Windscreen wiper control method and system based on AI vision

Abstract

The invention discloses a windscreen wiper control method and system based on AI vision, and belongs to the technical field of image processing and computer vision. The windscreen wiper control method comprises the steps that a real-time image of a front window of a vehicle and real-time vehicle state parameters are obtained; performing non-uniform dynamic weight grid division on the real-time image, generating a dynamic weight grid, extracting visual interference characteristics of each grid unit, generating a visual interference characteristic vector, fusing real-time vehicle state parameters, and generating a multi-dimensional characteristic vector; and generating a parameterized control instruction according to the multi-dimensional feature vector, and driving a windscreen wiper execution mechanism to act. According to the technical scheme, dynamic grid division is carried out on the real-time image of the front window of the vehicle, the visual interference features are extracted, and the real-time vehicle state parameters are fused, so that a refined parameterized control instruction can be generated, and more accurate, intelligent and self-adaptive control over the windscreen wiper system is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing and computer vision, in particular to a rain wiper control method and system based on AI vision. BACKGROUND

[0002] Image processing technology, as a core component of computer vision, plays a crucial role in intelligent transportation and vehicle assisted driving systems. Its main task is to analyze and understand the collected image information to achieve effective perception of the vehicle's surrounding environment. In the field of vehicle active safety and driving experience, using visual technology to automatically control the wiper to ensure that the driver always has a clear view in rainy and snowy weather is an important application direction of this technology.

[0003] In the prior art, the mainstream scheme to realize automatic wiper function usually adopts a rain sensor based on infrared optical principle. This type of sensor is installed on the inner side of the vehicle's front windshield. It works by emitting a beam of infrared light and detecting the intensity of the reflected light on the inner and outer surfaces of the glass. When the glass surface is dry, most of the light will be totally reflected back to the receiving end. When the glass surface is attached with raindrops, the raindrops will change the light path, causing part of the light to refract out, so that the receiving end receives a weakened reflected light intensity. The system determines the amount of rain according to the degree of light intensity attenuation, and then controls the start and stop of the wiper and the working frequency.

[0004] However, the above-mentioned technical scheme based on infrared optical sensor has several inherent technical defects. First, the sensing dimension of this scheme is single, and it can only roughly quantify the change of light reflectivity, so it is difficult to accurately distinguish different types of visual interference such as light rain, fog, mud or stains, often leading to non-starting in light rain or fog environment, or false starting when the glass is dirty. Secondly, its response has limitations, and it is slow to react to transient and local visual obstruction events such as water splashing from the car in front. In addition, its control logic is relatively fixed and cannot be intelligently adjusted in combination with the current driving speed, steering intention or the working state of the advanced driving assistance system, making it difficult to meet the dynamic needs of visual clarity in complex driving scenarios. SUMMARY

[0005] To solve the above problems, the present application provides a rain wiper control method and system based on AI vision, which adopts a technical scheme of dynamically dividing the real-time image of the vehicle's front window, extracting visual interference features and fusing real-time vehicle state parameters, which can generate fine parameterized control instructions, thereby realizing more accurate, intelligent and adaptive control of the wiper system.

[0006] The above-mentioned object can be achieved by the following scheme:

[0007] An AI vision-based wiper control method, comprising: acquiring a real-time image of a front window of a vehicle and a real-time vehicle state parameter; performing non-uniform dynamic weight grid division on the real-time image to generate a dynamic weight grid; extracting visual interference features of each grid cell in the dynamic weight grid to generate a visual interference feature vector; fusing the visual interference feature vector and the real-time vehicle state parameter to generate a multi-dimensional feature vector; generating a parameterized control instruction according to the multi-dimensional feature vector, the parameterized control instruction comprising a basic wiping frequency, a regional weight adjustment factor, a response sensitivity, and a wiping mode; and driving a wiper actuator to act based on the parameterized control instruction.

[0008] Optionally, the real-time vehicle state parameter comprises a real-time vehicle speed, a steering signal, an ambient light intensity, and a driving assistance system state.

[0009] Optionally, the dynamic weight grid generation comprises: identifying a visual core area defining a driver's main focus range using the real-time image; dynamically expanding the visual core area according to the real-time vehicle speed to obtain a dynamic core area; assigning a preset first density grid and a preset first weight value to the area within the dynamic core area, and assigning a preset second density grid and a preset second weight value to the area outside the dynamic core area; and integrating the first density grid and the second density grid to generate a dynamic weight grid; wherein the grid density of the first density grid is greater than that of the second density grid, and the first weight value is greater than the second weight value.

[0010] Optionally, the visual interference feature vector generation comprises: calculating a texture blur change value of each grid in the dynamic weight grid; detecting a brightness distribution abnormal value of each grid in the dynamic weight grid; analyzing a rainwater motion vector of each grid in the dynamic weight grid; and weighting and fusing the texture blur change value, the brightness distribution abnormal value, and the rainwater motion vector according to the first weight value or the second weight value to generate a visual interference feature vector; wherein the visual interference features comprise the texture blur change value, the brightness distribution abnormal value, and the rainwater motion vector.

[0011] Optionally, the multi-dimensional feature vector generation comprises: compensating the visual interference feature vector using the ambient light intensity parameter; jointly encoding the compensated visual interference feature vector, the real-time vehicle speed, and the steering signal to generate a driving scenario feature; generating a response priority based on the driving assistance system state; and fusing the driving scenario feature and the response priority to generate a multi-dimensional feature vector.

[0012] Optionally, the generating the parameterized control instruction according to the multi-dimensional feature vector comprises: calculating a basic wiping frequency based on the multi-dimensional feature vector; calculating an interference score of each grid in the dynamic weight grid based on the texture blur degree change value and the brightness distribution abnormal value; generating a regional weight adjustment factor based on the interference score and the multi-dimensional feature vector; calculating a response sensitivity by analyzing the change gradient of the multi-dimensional feature vector in a continuous time sequence; classifying and judging the current overall scene based on the rainwater motion vector and the multi-dimensional feature vector, and selecting a wiping mode according to the judgment result; and integrating the basic wiping frequency, the regional weight adjustment factor, the response sensitivity and the wiping mode to obtain the parameterized control instruction.

[0013] Optionally, the driving the wiper execution mechanism to act according to the parameterized control instruction comprises: smoothing the basic wiping frequency according to the response sensitivity to obtain a smoothed frequency; distributing a wiping area priority in executing the wiping mode according to the regional weight adjustment factor; and generating a motor control waveform by combining the smoothed frequency and the wiping area priority to drive the wiper execution mechanism.

[0014] Optionally, the method further comprises: when detecting a driver's manual intervention operation, recording the manual intervention operation and the corresponding multi-dimensional feature vector to form labeled data; and using the labeled data to correct the calculation process of the basic wiping frequency, the regional weight adjustment factor and the response sensitivity.

[0015] Optionally, the method further comprises: when the wiper is started, calculating an average value of the interference scores of each grid in the core area of the field of view to obtain an interference average value; when detecting that the interference average value is greater than a preset interference threshold value, correcting the regional weight adjustment factor and judging whether the interference average value is less than the preset interference threshold value within a preset time window; and if not, expanding the range of the core area of the field of view.

[0016] Based on the same inventive concept, the application also provides an AI vision-based wiper control system, which comprises: a basic parameter acquisition module for acquiring real-time images of a vehicle front window and real-time vehicle state parameters; a dynamic grid generation module for performing non-uniform dynamic weight grid division on the real-time images to generate a dynamic weight grid; a visual interference processing module for extracting visual interference features of each grid unit in the dynamic weight grid to generate a visual interference feature vector; a multi-dimensional fusion module for fusing the visual interference feature vector and the real-time vehicle state parameters to generate a multi-dimensional feature vector; a control decision module for generating a parameterized control instruction according to the multi-dimensional feature vector, the parameterized control instruction comprising a basic wiping frequency, a regional weight adjustment factor, a response sensitivity and a wiping mode; and a parameterized execution module for driving a wiper execution mechanism to act based on the parameterized control instruction.

[0017] Compared with the prior art, the application has the following advantages:

[0018] 1. The application can concentrate computing resources and analysis on the area that has the greatest impact on driving safety by performing non-uniform dynamic weight grid division on the vehicle front window images and identifying the core area of the driver's field of view. This dynamic area division based on the driver's attention, combined with the weighted processing of the visual interference features of different areas, enables the system to more accurately quantify the actual impact of rain and dirt on the key field of view, thereby significantly improving the accuracy and reliability of rain perception and effectively distinguishing complex situations such as light rain, fog, dirt or water splashed by the car in front;

[0019] 2. The application deeply fuses the extracted visual interference features with multi-dimensional vehicle parameters such as real-time vehicle speed, steering signal, ambient light intensity and driving assistance system state to build a comprehensive driving situation awareness. This fusion enables the wiper control decision to no longer rely solely on the amount of rain on the windshield, but to adapt to the specific needs of the vehicle under different working conditions such as high-speed driving, turning, night driving or activating advanced driving assistance systems, making the response of the wiper system more predictable and scenario-adaptive, thereby improving driving safety;

[0020] 3. The application generates a parameterized control instruction containing a basic wiping frequency, a regional weight adjustment factor, a response sensitivity and a wiping mode, realizing fine and personalized adjustment of wiper action. By smoothly adjusting the wiping frequency through the response sensitivity, the application avoids sudden changes in wiper speed and improves driving comfort. Through the regional weight adjustment factor, the application can focus on cleaning specific contaminated areas. This multi-parameter collaborative control makes the wiper action not a simple gear shift, but a truly intelligent and efficient dynamic response;

[0021] 4、The application also includes a closed-loop adaptive learning and self-correction mechanism; the system can record the driver's manual intervention operation and use it as labeled data to continuously optimize the internal decision model, so that the control strategy gradually fits the driver's personal habits; at the same time, the system will evaluate the cleaning effect of the core field of view in real time after the wiper works, and when the effect is not good, it can actively correct the control strategy or even expand the attention area, ensuring the continuous clarity of the key field of view, so that the system has the ability to evolve and cope with unknown situations.

[0022] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the application. The objects and other advantages of the application will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0024] Figure 1 is a flowchart of the AI vision-based wiper control method of the embodiment of the present application.

[0025] Figure 2 is a contrast schematic diagram of the core field of view of the embodiment of the present application.

[0026] Figure 3 is a structural schematic diagram of the AI vision-based wiper control system of the embodiment of the present application.

[0027] Figure 4 is a comparison chart of the system response accuracy under different rain conditions of the embodiment of the present application.

[0028] Figure 5 is a comparison chart of the system response time under sudden scenarios of the embodiment of the present application.

[0029] Figure 6 is a comparison chart of the average interference level of the core field of view when the ADAS is activated of the embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0031] With reference to Figure 1 An embodiment of the present application proposes an AI vision-based wiper control method, which adopts a technical solution of dynamically dividing a real-time image of a front window of a vehicle, extracting visual interference features and fusing real-time vehicle state parameters, and can generate fine parameterized control instructions, so as to realize more accurate, intelligent and adaptive control of a wiper system.

[0032] The method of the embodiment specifically includes:

[0033] obtaining a real-time image of a front window of a vehicle and real-time vehicle state parameters;

[0034] performing non-uniform dynamic weight grid division on the real-time image to generate a dynamic weight grid;

[0035] extracting visual interference features of each grid unit in the dynamic weight grid to generate a visual interference feature vector;

[0036] fusing the visual interference feature vector and the real-time vehicle state parameters to generate a multi-dimensional feature vector;

[0037] generating a parameterized control instruction according to the multi-dimensional feature vector, the parameterized control instruction including a basic wiping frequency, a regional weight adjustment factor, a response sensitivity and a wiping mode;

[0038] driving a wiper actuator to act based on the parameterized control instruction.

[0039] Specifically, the real-time image of the front window of the vehicle and the dynamic state parameters of the vehicle are obtained in parallel, laying a foundation for multi-source information input. Instead of uniform processing of the image, a non-uniform dynamic weight grid division technique is introduced to divide the image into zones and assign different weights, so as to simulate the visual focus of the driver and realize key analysis of the key visual field. On this basis, the system finely extracts quantized visual interference features from each grid unit and deeply fuses them with the vehicle state parameters to generate a multi-dimensional feature vector that can fully represent the current driving situation. Finally, the system generates a set of parameterized control instructions including frequency, mode and other dimensions according to the multi-dimensional feature vector to drive the wiper actuator.

[0040] Optionally, the real-time vehicle status parameters include real-time vehicle speed, steering signal, ambient light intensity, and driving assistance system status.

[0041] Specifically, when executing AI vision-based windshield wiper control, it not only analyzes the visual information of the vehicle's windshield but also integrates a series of key real-time vehicle status parameters to build a comprehensive understanding of the current driving situation. First, it obtains the vehicle's real-time speed data via the vehicle's controller area network (CAN bus) or other in-vehicle communication networks. Real-time speed directly relates to the relative speed and angle of raindrops hitting the windshield and affects the driver's visual focus range, serving as a core input for dynamically adjusting the image analysis area and wiping strategy. Second, also via the in-vehicle bus network, it obtains the vehicle's steering signals. These signals clearly indicate upcoming maneuvers requiring high driver concentration, such as turning or lane changes, making clear visibility crucial. Third, it obtains current ambient light intensity data through an in-vehicle integrated ambient light sensor. Ambient light intensity significantly affects the visual characteristics of raindrops and water stains captured by the image sensor. For example, in low-light environments such as at night or in tunnels, raindrops under headlights produce completely different highlights or distortions compared to daytime. This parameter must be used to compensate for the visual algorithm to ensure accurate recognition in all weather conditions. Furthermore, the method also needs to acquire the status information of the driver assistance system, which also comes from the vehicle bus network, specifically whether advanced driver assistance functions such as adaptive cruise control and lane keeping assist are active. When these systems are working, it usually means that the vehicle is in a relatively stable or high-speed driving state, requiring a higher level of continuous clarity of the forward visibility. The acquired real-time vehicle status parameters, such as vehicle speed, steering signal, ambient light intensity, and driver assistance system status, provide a solid data foundation for subsequently generating precise parameterized control commands.

[0042] Optionally, generating the dynamic weighted grid includes:

[0043] The real-time image recognition is used to define the core area of ​​the driver's field of vision, which is the area of ​​primary focus.

[0044] The range of the core visual area is dynamically expanded based on the real-time vehicle speed to obtain a dynamic core area.

[0045] A preset first density grid and a preset first weight value are assigned to the region within the dynamic core region, and a preset second density grid and a preset second weight value are assigned to the region outside the dynamic core region.

[0046] Integrate the first density grid and the second density grid to generate a dynamic weighted grid;

[0047] Wherein, the grid density of the first density grid is greater than the grid density of the second density grid; and the first weight value is greater than the second weight value.

[0048] Specifically, this is achieved by simulating the driver's visual attention mechanism and prioritizing computational resources for the most critical visual field area. First, the onboard image recognition algorithm analyzes the acquired real-time image of the vehicle's windshield to identify and define the area the driver primarily focuses on during normal straight-line driving; this area is defined as the core visual field region. This is typically the central portion of the image pointing towards the vanishing point of the road and containing the main lane lines. Subsequently, instead of using a fixed core visual field region, real-time vehicle speed parameters are introduced to dynamically adjust it. As real-time vehicle speed increases, the driver's gaze naturally extends and focuses further away. To simulate this behavior, the range of the core visual field region is dynamically expanded according to the real-time vehicle speed, generating a dynamic core region that changes with vehicle speed. This expansion process can be expressed by the following relationship:

[0049] H dc =H c ·[1+k v ·f(v)],

[0050] Among them, H dc H represents the height of the dynamic core region in the vertical direction of the image. c The reference height for the pre-defined static view core area; v represents the real-time vehicle speed obtained from the vehicle bus; k v is a dimensionless preset speed influence coefficient used to adjust the sensitivity of vehicle speed to area expansion; f(v) is a normalization function that processes real-time vehicle speed, for example, it can be processed as the ratio of the current vehicle speed to the maximum design speed. After determining the dynamic core region, the system performs non-uniform grid division on the real-time image. For the image range within the dynamic core region, the system uses a preset first density grid for division, characterized by small cell size (i.e., high grid density), and assigns a high preset first weight value to all grid cells within this region. For the surrounding area outside the dynamic core region, the system uses a preset second density grid for division, with larger grid cell size (i.e., lower grid density), and assigns a lower preset second weight value to these grid cells. Here, the first density grid has a significantly higher grid density than the second density grid, and the first weight value is also greater than the second weight value. Finally, the results of these two grid divisions with different densities and weights are integrated to form a single data structure covering the entire effective area of ​​the windshield, namely the dynamic weighted grid. This process is as follows: Figure 2As shown, the dynamic expansion of the core visual field region with vehicle speed is clearly demonstrated. Each cell of the grid contains its position, size, and weight information, providing a foundation for subsequent accurate extraction of visual interference features.

[0051] Optionally, the generation of the visual interference feature vector includes:

[0052] Calculate the texture blur change value of each grid in the dynamic weighted grid;

[0053] Detect abnormal values ​​in the brightness distribution of each grid in the dynamic weighted grid;

[0054] Analyze the rainwater movement vectors of each grid in the dynamic weighted grid;

[0055] Based on the first weight value or the second weight value, the texture blur change value, the brightness distribution anomaly value and the rain movement vector are weighted and fused to generate a visual interference feature vector.

[0056] The visual interference features include the texture blurring change value, the brightness distribution anomaly value, and the rain movement vector.

[0057] Specifically, the goal of this step is to quantitatively extract and integrate interference information affecting the driver's vision based on the dynamic weighted grid generated in the previous steps, ultimately forming a structured visual interference feature vector. This process first calculates three core visual interference features independently and in parallel for each grid cell in the dynamic weighted grid. The first feature is the texture blur change value, which aims to quantify the decrease in visual clarity caused by water film, smudges, or fine rain. In practice, the energy of the image gradient or high-frequency components of each grid cell within the current image frame is calculated and compared with a baseline state, such as a clear image without rain or an earlier clear frame. The texture blur change value represents the amount of attenuation of the current clarity relative to the baseline state; the greater the attenuation, the higher the blurriness of the area. The second feature is the brightness distribution anomaly value, which is mainly used to detect individual raindrops or water droplets that form localized highlights or shadows on the windshield, disrupting the uniformity of background brightness. To address this, the statistical distribution characteristics of pixel brightness within each grid cell are analyzed, such as calculating its variance, skewness, or deviation from the normal background brightness model. Brightness distribution anomalies are an indicator of the degree of deviation; higher values ​​indicate more or more significant discrete water droplets within the grid cell. The third feature is the rainwater motion vector. To capture the dynamic information of rainwater flow, optical flow analysis is used. By comparing several consecutive frames of images, the movement trajectory of raindrops or watermarks within the grid cell is tracked, thereby calculating their speed and direction, forming the rainwater motion vector. The magnitude of this vector reflects the intensity of rainfall or the speed at which airflow affects watermarks during vehicle movement, while its direction reveals the flow trend of the rainwater. After acquiring the texture blurring change value, brightness distribution anomalies, and rainwater motion vector for each grid cell, the system performs a weighted fusion operation to generate the final visual interference feature vector. The fusion weights here are not equal; instead, the first or second weight value assigned to each grid cell in the previous steps is directly used. Grid cells located in the core region of the field of view have their extracted features multiplied by a higher first weight value, while features in the outer regions are multiplied by a lower second weight value. The final visual interference feature vector is generated by integrating the weighted features of all grid cells. For the visual interference feature vector F... v ,have:

[0058] F v =[∑(W i ·T′ i ),∑(W i ·L′ i ),∑(W i ·M′ i )],

[0059] Among them, the visual interference feature vector F vThe three components represent the total texture blur, total brightness anomaly, and total rain motion intensity across the entire windshield after importance weighting. i represents the index of any grid cell in the dynamic weighted grid. W i T is the first or second weight value assigned to the i-th grid cell; it is a dimensionless weight coefficient. ′ i L ′ i With M ′ i These represent the normalized values ​​of the texture blur change, brightness distribution anomaly, and rainwater motion vector magnitude calculated for the i-th grid cell, respectively.

[0060] Optionally, generating the multidimensional feature vector includes:

[0061] The visual interference feature vector is compensated using the ambient light intensity parameter.

[0062] The compensated visual interference feature vector is jointly encoded with the real-time vehicle speed and the steering signal to generate driving context features;

[0063] Based on the state of the driving assistance system, a response priority is generated, and the driving situation features are fused with the response priority to generate a multi-dimensional feature vector.

[0064] Specifically, the pure visual information generated in the preceding steps is deeply integrated with the vehicle's dynamic and state information to generate a composite data structure that comprehensively represents the current driving environment and needs—a multi-dimensional feature vector. The process begins with the correction of visual features, followed by the incorporation of driving behavior, and finally, the overlay of system-level priorities. The first step involves compensating the visual interference feature vector using the acquired ambient light intensity parameters. The visual interference feature vector extracted in the preceding steps contains components such as outliers in brightness distribution that are highly sensitive to changes in ambient light intensity. For example, raindrops under headlights at night produce stronger highlights than those under daylight. To eliminate this difference and ensure consistent recognition in all weather conditions, the visual interference feature vector is compensated and corrected. The compensated visual interference feature vector F′ is then obtained. v ,have:

[0065]

[0066] Among them, F v This is the original visual interference feature vector generated in the previous steps. L represents the real-time ambient light intensity obtained through the onboard sensor. c (L) is a dimensionless compensation function that outputs based on the light intensity L. Its function is to generate one or a set of correction coefficients to offset the effect of light intensity changes on the magnitude of the eigenvalue. The first step involves applying a corrective transformation, such as element-wise multiplication of a vector, to adjust the original feature values ​​to a standardized, illumination-independent reference. The second step involves jointly encoding the compensated visual interference feature vector with real-time vehicle speed and steering signals to generate driving context features. To enable the system to understand the varying importance of visual interference under different driving behaviors, the system combines the corrected visual information with the vehicle's dynamic parameters. This joint encoding process takes the compensated visual interference feature vector, the normalized real-time vehicle speed, and the steering signal representing steering intent as input, and outputs a new feature vector—the driving context feature—through a pre-defined algorithm model, such as a small neural network or a feature cross-validation module. This new feature vector not only contains information about the amount of rainfall but also implies more complex driving scenarios such as heavy rain at high speeds or obstructed vision during turns. The third step generates response priorities based on the state of the driver assistance systems and fuses them with the driving context features. This involves acquiring the current operating status of advanced driver assistance systems such as adaptive cruise control and lane keeping assist. When these systems, which rely on visual sensors, are activated, the requirement for clear forward visibility increases significantly. Therefore, the system generates a response priority index based on its activation state. Finally, this response priority index is fused with the driving scenario features generated in the previous step, for example, through vector concatenation, to form a multi-dimensional feature vector that includes visual interference, vehicle dynamics, driving intent, and system safety requirements. This vector provides a comprehensive and in-depth decision-making basis for subsequent control command generation.

[0067] Optionally, the step of generating parameterized control instructions based on the multidimensional feature vector includes:

[0068] Based on the multidimensional feature vector, the basic brushing frequency is calculated;

[0069] Based on the texture blurring change value and the brightness distribution anomaly value, the interference score of each grid in the dynamic weighted grid is calculated;

[0070] Based on the interference score and the multidimensional feature vector, a region weight adjustment factor is generated;

[0071] The response sensitivity is calculated by analyzing the gradient of the multidimensional eigenvectors of a continuous time series.

[0072] Based on the rainwater movement vector and the multidimensional feature vector, the current overall scene is classified and judged, and the wiping mode is selected according to the judgment result;

[0073] By integrating the basic brushing frequency, the area weight adjustment factor, the response sensitivity, and the brushing mode, a parameterized control command is obtained.

[0074] Specifically, the core task of this step is to transform the highly integrated multidimensional feature vector generated in the previous steps into a set of explicit and executable wiper control parameters. First, the system calculates the base wiping frequency. The system treats the input multidimensional feature vector as a whole and feeds it into a pre-trained regression model or a complex mapping function. Based on the comprehensive information encoded in the multidimensional feature vector, such as rainfall, vehicle speed, and driver assistance status, the model directly outputs a quantified base wiping frequency value. This value represents the base wiping speed that the system determines is required under the current macroscopic environment. Second, the system generates a regional weight adjustment factor. This step aims to fine-tune the wiping action spatially. The system first calculates an interference score, which is calculated independently for each grid cell in the dynamic weighted grid. The calculation method is based on fusing the texture blur change value and brightness distribution anomaly value of that grid cell. For the interference score S of the i-th grid cell… i ,have:

[0075] S i =α·T′ i +β·L′ i ,

[0076] Among them, T ′ i and L ′ iThese are the normalized texture blur change value and brightness distribution anomaly value, respectively. α and β are preset dimensionless weighting coefficients. After obtaining the interference score distribution map of the entire windshield, the system combines the overall multidimensional feature vector to generate the final regional weight adjustment factor. This factor is not a single value, but a data structure describing the importance or priority of wiping in different areas of the windshield. Next, the response sensitivity is calculated. This parameter determines how quickly the system reacts to environmental changes. The system analyzes the gradient of the multidimensional feature vector over a continuous time series. Specifically, the system caches the multidimensional feature vector from several past moments and calculates its rate of change over time. When the rain suddenly intensifies or a vehicle suddenly accelerates into the rain area, the multidimensional feature vector changes drastically, and its gradient value increases accordingly, allowing the system to calculate a higher response sensitivity. Conversely, under stable rain conditions, this value remains at a lower level. Finally, the wiping mode is selected. This step aims to choose the most suitable wiping action type based on the specific form of the rain. The system utilizes a classification model whose inputs are the rainwater movement vector obtained from previous steps and the current multidimensional feature vector. For example, when the rainwater movement vector shows small, slow-moving raindrops, and the overall intensity of the multidimensional feature vector is low, the system classifies it as fog or drizzle and selects an intermittent wiping mode. When the rainwater movement vector shows a large-area, rapidly flowing pattern, and the multidimensional feature vector indicates a high severity level, a high-speed continuous wiping mode is selected. Finally, the system integrates four parameters—the calculated base wiping frequency, the area weight adjustment factor, the response sensitivity, and the selected wiping mode—to form a complete parameterized control command, which is then passed to the downstream execution module.

[0077] Optionally, driving the wiper actuator based on the parameterized control command includes:

[0078] The base brushing frequency is smoothly adjusted based on the response sensitivity to obtain a smooth frequency.

[0079] Based on the area weight adjustment factor, the priority of the scraping area is assigned when the scraping mode is executed;

[0080] The smoothing frequency and the priority of the wiping area are combined to generate a motor control waveform to drive the wiper actuator.

[0081] Specifically, this step details how to transform the abstract parameterized control commands generated in the previous steps into precise physical actions that drive the wiper actuators. This transformation process is not a simple direct command, but rather an organic integration of temporal dynamics, spatial priority, and basic modes through three hierarchical operations. The first step involves smoothing adjustments in the time dimension. The system first obtains the base wiping frequency and response sensitivity from the command. The base wiping frequency represents the ideal target frequency determined by the system based on the current macroscopic environment. However, directly using this frequency would lead to frequent abrupt changes in wiping speed, affecting the driving experience. Therefore, response sensitivity is introduced as an adjustment factor to smooth the wiping frequency, generating a dynamically changing smooth frequency. The smooth frequency F output for the current control cycle... smooth (t), we have:

[0082] F smooth (t)=F smooth (t-1)+R sens ·[F base -F smooth [(t-1)],

[0083] Among them, F smooth (t) is the smoothed frequency output of the current control cycle, which will serve as the actual reference frequency for driving the wipers. F smooth (t-1) is the smoothing frequency of the previous control cycle, representing the current operating state of the wipers. F base The target basic scraping frequency is calculated from multi-dimensional feature vectors. R sens R is a dimensionless adjustment coefficient derived from the response sensitivity, with a value between 0 and 1. When the response sensitivity is high, R... sens A larger R value allows the actual frequency to quickly approach the target frequency; conversely, when the response sensitivity is low, R... sensThe first step involves a smaller frequency response, resulting in smoother frequency changes. The second step is to prioritize spatial allocation. Based on the area weight adjustment factor in the parameterized control command, the system assigns a spatially differentiated strategy to the upcoming wiping action. When executing a wiping mode, this area weight adjustment factor guides the system to prioritize wiping resources, such as wiping pressure or wiping speed, to the driver's core field of vision or other areas with high interference scores. This means that the force or speed of the wiper on the windshield during a single reciprocating motion is not constant, but dynamically modulated according to its current position and the wiping area priority at that position. The third step generates and outputs the final motor control waveform. The system integrates the results of the first two steps, combined with the wiping mode selected in the parameterized control command, to generate the final motor control waveform driving the wiper actuator. This waveform uses the wiping mode, such as continuous or intermittent, as its basic framework. Its overall repetition period is determined by the smoothing frequency. Within the wiping stroke of each cycle, the amplitude or pulse width of the control waveform is adjusted in real time according to the wiping area priority. This finely modulated motor control waveform is sent to the motor controller, thereby precisely driving the wiper actuator to achieve complex and intelligent wiping actions.

[0084] Optionally, the method further includes:

[0085] When a driver's manual intervention is detected, the manual intervention and its corresponding multidimensional feature vector are recorded to form labeled data.

[0086] The calculation process of correcting the basic brushing frequency, the regional weight adjustment factor, and the response sensitivity using the labeled data.

[0087] Specifically, this method introduces a closed-loop adaptive learning mechanism, enabling the wiper control system to self-optimize based on the driver's actual behavior. The core of this process lies in treating the driver's manual actions as corrective instructions for the system's automatic decisions, and continuously improving the algorithm model based on this. The operation begins by monitoring the driver's manual interventions. The system detects whether the driver has made any manual interventions by monitoring the vehicle's control units in real time, such as the wiper control lever behind the steering wheel or the virtual buttons on the central control screen. These interventions include, but are not limited to, manually turning on the wipers, adjusting the wiper speed, or triggering a single wipe. Once such an operation is detected, the system immediately executes a data logging program. The data logging program simultaneously captures two key pieces of information. The first is the recorded manual intervention itself. The system converts this into a specific, quantifiable target value. For example, if the driver switches the wipers from automatic mode to medium-speed continuous mode, this operation is recorded as the expected medium-speed wiper frequency. The second is the environment and vehicle state perceived by the system at the exact moment the manual intervention occurred, i.e., the corresponding multi-dimensional feature vector. This vector fully encapsulates the visual information of rainfall, vehicle speed, steering intention, and driver assistance status at the moment of intervention. Subsequently, the system pairs the recorded target value of the manual intervention with the corresponding multi-dimensional feature vector, forming a new, high-quality labeled dataset. This labeled dataset can be understood as a learning sample, meaning "in this specific situation described by the multi-dimensional feature vector, the driver's desired control output is the result indicated by the manual operation." The final step is to use this newly generated labeled data to correct the system's internal calculation process. This data is fed into the system's background learning module to update the algorithm model responsible for generating parameterized control commands. Specifically, it corrects the calculation process for the base wiper frequency, area weight adjustment factor, and response sensitivity. For example, if the driver manually increases the frequency when the system determines it's light rain, the newly generated labeled data will be used to fine-tune the regression model that maps the multi-dimensional feature vector to the base wiper frequency, so that it can output a higher frequency value when encountering similar feature vectors in the future. Similarly, the driver's intervention can indirectly reflect their need for response speed or the cleanliness of a specific area, thus being used to correct the calculation logic of response sensitivity and area weight adjustment factor. This correction process can be executed periodically in the background, enabling continuous iteration and optimization of the model.

[0088] Optionally, the method further includes:

[0089] When the wipers are activated, the average value of the interference score of each grid in the core area of ​​the field of view is calculated to obtain the average interference value.

[0090] When the average interference value is detected to be greater than the preset interference threshold, the regional weight adjustment factor is corrected, and it is determined whether the average interference value is less than the preset interference threshold within the preset time window.

[0091] If not, then the range of the core region of the field of view is expanded.

[0092] Specifically, this method introduces a real-time, closed-loop self-correction and strategy upgrade mechanism to the wiper control system. This mechanism runs continuously after the wipers are activated to verify and optimize the actual cleaning effect of the wiping action. The core of this process is to assess the cleanliness of the most critical viewing area and dynamically adjust the control strategy based on the assessment results. First, after the wiper system is activated and performs the wiping action, the system immediately begins a verification process. It precisely locks the currently dynamically defined core viewing area and calls up the calculated interference scores for all grid cells within that area. Subsequently, the system performs an arithmetic average of these interference scores to obtain a single, quantitative indicator, namely the interference mean. have:

[0093]

[0094] Here, the Σ symbol represents the summation operation, and the summation object i is each grid cell within the core region of the field of view. i N is the dimensionless interference score calculated for the i-th grid cell in the previous step, which is a fusion of the texture blur variation value and the brightness distribution anomaly value. cThis is the total number of grid cells contained within the core visual field area, which is an integer. Next, the system compares the calculated mean interference value with a preset interference threshold. This threshold represents the maximum level of interference that the system can accept without significantly affecting driving. If the mean interference value is detected to be greater than this preset threshold, it indicates that even after the wipers have activated, the driver's primary focus area remains unclear. In this case, the system triggers a first-level correction, adjusting the area weight adjustment factor. This correction aims to enhance the cleaning of the core visual field area, for example, by increasing the priority value of that area in the area weight adjustment factor data structure, thereby allocating more wiping resources in the next wiping action. After executing the first-level correction, the system does not immediately end the process but initiates a preset time window to continuously monitor the mean interference value in the core visual field area. This is intended to give the corrected wiping strategy sufficient time to take effect. If, at the end of the time window, the mean interference value is still found to have failed to decrease below the preset interference threshold, the system considers the first-level correction strategy to have failed to effectively solve the problem. This may mean the problem isn't just insufficient wiping force, but rather that the system's initial definition of the driver's area of ​​focus is too narrow. Therefore, the system will trigger a second-level escalation strategy, directly expanding the range of the core visual field. This expansion defines a larger image area as a high-priority region, thus receiving greater attention and stronger cleaning in all subsequent calculations and control.

[0095] Based on the same inventive concept, such as Figure 3 As shown, the present invention also provides a windshield wiper control system based on AI vision, the system comprising:

[0096] The basic parameter acquisition module is used to acquire real-time images of the vehicle's windshield and real-time vehicle status parameters.

[0097] The dynamic mesh generation module is used to perform non-uniform dynamic weighted mesh division on the real-time image to generate a dynamic weighted mesh.

[0098] The visual interference processing module is used to extract the visual interference features of each grid cell in the dynamic weighted grid and generate a visual interference feature vector.

[0099] The multi-dimensional fusion module is used to fuse the visual interference feature vector with the real-time vehicle state parameters to generate a multi-dimensional feature vector;

[0100] The control decision module is used to generate parameterized control instructions based on the multidimensional feature vector. The parameterized control instructions include a basic brushing frequency, a region weight adjustment factor, a response sensitivity, and a brushing mode.

[0101] The parameterized execution module is used to drive the wiper actuator to move based on the parameterized control commands.

[0102] To verify the feasibility of this invention in practice, it was applied to an electric test vehicle from a certain brand. This test vehicle aims to provide an intelligent driving experience. However, traditional rain sensors based on infrared optics often suffer from misjudgments, missed detections, or slow responses when faced with complex rain conditions such as drizzle, fog, dirty glass, or splashes from other vehicles, affecting driving safety and experience. This invention aims to achieve precise, intelligent, and adaptive control of the windshield wiper system through AI vision methods.

[0103] To verify the effectiveness of the invention, the wiper control system of the invention was deployed on the test vehicle (experimental group), and another vehicle with the exact same configuration but using the original factory conventional infrared rain sensor (control group) was selected. The two vehicles underwent a three-month comparative test in the same real road environment, covering various conditions such as urban roads, highways, and tunnels, as well as precipitation weather of different intensities and forms.

[0104] In this embodiment, the experimental group vehicle's system first acquires real-time images of the windshield and vehicle status parameters, including vehicle speed, turn signal, ambient light intensity measured by the roof light sensor, and the activation status of advanced driver assistance systems (ADAS) such as adaptive cruise control (ACC) and lane keeping assist (LKA), through a front-mounted high-definition camera (1920x1080 resolution) and the vehicle's CAN bus.

[0105] During system operation, the real-time image is first divided into a non-uniform dynamic weighted grid. When the vehicle is traveling at 50 km / h on urban roads, the system identifies the core area of ​​the driver's field of vision and applies a fine 32x32 pixel grid (first density grid) and a weight value of 1.0 to it. For the surrounding area outside the core area, a sparse 128x128 pixel grid (second density grid) and a weight value of 0.3 are applied. When the vehicle accelerates to 120 km / h and enters a highway, the system dynamically expands the range of the core area of ​​vision according to the vehicle speed, increasing it by approximately 20% vertically to simulate the driver's gaze focusing on distant objects.

[0106] Subsequently, the system extracts visual disturbance features within each grid cell. In a scenario of light rain turning into moderate rain (15:20 on October 5, 2024), the system initially detected a large number of independent brightness distribution anomalies, i.e., discrete raindrops, within the core area of ​​the field of view, with low texture blur. As the rain intensified, the changes in texture blur and the values ​​of the rainwater motion vectors increased significantly, with the rainwater motion vectors manifesting as streaks of water forming on the windshield. These features, after being weighted according to their respective grids, are fused into a visual disturbance feature vector.

[0107] In a nighttime high-speed driving scenario (November 12, 2024, 20:45), ambient light intensity parameters were used to compensate for visual interference features. Due to the illumination from oncoming headlights, raindrops on the glass produced extremely strong glare. The system used low light intensity parameters to normalize and compensate for abnormal brightness distribution values, avoiding excessive swiping frequency caused by glare. Simultaneously, since the vehicle's ACC and LKA systems were active, the system generated a high response priority and fused it with visual, vehicle speed, and other features to generate a multi-dimensional feature vector containing high safety requirements.

[0108] Based on this multi-dimensional feature vector, the system generates parameterized control commands. In the aforementioned high-speed nighttime scenario, the system calculates a high base wiping frequency of 75 wipes per minute, selects the "high-speed continuous wiping" mode, and sets a high response sensitivity due to rapid changes in rain intensity. Simultaneously, based on the fact that the core area of ​​the field of vision has a higher interference score, a regional weight adjustment factor is generated to prioritize cleaning the central area. Finally, through a smooth adjustment algorithm, the system smoothly increases the wiper frequency from the previous intermittent mode to the target frequency within 1.2 seconds and drives the motor to execute the command.

[0109] This invention also demonstrates its self-learning and self-correction capabilities. In testing, when the system determined it to be drizzle and used low-frequency intermittent wiping, the driver manually switched the wipers to medium speed. The system immediately recorded this manual intervention and its corresponding multi-dimensional feature vector, reflecting a visually inconspicuous but significantly impactful fog-like vapor. This record constitutes new labeled data for offline correction of the decision model, enabling it to better identify such scenarios in the future. Furthermore, after one wipe, if the system detected that the average interference score in the core visual field was still higher than a preset threshold of 0.5, the system first adjusted the region weight adjustment factor, increasing the wiping intensity in that area. Within the subsequent 5-second time window, if the interference score still did not meet the threshold, the system automatically expanded the definition of the core visual field, ensuring absolute clarity in critical areas of vision.

[0110] Based on data comparison, the intelligent control system of this invention has significant advantages in response accuracy, timeliness, and driving experience.

[0111] Table 1 Comparison of Response Accuracy Rate and Number of Driver Manual Interventions

[0112]

[0113]

[0114] Table 2 Comparison of response times under different driving scenarios

[0115]

[0116] Table 3 Comparison of Core Visual Clarity Maintenance Rate When ADAS is Activated

[0117]

[0118] Tables 1-3 above record the comparative test data of this invention in practical applications. For example... Figure 4 As shown in Table 1, the experimental group, after applying this invention, demonstrated a significantly higher accuracy rate in responding to various complex rain conditions compared to the control group. This was particularly evident in scenarios where traditional sensors struggle to handle drizzle / fog and splashes from other vehicles, where the advantage was extremely pronounced. This directly led to a substantial reduction in the number of manual interventions required by the driver, decreasing the average number of interventions per 100 kilometers from nearly 10 in the control group to less than 1 in the experimental group, significantly improving driving convenience and focus. Figure 5 As shown in Table 2, the system response of this invention is much faster. Whether starting statically or entering a rainy area at high speed, the response time of the experimental group was reduced by more than 60%. Especially in sudden and highly dangerous scenarios such as water splashing from a vehicle ahead, this invention can achieve a sub-second rapid response, which traditional sensors can hardly handle effectively. This is crucial for ensuring clear instantaneous visibility and improving driving safety. Figure 6 As shown in Table 3, when the ADAS system, which has extremely high requirements for visual clarity, is in active state, the present invention can maintain the interference level of the driver's core visual field at an extremely low level, with an average interference value of less than 0.15 and a clarity maintenance rate of over 97%, providing a solid perception foundation for the reliable operation of the intelligent driving system.

[0119] In summary, the data results of this embodiment fully demonstrate the significant technological advancements of this invention in the field of windshield wiper control. Through refined perception based on AI vision, multi-dimensional contextual fusion decision-making, and adaptive learning capabilities, this invention effectively overcomes the limitations of traditional rain sensors, achieving more accurate, faster, and smarter windshield wiper control, greatly improving driving safety, comfort, and the intelligent driving experience.

[0120] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0121] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A windshield wiper control method based on AI vision, characterized in that, The method includes: Acquire real-time images of the vehicle's windshield and real-time vehicle status parameters; The real-time image is divided into a non-uniform dynamic weighted grid to generate a dynamic weighted grid; Visual interference features of each grid cell in the dynamic weighted grid are extracted to generate a visual interference feature vector; The visual interference feature vector is fused with the real-time vehicle state parameters to generate a multi-dimensional feature vector. Based on the multidimensional feature vector, parameterized control instructions are generated, which include a basic brushing frequency, a region weight adjustment factor, a response sensitivity, and a brushing mode. The wiper actuator is driven to move based on the parameterized control commands.

2. The windshield wiper control method based on AI vision according to claim 1, characterized in that, The real-time vehicle status parameters include real-time vehicle speed, steering signal, ambient light intensity, and driving assistance system status.

3. The windshield wiper control method based on AI vision according to claim 2, characterized in that, The generation of the dynamic weighted grid includes: The real-time image recognition is used to define the core area of ​​the driver's field of vision, which is the area of ​​primary focus. The range of the core visual area is dynamically expanded based on the real-time vehicle speed to obtain a dynamic core area. A preset first density grid and a preset first weight value are assigned to the region within the dynamic core region, and a preset second density grid and a preset second weight value are assigned to the region outside the dynamic core region. Integrate the first density grid and the second density grid to generate a dynamic weighted grid; Wherein, the grid density of the first density grid is greater than the grid density of the second density grid; and the first weight value is greater than the second weight value.

4. The windshield wiper control method based on AI vision according to claim 3, characterized in that, The generated visual interference feature vector includes: Calculate the texture blur change value of each grid in the dynamic weighted grid; Detect abnormal values ​​in the brightness distribution of each grid in the dynamic weighted grid; Analyze the rainwater movement vectors of each grid in the dynamic weighted grid; Based on the first weight value or the second weight value, the texture blurring change value, the brightness distribution anomaly value and the rain movement vector are weighted and fused to generate a visual interference feature vector. The visual interference features include the texture blurring change value, the brightness distribution anomaly value, and the rain movement vector.

5. The windshield wiper control method based on AI vision according to claim 4, characterized in that, The generation of multidimensional feature vectors includes: The visual interference feature vector is compensated using the ambient light intensity parameter. The compensated visual interference feature vector is jointly encoded with the real-time vehicle speed and the steering signal to generate driving context features; Based on the state of the driving assistance system, a response priority is generated, and the driving situation features are fused with the response priority to generate a multi-dimensional feature vector.

6. The windshield wiper control method based on AI vision according to claim 5, characterized in that, The generation of parameterized control instructions based on the multidimensional feature vector includes: Based on the multidimensional feature vector, the basic brushing frequency is calculated; Based on the texture blurring change value and the brightness distribution anomaly value, the interference score of each grid in the dynamic weighted grid is calculated; Based on the interference score and the multidimensional feature vector, a region weight adjustment factor is generated; The response sensitivity is calculated by analyzing the gradient of the multidimensional eigenvectors of a continuous time series. Based on the rainwater movement vector and the multidimensional feature vector, the current overall scene is classified and judged, and the wiping mode is selected according to the judgment result; By integrating the basic brushing frequency, the area weight adjustment factor, the response sensitivity, and the brushing mode, a parameterized control command is obtained.

7. The windshield wiper control method based on AI vision according to claim 6, characterized in that, The action of driving the wiper actuator based on the parameterized control command includes: The base brushing frequency is smoothly adjusted based on the response sensitivity to obtain a smooth frequency. Based on the area weight adjustment factor, the priority of the scraping area is assigned when the scraping mode is executed; The smoothing frequency and the priority of the wiping area are combined to generate a motor control waveform to drive the wiper actuator.

8. The windshield wiper control method based on AI vision according to claim 6, characterized in that, The method further includes: When a driver's manual intervention is detected, the manual intervention and its corresponding multidimensional feature vector are recorded to form labeled data. The calculation process of correcting the basic brushing frequency, the regional weight adjustment factor, and the response sensitivity using the labeled data.

9. The windshield wiper control method based on AI vision according to claim 6, characterized in that, The method further includes: When the wipers are activated, the average value of the interference score of each grid in the core area of ​​the field of view is calculated to obtain the average interference value. When the average interference value is detected to be greater than the preset interference threshold, the regional weight adjustment factor is corrected, and it is determined whether the average interference value is less than the preset interference threshold within the preset time window. If not, then the range of the core region of the field of view is expanded.

10. A windshield wiper control system based on AI vision, applied to the windshield wiper control method based on AI vision as described in any one of claims 1-9, characterized in that, The system includes: The basic parameter acquisition module is used to acquire real-time images of the vehicle's windshield and real-time vehicle status parameters. The dynamic mesh generation module is used to perform non-uniform dynamic weighted mesh division on the real-time image to generate a dynamic weighted mesh. The visual interference processing module is used to extract the visual interference features of each grid cell in the dynamic weighted grid and generate a visual interference feature vector. The multi-dimensional fusion module is used to fuse the visual interference feature vector with the real-time vehicle state parameters to generate a multi-dimensional feature vector; The control decision module is used to generate parameterized control instructions based on the multidimensional feature vector. The parameterized control instructions include a basic brushing frequency, a region weight adjustment factor, a response sensitivity, and a brushing mode. The parameterized execution module is used to drive the wiper actuator to move based on the parameterized control commands.

Images