road surface images at a preset frequency and transmits the captured images to an onboard computing platform for processing in real time.

[0029] As a preferred implementation, referring to the figure, obtaining road surface category information from the visual image specifically includes: inputting the visual image into an interactive road surface perception network (IRTP-Net) to obtain the road surface category information. The interactive road surface perception network includes a four-stage convolutional encoder, inspired by the downsampling path of U-Net, but omitting the decoder and skip connection structure to significantly reduce computational complexity. The encoder gradually reduces the spatial resolution of the feature map (e.g., from 56×56 to 7×7) through convolutional operations with a stride of 2, while increasing the channel dimension from 64 to 512. In each stage of the encoder, the proposed cross-scale enhancement module (CSE) is connected after the convolutional backbone network. The final feature map is input into a global average pooling layer and a fully connected classifier to complete the road surface type recognition task. This model design achieves multi-scale feature extraction of road surface texture and effectively enhances the ability to discriminate road surface categories related to friction characteristics. The loss function used in this embodiment is the cross-entropy loss function.

[0030] Furthermore, the cross-scale enhancement module (CSE) is a core component of IRTP-Net, designed to address the issue of road surface images exhibiting different visual patterns at different scales: local microtextures provide fine-grained information, while global macro-patterns provide contextual cues; traditional single-scale networks struggle to capture these two complementary features simultaneously. In one specific implementation, the cross-scale enhancement module processes the input feature map in the following manner.

[0031] First, the CSE module receives the input feature map. The input feature map is processed through two decoupled depthwise separable convolutional branches. The first branch uses a standard 3×3 depthwise separable convolution to capture local spatial patterns, obtaining the first branch features. This branch focuses on extracting subtle texture information from road surface images, such as the graininess and cracks in the road surface. The second branch uses cascaded 5×5 dilated convolutions to capture long-range dependencies in the input feature map, obtaining the second branch features. This branch focuses on extracting broad contextual information and macroscopic structural patterns, such as the boundary between the road surface and the background area, and large areas of water or snow accumulation. By setting dilated convolutions, the receptive field can be effectively expanded without increasing the number of additional parameters.

[0032] Then, channel concatenation and point-by-point convolution are performed on the first branch features and the second branch features to obtain multi-scale fused features. Specifically, the output feature maps of both branches are projected onto the C / 2 channel and then concatenated to form a unified multi-scale representation. Its mathematical expression is: in, _ indicates pointwise convolution operation, used to adjust the number of channels; Concat indicates concatenation operation along the channel dimension.

[0033] Next, the multi-scale fused features are processed through spatial attention branch and channel attention branch. Process to obtain an attention map Specifically, the spatial attention branch pairs Average pooling and max pooling operations are performed along the channel direction. Then, a spatial attention map is generated using a 7×7 convolution and a sigmoid activation function to identify which spatial locations in the feature map are more critical for road surface classification. Simultaneously, global average pooling is performed on the channel attention branch. After passing through a bottleneck structure (composed of 1×1 convolution, batch normalization, and an activation function), softmax normalization is applied to generate branch-specific coefficients for recalibrating features from different channels. The two sets of weights are multiplied element-wise to generate locally modulated weights, which are then applied to the features of the corresponding branches.

[0034] Finally, the input feature map is multiplied element-wise with the attention map to obtain the enhanced output feature map. : Here, ⊙ represents element-wise multiplication. Through the above processing, the CSE module can jointly model spatial structure and channel-level dependencies, making the network more focused on discriminative regions related to frictional properties, while suppressing interference from irrelevant background or noise.

[0035] Step S2: Determine the prior information of the visual friction coefficient of the road surface based on the road surface category information.

[0036] The prior information on the visual friction coefficient includes friction coefficient intervals and / or the center value of the friction coefficient. The output of IRTP-Net is a discrete semantic label indicating the road surface category. To bridge the gap between the semantic classification results and the physical control parameters, this embodiment establishes a quantitative mapping strategy.

[0037] Specifically, based on a preset mapping relationship, the road surface category information is mapped to a corresponding friction coefficient range, and then the friction coefficient range is determined as the prior information of the visual friction coefficient. Its mathematical expression is: in, For the category predicted by the model, This is a mapping function used to assign a specific range of friction coefficients to various road surface categories. In its implementation, this invention defines six typical road surface conditions. For each category, the center value of its corresponding physical friction range is used as the visual coefficient value. For example, the preset mapping relationship can be configured as follows: dry asphalt pavement is mapped to the friction coefficient range [0.7, 0.9], with a center value of 0.8; wet asphalt pavement is mapped to the range [0.4, 0.6], with a center value of 0.5; icy and snowy pavement is mapped to the range [0.1, 0.3], with a center value of 0.2; dry cement pavement is mapped to the range [0.6, 0.8], with a center value of 0.7; gravel pavement is mapped to the range [0.5, 0.7], with a center value of 0.6; and snow-covered pavement is mapped to the range [0.2, 0.4], with a center value of 0.3. It is understood that the above values ​​are only illustrative examples, and in actual applications, they can be adjusted according to the specific vehicle model, tire type, and experimental calibration results.

[0038] In this way, the discrete semantic categories output by the visual perception network can be effectively transformed into prior information of continuous friction coefficients with physical scale, providing a reliable feedforward benchmark for subsequent dynamic estimation and solving the problem of the lack of physical numerical scale in pure vision methods.

[0039] Step S3: Determine the braking pressure level based on the prior information of the visual friction coefficient, and apply braking pulses to the wheels of the vehicle according to the braking pressure level to excite the tire-road system.

[0040] To overcome the limitations of purely visual estimation methods and address the unobservable nature of purely dynamic methods under low-excitation conditions, this embodiment introduces a visually guided selective excitation strategy. This strategy directly selects the most suitable braking pressure level based on the visual road surface classification output, ensuring sufficient observability while avoiding over-excitation.

[0041] Specifically, the visual road surface classification output provides the range of friction coefficients of the road surface ahead, indicating the friction level of the road surface about to be entered. Based on this information, the strategy in this embodiment directly selects the most suitable braking pressure level: a light pulse corresponds to low-friction surfaces (μ<0.2), a medium pulse corresponds to medium-friction surfaces (0.2<μ<0.5), and a strong pulse corresponds to high-friction surfaces (μ>0.5). This selective method avoids over-excitation on low-friction surfaces, ensuring vehicle stability, while ensuring sufficient excitation on high-friction surfaces to activate dynamic observability.

[0042] In one specific implementation, a predetermined braking pressure level is applied to the wheels of the vehicle in the form of a trapezoidal braking pulse. The trapezoidal braking pulse has the characteristics of a smooth rise, sustained duration, and smooth descent. Compared to step braking, the trapezoidal pulse avoids causing severe impacts on the smoothness of vehicle operation and improves passenger comfort.

[0043] Meanwhile, to prevent safety risks caused by visual misclassification, such as mistakenly applying a stronger pulse on a low-friction surface, this embodiment applies a slip ratio safety constraint during pulse execution. Specifically, during the application of the braking pulse, the tire slip ratio of the vehicle is monitored in real time; if the tire slip ratio exceeds a preset safety threshold, the application of the braking pulse is immediately terminated, regardless of the selected level. This constraint ensures that the estimation process remains physically safe even in the adverse condition of visual misclassification. It is understood that the preset safety threshold can be calibrated according to the actual vehicle model and tire characteristics, for example, it can be set between 15% and 20% to prevent wheel lock-up.

[0044] Please see Figure 4 and Figure 5 The system applies target pressure in the form of trapezoidal braking pulses based on the visual road surface classification results. For example... Figure 4 As shown, with target braking pressure For example, this pulse from It rises at a preset slope at all times, at Reached in seconds ;exist to Maintain during seconds Constant; self Starting from the second, it will fall back at the preset slope, at... It drops to zero in seconds.

[0045] To prevent safety risks caused by visual classification misjudgment, this embodiment introduces a slip ratio safety constraint mechanism during braking pulse application. The system monitors the absolute value of transient tire slip ratio in real time. ,once Exceeding the preset safe slip rate threshold Regardless of the current situation Figure 5 What braking pressure level is indicated (e.g.) , or In all cases, the applied braking pressure is immediately and forcibly terminated. This mechanism effectively avoids excessive tire slippage or improper intervention of the anti-lock braking system (ABS), ensuring the physical safety of the excitation estimation process. Figure 5 As shown, after each pulse ends (including complete execution or safe termination), the system will enter the "slip rate rule" (evaluation phase) between two pulses. The system will then evaluate the actual slip rate. The system compares the results with a preset slip ratio threshold corresponding to a specific level to verify whether the constrained unscented Kalman filter has reached convergence. The preset slip ratio threshold is derived based on the Dugoff tire model; in this embodiment, 0.007, 0.013, and 0.023 are selected as critical slip ratio boundaries to distinguish between low, medium, and high friction ranges. Through the above evaluation, the system ultimately generates discrete convergent state variables. This is used to identify the current actual tire-road friction zone.

[0046] Meanwhile, regardless of whether it is a strong pulse or a medium pulse, the tire slip ratio is monitored in real time during execution. Once the slip ratio exceeds the preset safety threshold, the pulse application is terminated immediately to ensure the safety of the braking process under any road surface conditions.

[0047] As a further improvement to this embodiment, the method further includes a convergence state determination step. Specifically, after the braking pulse is applied or completes normally (or terminates prematurely due to exceeding a safety threshold), convergence state information is determined based on the comparison results of the tire slip ratio with multiple slip ratio thresholds. The convergence state information is used to indicate the friction range in which the estimation result of the constrained unscented Kalman filter is located.

[0048] The aforementioned slip ratio thresholds are derived from theoretical analysis of the Dugoff tire model: under a given braking pressure, tire slip ratio exhibits significant characteristics corresponding to different friction levels. For example, slip ratio thresholds can be set to 1% (distinguishing between extremely low friction), 3% (distinguishing between low and medium friction), 8% (distinguishing between medium and high friction), etc. This evaluation generates discrete convergent states. This is used to identify the current friction interval, such as a low friction interval [0, 0.2], a medium friction interval [0.2, 0.5], or a high friction interval [0.5, 1.0]. Then, based on the convergence state information, the kinetic prior interval and prior confidence level are determined. The kinetic prior interval is determined based on the friction interval corresponding to the convergence state information, and the prior confidence level characterizes the reliability of this prior information and plays a role in the subsequent fusion process.

[0049] Step S4: Obtain the dynamic response information of the vehicle during the application of the braking pulse, and obtain an estimated value of the dynamic friction coefficient based on the dynamic response information.

[0050] The dynamic response information includes wheel speed, vehicle speed, and / or longitudinal acceleration. The vehicle is equipped with wheel speed sensors (typically located at each wheel), an inertial measurement unit (IMU for measuring longitudinal acceleration), and a vehicle speed estimation module (obtained via GPS or wheel speed fusion). These sensors acquire data at high frequencies (e.g., 100 Hz) during the application of braking pulses.

[0051] Please see Figure 6 It shows the vehicle dynamics principle diagram used in this invention, namely a simplified bicycle model. This model focuses on the straight braking condition, and retains the key longitudinal load transfer dynamic characteristics while ignoring lateral dynamics, in order to describe the longitudinal dynamic behavior of the vehicle.

[0052] In a preferred embodiment, this step specifically includes the following sub-steps.

[0053] First, the vertical load on the front axle of the vehicle is dynamically corrected based on the vehicle's longitudinal acceleration to obtain the corrected vertical load. Since the vehicle's mass shifts forward during braking, the vertical load on the front axle increases; ignoring this effect would lead to an inaccurate estimation of the friction coefficient. Specifically, the vertical load is dynamically estimated based on the longitudinal acceleration, taking into account the load shift effect caused by braking; its calculation formula is as follows: in This represents the static load distribution, where m is the vehicle mass. For the height of the center of mass, Wheelbase For longitudinal acceleration. This load transfer estimate ensures accurate normalization of the vertical force under braking conditions, making the friction coefficient estimate unaffected by load variations.

[0054] Then, the tire longitudinal slip ratio is determined based on the corrected vertical load, the wheel speed, and the vehicle speed. Longitudinal slip ratio is a physical quantity that describes the degree of tire slippage. Its calculation method is well known to those skilled in the art; for example, it can be calculated using formulas. , where v is the vehicle speed, ω is the wheel speed, and r is the tire rolling radius.

[0055] Finally, based on the tire longitudinal slip ratio and the corrected vertical load, the friction coefficient of the vehicle is estimated using a constrained unscented Kalman filter (CUKF) to obtain the estimated dynamic friction coefficient. The unscented Kalman filter (UKF) is a state estimation method for nonlinear systems. It approximates the posterior distribution of the state by sampling sigma points, without requiring the calculation of the Jacobian matrix, making it suitable for the nonlinear characteristics of the tire model in this embodiment.

[0056] Specifically, the CUKF framework is formulated as a parameter estimation problem, with state vectors... This framework is used to estimate the friction coefficients of the front and rear axles. It utilizes wheel speed, vehicle speed, and tire vertical force as inputs, with the observation vector consisting of longitudinal forces. CUKF adapts to environmental changes while recursively estimating the state vector. During the time update phase, sigma points are generated and propagated through a random walk model; during the measurement update phase, the sigma points are mapped to the observation space using the Dugoff tire model, and the desired longitudinal forces are calculated.

[0057] Furthermore, the constrained unscented Kalman filter skips the measurement update step and keeps the state frozen when the tire longitudinal slip ratio is lower than a preset slip ratio threshold. In other words, this embodiment introduces a low-speed / low-excitation locking mechanism: if the vehicle speed or slip ratio is lower than the corresponding threshold (e.g., ...), the system will remain frozen. or If the UKF update step is skipped and the state remains frozen, this mechanism avoids the divergence problem caused by the filter's numerical sensitivity when there is a lack of effective stimulus.

[0058] In addition, CUKF introduces physical boundary constraints: constraining the state estimates to a physically reasonable range, for example... This is to ensure that the estimation results are always within a reasonable range of [0.05, 1.0].

[0059] In a preferred embodiment, the constrained unscented Kalman filter uses the Dugoff tire model as the observation model. The Dugoff tire model describes the longitudinal force generation characteristics through the analytical relationship between slip ratio and friction coefficient. Its advantages lie in its high computational efficiency and ability to accurately capture the force-slip relationship in the low slip ratio region. The Dugoff tire model describes the longitudinal force according to the following formula. The relationship between the longitudinal slip ratio κ and the coefficient of friction μ of the tire: The normalized slip parameter λ is defined as follows: The function f(λ) is a piecewise function, and its expression is: in For longitudinal stiffness, For normalized slip parameters, For vertical loads, This is the tire-road friction coefficient. When... When the tire is in a state of partial slippage, the longitudinal force and the slip ratio have an approximately linear relationship; when When the tire enters a state of complete slippage, the longitudinal force reaches the friction limit.

[0060] Please see Figure 7 This diagram illustrates the fusion flowchart of the adaptive fuzzy fusion framework. The vision module provides the visual friction coefficient interval, while the dynamics module decodes the prior dynamic interval and prior confidence based on the convergence state information. The two information sources are first fused using interval intersection logic: if the two intervals intersect, the target interval is the intersection; if the intersection is empty, the interval corresponding to the higher confidence is retained. The target center value is then calculated as the fusion reference point. Simultaneously, the visual interval width, the posterior dynamic covariance, and the deviation between the dynamic estimate and the target interval are obtained. These three variables are input into the fuzzy inference system, and after fuzzification, rule-based inference, and defuzzification, the basic fusion weights are output. After prior confidence correction and deviation suppression, the final fusion weights are obtained. Finally, the final adhesion coefficient estimate is calculated and output using a soft-pull mechanism.

[0061] As a further improvement to this embodiment, the method further includes: adaptively resetting the process noise covariance and state error covariance of the constrained unscented Kalman filter when a sudden change in the road surface category information is detected. Specifically, if the road surface category detected by the visual perception network changes over multiple consecutive frames (e.g., three consecutive frames), it is determined that the vehicle is experiencing a sudden change in road surface (e.g., moving from a dry asphalt road surface to an icy or snowy road surface). At this time, the process noise covariance Q and state error covariance P are adaptively reset, enabling the filter to respond quickly to road surface changes and accelerate convergence to a new friction coefficient value.

[0062] Step S5: Determine the fusion weight based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient, and obtain the final estimated value of the adhesion coefficient of the road surface based on the fusion weight.

[0063] To synergistically integrate the forward-looking capabilities of visual estimation with the physical accuracy of dynamic estimation, this embodiment proposes an adaptive fuzzy fusion framework. The basic design principle of this framework is to dynamically allocate weights based on the real-time reliability and confidence level of each information source, thereby leveraging their respective strengths while suppressing their weaknesses, achieving a balanced estimation of the friction coefficient.

[0064] In one specific implementation, this step includes the following sub-steps.

[0065] First, the width of the visual interval is determined based on the prior information of the visual friction coefficient. ,in This represents the friction coefficient range output by the vision module. A narrower range indicates a more confident visual recognition result, while a wider range indicates greater uncertainty in the visual recognition. Understandably, the width of the visual range can be derived from the information entropy of the Softmax probability distribution output by IRTP-Net; the more concentrated the probability distribution, the narrower the range, and the higher the confidence level.

[0066] Simultaneously, the posterior covariance of the estimated dynamic friction coefficient is obtained. The posterior covariance directly reflects the confidence level of the dynamic estimator—low covariance indicates that the filter has converged to a stable estimate, while high covariance indicates that adaptive adjustments are underway or that the excitation is insufficient. The posterior covariance is output by CUKF after each measurement update.

[0067] Then, the estimated value of the dynamic friction coefficient is determined. Deviation from the target interval The target interval is determined based on the prior information of the visual friction coefficient and the prior interval of the dynamics. Specifically, it is derived from the convergence state of the selective excitation inference module. Decoded as a priori interval of dynamics The vision module also provides the friction zone obtained from road surface image recognition. The two information sources are fused through intersection logic to construct a target interval representing the most reliable friction range. : When the intersection is valid When both information sources agree on the feasible range, the target interval is directly adopted. When the intersection is empty... When this indicates a conflict between vision and dynamics, this invention employs a confidence-based conflict resolution strategy: retaining intervals with high correlation confidence. This design ensures that the fusion framework intelligently handles discrepancies, rather than blindly averaging conflicting information. The target center is then calculated as a reference point for subsequent fusion. Then, the estimated value of the dynamic friction coefficient is determined. Deviation from the target interval : Large deviations indicate that the dynamic estimates may have drifted or not yet adapted to recent road surface changes, and their weight in the fusion should be reduced accordingly.

[0068] Next, the visual range The posterior covariance and the aforementioned deviation Input into the fuzzy inference system to obtain basic fusion weights. Specifically, the three inputs mentioned above are mapped to fuzzy sets {low, medium, high} through triangular membership functions, forming the input space for rule-based inference. The design of the fuzzy rule base has a clear physical logic: when the uncertainty of both CUKF and vision is low, the system should strongly trust the CUKF estimate; when the uncertainty of both is high, the system should favor the target center; and intermediate cases receive appropriate weights. Inference uses a product-minimum composition method to generate basic weights. .

[0069] Then, based on the aforementioned prior confidence level The basic fusion weights are then corrected to obtain the corrected fusion weights. The specific method for prior correction is as follows: a lower bound is set for the weights to prevent the fuzzy system from underestimating valuable physical priors. Further, the corrected fusion weights are subjected to deviation suppression based on the deviation amount to obtain the fusion weights. : in, Let δ be the membership value of the fuzzy set "high" in relation to the deviation. The purpose of deviation suppression correction is to reduce the fusion weights when the dynamic estimate deviates significantly from the target interval, thus preventing the filter from locking onto erroneous predictions.

[0070] Finally, based on the fusion weight w, the final adhesion coefficient estimate is calculated according to the following soft-pull mechanism formula. : Finally, the final adhesion coefficient estimate is constrained to a preset physical boundary range [0.01, 1] to ensure that the output result is always within a physically reasonable range. This final estimate can be output to the vehicle's ABS, ESC, or autonomous driving decision control system, providing a key input for the vehicle's active safety control.

[0071] Through the steps of the first embodiment described above, this embodiment achieves a deep fusion of visual perception and dynamic estimation: the visual module provides a forward-looking prior interval for the friction coefficient, the dynamic module achieves observable physical estimation under visually guided hierarchical excitation, and the fusion module dynamically integrates the two information sources through multi-dimensional fuzzy inference. The entire method solves the problem of unobservable dynamics under low-excitation conditions without requiring continuous large-amplitude braking, while overcoming the lack of physical numerical accuracy in pure visual methods, thus achieving accurate estimation of the adhesion coefficient under all operating conditions.

[0072] Second Embodiment Please see Figure 8 This embodiment proposes a road surface adhesion coefficient estimation system based on visual dynamics adaptive fusion. The system is used to implement the road surface adhesion coefficient estimation method as described in the first embodiment. Since the system embodiment in this embodiment corresponds to the aforementioned method embodiment, any details not described in this embodiment can be found in the relevant descriptions in the aforementioned method embodiments, and will not be repeated here.

[0073] The road surface adhesion coefficient estimation system provided in this embodiment includes a visual perception module, a dynamic estimation module, and a fusion module.

[0074] The visual perception module acquires a visual image of the road surface ahead of the vehicle and obtains road surface category information based on the visual image. This road surface category information indicates the semantic category of the road surface the vehicle is currently traveling on. Furthermore, based on the road surface category information, it determines prior information about the visual friction coefficient of the road surface, where the prior information includes a friction coefficient range and / or a friction coefficient center value. Specifically, the visual perception module may include an onboard camera (such as a forward-facing camera) and an image processing unit. The image processing unit deploys the aforementioned Interactive Road Surface Perception Network (IRTP-Net), which includes a four-level convolutional encoder and a cross-scale enhancement module (CSE) for extracting multi-scale texture features from the road surface image and performing classification. The specific structure and processing flow of the IRTP-Net and CSE modules have been described in detail in the first embodiment and will not be repeated here.

[0075] The dynamic estimation module is used to determine the braking pressure level based on the prior information of the visual friction coefficient, and apply braking pulses to the wheels of the vehicle according to the braking pressure level to excite the tire-road system; and to acquire the dynamic response information of the vehicle during the application of the braking pulses, and obtain an estimated value of the dynamic friction coefficient based on the dynamic response information, wherein the dynamic response information includes wheel speed, vehicle speed and / or longitudinal acceleration. Specifically, the dynamic estimation module may include a braking execution unit (such as an electro-hydraulic braking system EHB or an electro-mechanical braking system EMB), wheel speed sensors, an inertial measurement unit (IMU), and a constrained unscented Kalman filter (CUKF) calculation unit. The braking execution unit selects the corresponding braking pressure level according to the prior information of the friction coefficient output by the visual perception module, and applies braking in the form of trapezoidal pulses. The wheel speed sensors and IMU acquire wheel speed and longitudinal acceleration data during braking. The CUKF calculation unit processes the acquired dynamic data based on the Dugoff tire model and outputs an estimated value of the dynamic friction coefficient. The selective excitation strategy for visual guidance, the slip ratio safety constraint, the specific filtering process of CUKF, and the mathematical expression of the Dugoff tire model have been described in detail in the first embodiment, and will not be repeated here.

[0076] The fusion module is used to determine fusion weights based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient, and to obtain the final estimated value of the road surface adhesion coefficient based on the fusion weights. Specifically, the fusion module may include a fuzzy inference unit and a soft-pull calculation unit. The fuzzy inference unit uses visual intervals... The posterior covariance and the aforementioned deviation As input, the basic fusion weights are derived through inference using triangular membership functions and a predefined fuzzy rule base. After prior confidence correction and bias suppression correction, the final fusion weight w is output. The soft-pull calculation unit calculates according to the formula... The final estimated adhesion coefficient is calculated and constrained to the physical boundary range of [0.01, 1]. The method for constructing the target interval, the design logic of the fuzzy rule base, and the specific calculation methods for prior correction and bias suppression have been described in detail in the first embodiment and will not be repeated here.

[0077] Through the collaborative work of the aforementioned visual perception module, dynamics estimation module, and fusion module, the system in this embodiment can estimate the tire-road adhesion coefficient in real time and accurately, providing reliable input parameters for vehicle active safety control and autonomous driving decisions.

[0078] Third Embodiment This embodiment provides a vehicle including a processor and a memory for storing processor-executable instructions. The processor is configured to perform the steps of the road adhesion coefficient estimation method described in the first embodiment. The vehicle may be a hybrid vehicle, an electric vehicle, a fuel cell vehicle, or other types of vehicle, and may also be an autonomous or semi-autonomous vehicle. For the specific steps of the method, please refer to the relevant description in the first embodiment; they will not be repeated here.

[0079] Specifically, the vehicle may include various subsystems, such as an infotainment system, a perception system, a decision control system, a drive system, and a computing platform. The perception system includes cameras for acquiring road surface images, wheel speed sensors for acquiring wheel speeds, and inertial measurement units for acquiring longitudinal acceleration. The computing platform includes at least one processor and a memory. The processor executes instructions stored in the memory to implement all or part of the steps described in the first embodiment. The processor can be any conventional processor, such as a commercially available CPU, and may also include graphics processing units (GPUs), field-programmable gate arrays (FPGAs), system-on-a-chip (SoCs), application-specific integrated circuits (ASICs), or combinations thereof. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0080] Fourth embodiment This embodiment provides a computer-readable storage medium storing computer program instructions thereon. When executed by a processor, these program instructions implement the steps of the road surface adhesion coefficient estimation method described in the first embodiment. The computer-readable storage medium can be any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0081] Fifth embodiment This embodiment provides a computer program product, including a computer program. When executed by a processor, the computer program implements the steps of the road adhesion coefficient estimation method described in the first embodiment. This computer program product can be included in the vehicle's computing platform or stored independently in an external storage medium for vehicle loading and use.

[0082] Experimental verification I. High-Low-High Friction Transition Scenarios To verify the effectiveness and robustness of the adaptive fuzzy vision-prior fusion framework described in this invention, a joint simulation platform of CarSim and MATLAB / Simulink was used for implementation and verification. CarSim provided a high-fidelity vehicle dynamics model (vehicle mass m = 1650 kg), while the fusion algorithm described in this invention was implemented in Simulink. The two software programs interacted in real time using data such as vehicle speed, wheel speed, and braking pressure. The test vehicle entered the track at an initial speed of 100 km / h, and the estimation logic was triggered by intermittent braking pulses. This embodiment simulated the scenario of a vehicle driving from a dry asphalt road onto an ice surface and then returning to a dry road surface, with the actual friction coefficient varying according to a step trajectory of 0.8 → 0.3 → 0.8.

[0083] Please see Figure 9 and Figures 10 to 15 The specific execution process is as follows: (1) In the initial high friction stage (0~5s), the vision module continuously identifies the road surface as dry asphalt and provides the friction range [0.6, 0.7]; the dynamics module applies braking pressure through a selective excitation strategy and estimates that the curve converges to 0.8; (2) At t≈5s, the vehicle enters the low-friction road surface, and the friction coefficient suddenly drops to 0.3. The vision module identifies the change in road surface type and provides the friction range [0.2, 0.35]; the event-driven covariance reset module detects the sudden change in road surface and adaptively resets the noise covariance Q and the state error covariance P; with the prior assistance of the vision range, the algorithm quickly detects the change, and the estimated value converges to 0.3 within about 1.0 seconds without significant overshoot; (3) At t≈10s, the vehicle returns to the high-friction surface, and the estimated value recovers to 0.8; (4) During the coasting phase (braking interval), the dynamic estimate freezes the previous convergence value, while the visual estimate maintains continuous estimation to ensure the stability of the system when the dynamic excitation is insufficient.

[0084] At the same time, estimate the longitudinal force (F) x Closely tracks the true value, estimating the vertical load F z Dynamically capture the load transfer caused by braking, from static load to braking peak load, to ensure that friction calculations are based on correct vertical force normalization.

[0085] Implementation Case 2: Medium-High-Low Friction Transition Scenario This embodiment further verifies the algorithm's adaptability to different initial conditions and continuous transition scenarios. The actual friction coefficient changes in a sequence from 0.5 to 0.8 to 0.3 to simulate the transition from wet asphalt to dry asphalt and then to icy and snowy roads.

[0086] Please see Figure 16 and Figures 17 to 22 The specific execution process is as follows: (1) In the initial stage, the vision module identifies the road surface as wet asphalt and provides the friction range [0.4, 0.5]; the dynamics module applies moderate braking pressure through a selective excitation strategy and estimates that the curve converges to 0.5; (2) At t≈5s, the friction coefficient jumps to 0.8, and the estimator accurately tracks the rising step; at this time, spatial interval fusion uses the high confidence prior derived from dynamics to prevent estimation drift. (3) At t≈10s, the friction coefficient drops sharply to 0.3. Despite the drastic change, the estimated value converges to the true value within a short time window. The slip ratio curve confirms that the braking operation effectively stimulates the system dynamics, providing sufficient observability for CUKF. The fusion framework successfully filters out noise and potential misjudgments during the critical transition. (4) Vertical load F at each stage z The consistency of the estimates further validates the reliability of the observation equations based on the bicycle model.

[0087] The two implementation examples above demonstrate that the fusion framework described in this invention achieves rapid convergence and stable estimation under abrupt changes in the friction coefficient, and maintains continuous estimation through visual look-ahead during the steady-state driving phase, effectively balancing the responsiveness during the transition period and the smoothness during the steady-state period.

[0088] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this disclosure.

[0089] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0090] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for estimating road surface adhesion coefficient based on visual-dynamic adaptive fusion, characterized in that, The method includes: A visual image of the road surface in front of the vehicle is acquired, and road surface category information is obtained based on the visual image, wherein the road surface category information is used to indicate the semantic category of the road surface on which the vehicle is currently traveling; Based on the road surface category information, the prior information of the visual friction coefficient of the road surface is determined, wherein the prior information of the visual friction coefficient includes the friction coefficient range and / or the center value of the friction coefficient. Based on the prior information of the visual friction coefficient, the braking pressure level is determined, and braking pulses are applied to the wheels of the vehicle according to the braking pressure level to excite the tire-road system. The vehicle's dynamic response information during the application of the braking pulse is obtained, and an estimated value of the dynamic friction coefficient is obtained based on the dynamic response information, wherein the dynamic response information includes wheel speed, vehicle speed and / or longitudinal acceleration; Based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient, the fusion weight is determined, and based on the fusion weight, the final estimated value of the adhesion coefficient of the road surface is obtained.

2. The method for estimating the road surface adhesion coefficient according to claim 1, characterized in that, The step of obtaining road surface category information based on the visual image includes: The visual image is input into an interactive road surface perception network to obtain the road surface category information. The interactive road surface perception network includes a four-level convolutional encoder, and each stage of the four-level convolutional encoder is connected to a cross-scale enhancement module.

3. The method for estimating the road surface adhesion coefficient according to claim 2, characterized in that, The cross-scale enhancement module processes the input feature map in the following way: The input feature map is captured by a first depthwise separable convolutional branch to obtain the first branch features. The input feature map is captured by a second depthwise separable convolutional branch to obtain the second branch features, wherein the second depthwise separable convolutional branch includes cascaded 5×5 dilated convolutions; Channel splicing and point-by-point convolution are performed on the first branch features and the second branch features to obtain multi-scale fused features; The multi-scale fusion features are processed by spatial attention branch and channel attention branch to obtain an attention map; The input feature map is multiplied element-wise with the attention map to obtain the enhanced output feature map.

4. The method for estimating the road surface adhesion coefficient according to claim 1, characterized in that, The step of obtaining an estimated value of the dynamic friction coefficient based on the dynamic response information includes: Based on the longitudinal acceleration of the vehicle, the vertical load on the front axle of the vehicle is dynamically corrected to obtain the corrected vertical load. The tire longitudinal slip ratio is determined based on the corrected vertical load, the wheel speed, and the vehicle speed. Based on the tire longitudinal slip ratio and the corrected vertical load, the friction coefficient of the vehicle is estimated using a constrained unscented Kalman filter to obtain the estimated value of the dynamic friction coefficient. Specifically, when the longitudinal slip ratio of the tire is lower than a preset slip ratio threshold, the constrained unscented Kalman filter skips the measurement update step and keeps the state frozen.

5. The method for estimating the road surface adhesion coefficient according to claim 4, characterized in that, Applying braking pulses to the wheels of the vehicle according to the braking pressure level includes: Braking pulses of the braking pressure level are applied to the wheels of the vehicle in the form of trapezoidal braking pulses; During the application of the braking pulse, the tire slip ratio of the vehicle is monitored in real time; If the tire slip ratio exceeds a preset safety threshold, the application of the braking pulse will be terminated immediately.

6. The method for estimating the road surface adhesion coefficient according to claim 5, characterized in that, The method further includes: After the braking pulse is applied or terminated, convergence state information is determined based on the comparison results of the tire slip ratio and multiple slip ratio thresholds, wherein the convergence state information is used to indicate the friction range in which the estimation result of the constrained unscented Kalman filter is located. Based on the convergence state information, the dynamic prior interval and prior confidence level are determined.

7. The method for estimating the road surface adhesion coefficient according to claim 6, characterized in that, The step of determining the fusion weight based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient includes: The visual interval width is determined based on the prior information of the visual friction coefficient. Obtain the posterior covariance of the estimated dynamic friction coefficient; Determine the deviation between the estimated value of the dynamic friction coefficient and the target interval, wherein the target interval is determined based on the prior information of the visual friction coefficient and the prior dynamic interval; The visual interval width, the posterior covariance, and the bias are input into the fuzzy inference system to obtain the basic fusion weights. The basic fusion weights are corrected based on the prior confidence level to obtain the fusion weights.

8. The method for estimating the road surface adhesion coefficient according to claim 7, characterized in that, The step of obtaining the final adhesion coefficient estimate of the road surface based on the fusion weights includes: The target center value is determined based on the prior information of the visual friction coefficient and the prior dynamic interval. The final estimated adhesion coefficient is calculated using the following formula: ； in, This is the estimated final adhesion coefficient. The fusion weights are... This is the estimated value of the dynamic friction coefficient. The target center value; The final adhesion coefficient estimate is constrained to a preset physical boundary range.

9. The method for estimating the road surface adhesion coefficient according to claim 4, characterized in that, The prior information on the visual friction coefficient includes a friction coefficient range. Determining the prior information on the visual friction coefficient of the road surface based on the road surface category information includes: According to the preset mapping relationship, the road surface category information is mapped to the corresponding friction coefficient range; The range of friction coefficients is determined as the prior information of the visual friction coefficient; The method further includes: When a sudden change in the road surface category information is detected, the process noise covariance and state error covariance of the constrained unscented Kalman filter are adaptively reset.

10. A road surface adhesion coefficient estimation system based on visual dynamics adaptive fusion, used to implement the road surface adhesion coefficient estimation method as described in any one of claims 1-9, characterized in that, The system includes: A visual perception module is used to acquire a visual image of the road surface in front of the vehicle, and obtain road surface category information based on the visual image, wherein the road surface category information is used to indicate the semantic category of the road surface on which the vehicle is currently traveling; and determine the visual friction coefficient prior information of the road surface based on the road surface category information, wherein the visual friction coefficient prior information includes friction coefficient range and / or friction coefficient center value. The dynamic estimation module is used to determine the braking pressure level based on the prior information of the visual friction coefficient, and apply a braking pulse to the wheels of the vehicle according to the braking pressure level to excite the tire-road system; and to acquire the dynamic response information of the vehicle during the application of the braking pulse, and obtain an estimated value of the dynamic friction coefficient based on the dynamic response information, wherein the dynamic response information includes wheel speed, vehicle speed and / or longitudinal acceleration; The fusion module is used to determine the fusion weight based on the prior information of the visual friction coefficient and the estimated value of the dynamic friction coefficient, and to obtain the final estimated value of the adhesion coefficient of the road surface based on the fusion weight.

Images