Intelligent headlamp control method and system based on driving intention prediction

Abstract

The invention relates to the technical field of vehicle intelligent lighting, in particular to an intelligent headlamp control method and system based on driving intention prediction. The method comprises the steps that environment data, driver biological signals and vehicle dynamic data are collected through a multi-source sensor; predicting a future short-time driving intention and confidence by adopting a space-time diagram convolutional network; correcting illumination parameters in real time according to the eye movement gazing area of the driver and the heart rate variability; initiating a conservative wide-angle lighting mode when the intention confidence is below a threshold; a stroboscopic warning light pattern is triggered in response to an ADAS system pre-collision signal. The system comprises a data acquisition unit, a processing unit and an execution layer, and cooperative adjustment of the deflection angle, the brightness and the irradiation range of the headlamp is achieved. The problems of response lag and blind area risk of a traditional self-adaptive headlamp are solved, and the illumination timeliness of a blind area of a curve and the safety of a low-visibility working condition are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle lighting technology, and in particular to an intelligent headlight control method and control system based on driver intention prediction. Background Technology

[0002] Current Advanced Front-lighting Systems (AFS) primarily rely on vehicle dynamic parameters such as steering wheel angle and vehicle speed to passively adjust the beam pattern, which has the following fundamental drawbacks: 1. Response lag issue: Traditional solutions require the detection of actual steering action (such as steering wheel rotation > 15°) before adjusting the lighting, resulting in a 0.8-1.2 second blind spot in sharp bends, increasing the risk of nighttime driving; 2. Lack of human factor adaptation: The driver's eye movement focus distribution and physiological state (such as distraction and fatigue) are not taken into account, and it is impossible to enhance the lighting for the visual focus area; 3. Insufficient safety redundancy: There is a lack of conservative lighting strategies based on prediction confidence under low visibility conditions, and there is no light pattern coordination with the ADAS emergency braking system.

[0003] Although some solutions attempt to combine navigation maps to predict curves, they still cannot solve the lighting response needs of drivers who suddenly intend to change lanes. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the existing lighting steering changes respond after the actual operation, lack the analysis of the driver's driving intention and driving state, and lack lighting safety settings in some special scenarios.

[0005] Therefore, this invention provides an intelligent headlight control method and control system based on driving intention prediction, which achieves: Intention-driven pre-lighting mechanism: By predicting driving intentions through neural networks, the light spot deflection is initiated before physical steering occurs, eliminating blind spots; Biosignal fusion control: mapping eye-tracking gaze coordinates to road surface area weights, dynamically adjusting local brightness, and simultaneously extending the illumination distance based on heart rate variability; The confidence-adaptive safety strategy automatically switches to wide-angle lighting mode (e.g., ≥120° illumination angle) when the prediction confidence is lower than a set threshold, and triggers a flashing warning in response to the ADAS collision warning signal.

[0006] The technical solution adopted by this invention to solve its technical problem is: A method for intelligent headlight control based on driver intention prediction includes the following steps: Step 1: Collect lane line and obstacle information, vehicle information and surrounding vehicle information, and also collect driver biosignals; The driver's biosignals include the driver's gaze coordinates, heart rate variability, and grip strength; Step 2: Preprocess the collected multi-source data; Step 3: Use a neural network to process the preprocessed data obtained in Step 2, and output the driving intention prediction result and confidence level; Step 4: Based on the dynamic light pattern control logic, multi-dimensional light pattern adjustment commands are generated according to the driving intention prediction results, confidence data and driver biosignals to dynamically adjust the headlight light pattern; The light pattern includes the horizontal deflection angle, pitch angle, and regional brightness distribution of the headlights; Step 5: When the confidence level is below the threshold, expand the headlight illumination angle to wide-angle lighting mode; Step 6: Trigger warning light pattern in response to ADAS pre-collision signal.

[0007] Furthermore, in step 2, the preprocessing of the multi-source data includes time synchronization processing, outlier processing, data quality checking, and feature fusion.

[0008] Further, in step 3, the neural network is a spatiotemporal graph convolutional network, used to construct a spatiotemporal interaction graph containing the vehicle, surrounding vehicles, and lane lines. The vehicle nodes in the spatiotemporal interaction graph structure have feature vectors composed of their position coordinates, velocity components, and heading angles. The lane line nodes contain lane line curvature radius, lane type, and lane width feature parameters. When the centroid Euclidean distance between the vehicle and surrounding vehicles is less than 50 meters, an undirected connection edge is established between the vehicles. The vehicle always establishes an undirected connection edge between itself and the centerline of its lane. The edge weights are calculated using a dynamic function (Formula 3) in a learnable form. This function comprehensively considers the absolute value of the speed difference (relative speed) between the vehicle and surrounding vehicles and the Euclidean distance. Specific examples include: Basic form (vehicle-to-vehicle connection): used to quantify the dynamic interaction risk between the vehicle and surrounding vehicles (such as a following vehicle approaching at high speed), the function formula is: (1) in, Let be the Euclidean distance (in meters) between the centroids of vehicles i and j. The absolute value of the speed difference between the vehicle and surrounding vehicles. This is the distance attenuation coefficient; Extended form (vehicle and lane connection): Used to quantify the relationship between a vehicle and its lane (e.g., vehicle deviation from lane center), the function formula is: (2) in, It is the vertical distance (in meters) from the vehicle's center of gravity to the center line of its lane. This is the distance attenuation coefficient. Lane type weights; Learnable Form (Complex Scenarios): Adaptive weight adjustment for complex traffic scenarios (such as high traffic volume and frequent lane changes), the function formula is: (3) in, The activation function maps the output to the 0-1 range (0 represents no correlation, and 1 represents strong correlation). It is a 2-layer neural network (64-dimensional input layer, 1-dimensional output layer), with the input being the absolute value of the speed difference. Euclidean distance from the center of mass .

[0009] The spatial topology of vehicles and roads is modeled by spatiotemporal interaction graph structure, and combined with the changes in time series (such as 5 consecutive frames), a two-layer graph convolutional network is used to process the topology. The graph convolutional network can capture dynamic high-risk interactions (such as vehicles approaching from a distance, changing lanes and cutting in).

[0010] An LSTM-CRF intent prediction model is constructed, consisting of LSTM and CRF layers, to process the feature sequences output by the spatiotemporal graph convolutional network. The spatiotemporal graph convolutional network outputs a 128-dimensional feature sequence (e.g., 5 consecutive frames), which is input into the LSTM-CRF intent prediction model, and the output is a driving intent label and probability distribution.

[0011] Furthermore, in step 3, the confidence calculation process is as follows: fifty forward propagation samplings are performed using the Monte Carlo Dropout mechanism, with twenty percent of neurons randomly discarded each time, and the mean probability μ and standard deviation σ of each driving intention label are calculated respectively.

[0012] When the mean of the maximum probability is below 0.6 or the standard deviation exceeds 0.25, it is judged as a low-confidence scene and the wide-angle lighting mode is triggered.

[0013] Furthermore, in step 4, the dynamic control logic for the light pattern includes: a first layer based on the intent prediction of a spatiotemporal graph convolutional network to drive pre-illumination actions, completing the spot deflection before physical turning occurs, eliminating the 0.8-second illumination blind spot caused by mechanical delay in traditional solutions; a second layer that integrates driver eye movement vectors and heart rate variability data, such as automatically increasing the illumination distance by 8 meters when the heart rate standard deviation is >50ms to cope with sudden operational hesitation; and a third layer that establishes a confidence-light pattern mapping rule, which provides a large coverage area for the wide-angle lighting mode in low-confidence scenarios and improves the ADAS-assisted flicker response speed.

[0014] Further, in step 4, the calculation process for regional brightness adjustment is as follows: the gaze point coordinates are mapped to road surface grid weight values ​​using a vector projection algorithm, and the calculation formula is: (4) in, The road surface grid weight value indicates the priority of brightness adjustment in the eye-tracking fixation area; , The driver's eye-tracking gaze coordinates (normalized values, ranging from 0 to 1) correspond to the relative position of the road surface. , The road surface is divided into 1024 independent grids using the center coordinates of the road surface grid (normalized value, range 0~1); This is the heart rate variability coefficient, reflecting the driver's condition.

[0015] The road surface grid weight value is used to adjust the brightness of the corresponding area, so as to achieve adaptive matching between physiological state and optical parameters.

[0016] Furthermore, in step 5, the wide-angle lighting mode has a horizontal illumination angle of ≥120° and the brightness is increased to more than 80% of the maximum level.

[0017] The goal is to ensure that drivers have sufficient safe visibility by maximizing coverage and high brightness redundancy without relying on intent prediction, in low-confidence scenarios such as complex environments (fog, rain), poor driver condition (distraction, fatigue) or sensor failure.

[0018] Furthermore, in step 6, the warning light pattern is generated by independent partitions within the main beam of the headlight and focused on the potential collision area.

[0019] A control system, comprising, The data acquisition unit is used for multi-source data acquisition, including acquiring driver biosignals, environmental and vehicle dynamic information; and then preprocessing the multi-source data through the data fusion layer in the data acquisition unit. The processing unit has a built-in intent prediction module and confidence assessment module, which outputs intent prediction results and prediction uncertainty based on the collected data. The uncertainty in prediction is characterized by a confidence level. The execution layer is connected to the vehicle deflection controller, modulator, and ADAS. The execution layer calls the light pattern dynamic control logic through the light pattern dynamic control module, sends dynamic adjustment of the headlight illumination direction and illumination range, and responds to the ADAS pre-collision signal to trigger the warning light pattern.

[0020] Compared to traditional headlight control schemes, the advantages of this invention are: 1. Eliminate blind spots due to lighting delay: Adjust lighting before steering actions by predicting intent, solving the problem of blind spots in corners caused by mechanical response delay in traditional AFS, and significantly improving nighttime cornering safety; 2. Enhanced human adaptability: The system dynamically allocates lighting resources based on the driver's eye focus, automatically enhancing the brightness of key areas when the driver is distracted or fatigued, thus reducing the probability of accidents caused by visual negligence. 3. Systematized safety redundancy: The confidence-driven wide-angle lighting mode provides backup lighting for low visibility conditions, and the ADAS collaborative warning mechanism constructs an active safety closed loop. Actual simulations show that it can reduce lighting-related accidents by 23%-41%. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is an architecture diagram of the control system of the present invention; Figure 2 This is a logic block diagram of the control method of the present invention; Figure 3 This is a flowchart of the multi-source data acquisition and fusion process of this invention; Figure 4 This is a schematic diagram illustrating the principle of predicting intent and assessing confidence levels in this invention. Figure 5 This is the flowchart of the dynamic control logic for the optical pattern of this invention. Detailed Implementation

[0023] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and therefore showing only the components relevant to the invention. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of this disclosure.

[0024] It should also be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use.

[0025] Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and apparatus should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0026] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0027] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0028] A method and control system for intelligent headlight control based on driver intention prediction, applicable to, but not limited to, motor vehicles equipped with advanced driver assistance systems (ADAS), such as passenger cars and commercial vehicles.

[0029] Reference Figure 1 , Figure 3 A smart headlight control system based on driving intention prediction includes a data acquisition unit, a processing unit, and an execution layer.

[0030] The data acquisition unit is used for multi-source data acquisition, including environmental information, vehicle information, and driver biosignals. Specifically, the data acquisition unit acquires lane line information (lane line curvature radius, lane type, and lane width, etc.) and obstacle information through the vehicle's forward-facing camera; detects vehicle distance and relative speed through the vehicle's millimeter-wave radar; captures gaze point coordinates through an eye tracker installed in the vehicle; monitors heart rate variability and grip strength through biosensors on the vehicle's steering wheel; and reads vehicle dynamic parameters (position coordinates, speed components, and heading angle, etc.) through the vehicle's CAN bus. Simultaneously, the data acquisition unit integrates a data fusion layer to perform time synchronization processing, outlier handling, data quality checks, and feature fusion on the acquired multi-source data. Time synchronization processing: Timestamps are unified for multi-source data using linear interpolation to ensure data time sequence consistency; Outlier handling: Outlier data is identified and removed, and missing values ​​are filled with sliding window mean to avoid outlier data interfering with subsequent predictions; Data quality check: Data is checked to see if it meets the standards; Feature fusion: Multi-source data is converted into high-dimensional feature vectors used by the processing unit.

[0031] The spatiotemporal graph construction module in the processing unit is based on a spatiotemporal graph convolutional network to construct a spatiotemporal interaction graph containing the vehicle, surrounding vehicles, and lane lines. The processing unit also includes an intent prediction module and a confidence evaluation module. The intent prediction module obtains the intent prediction results, namely driving intent labels and probability distributions, through an LSTM-CRF intent prediction model. The confidence evaluation module uses Monte Carlo Dropout sampling to calculate the probability mean μ and standard deviation σ of each driving intent label (intent type), which reflects the prediction uncertainty.

[0032] The execution layer is connected to the headlight spot deflection controller and modulator of the vehicle, and also to the vehicle's ADAS system. The execution layer calls the light pattern dynamic control logic through the light pattern dynamic control module, and combines the driving intention prediction results, confidence data and driver biosignals to generate multi-dimensional light pattern adjustment commands to dynamically adjust the light pattern. The light pattern includes the headlight illumination direction, illumination range, brightness, etc. The brightness adjustment is calculated by the eye-tracking-light spot linkage module in the light pattern dynamic control module. When the confidence level is below the threshold, the state machine decision-making mechanism built into the execution layer triggers a safety strategy to automatically switch to the wide-angle lighting mode executed by the safety redundancy module in the execution layer when σ > 0.25. At the same time, the current scene features are stored and uploaded to the edge server of the control system through the vehicle communication system (V2X) to optimize the spatiotemporal graph convolutional network loaded therein. In this mode, the illumination angle is expanded to a horizontal illumination angle ≥ 120° and the brightness is increased to more than 80% of the maximum level. Specifically, the light pattern dynamic control module is also connected to the vehicle's ADAS system through the ADAS collaboration module. The light pattern dynamic control module receives the pre-collision signal and triggers the warning light pattern. The warning light pattern is generated by independent partitions inside the headlight main beam. It achieves 8-12Hz high-frequency flashing through digital micromirror devices (DMD, which can realize millisecond-level light pattern reconstruction, supports 1024 independent lighting zone control, and its wavelength adaptation range covers 450-750nm, ensuring penetration performance in rainy and foggy weather) or matrix LED modules, and focuses on the potential collision area. The ADAS collaborative strobe response speed is 5 times faster than CAN bus transmission. In nighttime curve tests in mountainous areas, this solution increases the effective lighting coverage from 67% to 92% of the baseline solution; the warning light pattern can also include a dynamic warning light strip, which is projected onto vehicles behind to alert them to the danger.

[0033] The control system is compatible with mainstream vehicle sensors via CAN / Ethernet interface, supports matrix LED and DLP projection light sources, and has a retrofit cost that is 30% lower than industry solutions.

[0034] Reference Figures 2-5 A method for intelligent headlight control based on driver intention prediction includes the following steps: S1, Multi-source data acquisition: The data acquisition unit collects lane lines and obstacle information, vehicle information and surrounding vehicle information, and also collects driver biosignals including driver gaze coordinates, heart rate variability and grip strength.

[0035] S2, Data Preprocessing: The collected multi-source data is preprocessed through the data fusion layer, specifically including the following steps: S21 Time Synchronization Processing: Differences in sampling frequencies of different sensors (including vehicle sensors, biosensors, etc.) (e.g., 30fps for cameras, 100fps for radar, and 50fps for biosensors) can lead to timing misalignment. By using linear interpolation, all data is unified to a 100ms timestamp (matching the radar sampling frequency) to ensure data timing consistency. S22 Outlier Handling: By identifying and removing outlier data (such as "ghost cars" falsely detected by radar or "heart rate mutations" from biosensors), missing values ​​are filled with the mean of a sliding window (e.g., a window size of 5 frames) to avoid outlier data interfering with subsequent predictions. S23 Data Quality Check: Check for missing values ​​(≤5%), outliers (≤2%), and time sequence consistency. If the data meets the standard, proceed to the feature extraction step; otherwise, re-collect the data. S24 Feature Fusion: Multi-source data is converted into high-dimensional feature vectors, which are then weighted and concatenated (e.g., environmental features weighted at 0.4, biological features weighted at 0.3, and vehicle features weighted at 0.3) and input into the spatiotemporal graph construction module of the processing unit. The fused features include: environmental features (lane line curvature, vehicle distance, lane type, lane width, relative speed between vehicles, etc.), biological features (gaze coordinates, grip strength, heart rate variability HRV_index=SDNN / 50, etc.), and vehicle features (position coordinates, velocity components, lateral angle, steering wheel angle, vehicle speed, yaw rate, etc.).

[0036] S3, Intent Prediction and Confidence Assessment: The preprocessed data obtained in step 2 is processed using a neural network to output the driving intent prediction result and confidence level; specifically, it includes the following steps: S31 Spatiotemporal Graph Convolutional Network Construction: The spatiotemporal graph construction module is based on a spatiotemporal graph convolutional network to construct a spatiotemporal interaction graph containing the vehicle, surrounding vehicles, and lane lines. Vehicle nodes have feature vectors composed of position coordinates, velocity components, and heading angles. Lane line nodes have feature vectors composed of lane line curvature radius, lane type, and lane width. When the centroid Euclidean distance between the vehicle and surrounding vehicles is less than 50 meters, an undirected connection edge is established between the vehicles. The vehicle always has an undirected connection edge between itself and the centerline of its lane. The edge weights are calculated using a learnable dynamic function (Formula 3). This function comprehensively considers the absolute value of the speed difference between the vehicle and surrounding vehicles (relative speed between vehicles) and the Euclidean distance. Specific examples include: Basic form (vehicle-to-vehicle connection): used to quantify the dynamic interaction risk between the vehicle and surrounding vehicles (such as a following vehicle approaching at high speed), the function formula is: (1) in, Let be the Euclidean distance (in meters) between the centroids of vehicles i and j. This is the absolute value of the speed difference between the vehicle and surrounding vehicles (e.g., if the vehicle's speed is 60 km / h and the vehicle behind it is 80 km / h, then...). ), The distance attenuation coefficient (obtained through experimental optimization - that is, by averaging multiple experiments); Extended form (vehicle and lane connection): Used to quantify the relationship between a vehicle and its lane (e.g., vehicle deviation from lane center), the function formula is: (2) in, It is the vertical distance (in meters) from the vehicle's center of gravity to the center line of its lane. This is the distance attenuation coefficient. Lane type weights (e.g., fast lanes) slow lane ); Learnable Form (Complex Scenarios): Adaptive weight adjustment for complex traffic scenarios (such as high traffic volume and frequent lane changes), the function formula is: (3) in, The activation function maps the output to the 0-1 range (0 represents no correlation, and 1 represents strong correlation). It is a 2-layer neural network (64-dimensional input layer, 1-dimensional output layer), with the input being the absolute value of the speed difference. Euclidean distance from the center of mass .

[0037] It is important to note that edge weights can quantify the strength of the association between nodes, allowing spatiotemporal graph convolutional networks to pay more attention to vehicles that are close to each other and have high relative speeds, or vehicles that are off the center of the lane, thereby improving the accuracy of intent prediction and solving the problem of extracting high-risk interaction features.

[0038] For example: Edge weights between vehicles: Combining relative speed and Euclidean distance, vehicles that are close to each other and have high relative speeds (such as a car accelerating to overtake) have larger edge weights. Spatiotemporal graph convolutional networks will enhance the feature extraction of such high-risk interactions. Vehicle edge weights relative to the lane centerline: Based on the distance of the vehicle from the lane centerline, vehicles that deviate from the lane centerline (such as distracted vehicles) have smaller edge weights, indicating their possible lane change intentions.

[0039] S32 Feature Extraction: The spatial topology of vehicle-to-vehicle and vehicle-to-road interactions is modeled using a spatiotemporal interaction graph, combined with changes over a time series (5 consecutive frames). A two-layer graph convolutional network is then used to process this topology. Specifically: The first layer outputs a 128-dimensional feature matrix and processes it using the ReLU activation function; the second layer introduces a four-head attention mechanism to enhance the feature extraction of high-risk interaction relationships.

[0040] Graph convolutional networks can capture dynamic, high-risk interactions (such as vehicles approaching from a distance or changing lanes).

[0041] S33 Intent Prediction: An LSTM-CRF intent prediction model is established. The LSTM-CRF intent prediction model consists of LSTM layers and CRF layers, specifically: LSTM layer (Long Short-Term Memory Network): It adopts a 2-layer stacked structure (128 memory units per layer), inputs the graph convolutional features of 5 consecutive frames for temporal modeling, and extracts temporal features (such as "the steering wheel angle gradually increases" to indicate the steering intention). CRF layer (Conditional Random Field): It handles the dependencies between labels (e.g., the probability of "turning left" is higher after "going straight"), selects the label sequence with the highest probability through an algorithm (existing technology in this layer), and outputs the driving intention label (e.g., "turning right") and probability distribution (e.g., "probability of turning right is 92.7%)).

[0042] The LSTM-CRF intent prediction model is used to process the feature sequence output by the spatiotemporal graph convolutional network. Specifically, the spatiotemporal graph convolutional network outputs a 128-dimensional feature sequence (5 consecutive frames), which is input into the LSTM-CRF model, and the output is the driving intent label and probability distribution.

[0043] S34 Confidence Calculation: Fifty forward propagation samples are taken using the Monte Carlo Dropout mechanism, with twenty percent of neurons randomly discarded each time. The mean probability μ and standard deviation σ of each driving intention label (intention type) are calculated. When the mean maximum probability is lower than the threshold of 0.6 or the standard deviation exceeds the threshold of 0.25, it is judged as a low-confidence scenario and S5 is triggered.

[0044] This confidence quantification method effectively characterizes prediction uncertainty based on Bayesian principles, achieving high accuracy in high-speed lane-changing scenario tests while keeping the average inference time within 18 milliseconds.

[0045] S4, the light pattern dynamic control module, based on the light pattern dynamic control logic, generates multi-dimensional light pattern adjustment commands according to the driver's intention prediction results, confidence assessment data, and driver's biosignals. It dynamically adjusts the horizontal deflection angle, pitch angle, and regional brightness distribution of the headlights. The brightness of the corresponding area is mapped to road surface grid weight values ​​by the eye-tracking-light spot linkage module using a vector projection algorithm. The calculation formula is as follows: (4) in, The road grid weight value (unitless, range 0~1) represents the brightness adjustment priority of the eye-tracking fixation area (the larger the value, the higher the brightness of the corresponding area). , The driver's eye-tracking gaze coordinates (normalized values, ranging from 0 to 1) correspond to the relative position on the road surface (e.g., x=0.35 represents the 35% area on the left, y=0.72 represents the 72% area in front). , The road surface is divided into 1024 independent grids using the center coordinates of the road surface grid (normalized value, range 0~1); The heart rate variability coefficient (unitless, range 0-3), detected by a biosensor. Where SDNN is the standard deviation of heart rate, in milliseconds, it reflects the driver's state (e.g., <1.0 indicates a stable state. >2.0 indicates fatigue state). The road surface grid weight value calculated by the above formula is directly used to adjust the brightness of the corresponding area, achieving adaptive matching between physiological state and optical parameters. For example, when the driver's gaze point deviates from the center of the curve ( , The curve exit area was calculated. Based on this, the system increases the brightness of the area by 35%, covering the driver's visual focus.

[0046] The dynamic light pattern control logic achieves precise light pattern adjustment through a three-layer closed-loop mechanism: The first layer, based on a spatiotemporal graph convolutional network, drives pre-illumination actions based on intent prediction, completing the light spot deflection before physical turning occurs, eliminating the 0.8-second blind spot caused by mechanical delay in traditional solutions; the second layer integrates driver eye movement vectors and heart rate variability data, for example, automatically increasing the illumination distance by 8 meters when the heart rate standard deviation is >50ms to cope with sudden operational hesitation; the third layer establishes a confidence-light pattern mapping rule, achieving a 2.3-fold wider-angle lighting mode coverage in low-confidence scenarios compared to traditional solutions, and improving the ADAS collaborative flicker response speed by 5 times compared to CAN bus transmission. In mountainous nighttime curve tests, this solution increased the effective lighting coverage from 67% to 92% of the baseline solution, while reducing driver reaction time in dangerous scenarios by 0.4 seconds.

[0047] In the dynamic control logic of the light pattern, the dynamic control module of the control system generates multi-dimensional light pattern adjustment commands based on the prediction results of driving intention, confidence calculation data, and driver biosignals. Specifically, when a left turn intention is predicted and the confidence level is ≥80%, the light spot deflection controller drives the headlights to rotate 20° to the left and increases the brightness of the left area by 30%, illuminating the inner curve trajectory 1.2 seconds earlier. Simultaneously, the coordinates of the gaze point obtained by the eye tracker are analyzed. If the driver's gaze focus is detected to deviate from the center of the curve by more than 15° for more than 2 seconds, non-uniform brightness analysis is used to adjust the light pattern. The algorithm enhances the illumination intensity in the offset direction by 25%-40%, compensating for the loss of vision caused by distraction. The uncertainty index output by the confidence assessment module triggers the safety strategy adjustment. When the standard deviation σ > 0.25, the illumination range modulator expands the horizontal beam angle to 130° and maintains 85% of the maximum brightness, constructing a conservative illumination redundancy under low visibility conditions - thus triggering S5. At the same time, the ADAS collaborative interface monitors the braking system (AEB) signal in real time through a hard-wired direct connection channel. Within 20 milliseconds after receiving the AEB pre-pressurization command, it activates the high-frequency flashing mode, thus triggering S6.

[0048] S5, when the confidence level is below the threshold, expand the headlight illumination angle to a wide-angle lighting mode. In the wide-angle lighting mode, the illumination angle is expanded to a horizontal illumination angle of ≥120°, preferably 130°-140°, and the brightness is increased to more than 80% of the maximum level, preferably 85%-90%. At the same time, it can also add an additional 8-12 meters of illumination distance. The aim is to ensure that the driver has sufficient safe visibility in low-confidence scenarios such as complex environments (fog, rain), poor driver condition (distraction, fatigue) or sensor failure, by maximizing coverage and high brightness redundancy without relying on intention prediction.

[0049] The S6 responds to the ADAS pre-collision signal and triggers a warning light pattern. The warning light pattern focuses on the potential collision area with the main beam of the headlights at a strobe rhythm of 8-12Hz, while simultaneously projecting a dynamic warning light strip to vehicles behind.

[0050] Example 1 Scene setting: Mountain night road curve scene Specific implementation process: In a mountainous nighttime curve scenario, a vehicle enters a right curve with a lane curvature radius of 85m at a speed of 60km / h. The system identifies the curve's geometric features (8% gradient, lane curvature radius of 85m) using a forward-facing camera. Simultaneously, millimeter-wave radar detects an oncoming vehicle approaching at 25m / s relative speed from 62m away. An eye tracker monitors the driver's gaze point coordinates (0.35, 0.72) in real time, showing a continuous shift from the curve's center for 1.8 seconds. A steering wheel biosensor detects the driver's heart rate variability (SDNN) at 38ms (HRV_index = 0.76). A spatiotemporal graph convolutional network constructs a spatiotemporal interaction graph including the vehicle, three surrounding vehicles, and two lane lines. LSTM temporal analysis outputs a right turn intention probability of 92.7%. Monte Carlo Dropout sampling with 50 iterations and a standard deviation σ = 0.08 confirms a high-confidence scenario. Based on this, the dynamic beam pattern control module drives the Texas Instruments DLP5531-Q1 digital micromirror device to perform triple adjustments: the beam is deflected 18° to the right to cover the curve trajectory; the road surface grid weight value in the curve exit area is calculated to 0.83 based on the gaze point coordinates, and the brightness of this area is increased by 35%; at the same time, the illumination distance is increased by 6m forward because HRV_index < 1.0. The final result is a beam pattern distribution covering the area from the curve's center to the exit, with a maximum illuminance of 82 lux. Compared to traditional AFS systems, this reduces the duration of the blind spot from 1.1 seconds to 0.2 seconds, advances the driver's braking point by 0.7 seconds, and increases the curve passage speed by 12 km / h.

[0051] Example 2 Scene setting: Foggy highway scene Specific implementation process: In a foggy highway scenario with visibility below 50m, the vehicle travels at 105km / h with an ambient humidity of 90%. Due to fog interference, the confidence level of the forward-facing camera drops to 45%. The vehicle communication system (V2X) receives an accident warning signal 2km ahead. The biosensor detects a driver heart rate variability (SDNN) of 112ms (HRV_index=2.24). The intent prediction module outputs a straight-ahead probability of 64.3%, and the standard deviation of the Monte Carlo Dropout calculation is σ=0.31 (exceeding the 0.25 threshold). The system activates a three-level safety response: First, it triggers a wide-angle illumination mode to expand the horizontal beam to 140° and increase the brightness to the maximum level of 90%. Second, based on the physiological state of HRV_index>2.0, it adds an additional 12m of illumination distance. Finally, within 20ms after receiving the AEB pre-pressurization signal, it activates a 10Hz high-frequency flashing mode through a hard-wired direct connection channel and focuses on the potential collision area 35m to the right front. This solution achieves 2.1 times the lighting coverage of traditional solutions, reduces the risk of rear-end collisions by 41%, and keeps the system's end-to-end latency within 50ms, meeting the functional safety requirements of Automotive Safety Integrity Level (ASIL-B). Simulations showed that during a 500ms emergency avoidance maneuver, the warning light pattern successfully guided following vehicles to decelerate by 0.7 m / s², preventing a rear-end collision.

[0052] Example 3 Scenario setting: Distracted driving scenario at an urban intersection Specific implementation process: In an urban intersection scenario, a vehicle approaches the intersection at 40 km / h to switch traffic lights. The forward-facing camera detects a pedestrian suddenly crossing the zebra crossing, and millimeter-wave radar detects the pedestrian at a distance of 15 m and a relative speed of 3 m / s. The eye tracker detects that the driver's gaze point continuously deviates from the road focus for 3.2 seconds (coordinates (0.15, 0.35) are located in the vehicle information screen area). The biosensor detects a heart rate variability (SDNN) of 65 ms (HRV_index=1.3) and a sudden drop in steering wheel grip force to 8 N. A spatiotemporal graph convolutional network constructs a spatiotemporal interaction graph including pedestrians and adjacent vehicles, predicting an emergency braking intent probability of 88.4% with σ=0.12. The system triggers a composite response: based on the gaze point offset, a pedestrian area weight value of 0.92 is calculated, and the illumination in that area is increased to 120%. Based on the grip force data, the system determines the risk of operation delay and activates automatic braking pre-pressurization. At the same time, the confidence assessment unit maintains the standard light pattern because σ<0.15. Within a 200ms response time, the headlights successfully increased pedestrian visibility by 2.3 times in the focused area, enabling the AEB system to brake within 0.8 seconds and avoid a collision. Subsequent data analysis showed that in this scenario, the system improved hazard recognition time by 1.2 seconds through biosignal fusion, achieving a 67% performance improvement compared to a purely environmental perception-based solution.

[0053] In summary, this invention uses a neural network to predict driving intentions and initiates light spot deflection 0.5-1 seconds before physical steering occurs to eliminate blind spots; it maps eye-tracking coordinates to road area weights, dynamically adjusts local brightness, and simultaneously expands the illumination distance based on heart rate variability; when the prediction confidence is <60%, it automatically switches to wide-angle mode (≥120° illumination angle) and triggers an 8-12Hz strobe warning in response to ADAS collision warning signals.

[0054] This concludes the detailed description of the intelligent headlight control method and control system based on driving intention prediction according to this disclosure. To avoid obscuring the concept of this disclosure, some details known in the art have not been described. Those skilled in the art will fully understand how to implement the technical solutions disclosed herein based on the above description.

[0055] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined by the scope of the claims.

Claims

1. A method for intelligent headlight control based on driver intention prediction, characterized in that, Includes the following steps: S1 collects environmental, vehicle, and surrounding vehicle information, as well as driver biosignals. S2, preprocesses the collected multi-source data; S3, use a neural network to process the data obtained in step 2, and output the driving intention prediction result and confidence level; S4, based on the dynamic light pattern control logic, generates multi-dimensional light pattern adjustment commands according to the driving intention prediction results, confidence level and driver biosignals, and dynamically adjusts the headlight light pattern; The light pattern includes the horizontal deflection angle, pitch angle, and regional brightness distribution of the headlights; S5, in low-confidence scenes, the light pattern is adjusted to wide-angle lighting mode; S6 is a warning light pattern triggered in response to ADAS pre-collision signals.

2. The intelligent headlight control method based on driving intention prediction according to claim 1, characterized in that, In S1, the environmental information includes lane lines and obstacle information; the driver's biosignals include the driver's gaze coordinates, heart rate variability, and grip strength.

3. The intelligent headlight control method based on driving intention prediction according to claim 2, characterized in that, Specifically, S3 includes: S31 Spatiotemporal Graph Convolutional Network Construction: The neural network is a spatiotemporal graph convolutional network, constructing a spatiotemporal interaction graph containing the vehicle, surrounding vehicles, and lane lines. Vehicle nodes have feature vectors composed of their position coordinates, velocity components, and heading angles. Lane line nodes contain lane line curvature radius, lane type, and lane width feature parameters. When the centroid Euclidean distance between the vehicle and surrounding vehicles is less than 50 meters, undirected connections are established between vehicles. The vehicle always establishes undirected connections between itself and the centerline of its lane. Edge weights are calculated using a learnable function that comprehensively considers the absolute value of the speed difference between the vehicle and surrounding vehicles, as well as the Euclidean distance. Specifically, the weights include: Basic form: Used to quantify the dynamic interaction risk between the vehicle and surrounding vehicles, the function formula is: in, Let be the Euclidean distance between the centroids of vehicles i and j. The absolute value of the speed difference between the vehicle and surrounding vehicles. This is the distance attenuation coefficient; Extended form: Used to quantify the affiliation of a vehicle with its lane; the function formula is: in, It is the vertical distance from the vehicle's center of gravity to the center line of its lane. This is the distance attenuation coefficient. Lane type weights; Learnable form: Adaptive weight adjustment for complex traffic scenarios, the function formula is: in, As an activation function, it maps the output to the 0-1 range. It is a 2-layer neural network, and the input is the absolute value of the speed difference. Euclidean distance from the center of mass ; S32 Feature Extraction: The spatial topology of vehicle-vehicle and vehicle-road is modeled by spatiotemporal interaction graphs, and the topology is processed by a two-layer graph convolutional network to extract features, taking into account the changes in time series. S33 Constructing an LSTM-CRF Intent Prediction Model: The LSTM-CRF intent prediction model includes an LSTM layer and a CRF layer; The input to the LSTM-CRF intent prediction model is the feature sequence output by the spatiotemporal graph convolutional network, and the output is the driving intent label and probability distribution. S34 confidence calculation: Sampling is performed using the Monte Carlo Dropout mechanism to calculate the probability mean μ and standard deviation σ of each driving intention label. Low confidence scenarios are defined when the maximum probability mean is lower than the threshold or the standard deviation exceeds the threshold.

4. The intelligent headlight control method based on driving intention prediction according to claim 3, characterized in that, In S4, the dynamic control logic for light pattern achieves precise light pattern adjustment through a three-layer closed-loop mechanism: The first layer uses intent prediction based on spatiotemporal graph convolutional network to drive pre-illumination action, completing spot deflection before physical turning occurs, thus eliminating the brief illumination blind spot caused by mechanical delay in traditional schemes. The second layer integrates driver eye movement vector and heart rate variability data to adjust the light pattern to cope with sudden operational hesitation; The third layer establishes confidence-light pattern mapping rules, adopts wide-angle lighting mode to improve coverage in low-confidence scenarios, and uses ADAS-assisted flicker.

5. The intelligent headlight control method based on driving intention prediction according to claim 4, characterized in that, The calculation process for regional brightness adjustment in S4 is as follows: A vector projection algorithm is used to map the gaze point coordinates to road surface grid weight values. The brightness of the corresponding region is then adjusted using these road surface grid weight values. The calculation formula is as follows: in, The road surface grid weight value indicates the priority of brightness adjustment in the eye-tracking fixation area; , The coordinates for the driver's eye movement gaze correspond to the relative position on the road surface. , The coordinates of the center of the road surface grid; This is the heart rate variability coefficient, reflecting the driver's condition.

6. The intelligent headlight control method based on driving intention prediction according to claim 1, characterized in that, In S5, the wide-angle lighting mode has a horizontal illumination angle ≥120° and the brightness is increased to more than 80% of the maximum level.

7. The intelligent headlight control method based on driving intention prediction according to claim 1, characterized in that, The preprocessing of multi-source data in S2 includes time synchronization processing, outlier handling, data quality checking, and feature fusion, wherein: S21 Time Synchronization Processing: Linear interpolation is used to handle timing misalignment caused by differences in sampling frequencies between different sensors, ensuring data timing consistency; S22 Outlier Handling: Identify and remove outlier data, and fill in missing values ​​with the sliding window mean; S23 Data Quality Check: Check for missing values, outliers, and time series consistency. If the data meets the standards, proceed to the feature extraction step; otherwise, re-collect the data. S24 Feature Fusion: Converts multi-source data into high-dimensional feature vectors, and uses environmental feature weights, biological feature weights, and vehicle feature weights to weight and concatenate them to construct a spatiotemporal interaction graph.

8. The intelligent headlight control method based on driving intention prediction according to claim 1, characterized in that, The warning light pattern in S6 is generated by independent partitions within the main beam of the headlight and focused on the potential collision area.

9. A smart headlight control system based on driving intention prediction according to any one of claims 1-8, comprising: The data acquisition unit is used for multi-source data acquisition, including acquiring driver biosignals, environmental and vehicle dynamic information; The multi-source data is then preprocessed through the data fusion layer in the data acquisition unit. The processing unit has a built-in intent prediction module and confidence assessment module, which outputs intent prediction results and prediction uncertainty based on the collected data. The execution layer is connected to the vehicle deflection controller, modulator, and ADAS. The execution layer calls the light pattern dynamic control logic through the light pattern dynamic control module to generate multi-dimensional light pattern adjustment commands to dynamically adjust the headlight illumination direction and illumination range, and trigger warning light patterns in response to ADAS pre-collision signals.

10. The intelligent headlight control system based on driving intention prediction according to claim 9, characterized in that, The processing unit also includes a spatiotemporal graph construction module, which constructs a spatiotemporal interaction graph containing the vehicle, surrounding vehicles, and lane lines based on a spatiotemporal graph convolutional network.

Images