Wet road condition recognition method and vehicle

Abstract

This application discloses a method and vehicle for identifying slippery road conditions. The method includes: acquiring sensor data of the target vehicle for identifying slippery road conditions; processing the sensor data with different inputs in parallel using at least two slippery road condition identification models to obtain slippery risk scores output by each model; generating dynamic weight allocation results for each identification model based on real-time sensor data using a reinforcement learning model; and weighting and fusing the slippery risk scores output by each model according to the weight allocation results to obtain the target slippery risk score. This method provides multi-dimensional judgment criteria through parallel processing of multiple models and utilizes reinforcement learning to dynamically optimize the fusion weights based on the environment, effectively overcoming the limitations of single-model or fixed-weight methods, and significantly improving the system's identification accuracy, robustness, and adaptability in complex weather and road scenarios.

Description

Technical Field

[0001] This application relates to the field of automotive technology, specifically to a method for identifying slippery conditions and a vehicle. Background Technology

[0002] Currently, vehicle slippery condition identification mainly relies on a single type of sensor or a fixed threshold judgment strategy. These methods are applicable in scenarios with simple weather and road conditions, but in complex environments such as rain, snow, fog, day and night cycles, and varying road surface materials, they are difficult to fully and accurately perceive the road surface condition, leading to frequent misjudgments and missed judgments. Summary of the Invention

[0003] In view of this, this application aims to provide a method and vehicle for identifying slippery working conditions, which can improve the accuracy of identifying slippery working conditions.

[0004] According to a first aspect of this application, a method for identifying slippery driving conditions is provided, comprising: acquiring sensor data of a target vehicle for identifying slippery driving conditions; processing the sensor data respectively using at least two slippery driving condition identification models to obtain slippery risk scores output by each identification model; generating a weight allocation result based on the sensor data using a reinforcement learning model; the weight allocation result including weight values ​​corresponding to each identification model; and performing weighted fusion of the slippery risk scores output by each identification model according to the weight allocation result to obtain a target slippery risk score for the target vehicle. Therefore, a technical framework for multi-model parallel processing and reinforcement learning dynamic weighted fusion was constructed: multiple heterogeneous recognition models independently analyze sensor data from different dimensions, providing a more comprehensive judgment basis and overcoming the limitation of a single model's perception perspective; the reinforcement learning model continuously learns the mapping relationship between environmental features and the recognition performance of each model, dynamically evaluating the credibility of each model under the current conditions, and thus autonomously adjusting its weight in the final decision. This allows the system to maintain stable and accurate judgment capabilities even in complex scenarios such as changes in lighting and weather, thereby comprehensively improving the robustness and generalization of the recognition system in complex environments. Actual tests show that compared to traditional fixed threshold or single-model methods, the multi-model parallel processing and reinforcement learning dynamic weighted fusion framework constructed in this application achieves significant performance improvements. Especially in real-vehicle tests in complex scenarios such as rainy nights, the false alarm rate is significantly reduced, and the accuracy of wet and slippery condition recognition is significantly improved.

[0005] According to a second aspect of this application, a device for identifying slippery working conditions is provided, comprising: The acquisition module is used to acquire sensor data of the target vehicle for wet and slippery conditions identification. The identification module is used to process the sensor data through at least two identification models for slippery working conditions to obtain slippery risk scores output by each identification model. The allocation module is used to generate weight allocation results based on sensor data and a reinforcement learning model; the weight allocation results include the weight values ​​corresponding to each of the recognition models. The fusion module is used to perform weighted fusion of the slippery risk scores output by each of the recognition models according to the weight allocation results, so as to obtain the target slippery risk score of the target vehicle.

[0006] According to a third aspect of this application, a computer-readable storage medium is provided, the storage medium storing a computer program for performing the methods described in any of the above embodiments.

[0007] According to a fourth aspect of this application, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; the processor being configured to perform the method described in any of the above embodiments.

[0008] According to a fifth aspect of this application, a vehicle is provided, including the aforementioned electronic equipment.

[0009] This application discloses a method and vehicle for identifying slippery road conditions. The method includes: acquiring sensor data of the target vehicle for identifying slippery road conditions; processing the sensor data with different inputs in parallel using at least two slippery road condition identification models to obtain slippery risk scores output by each model; generating dynamic weight allocation results for each identification model based on real-time sensor data using a reinforcement learning model; and weighting and fusing the slippery risk scores output by each model according to the weight allocation results to obtain the target slippery risk score. This method provides multi-dimensional judgment criteria through parallel processing of multiple models and utilizes reinforcement learning to dynamically optimize the fusion weights based on the environment, effectively overcoming the limitations of single-model or fixed-weight methods, and significantly improving the system's identification accuracy, robustness, and adaptability in complex weather and road scenarios. Attached Figure Description

[0010] Figure 1 The diagram shown is a flowchart illustrating a method for identifying slippery working conditions according to an embodiment of this application.

[0011] Figure 2 The diagram shown is a block diagram of a wet and slippery working condition identification device provided in one embodiment of this application.

[0012] Figure 3 The diagram shown is a structural block diagram of an electronic device provided in one embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Application Overview Currently, the identification of vehicles in slippery conditions mainly relies on specific types of sensors, but these methods have significant limitations in complex and ever-changing real-world environments. For example, raindrop sensors can monitor rainfall intensity but cannot directly reflect the actual coefficient of adhesion on the road surface; identification methods based on wheel speed difference logic are easily affected by instantaneous dynamics such as vehicle acceleration, braking, or turning, resulting in false alarms; and vision-based algorithms mostly rely on static features of image grayscale or texture, and their recognition performance deteriorates significantly in poor lighting conditions such as nighttime, backlighting, or road shadows.

[0015] To address the aforementioned issues, this application provides a method for identifying slippery road conditions based on multi-model dynamic fusion. This method first acquires various sensor data related to wheel slippage. Then, it uses at least two different identification models to process this data in parallel, generating preliminary slippery risk scores from different dimensions. A reinforcement learning model dynamically generates the weights of each identification model in the final decision based on real-time sensor data. Finally, based on this dynamic weight allocation, the risk scores output by each model are weighted and fused to obtain the final target slippery risk score. This effectively improves the accuracy and robustness of slippery road condition identification in complex environments, reducing the probability of false positives and false negatives.

[0016] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0017] Exemplary methods Figure 1 This is a flowchart illustrating a method for identifying slippery working conditions according to an embodiment of this application. Figure 1 The method described is executed by an onboard computing device, which may be a central domain controller, an advanced driver assistance system controller, or a dedicated vehicle computing unit. This application does not limit the specific device to this type. Figure 1 As shown, the method includes the following: Step S110: Acquire at least two types of sensor data related to the target vehicle and wheel slippage.

[0018] In this embodiment of the application, the target vehicle refers to the vehicle to which this wet and slippery condition identification method is applied. As the subject of sensor data collection and the object of wet and slippery condition identification, it is equipped with a variety of sensors for collecting environmental and vehicle status information in different dimensions.

[0019] In this embodiment, the sensor data refers to raw or pre-processed information collected by various sensors mounted on the vehicle for sensing and judging road surface slipperiness. This data characterizes the vehicle's motion state, environmental state, or visual road surface condition. For example, the sensor data may include, but is not limited to: rainfall data collected by a raindrop sensor, wheel speed data collected by a wheel speed sensor, vehicle acceleration data collected by an inertial measurement unit, and image data of the road ahead collected by an onboard camera.

[0020] Step S120: Process the sensor data using at least two wet and slippery working conditions identification models to obtain wet and slippery risk scores output by each identification model.

[0021] In this embodiment, the identification model refers to a dedicated computational model built based on different artificial intelligence algorithms, used to analyze and judge the risk of slippery conditions from specific types of data. These models can be designed to model slippery conditions from complementary dimensions. For example, some models excel at mining complex correlations between static features of multiple sensors, some models focus on analyzing the temporal evolution patterns of vehicle dynamic signals, and others are used to parse the visual semantic information of road images. By combining these functionally diverse models, the system can obtain a more comprehensive and reliable risk perception capability. Specific models may include, but are not limited to, feature cross-networks, temporal neural networks, and instance segmentation neural networks.

[0022] In this embodiment, the difference in input data refers to the difference in data type or data format between the subset of data received and processed by each recognition model. For example, the feature interaction model mainly receives numerical state data such as rainfall and wheel speed difference; the time series analysis model receives wheel speed and acceleration sequences organized in chronological order; and the image recognition model receives image data of the road ahead.

[0023] In this embodiment, the slippery risk score is a numerical value generated by various identification models to quantify the likelihood of slippery conditions in the current environment or vehicle state. These scores are typically normalized to a continuous value or discrete level within the same range (e.g., between 0 and 1), with higher values ​​indicating a greater slippery risk. For example, a score of 0.85 indicates that the model determines there is a very high probability of slipping, while a score of 0.10 indicates an extremely low risk.

[0024] Step S130: Based on the reinforcement learning model, generate weight allocation results according to the sensor data; the weight allocation results include the weight values ​​corresponding to each of the recognition models.

[0025] In this application embodiment, the reinforcement learning model refers to a type of machine learning model that can autonomously learn the optimal decision-making strategy through continuous interaction with the environment. In this model, a computing unit called an agent observes the environmental state reflected by sensor data, takes actions to adjust the weights of each recognition model, and continuously optimizes its decision-making strategy based on the feedback (i.e., reward) of the overall recognition accuracy and control stability of the system after the actions. For example, in the wet and slippery working condition recognition scenario of this application, the goal of this model is to learn how to assign appropriate weights to the feature interaction model, time series analysis model, and image recognition model in different environments such as rainy days, nights, and high speeds, so that the final comprehensive judgment is the most accurate and stable.

[0026] In this embodiment, the weight allocation result refers to a series of coefficients output by the reinforcement learning model, used to indicate the importance of the output results of each recognition model in subsequent weighted fusion calculations. Each recognition model corresponds to a weight value, which is typically a non-negative real number. The higher the weight value, the more reliable and important the judgment result of the corresponding recognition model is in the current environment. All weight values ​​together constitute the decision basis for multi-model fusion.

[0027] In this embodiment of the application, generating weight allocation results based on sensor data can refer to: generating weight allocation results based on sensor data using a reinforcement learning model, including: inputting an environmental state vector, which is constructed based on real-time sensor data and characterizes the comprehensive environmental conditions of the vehicle, into a pre-trained reinforcement learning model; and having the reinforcement learning model output a weight allocation result indicating the importance of each recognition model in the weighted fusion based on the environmental state vector.

[0028] The specific training process is as follows: First, the basic elements of the reinforcement learning problem are defined. The state space is defined as an environmental state vector encoded from real-time sensor data. This vector comprehensively represents the overall environmental conditions in which the vehicle operates, and its dimensions include, but are not limited to, rainfall level, ambient light intensity, external humidity, vehicle speed, and signal quality evaluation values ​​of each sensor channel. The action space is defined as all possible combinations of weight assignments.

[0029] The reward function is designed to consider two core objectives: the accuracy of slippery surface recognition and vehicle stability after implementing control strategies. Specifically, at each training step, the agent outputs a set of weights based on its current state. The system uses these weights to weightedly fuse the outputs of the three recognition models, obtaining a final target slippery surface risk score. This score is used to trigger the corresponding vehicle control strategy. The reward value is calculated based on the matching accuracy between the fused score and the real slippery surface label, and the vehicle's dynamic stability index. The weighted combination of accuracy and stability index constitutes the immediate reward for that step. The training objective of the agent is to maximize the long-term cumulative discounted reward.

[0030] Training employs deep reinforcement learning algorithms, such as the deep deterministic policy gradient algorithm. Training data comes from large-scale real-vehicle road datasets accumulated by automakers and synthetic data generated by high-fidelity simulation platforms, covering various weather conditions, road types, and driving scenarios. During training, the agent gradually learns the mapping relationship from environmental states to the optimal weight allocation strategy through extensive trial-and-error interactions. Once training is complete, the reinforcement learning model deployed on the vehicle can dynamically output a weight allocation scheme that maximizes long-term cumulative rewards based on the real-time perceived environmental states.

[0031] The trained reinforcement learning model is deployed to the vehicle's central computing unit. During actual driving, the model infers and outputs the optimal weight allocation result in real time based on the environmental state vector generated from real-time sensor data. Furthermore, the system is designed with an online adaptive mechanism: when a persistent deviation is detected between the fused recognition result and the actual vehicle dynamic feedback over a period of time, a short-cycle relearning process can be triggered. Using this new onboard data, the model is rapidly fine-tuned while preserving existing knowledge, enabling it to adapt to new vehicle conditions, new environments, or aging vehicle components, thereby achieving continuous evolution and personalized adaptation of model performance.

[0032] Step S140: Based on the weight allocation results, the wet skid risk scores output by each of the identification models are weighted and fused to obtain the target wet skid risk score of the target vehicle.

[0033] In this embodiment, the weighted fusion refers to a mathematical operation process that linearly combines the outputs of multiple recognition models according to their weight values. Specifically, this process multiplies the dynamic weight coefficients generated by the reinforcement learning model with the slippery risk scores of the corresponding recognition models, and then sums all the product results to obtain a comprehensive risk assessment value, namely the target slippery risk score.

[0034] In this embodiment, the target slippery risk score refers to a final quantified slippery risk value calculated through the weighted fusion step, which integrates the judgment results of all identification models. It represents the system's optimal estimate of the overall slippery risk of the road surface where the vehicle is currently located.

[0035] For example, if the scores of the feature interaction model, the time series analysis model, and the image recognition model are S1, S2, and S3, respectively, and their dynamic weights are w1, w2, and w3, respectively, then the target slippery risk score is S = w1 × S1 + w2 × S2 + w3 × S3.

[0036] In this embodiment of the application, the target slippery risk score can be used to directly adjust the vehicle's control strategy parameters; it can also be used to determine a comprehensive slippery risk level, thereby determining the specific control modes of the chassis, drive, braking and other systems, as well as the warning prompts of the vehicle's human-machine interface.

[0037] In this embodiment, multiple sensor data related to wheel slippage are first acquired, and these data are processed in parallel using at least two different recognition models to generate preliminary wet slip risk scores from different dimensions. A reinforcement learning model dynamically generates the weights of each recognition model in the final decision based on real-time sensor data. Finally, based on this dynamic weight allocation, the risk scores output by each model are weighted and fused to obtain the final target wet slip risk score. This constructs a technical framework of multi-model parallel processing and dynamic weighted fusion of reinforcement learning: multiple heterogeneous recognition models independently analyze sensor data from different dimensions, providing a more comprehensive judgment basis and overcoming the limitation of a single model's perception perspective; the reinforcement learning model continuously learns the mapping relationship between environmental features and the recognition performance of each model, dynamically evaluating the credibility of each model under current conditions, thereby autonomously adjusting its weight in the final decision. This ensures the system maintains stable and accurate judgment capabilities even in complex scenarios such as changes in lighting and weather, thus comprehensively improving the robustness and generalization of the recognition system in complex environments.

[0038] Actual testing demonstrates that, compared to traditional fixed threshold or single-model methods, the multi-model parallel processing and reinforcement learning dynamic weighted fusion technical framework constructed in this application achieves significant performance improvements. Particularly in real-vehicle tests under complex scenarios such as rainy nights, the false alarm rate is significantly reduced, and the accuracy of slippery condition recognition is significantly improved.

[0039] based on Figure 1 In addition to the method described in the embodiments of this specification, some specific implementation schemes of the method are also provided, which will be described below.

[0040] Optionally, the recognition model includes at least two of the following: a feature interaction model, a time series analysis model, and an image recognition model; Correspondingly, the sensor data is processed using at least two wet-slip condition identification models to obtain wet-slip risk scores output by each identification model, including: When the recognition model includes the feature interaction model, multi-source feature data is input into the pre-trained feature interaction model to obtain a first slippery risk score; the multi-source feature data is obtained by encoding sensor data from multiple data channels related to slippery working conditions at the current moment. When the recognition model includes the time-series analysis model, the time-series data in the sensor data is input into the pre-trained time-series analysis model to obtain a second slippery risk score; the time-series data includes vehicle dynamics state data at multiple time steps; the dynamics state data is used to describe the real-time motion state of the target vehicle; When the recognition model includes the image recognition model, the target image in the sensor data is input into the image recognition model to obtain a third slippery risk score; the target image includes a road image in front of the target vehicle.

[0041] In this embodiment, the feature interaction model can be used to process multi-source feature data, mine the correlation between features from different sources, and output slippery risk assessment results. The feature interaction model captures a comprehensive representation of slippery working conditions from the structured data level by modeling the cross-correlation information of multi-dimensional features. The feature interaction model can include various models based on the factorization machine architecture, such as Factorization Machine (FM), Field-aware Factorization Machine (FM), Deep Factorization Machine (DeepFM), Deep & Cross Network (DCN), and eXtreme Deep Factorization Machine (xDeepFM).

[0042] In this embodiment, the data channel refers to an independent path or data source for transmitting specific types of sensor data. Each data channel is responsible for transmitting a specific type of physical quantity measurement value. For example, a dedicated wheel speed data channel transmits the rotational speed information of the four wheels, while a rainfall data channel transmits the intensity signal of raindrops on the windshield.

[0043] In this embodiment, the multi-source feature data refers to a set of features related to slippery working conditions, provided by various sensors or data channels. These features originate from different perception modalities and are typically preprocessed, such as standardized and filtered, before being input into the model. They are then converted into a unified numerical vector form and concatenated or combined to form the input of the feature interaction model. For example, it can simultaneously include rainfall level from a raindrop sensor, wheel speed difference from a wheel speed sensor, longitudinal acceleration from an inertial measurement unit, and the average grayscale value of the ground extracted from an image.

[0044] In this embodiment, the first slippery risk score specifically refers to the quantified risk assessment result output by the feature interaction model after processing the multi-source feature data. This score reflects the slippery probability obtained based on the cross-analysis of multiple state features at the current moment.

[0045] In this embodiment, the time-series analysis model is a machine learning model specifically designed for processing time-series data. Its core capability is to capture the dynamic changes and dependencies of data over time. In wet-slip condition identification, this model identifies dynamic trends related to slippage by analyzing the patterns of sensor data on the vehicle chassis over time, such as abnormal fluctuations in wheel speed after rapid acceleration or changes in slip ratio during braking. The time-series analysis model may include, but is not limited to, recurrent neural networks (RNNs) and their variants, such as long short-term memory networks (LSTMs), gated recurrent units (GRUs) and their bidirectional structures (BiGRUs), as well as other network structures suitable for modeling sequential data, such as temporal convolutional networks (TCNs).

[0046] In this embodiment, the vehicle dynamics state data refers to various types of data that can reflect the motion state of the target vehicle and are related to wet and slippery conditions and wheel slippage. This vehicle dynamics state data is used to describe the dynamic operating characteristics of the vehicle during driving. These data directly or indirectly reflect the interaction mechanical relationship between the vehicle chassis, tires, and the road surface. Typical vehicle dynamics state data includes, but is not limited to: wheel speed, vehicle longitudinal and lateral acceleration, yaw rate, steering wheel angle, and wheel speed difference signals derived from these raw data.

[0047] In this embodiment, the second slippery risk score specifically refers to the quantified risk assessment result output by the time-series analysis model after processing the time-series data. This score reflects the likelihood of slippage or loss of control based on the analysis of the vehicle's recent dynamic behavior sequence.

[0048] In this embodiment, the image recognition model refers to a type of computer vision model based on deep learning, which can automatically learn and extract visual features related to slippery road surfaces from input image data and output a quantitative risk assessment. The image recognition model may include various image classification, segmentation, or detection models based on Convolutional Neural Networks (CNNs), whose essential function is to transform visual information into a quantitative score that can be used for fusion decision-making.

[0049] In this embodiment, the target image refers to a visual data frame captured by a vehicle-mounted forward-facing camera, containing the road scene in front of the target vehicle. This image serves as the direct input to the image recognition model, and its content directly reflects the visual condition of the road surface the vehicle is about to pass. For example, in rainy weather, a target image may contain areas of asphalt road surface with specular reflections due to water accumulation. The target image typically undergoes preprocessing such as exposure compensation and brightness normalization to unify resolution and enhance feature visibility, ensuring the model can operate stably under different lighting conditions.

[0050] In this embodiment, the third slippery road risk score specifically refers to the quantified risk assessment result output by the image recognition model after analyzing and processing the target image. This score reflects the likelihood of a slippery road surface determined purely based on visual information.

[0051] In this embodiment, the first slippery road risk score, the second slippery road risk score, and the third slippery road risk score can all be continuous values ​​between 0 and 1. For any score, the higher the value, the greater the risk of slippery road surface determined by the corresponding model based on the specific type of data it processes.

[0052] In this embodiment, by deploying at least two heterogeneous models from feature interaction, temporal analysis, and image recognition in parallel, the risk of slippery conditions can be independently and deeply assessed from multiple complementary dimensions, such as multi-source feature intersection, vehicle dynamic temporal evolution, and road surface visual semantics. The feature interaction model focuses on mining the deep correlation between multi-source state features at the current moment and assesses the risk from a static comprehensive level. It has high accuracy when sensor data is complete and feature relationships are stable, but its judgment will be limited when some sensors fail or data quality fluctuates. The temporal analysis model captures the dynamic behavior patterns of the vehicle over a continuous period of time and identifies potential slip trends. It is particularly effective in identifying progressive slippage caused by acceleration, braking, and other operations. However, its sensitivity will decrease in scenarios where the vehicle is traveling at a constant speed and there is a lack of obvious dynamic stimulation. The image recognition model directly analyzes the visual semantic information of the road ahead, providing intuitive evidence of the road surface condition. It has high recognition accuracy under daytime conditions with good lighting and clear visibility, but its reliability will decrease significantly in scenarios with limited vision, such as at night, strong backlight, or when the camera is obstructed by dirt. Based on this, a reinforcement learning model is introduced to dynamically allocate decision weights for each model according to the real-time environment, and the risk scores output by each model are adaptively weighted and fused to generate the final target slippery risk score.

[0053] Optionally, the feature interaction model includes a deep decomposition machine model; the step of inputting multi-source feature data into a pre-trained feature interaction model to obtain a first slippery risk score includes: The sensor data for each data channel is encoded into a sparse feature vector corresponding to that data channel. The sparse feature vectors corresponding to each data channel are concatenated to obtain the multi-source feature data. The multi-source feature data is input into the deep decomposition machine model to obtain the first slippery risk score.

[0054] In this embodiment, the sparse feature vector refers to a numerical vector in a higher-dimensional vector representation where only a few elements are non-zero, while the vast majority of elements are zero. This representation method is often used to efficiently encode feature data with a large number of possible values ​​or a highly discrete distribution. For example, for a feature representing "rainfall level", if it is divided into four categories: "no rain", "light rain", "moderate rain", and "heavy rain", one-hot encoding will result in a four-dimensional sparse feature vector, such as "moderate rain" encoded as [0,0,1,0].

[0055] In the embodiments of this application, the sparse feature vector encoding refers to the data processing process of converting raw, non-numerical or unstructured sensor data into the sparse feature vector through encoding rules such as one-hot encoding, thereby transforming the heterogeneous information of different data channels into a unified mathematical form that is suitable for direct processing by machine learning models.

[0056] In this embodiment, the multi-source feature data specifically refers to the feature vector formed by sequentially concatenating the sparse feature vectors corresponding to each data channel along the feature dimension. This vector integrates all sparse encoded information from different sensors or data sources, serving as the unified input to the feature interaction model.

[0057] In this embodiment, the Deep Factorization Machine (DeepFM) model is a feature interaction model that combines factorization machines and deep neural networks. This model, through parallel factorization network pathways and deep neural network pathways, simultaneously models low-order second-order cross-relationships and high-order complex combinational relationships between input features, thereby achieving comprehensive mining of deep interaction information from multiple feature sources.

[0058] Specifically, one pathway of the model is the factorization network, which explicitly represents the strength of second-order interactions between features by learning a low-dimensional embedding vector for each pair of features and calculating the dot product between the vectors. For example, it learns the contribution of the combination of "raindrop level" and "wheel speed difference" to the risk of slippery conditions. The other pathway is the deep neural network, which automatically performs nonlinear transformations and combinations on the features through a multi-layer fully connected network to extract and represent high-order feature interaction patterns that are difficult to enumerate directly. The outputs of the two pathways are finally combined to generate the final risk score. For example, the interaction between raindrop level and wheel speed difference may reveal a slight tendency for vehicles to slip in heavy rain. In the deep channel, a three-layer fully connected network structure further combines these feature vectors to generate a high-order representation of slippery environments.

[0059] In this embodiment, the training process of the deep decomposition machine model is based on real-vehicle test datasets accumulated by automakers. This dataset contains multi-source sensor data collected under different weather conditions, road conditions, and driving scenarios. Each training sample contains complete multimodal input features and corresponding ground truth labels. The labels can be binary classification labels, clearly indicating whether each training sample is in a slippery driving condition. Cross-entropy is used as the loss function during training, and the Adam optimizer is used to iteratively update the model parameters.

[0060] During the hyperparameter tuning phase, to balance model generalization ability and deployment efficiency, a strategy combining grid search and Bayesian optimization is employed to systematically adjust key parameters such as feature embedding dimension, number of deep network layers, number of neurons, and Dropout ratio. This tuning process can be performed independently for each vehicle platform. Each vehicle project can be independently tuned according to its own characteristics, and the saved optimal configuration includes model structure parameters, weight parameters, and corresponding performance metrics. When deploying the system for a new vehicle model, configurations for similar platforms can be found in the parameter repository as a base, followed by lightweight fine-tuning, significantly shortening the model adaptation cycle.

[0061] To further improve the model's reliability in extreme conditions, the system also incorporates historical accident data for reverse verification. By comparing the model's output during the accident period with the actual situation, perceptual deviations in extreme scenarios are identified, and the model is fine-tuned accordingly to continuously optimize its ability to capture and warn of high-risk slippery scenarios.

[0062] In this embodiment, sensor data from each data channel is encoded into sparse feature vectors and concatenated to form a unified multi-source feature representation, which is then input into a deep decomposition machine (DDM) model for processing. The DDM model can automatically learn and fuse the complex low-order and high-order feature interactions within the sparse feature vectors, without relying on manually designed complex feature crossover rules. This model explicitly models second-order interactions between features through its factorization network, while implicitly capturing higher-order nonlinear combination patterns through deep neural networks, thereby more comprehensively uncovering deep correlation information related to slippery conditions in multi-source data. This mechanism enables the model to more accurately characterize the risk state under the combined influence of multiple factors, thereby improving the accuracy and generalization ability of slippery condition identification.

[0063] Optionally, the sensor data of the multiple data channels includes various parameters such as raindrop level, external humidity, wheel speed, wheel speed difference, lateral acceleration, longitudinal acceleration, driving speed, and average gray value and reflective area ratio of the ground area in the target image; the target image includes a road image in front of the target vehicle.

[0064] In this embodiment of the application, the raindrop level refers to the rainfall intensity index detected and quantified by the raindrop sensor, which is used to reflect the amount of rainfall in the current environment. For example, it can be divided into 1-5 levels, with level 1 corresponding to light rain and level 5 corresponding to heavy rain. The higher the level, the greater the risk of road flooding.

[0065] In this embodiment, the external humidity refers to the percentage of relative humidity in the ambient air outside the vehicle, measured by a humidity sensor located outside the vehicle. This data is used to help determine whether the atmospheric environment is humid and is an indirect indicator that the road surface may become wet or foggy in advance.

[0066] In this embodiment, the wheel speed refers to the rotational speed of each individual wheel, which is collected in real time by wheel speed sensors installed on each wheel. This data is the basis for calculating vehicle speed, wheel speed difference, and determining whether a single wheel is locked or spinning freely.

[0067] In this embodiment, the wheel speed difference refers to the difference in rotational speed between different wheels at the same moment, typically focusing on the difference in rotational speed between left and right wheels on the same axle or between the front and rear axles. On slippery surfaces, drive wheels or steering wheels may slip excessively due to insufficient traction, resulting in their wheel speed being significantly higher than other wheels, thus creating an abnormal wheel speed difference. This value is a key dynamic feature for identifying slippage behavior.

[0068] In this embodiment, the lateral acceleration refers to the acceleration component of the vehicle in the horizontal plane perpendicular to the longitudinal axis of the vehicle body (i.e., the left-right direction). It is mainly generated when the vehicle turns or is subjected to lateral forces, and its magnitude and rate of change can reflect the lateral stability of the vehicle and whether there is a tendency to sideslip.

[0069] In this embodiment, the longitudinal acceleration refers to the acceleration component of the vehicle along its longitudinal axis (i.e., the forward or backward direction). It is generated by vehicle acceleration or braking, and abnormal changes in longitudinal acceleration may indicate drive wheel slippage or brake wheel lockup.

[0070] In this embodiment of the application, the driving speed refers to the overall driving speed of the vehicle, which is obtained by fusion of wheel speed sensor signals or direct measurement by vehicle speed sensor, and is used to determine the slippery risk level in combination with other features. For example, the risk of slippage on a slippery road surface is higher when driving at high speed than when driving at low speed.

[0071] In this embodiment, the ground area refers to the pixel area belonging to the road paving surface, segmented from the image captured by the vehicle-mounted forward-facing camera using image recognition technology. This ground area is the foundational region for extracting visual features related to road surface slipperiness, and it defines the scope for subsequent analysis of road surface features such as grayscale calculation or reflection area identification. Generally, irrelevant objects such as the sky, vehicles, and buildings in the target image are excluded.

[0072] In this embodiment, the average gray value refers to the value obtained by arithmetically averaging the gray values ​​of all pixels within the ground area. It roughly reflects the average brightness of the area, and the average gray value of dry asphalt pavement and wet, reflective pavement usually differs.

[0073] In this embodiment, the reflective area ratio refers to the percentage of the area occupied by pixels identified as having specular reflective properties in the segmented ground area. This parameter is obtained by analyzing the distribution characteristics of bright areas in the image; for example, large, continuous, highly reflective areas usually indicate a higher probability of road surface water accumulation.

[0074] In this embodiment, the sensor data from the multiple data channels can be broadly categorized into three types: environmental state, vehicle motion state, and visual features. These reflect the complex information of the interaction between the vehicle and the road surface from different perspectives. By mining the inherent correlations and collaborative change patterns among these three types of data, especially among cross-category data, a more comprehensive and three-dimensional judgment basis for slippery conditions can be constructed. Compared to methods that rely solely on a single category of data, this multi-dimensional, cross-modal feature fusion mechanism can effectively compensate for the perception limitations or reliability degradation of any single data source in a specific scenario, thereby significantly improving the overall accuracy, system robustness, and scenario generalization ability of slippery risk identification under a wider range of actual working conditions.

[0075] Optionally, the sensor data includes dynamic state data; the step of inputting the time-series data from the sensor data into a pre-trained time-series analysis model to obtain a second slippery risk score includes: For each sampling moment within the time sliding window, the dynamic state data of each sampling moment is transformed into a state feature vector for that sampling moment; the time sliding window is a time window that includes a preset number of sampling moments with the current moment as the endpoint; The time series data is obtained based on the state feature vectors corresponding to each sampling time. The time series data is input into the time series analysis model to obtain the second slippery risk score.

[0076] In this embodiment, the time window refers to a continuous time period that extends backward from the current processing time. This window is used to capture a segment of historical sensor data for time-series model analysis to determine its changing patterns over time. For example, a time window of 5 seconds, ending at the current time, contains all sampling data from 5 seconds ago to the present.

[0077] In this embodiment, the sampling time refers to each specific time point within the time window where data is collected at fixed time intervals. Each sampling time corresponds to a complete set of sensor readings.

[0078] In this embodiment, the dynamic state data refers to a set of physical quantity data used to describe the motion state of the target vehicle at a specific instant during a sampling moment. It typically includes dynamic parameters related to the interaction between the vehicle and the road surface. For example, the dynamic state data at a certain sampling moment may include the rotational speed of each wheel at that moment, the longitudinal acceleration, lateral acceleration, yaw rate of the vehicle, and derived quantities such as wheel speed difference and slip ratio calculated from these basic data.

[0079] In this embodiment, the state feature vector refers to a numerical vector obtained by preprocessing and formatting the dynamic state data at a certain sampling moment. This vector represents the complete dynamic state of the vehicle at that moment in a standardized and compact mathematical form, and is the basic input unit that the time series analysis model can process.

[0080] In this embodiment, the time-series data refers to a data structure formed by arranging the state feature vectors corresponding to each sampling moment within the time window in chronological order. It completely represents the evolution of the vehicle state over a continuous period of time. For example, a state feature vector containing 50 consecutive sampling moments arranged in sequence constitutes a 50-step time-series data.

[0081] In this embodiment, by transforming vehicle dynamics state data within a fixed time window into a time-series sequence of state feature vectors and inputting it into a time-series analysis model for processing, it is possible to extract time-series evolution patterns related to slippery risk from the continuous dynamic behavior of vehicles. Unlike methods that only analyze the current instantaneous state, the slippage behavior of vehicles on low-adhesion surfaces often manifests as a process that evolves over time, such as abnormal fluctuations in wheel speed or specific trends in acceleration. By capturing the dependencies between different sequences of data, the time-series analysis model can identify this potential, gradual slippage trend, thereby enabling earlier risk warnings. This analysis method based on dynamic behavior sequences significantly improves the sensitivity and accuracy of identifying slippery conditions, especially their initial stages or inducing processes.

[0082] In the embodiments of this application, the time series analysis model may include, but is not limited to, recurrent neural networks (RNNs) and their variants, such as long short-term memory networks (LSTM), gated recurrent units (GRUs) and their bidirectional gated recurrent units (BiGRUs), as well as other network structures suitable for modeling sequence data, such as temporal convolutional networks (TCNs).

[0083] The following explanation uses the Bidirectional Gated Recurrent Unit (BiGRU) network as an example. The training of this network is based on real-vehicle time-series datasets containing annotations of slippery driving conditions accumulated by automakers. Training data is generated by extracting data from continuous driving data using a sliding time window approach. Each sample is a fixed-length time-series sequence, and its label is determined based on whether a slippage event actually occurred within that time window. During training, the time-series data constructed using the aforementioned method is used as input, with the objective of minimizing the difference between the model's output risk score and the true label. Binary cross-entropy is used as the loss function, and RMSprop (Root Mean Square Propagation) is selected as the optimizer for parameter updates.

[0084] In terms of model tuning and optimization, key hyperparameters adjusted include the number of units in the hidden layers, the length of the input time window, the Dropout ratio to prevent overfitting, and the number of layers in the recurrent network. The tuning process is performed using an automated hyperparameter search system, with the accuracy of slippery surface recognition on the validation set as the primary evaluation metric for selection. To balance model performance with the limitations of onboard computing resources, the final deployment version typically employs a single-layer or two-layer network structure, keeping the number of hidden units within a suitable range to ensure efficient real-time inference on the onboard controller.

[0085] To improve the model's adaptability to different vehicle models, vehicle platform attributes were introduced as contextual features during the training phase. Specifically, architectural information such as the vehicle's drive type and powertrain type was encoded into vectors and concatenated with the state feature vectors at each time step before being input into the network. This allows the model to learn the differentiated dynamic response patterns that different vehicle models may exhibit under the same operating conditions, such as the difference in slippage characteristics between front-wheel drive and rear-wheel drive vehicles on slippery roads.

[0086] To enhance the model's robustness under extreme or complex conditions, enhanced samples for specific scenarios such as rainy nights, flooded roads, and high-speed lane changes were introduced into the training data. These samples were partly derived from real-vehicle data collected under specific conditions and partly generated through a high-fidelity simulation platform, aiming to expand the model's coverage and generalization ability for boundary scenarios. Furthermore, in the model's post-processing stage, short-time filtering and temporal smoothing were applied to the continuously output risk scores to suppress misjudgments caused by instantaneous noise or abnormal fluctuations, ensuring the temporal stability of the output results.

[0087] In this embodiment, a bidirectional gated recurrent unit network (BiGRU) is used as the time-series analysis model, which can simultaneously process the forward and reverse information flows of the time series, thereby capturing the contextual dependencies of vehicle dynamic behavior more completely. This bidirectional modeling mechanism allows the system to not only make judgments based on the current and past states, but also refer to the contextual information of future moments, which helps to identify delayed slip signs caused by historical operations that appear in the current or future moments. Compared with the unidirectional recurrent model, this structure provides a more comprehensive modeling of behavioral sequence patterns, and can discover potential progressive slip trends and slip behaviors earlier and more sensitively, thereby achieving early warning of slip risk and improving the system's timeliness in perceiving dynamic risks.

[0088] Optionally, the dynamic state data may include multiple parameters of the target vehicle, such as wheel speed, lateral acceleration, longitudinal acceleration, steering wheel angle, driving speed, and wheel speed difference.

[0089] In this embodiment, the wheel speed refers to the rotational speed of each individual wheel, which is collected in real time by wheel speed sensors installed on each wheel. This data is the basis for calculating vehicle speed, wheel speed difference, and determining whether a single wheel is locked or spinning freely.

[0090] In this embodiment, the lateral acceleration refers to the acceleration component of the vehicle in the horizontal plane perpendicular to the longitudinal axis of the vehicle body (i.e., the left-right direction). It is mainly generated when the vehicle turns or is subjected to lateral forces, and its magnitude and rate of change can reflect the lateral stability of the vehicle and whether there is a tendency to sideslip.

[0091] In this embodiment, the longitudinal acceleration refers to the acceleration component of the vehicle along its longitudinal axis (i.e., the forward or backward direction). It is generated by vehicle acceleration or braking, and abnormal changes in longitudinal acceleration may indicate drive wheel slippage or brake wheel lockup.

[0092] In this embodiment, the steering wheel angle refers to the angle value input by the driver to the vehicle steering system through steering wheel operation, and is a key parameter characterizing the driver's steering intention and the instantaneous steering state of the vehicle. This data can be used to determine whether the vehicle is in a steering condition and, in conjunction with signals such as lateral acceleration, to analyze steering stability.

[0093] In this embodiment of the application, the driving speed refers to the overall driving speed of the vehicle, which is obtained by fusion of wheel speed sensor signals or direct measurement by vehicle speed sensor, and is used to determine the slippery risk level in combination with other features. For example, the risk of slippage on a slippery road surface is higher when driving at high speed than when driving at low speed.

[0094] In this embodiment, the wheel speed difference refers to the difference in rotational speed between different wheels at the same moment, typically focusing on the difference in rotational speed between left and right wheels on the same axle or between the front and rear axles. On slippery surfaces, drive wheels or steering wheels may slip excessively due to insufficient traction, resulting in their wheel speed being significantly higher than other wheels, thus creating an abnormal wheel speed difference. This value is a key dynamic feature for identifying slippage behavior.

[0095] In this embodiment, various dynamic state data of the target vehicle, such as wheel speed, lateral acceleration, longitudinal acceleration, steering wheel angle, driving speed, and wheel speed difference, are collected. After processing, time-series data are input into a time-series analysis model to mine the time dependencies between data and output a second slippery risk score. Therefore, by comprehensively utilizing various dynamic state data such as wheel speed, lateral acceleration, longitudinal acceleration, steering wheel angle, driving speed, and wheel speed difference, the real-time interaction between the vehicle and the road surface can be comprehensively depicted from multiple independent and complementary physical dimensions.

[0096] Optionally, obtaining time-series sequence data based on the state feature vectors corresponding to each sampling time includes: Obtain vehicle feature information of the target vehicle; the vehicle feature information is used to characterize vehicle platform information that affects the dynamic state data of the target vehicle; The vehicle feature information is encoded into a context vector; The context vector and the state feature vectors at each sampling time within the time window are concatenated to obtain the time-series data.

[0097] In this embodiment, the vehicle feature information refers to a set of parameters directly related to the target vehicle platform architecture and capable of influencing its dynamic response characteristics under wet conditions. This vehicle feature information characterizes the inherent differences in the slippage behavior of different vehicles. These are relatively fixed static features after the vehicle design is finalized; however, some attributes can be dynamically adjusted according to driving modes or system commands. For example, a four-wheel drive vehicle may also have front-wheel drive or rear-wheel drive modes. The purpose of introducing vehicle feature information is to provide a crucial vehicle state context for the time-series model, enabling it to accurately adapt to the physical response characteristics of different platforms or the same platform under different operating modes, avoiding identification bias caused by differences in dynamic characteristics between vehicles on different platforms or under different operating modes.

[0098] Specifically, the vehicle characteristic information may include, but is not limited to: drive type, powertrain type, powertrain parameters, braking system configuration, and even tire specifications and chassis tuning-related attributes. For example, drive type may include front-wheel drive, rear-wheel drive, four-wheel drive, and switchable modes (such as on-demand four-wheel drive and part-time four-wheel drive). Front-wheel drive vehicles are more prone to front wheel slippage on slippery surfaces, while rear-wheel drive vehicles may exhibit slight rear-end swaying. Powertrain type may include conventional gasoline, pure electric, and hybrid powertrains; the torque output characteristics of different powertrain types directly affect the slippage behavior. Powertrain parameters may include peak torque, torque response time constant, and the front-to-rear axle torque distribution logic for hybrid or dual-motor vehicles. Braking system configuration may include brake type (such as disc / drum), whether a high-intensity energy recovery system is integrated, and the calibration style of the anti-lock braking system and electronic stability control system.

[0099] In this embodiment, the context vector is a standardized vector form formed by digitally encoding vehicle feature information, which transforms unstructured vehicle attributes into input data that can be recognized by the time-series analysis model.

[0100] In this embodiment, the time-series data may include state feature vectors at each sampling moment within a time window, and context vectors for characterizing vehicle features. That is, the time-series data includes both the static attributes of the vehicle platform and the dynamic changes of the vehicle over continuous time.

[0101] Similarly, each training sample vector used in the training phase of the time series analysis model is composed of time series state data and its corresponding vehicle context, which enables the time series analysis model to learn differentiated slip patterns for different vehicle platforms.

[0102] In this embodiment, vehicle feature information that characterizes the platform information of the target vehicle and affects its dynamic state data is obtained. This vehicle feature information is encoded into a context vector, and the context vector is concatenated with the state feature vectors at each sampling time within the time window to obtain time series data. This allows the time series analysis model to identify the slippage mode of the target vehicle by combining the inherent attributes of the target vehicle's platform when processing time series data. This effectively avoids identification bias caused by the different dynamic characteristics of different vehicle models and platforms. At the same time, it significantly improves the generalization ability and deployment efficiency of the algorithm in multi-model projects of car companies and reduces the cost of calibrating and training models separately for each new car.

[0103] Optionally, the image recognition model includes an image segmentation model; The step of inputting the target image from the sensor data into the image recognition model to obtain a third slippery risk score includes: Based on the trained image segmentation model, the target image is segmented to determine the ground area information, reflective area information, and water accumulation area information in the target image. Based on the ground area information, the reflective area information, and the water accumulation area information, a third slippery risk score is generated; the third slippery risk score is used to characterize the slipperiness of the road on which the target vehicle is to travel.

[0104] In this embodiment, the image segmentation model is a deep learning-based computer vision model specifically designed for pixel-level semantic segmentation of input images. This model can classify each pixel in an image into a specific semantic category (such as ground, reflection, or puddles), achieving precise boundary delineation of different objects. The segmentation process refers to the computational process by which the image segmentation model performs pixel-level classification and region division of the input image. This process extracts image features layer by layer through the model's neural network structure, ultimately outputting the semantic category label for each pixel.

[0105] In this embodiment, the ground area information refers to the set of pixels in the target image that belong to the road surface area, obtained through segmentation processing, and their spatial location information. This information is used to focus subsequent analysis on the road surface where the vehicle is actually traveling, eliminating interference from irrelevant backgrounds such as the sky, vehicles, and buildings.

[0106] In this embodiment, the reflective area information refers to the set of pixels and their feature information that identify high-brightness areas in the ground area formed by the reflection of light from smooth surfaces such as water, obtained through segmentation processing. These areas typically exhibit significantly higher brightness than the surrounding road surface and lack texture details. It should be noted that simple reflective areas may be caused by various factors, such as temporary reflections from other vehicle headlights or the local reflective characteristics of dry road surfaces; therefore, the presence of reflective areas alone cannot be used to directly determine the risk of slipperiness.

[0107] In this embodiment, the water accumulation area information refers to the set of pixels and their feature information that identify actual water accumulation areas in the ground region, obtained through segmentation processing. These water accumulation areas in the target image generally appear as bright, textureless regions with continuous distribution characteristics, conforming to hydrodynamic morphology. Only when the system simultaneously identifies physically plausible water accumulation areas can the existence of a slippery risk be confirmed by combining reflective features.

[0108] In this embodiment of the application, the ground area information, reflective area information, and water accumulation area information can be represented as pixel sets, mask images, area boundary contours, and other commonly used area representation methods in computer vision.

[0109] In this embodiment, the third slippery risk score refers to a comprehensive numerical score generated through quantitative calculation based on the ground area information, reflective area information, and water accumulation area information. This score is specifically used to characterize the degree of slipperiness risk of the road ahead of the target vehicle, which is obtained purely based on visual information analysis; the higher the value, the greater the risk determined by visual semantic analysis.

[0110] In this embodiment of the application, the third slippery risk score can be calculated based on one or more quantitative indicators of the water accumulation area information and the reflective area information. Specifically, the water accumulation area ratio is the ratio of the number of pixels in the segmented water accumulation area to the total number of pixels in the ground area, which is used to directly characterize the coverage of water on the road surface. The reflective area distribution characteristics include the distribution range, concentration pattern and whether the reflective area is distributed along the lane line in the ground area, etc. The reflective intensity quantitative indicators cover the average brightness value, peak brightness value and brightness variance of the reflective area.

[0111] In this embodiment, the target image from sensor data is input into a trained image segmentation model. Segmentation processing determines ground area information, reflective area information, and water accumulation area information in the target image, thereby generating a third slippery risk score characterizing the slipperiness of the road surface to be driven by the target vehicle. By extracting ground area information, the actual extent of the road surface is accurately defined, effectively eliminating interference from non-road areas such as vehicles, buildings, and vegetation. By extracting reflective area information, specular reflection caused by water accumulation or ice on the road surface is captured. By extracting water accumulation area information, direct evidence of liquid water on the road surface is provided, distinguishing areas with actual water accumulation risk from temporary, localized reflections. The third slippery risk score generated based on these three types of information can accurately quantify the slippery risk of the road surface from a visual perspective, improving the reliability of the third slippery risk score and providing a high-quality visual dimension evaluation basis for subsequent multi-model weighted fusion. This effectively enhances the accuracy, robustness, and adaptability to complex road surface scenarios of the slippery condition recognition system.

[0112] Optionally, generating the third slippery risk score based on the ground area information, the reflective area information, and the water accumulation area information includes: Determine whether there is a water accumulation area indicated by the water accumulation area information within the ground area area indicated by the ground area information; If the water accumulation area exists in the ground area, the third slippery risk score is generated based on the distribution information of the reflective area in the ground area indicated by the reflective area information. If there is no water accumulation area in the ground area, then the third slippery risk score is obtained, which indicates that the target vehicle does not have a slippery risk.

[0113] In this embodiment, the distribution information may include at least one of the following: the area ratio of the water accumulation area to the ground area, the distribution characteristics of the reflective area, and the reflective intensity characteristics. The area ratio can be obtained by calculating the ratio of the number of pixels in the water accumulation area to the total number of pixels in the ground area; the distribution characteristics of the reflective area may include its spatial distribution density within the road surface area, its morphological continuity index, and its positional relationship relative to lane lines; the reflective intensity characteristics may include optical parameters such as the mean, variance, and peak brightness of pixels within the reflective area.

[0114] In this embodiment, the presence of a water accumulation area on the ground is used as the primary criterion for generating the third slippery risk score. The method first determines whether there are clear physical signs of water accumulation in the image. Only after confirming the existence of a water accumulation area does it further assess the risk level based on the distribution characteristics of reflective areas. If there is no water accumulation on the ground, it is directly determined that there is no slippery risk. This directly links the slippery risk assessment to the presence or absence of a water accumulation area, effectively eliminating reflective interference from non-water accumulation areas and avoiding the problem of misjudging slippery conditions based solely on image grayscale features or single reflective information. This significantly improves the accuracy and reliability of the third slippery risk score.

[0115] In this embodiment, the image recognition model is a Mask Region-based Convolutional Neural Network (Mask R-CNN). This model can accurately segment the ground region in the forward-facing camera image and extract low reflectivity features related to wet and slippery conditions, which helps determine whether the vehicle's current environment poses a risk of slipperiness. Unlike traditional methods that judge ground conditions based on the average grayscale value of an image, Mask R-CNN has instance-level segmentation capabilities, accurately identifying ground regions, water accumulation areas, and reflective areas, providing reliable visual input for slippery surface recognition.

[0116] The data processing begins with preprocessing the raw images captured by the vehicle-mounted camera. The input image is first scaled to a uniform preset resolution, and then processed by an image enhancement module. This module performs operations such as exposure compensation, brightness normalization, and color space standardization to improve the stability and consistency of the image under different lighting conditions. The enhanced image serves as the direct input to the model.

[0117] The Mask R-CNN architecture comprises a base feature extraction network, a Region Proposal Network (RPN), a Region of Interest (ROI) Alignment layer, and a branch prediction head. The base feature extraction network uses a Residual Network-50 (ResNet-50) as its backbone, combined with a Feature Pyramid Network (FPN) to enhance its ability to recognize multi-scale features. The RPN generates ground candidate regions based on the feature maps, and ROI Alignment performs fine-tuning of these candidate regions. Finally, the output branches into two branches: one for bounding box regression and the other for generating pixel-level segmentation masks for the ground regions.

[0118] After the preprocessed image is input into the Mask R-CNN model, a backbone feature extraction network consisting of a residual network Network-50 and a feature pyramid network FPN first performs multi-scale feature extraction on the input image, generating feature maps with rich semantic information. Subsequently, a Region Proposal Network (RPN) slides across these feature maps, generating a series of candidate boxes that may contain the target region. These candidate boxes undergo fine spatial alignment through a Region of Interest Align (ROI Align) layer, extracting fixed-size region features from the feature maps. Finally, these region features are fed into two parallel prediction head branches: one branch is responsible for classifying the candidate boxes and performing fine-grained regression of the bounding box coordinates; the other branch performs pixel-level semantic segmentation, generating accurate masks for each candidate region, thus outputting pixel-level segmentation results for ground areas, water accumulation areas, and reflective areas.

[0119] Based on the ground area information, the reflective area information, and the water accumulation area information, morphological post-processing is performed on each ground sub-region. Effective region boundaries are determined through connected component analysis, and interference regions with excessively small areas are filtered out. Subsequently, the feature aggregation module built into the prediction head extracts three key quantitative indicators: first, the area ratio of water accumulation areas to the total number of pixels in the ground area, calculated by comparing the number of pixels in water accumulation areas to the total number of pixels in the ground area; second, the distribution characteristics of reflective areas, including their spatial distribution density within the road surface area, morphological continuity indicators, and positional relationship relative to lane lines; and third, reflective intensity characteristics, obtained by statistically analyzing optical parameters such as the mean, variance, and peak brightness of pixels within the reflective area.

[0120] The fully connected layer or lightweight regression network inside the scoring prediction head fuses and nonlinearly maps the area ratio, distribution pattern, and brightness concentration of the reflective area, and outputs a continuous value between 0 and 1, which is the third slippery risk score. This score comprehensively reflects the degree of road slipperiness determined based on the visual semantic segmentation results.

[0121] The training and optimization of the Mask R-CNN model are based on road image datasets accumulated by automakers, covering various weather and lighting conditions. These datasets include pixel-level semantic masks manually annotated for ground surfaces, reflective areas, and puddles. During training, data augmentation techniques are employed to improve the model's generalization ability, and a multi-task loss function is used to jointly optimize the accuracy of bounding box localization and mask segmentation. The optimizer typically uses stochastic gradient descent (SGD) with a learning rate decay strategy. In the optimization phase, hyperparameters such as the backbone network structure, feature pyramid settings, and anchor point size are adjusted, and iterative optimization is performed based on validation set performance to achieve a balance between segmentation accuracy and computational efficiency.

[0122] To adapt to the camera parameters and image styles of different vehicle models, the model supports incremental fine-tuning of the pre-trained model using small batches of data specific to each vehicle model. Furthermore, the system can perform appropriate pruning or quantization of the model based on the computing power of the deployment platform to ensure that real-time requirements are met on in-vehicle hardware. The MaskR-CNN model trained and optimized through the above process can stably and accurately extract key visual semantic information from complex road scenes, providing reliable visual evidence for the quantitative assessment of slippery road risks.

[0123] Optionally, the step of generating weight allocation results based on sensor data using a reinforcement learning model includes: A state vector is constructed based on the sensor data; the state vector is used to characterize the working environment information that affects the reliability of the output results of each of the recognition models. The state vector is input into a pre-trained reinforcement learning model to obtain the weight allocation result; The reinforcement learning model is trained using the state vector as the state space, the weight combination of each recognition model as the action space, and the accuracy of the target slippery risk score and / or the vehicle stability of the target vehicle as the reward function; the vehicle stability is used to characterize the stability of the dynamic state of the target vehicle when controlled based on the target slippery risk score.

[0124] In this embodiment, the state vector may include environmental state information and vehicle state information related to model credibility assessment from sensor data, used to characterize real-time working environment information that affects the credibility of the output results of each recognition model. Specifically, the state vector may include, but is not limited to: rainfall level, ambient light intensity, ambient humidity, vehicle speed, and the health status of various sensors.

[0125] For example, rainfall levels, obtained through raindrop sensors, reflect the density of raindrops on the windshield. High rainfall levels usually indicate an increased likelihood of road flooding, but they can also reduce the image quality and reliability of the visual model due to raindrop obstruction. Ambient light intensity directly affects camera imaging; insufficient light weakens the image recognition model's ability to identify reflective areas on the ground. Ambient humidity is related to phenomena such as fog outside the vehicle and fogging of the windows. Vehicle speed refers to the real-time speed of the vehicle; the temporal characteristics of wheel speed signals are more valuable for identification at high speeds, and the weights of the temporal analysis model need to be adjusted accordingly. Sensor health is used to assess the reliability of each data source itself; for example, when the camera lens is obscured by rain or fog, its health score decreases, and the weights of the image recognition model need to be reduced.

[0126] In this embodiment, the reinforcement learning model refers to a type of machine learning model that can autonomously learn the optimal decision-making strategy through continuous interaction with the environment. In this model, a computing unit called an agent observes the environmental state reflected by sensor data, takes actions to adjust the weights of each recognition model, and continuously optimizes its decision-making strategy based on the feedback (i.e., reward) of the overall recognition accuracy and control stability of the system after the actions.

[0127] In this embodiment, the reinforcement learning model is an intelligent decision-making agent that interacts with the environment and optimizes its decision-making strategy based on reward signals. The reinforcement learning model determines the optimal weight combination for the wet / slippery condition recognition model during fusion based on real-time perceived environmental and vehicle states.

[0128] During the training phase, the reinforcement learning model observes the environmental state (state vector), attempts to perform different weight allocation actions, and iteratively adjusts its internal parameters based on the reward value calculated from the effect of the action, thereby automatically learning the optimal weight allocation strategy under different environmental states.

[0129] In this embodiment, the state space refers to the set of state vectors corresponding to all possible working environments that the reinforcement learning model can perceive. The dimension of the state space corresponds to the number of feature types contained in the state vectors.

[0130] In the embodiments of this application, the action space refers to the set of all weight allocation schemes that the reinforcement learning model can execute. It covers all feasible combinations of weight values ​​corresponding to each recognition model. Each combination corresponds to a specific weight allocation action. The weight values ​​range from 0 to 1, and the sum of the weight values ​​of all recognition models is 1.

[0131] In this embodiment, the reward function is a mathematical function used to quantify the "gain" or "penalty" a reinforcement learning model receives after performing an action in a certain state. It is the core guiding signal that drives the model to learn and optimize its strategy. The goal of the reward function design in this application is to guide the model to learn a weight allocation strategy that simultaneously optimizes the accuracy of slippery risk identification and vehicle driving stability. Therefore, the reward function is typically constructed as a composite function based on the accuracy of the target slippery risk score and / or the vehicle stability of the target vehicle. For example, if the weights assigned by the model result in a fused risk score that highly matches the actual road conditions, and vehicle control implemented based on this score does not cause severe shaking or instability of the vehicle body, the model will receive a positive reward; conversely, if misjudgment or control leads to a deterioration in vehicle dynamics, a negative reward will be received. By continuously trying and adjusting the strategy based on the rewards, the model eventually learns to automatically output the optimal fused weights in complex and ever-changing environments.

[0132] In this embodiment, vehicle stability is used to specifically quantify the smoothness and controllability of the vehicle's dynamic state after implementing a control strategy based on a wet skid risk score. It is not a single indicator, but a comprehensive evaluation of whether the vehicle's yaw motion, lateral slip, pitch, and other combined dynamic characteristics are within a safe and expected range.

[0133] In this embodiment, a state vector representing the environmental state is dynamically constructed based on real-time sensor data. The optimal fusion weights for each model are output using a pre-trained intelligent decision-making body, achieving adaptive dynamic allocation of model weights. This ensures the robustness of overall recognition even when the reliability of specific sensors or models declines. Furthermore, since the training process aims to optimize both recognition accuracy and vehicle control stability, this dynamic weight strategy not only improves the accuracy of risk scoring but also ensures the smoothness of downstream control system actions, achieving closed-loop collaborative optimization from perception to control.

[0134] Optionally, after obtaining the target slipperiness risk score of the target vehicle, the method further includes: Obtain the target slippery risk score at multiple consecutive moments, including the current moment; Based on the target slippery risk scores at the multiple consecutive time points, the slippery risk level of the target vehicle is determined; The slippery risk level is used to trigger the corresponding vehicle control strategy.

[0135] In the embodiments of this application, the multiple consecutive moments refer to multiple moments that are sequential in time and constitute a decision window. For example, the system can set the decision window to 5 consecutive sampling moments (e.g., a total of 0.5 seconds), and only when the scores of all moments within this window indicate high risk will the high-risk state be finally confirmed.

[0136] In this embodiment of the application, the slippery road risk level refers to a quantitative classification and evaluation of the degree of slipperiness of the road surface where the target vehicle is currently located. Typically, the level can be divided into several levels from low to high, such as "no risk", "low risk", "medium risk", and "high risk".

[0137] In this embodiment, the vehicle control strategy refers to a set of preset vehicle actuator parameter adjustment instructions bound to a specific wet skid risk level. When the system determines that it is currently at a certain risk level, it automatically triggers and executes the complete control strategy corresponding to that level. The coverage of the vehicle control strategy may include the chassis domain, power domain, braking domain, etc. For example, a "low risk" level may only provide a visual prompt on the instrument panel; for a "medium risk" level, the control strategy may include appropriately reducing the energy recovery intensity and advancing the intervention threshold of the electronic stability program; while for a "high risk" level, the strategy may further include measures such as forcibly limiting the torque output of the drive wheels and adjusting the operating frequency of the anti-lock braking system. In this embodiment, by obtaining target wet skid risk scores at multiple consecutive moments, including the current moment, and determining the final wet skid risk level based on the temporal continuity of these scores, the problem of false triggering caused by instantaneous sensor anomalies or brief environmental interference is effectively avoided.

[0138] Based on the same technical concept, this application also provides another method for identifying slippery working conditions, the method comprising: First, sensor data from the target vehicle for identifying slippery conditions is acquired. This sensor data refers to various types of perception data that reflect the vehicle's own motion state, the surrounding environment, and the visual state of the road ahead. Acquisition of this data relies on multiple sensor devices mounted on the target vehicle, including but not limited to rain sensors, ambient humidity sensors, wheel speed sensors, inertial measurement units (IMUs), vehicle speed sensors, and forward-facing cameras. Specifically, the rain sensor collects raindrop level data, the ambient humidity sensor collects external humidity data, the wheel speed sensor collects wheel speed and wheel speed difference data, the inertial measurement unit collects lateral acceleration, longitudinal acceleration, and steering wheel angle data, the vehicle speed sensor or wheel speed signal fusion module collects driving speed data, and the forward-facing camera captures an image of the road ahead of the target vehicle (i.e., the target image).

[0139] Subsequently, the aforementioned sensor data is processed in parallel using at least two heterogeneous wet slip condition recognition models. This embodiment illustrates this by simultaneously deploying a feature interaction model, a time series analysis model, and an image recognition model.

[0140] From the sensor data at the current moment, raindrop level, external humidity, wheel speed of the four wheels, calculated wheel speed difference, lateral acceleration, longitudinal acceleration, driving speed, and the average gray value and reflective area ratio of the ground area obtained from the preliminary analysis of the current frame road image are extracted. These features from different data channels are standardized and encoded, transformed into corresponding sparse feature vectors. All sparse feature vectors are concatenated to form a unified set of multi-source feature data, which is then input into a pre-trained deep factorization machine model. This model, through its internal factorization network and deep neural network, automatically learns the low-order and high-order cross relationships between these multi-source features and outputs a first slippery risk score between 0 and 1, reflecting the comprehensive risk derived from the cross analysis of multi-source state features.

[0141] A 5-second time window ending at the current moment is defined, and dynamic state data for each sampling moment within this window is acquired at a frequency of 10 times per second, including wheel speeds, lateral acceleration, longitudinal acceleration, steering wheel angle, vehicle speed, and wheel speed difference. The data for each sampling moment is converted into a state feature vector and arranged chronologically to form a 50-step time-series data sequence. This sequence data is input into a pre-trained bidirectional gated recurrent unit network. This network, through its bidirectional structure, analyzes the evolution pattern of the vehicle state over time, captures potential slippage trends such as abnormal wheel speed fluctuations, and outputs a second wet skid risk score between 0 and 1, reflecting the risk derived from recent dynamic behavior sequence analysis.

[0142] The target image captured by the vehicle-mounted camera at the current moment is input into a pre-trained, mask-based region convolutional neural network. This model performs pixel-level instance segmentation of the image, outputting a precise semantic segmentation mask to identify ground areas, reflective areas, and water accumulation areas. The system further analyzes these segmentation results, calculating the area ratio of water accumulation areas within the ground area and analyzing the distribution and intensity characteristics of reflective areas. Based on this quantified visual information, a third slip risk score between 0 and 1 is generated through a mapping function. This score is specifically used to characterize the degree of road slippage obtained based on visual semantic analysis.

[0143] Next, dynamic fusion decision-making is performed: the system inputs real-time sensor data (such as rainfall, illumination, vehicle speed, etc.) into a pre-trained reinforcement learning model. This model dynamically infers and outputs a set of weight allocation results based on the current environmental state, assigning corresponding weight values ​​to the first, second, and third slippery risk scores. For example, in nighttime scenarios, the system may automatically assign a lower weight to the third slippery risk score; while during rapid vehicle acceleration, a higher weight is assigned to the second slippery risk score. Based on this dynamic weight allocation result, the first, second, and third slippery risk scores are weighted and summed to calculate the final target slippery risk score for the target vehicle. This comprehensive score is more comprehensive and stable than the output of any single model, providing accurate and robust condition judgment for subsequent vehicle stability control systems.

[0144] The target slippery risk score can be mapped to a preset slippery risk level. As one embodiment, this level can be divided into four levels: Level 0 is normal, requiring no intervention; Level 1 is slightly slippery; Level 2 is moderately slippery; and Level 3 is severely slippery. This level label is continuously output by the fusion module, and after it remains stable for a certain period, it triggers the subsequent control strategy switching process.

[0145] The central control module has a built-in predefined strategy mapping rule table that maps different levels of slippery road risk to parameter adjustment commands for each control domain of the vehicle. The main control domains include the chassis stability control domain, braking control domain, energy recovery system, drive domain, and driver assistance domain.

[0146] For the chassis stability control domain, when the wet slip level reaches level 2 or above, the Electronic Stability Program (ESP) can automatically increase its sensitivity to key state variables such as vehicle sideslip rate and yaw rate, and reduce its intervention delay, thus intervening earlier in the initial stage of slippage. When the level reaches level 3, the active traction control function is further activated to limit excessive slippage of the drive wheels.

[0147] For the braking control domain, it can communicate with the anti-lock braking system (ABS) controller through the Controller Area Network (CAN) to dynamically adjust the ABS operating frequency and braking pressure application curve according to the wet skid level, so as to adapt to the characteristics of low-adhesion road surfaces and optimize braking stability.

[0148] For energy recovery systems, when the slipperiness level is 2 or higher, the energy recovery intensity level can be automatically reduced, and the recovered braking force can be smoothed to avoid loss of drive wheel traction due to transient high torque feedback from the motor. This command is sent from the central controller to the vehicle control unit (VCU), and then the VCU coordinates the motor controller to execute it.

[0149] For drive domain control strategies, in four-wheel drive or dual-motor models, when the slipperiness level reaches level 3, the torque distribution strategy can be temporarily adjusted, such as increasing the rear axle drive force or limiting torque on wheels prone to slippage, in order to optimize the overall vehicle traction distribution.

[0150] For the driver assistance domain, under medium to high slippery risk levels, a control policy downgrade request can be sent to the Advanced Driver Assistance Systems (ADAS) domain controller, such as limiting the maximum acceleration of the adaptive cruise system or adjusting the steering intervention of lane keeping assist, to improve the system's operational safety margin under low adhesion conditions.

[0151] Simultaneously, the vehicle's head-up display (HUD) can provide drivers with graded and progressive alerts regarding the risk of slippery conditions. The alert logic is linked to the risk level: Level 1 receives a non-intrusive icon alert; Level 2 and above receive enhanced graphic and audio alerts; Level 3, accompanied by severe vehicle dynamic anomalies, triggers a high-intensity warning requiring driver confirmation. The alert timing is intelligently buffered in conjunction with current driving operations to avoid interference. All alert logic can be customized according to vehicle model and user preferences via configuration files and can be remotely updated via Over-The-Air (OTA) technology.

[0152] Exemplary device The apparatus embodiments of this application can be used to execute the method embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments of this application.

[0153] Figure 2 The diagram shown is a block diagram of a wet / slippery condition identification device according to an embodiment of this application. Figure 2 As shown, the device 200 includes: The acquisition module 210 is used to acquire sensor data of the target vehicle for wet and slippery condition identification. The identification module 220 is used to process the sensor data through at least two identification models for slippery working conditions to obtain slippery risk scores output by each identification model. The allocation module 230 is used to generate a weight allocation result based on sensor data according to a reinforcement learning model; the weight allocation result includes the weight values ​​corresponding to each of the recognition models. The fusion module 240 is used to perform weighted fusion of the slippery risk scores output by each of the recognition models according to the weight allocation results, so as to obtain the target slippery risk score of the target vehicle.

[0154] Exemplary electronic devices Below, for reference Figure 3 This describes an electronic device according to embodiments of the present application. Figure 3 A block diagram of an electronic device according to an embodiment of this application is illustrated.

[0155] like Figure 3 As shown, the electronic device 300 includes one or more processors 310 and memory 320.

[0156] The processor 310 may be another form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 300 to perform desired functions.

[0157] Specifically, processor 310 can be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, etc., or an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. It can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Processor 310 may also include a main processor, and may also include a baseband chip, a modem, etc.

[0158] The memory 320 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 310 may execute the program instructions to implement the wet slip condition identification method and / or other desired functions of the various embodiments of this application described above. Various contents, such as category correspondence, may also be stored in the computer-readable storage medium.

[0159] In one example, the electronic device 300 may also include an input device 330 and an output device 340, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0160] In addition, the input device 330 can also be a device that receives user input data and information, such as a keyboard, mouse, camera, scanner, light pen, voice input touchscreen, pedometer, or gravity sensor. The output device 340 can output various information to the outside. The output device 340 may include, for example, a display, speaker, printer, and communication networks and their connected remote output devices.

[0161] Of course, for the sake of simplicity, Figure 3 Only some of the components of the electronic device 300 relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device 300 may include any other suitable components depending on the specific application.

[0162] Exemplary vehicle In addition to the methods and devices described above, embodiments of this application may also include a vehicle, comprising a vehicle body and the electronic equipment.

[0163] Exemplary computer program products and computer-readable storage media In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the wet slip condition identification methods according to various embodiments of this application described in the "Exemplary Methods" section of this specification.

[0164] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0165] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the wet slip condition identification method according to various embodiments of this application described in the "Exemplary Methods" section of this specification.

[0166] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0167] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0168] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0169] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0170] The steps in the methods of the various embodiments of this application can be adjusted, merged, or deleted in order according to actual needs, and the technical features described in each embodiment can be replaced or combined.

[0171] The block diagrams of devices, equipment, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, equipment, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0172] It should also be noted that in the apparatus and method of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0173] The modules or submodules described as separate components may or may not be physically separate. The components that constitute a module or submodule may or may not be physical modules or submodules; that is, they may be located in one place or distributed across multiple network modules or submodules. Some or all of the modules or submodules can be selected to achieve the purpose of this embodiment's solution, depending on actual needs.

[0174] Furthermore, the functional modules or sub-modules in the various embodiments of this application can be integrated into one processing module, or each module or sub-module can exist physically separately, or two or more modules or sub-modules can be integrated into one module. The integrated modules or sub-modules described above can be implemented in hardware or in the form of software functional modules or sub-modules.

[0175] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0176] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software unit executed by a processor, or a combination of both. The software unit can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0177] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0178] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying slippery working conditions, characterized in that, include: Acquire sensor data from the target vehicle for identifying slippery conditions; The sensor data is processed using at least two identification models for slippery working conditions to obtain slippery risk scores output by each identification model. Based on a reinforcement learning model, a weight allocation result is generated according to sensor data; the weight allocation result includes the weight value corresponding to each of the recognition models. The wet skid risk scores output by each of the identification models are weighted and fused according to the weight allocation results to obtain the target wet skid risk score of the target vehicle.

2. The method according to claim 1, characterized in that, The recognition model includes at least two of the following: a feature interaction model, a time series analysis model, and an image recognition model. Correspondingly, the sensor data is processed using at least two wet-slip condition identification models to obtain wet-slip risk scores output by each identification model, including: When the recognition model includes the feature interaction model, multi-source feature data is input into the pre-trained feature interaction model to obtain a first slippery risk score; the multi-source feature data is obtained by encoding sensor data from multiple data channels related to slippery working conditions at the current moment. And / or, When the recognition model includes the time-series analysis model, the time-series data in the sensor data is input into the pre-trained time-series analysis model to obtain a second slippery risk score; the time-series data includes vehicle dynamics state data at multiple time steps; the dynamics state data is used to describe the real-time motion state of the target vehicle; And / or, When the recognition model includes the image recognition model, the target image in the sensor data is input into the image recognition model to obtain a third slippery risk score; the target image includes a road image in front of the target vehicle.

3. The method according to claim 2, characterized in that, The feature interaction model includes a deep decomposition machine model; the step of inputting multi-source feature data into a pre-trained feature interaction model to obtain a first slippery risk score includes: The sensor data for each data channel is encoded into a sparse feature vector corresponding to that data channel. The sparse feature vectors corresponding to each data channel are concatenated to obtain the multi-source feature data. The multi-source feature data is input into the deep decomposition machine model to obtain the first slippery risk score.

4. The method according to claim 2, characterized in that, The sensor data includes dynamic state data; the step of inputting the time-series data from the sensor data into a pre-trained time-series analysis model to obtain a second slippery risk score includes: For each sampling moment within the time sliding window, the dynamic state data of each sampling moment is transformed into a state feature vector for that sampling moment; the time sliding window is a time window that includes a preset number of sampling moments with the current moment as the endpoint; The time series data is obtained based on the state feature vectors corresponding to each sampling time. The time series data is input into the time series analysis model to obtain the second slippery risk score.

5. The method according to claim 4, characterized in that, The process of obtaining time-series sequence data based on the state feature vectors corresponding to each sampling time includes: Obtain vehicle feature information of the target vehicle; the vehicle feature information is used to characterize vehicle platform information that affects the dynamic state data of the target vehicle; The vehicle feature information is encoded into a context vector; The context vector and the state feature vectors at each sampling time within the time window are concatenated to obtain the time-series data.

6. The method according to claim 2, characterized in that, The image recognition model includes an image segmentation model; The step of inputting the target image from the sensor data into the image recognition model to obtain a third slippery risk score includes: Based on the trained image segmentation model, the target image is segmented to determine the ground area information, reflective area information, and water accumulation area information in the target image. Based on the ground area information, the reflective area information, and the water accumulation area information, a third slippery risk score is generated; the third slippery risk score is used to characterize the slipperiness of the road on which the target vehicle is to travel.

7. The method according to claim 6, characterized in that, The process of generating the third slippery risk score based on the ground area information, the reflective area information, and the water accumulation area information includes: Determine whether there is a water accumulation area indicated by the water accumulation area information within the ground area area indicated by the ground area information; If the water accumulation area exists in the ground area, the third slippery risk score is generated based on the distribution information of the reflective area in the ground area indicated by the reflective area information. If there is no water accumulation area in the ground area, then the third slippery risk score is obtained, which indicates that the target vehicle does not have a slippery risk.

8. The method according to claim 1, characterized in that, The generation of weight allocation results based on sensor data using a reinforcement learning model includes: A state vector is constructed based on the sensor data; the state vector is used to characterize the working environment information that affects the reliability of the output results of each of the recognition models. The state vector is input into a pre-trained reinforcement learning model to obtain the weight allocation result; The reinforcement learning model is trained using the state vector as the state space, the weight combination of each recognition model as the action space, and the accuracy of the target slippery risk score and / or the vehicle stability of the target vehicle as the reward function; the vehicle stability is used to characterize the stability of the dynamic state of the target vehicle when controlled based on the target slippery risk score.

9. The method according to claim 1, characterized in that, After obtaining the target slipperiness risk score of the target vehicle, the process further includes: Obtain the target slippery risk score at multiple consecutive moments, including the current moment; Based on the target slippery risk scores at the multiple consecutive time points, the slippery risk level of the target vehicle is determined; The slippery risk level is used to trigger the corresponding vehicle control strategy.

10. A vehicle, characterized in that, Includes the vehicle body and electronic equipment; the electronic equipment includes: processor; Memory used to store the processor's executable instructions; The processor is configured to perform the method according to any one of claims 1 to 9.

Images