Active light compensation lamp glare suppression method

Abstract

The application discloses a kind of active light supplement lamp glare suppression methods, by obtaining the traffic scene perception data containing pre-analysis image and vehicle point cloud trajectory data, scene semantic analysis is carried out to pre-analysis image, target vehicle type, contour boundary and driver's cabin key area position are identified;Fusion vehicle point cloud trajectory data and driver's cabin key area position, generate driver's cabin three-dimensional motion trajectory prediction;Based on target vehicle type and driver's cabin three-dimensional motion trajectory prediction, call glare evaluation physical model to generate anti-glare control parameter set;Anti-glare control parameter set and driver's cabin three-dimensional motion trajectory prediction are input into multi-agent collaborative decision algorithm, generate pixel-level light spot control instruction, according to this drive multiple light supplement units to execute collaborative light supplement operation, so that light spot contains and driver's cabin three-dimensional motion trajectory prediction synchronous movement dynamic shadow area.The application realizes the effective glare suppression of target vehicle driver, while guaranteeing the illumination required for license plate recognition.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method for suppressing glare from active supplemental lighting. Background Technology

[0002] In intelligent transportation systems, roadside lighting equipment is widely used to illuminate vehicles in low-light environments to support visual perception tasks such as license plate recognition. Traditional lighting solutions typically employ fixed intensity or global adjustment based on simple photosensitive triggers. Their beams cover a wide area and have a uniform intensity distribution, making it difficult to prevent strong light from directly hitting the driver's cabin area, causing momentary glare and posing a safety hazard. To mitigate this problem, existing technologies attempt to detect vehicle positions using sensors and control the lighting to perform simple on / off switching or brightness adjustments, such as reducing overall brightness or adjusting the beam angle when a vehicle approaches. While these methods alleviate glare to some extent, their avoidance strategies are mostly based on static or delayed responses to the vehicle's real-time two-dimensional position in the image. They fail to fully consider the vehicle's continuous movement trajectory in three-dimensional space and the dynamic changes in the driver's viewpoint. While suppressing glare, they often struggle to maintain the optimal recognition illumination for the license plate area in dynamic traffic scenarios, resulting in a trade-off between the two core objectives of "ensuring safety" and "meeting recognition requirements." Summary of the Invention

[0003] In view of this, the purpose of this invention is to propose an active fill light glare suppression method, which integrates scene semantic parsing and vehicle motion trajectory prediction, and coordinates the control of multiple fill light units to generate a dynamic shadow area that moves synchronously with the cockpit, thereby achieving a dynamic balance between glare suppression and license plate recognition illumination.

[0004] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is: an active fill light glare suppression method, comprising: Acquire traffic scene perception data collected by roadside perception units. The traffic scene perception data includes at least pre-analyzed images with low illumination and high dynamic range and vehicle point cloud trajectory data. Scene semantic parsing is performed on the pre-analyzed image to identify the type of the target vehicle, its outline boundary, and the location of key areas of the cockpit in the image coordinate system; By integrating vehicle point cloud trajectory data with the location of key areas in the cockpit, a spatiotemporal trajectory prediction algorithm is used to generate a prediction of the target vehicle's three-dimensional motion trajectory in the cockpit within a future time window. Based on the type of the target vehicle and the prediction of the three-dimensional motion trajectory of the cockpit, a pre-set glare evaluation physical model is called to perform calculations to generate an anti-glare control parameter set that meets the license plate recognition illuminance constraints and has the best glare evaluation index. The anti-glare control parameter set includes at least the supplementary light intensity benchmark value and the dynamic shadow area definition. The anti-glare control parameter set and the cockpit's three-dimensional motion trajectory prediction are input into the multi-agent collaborative decision-making algorithm. The multi-agent collaborative decision-making algorithm assigns collaborative working modes and parameters to multiple supplementary lighting units deployed on the roadside, and generates pixel-level light spot control instructions containing dynamic shadow area location information. Based on pixel-level light spot control commands, multiple supplementary lighting units are driven to perform coordinated supplementary lighting operations. Among them, at least one supplementary lighting unit projects a light spot containing a dynamic shadow area that moves synchronously with the three-dimensional motion trajectory prediction of the cockpit, so as to suppress glare for the driver of the target vehicle.

[0005] In some embodiments, scene semantic parsing is performed on the pre-analyzed image to identify the type of the target vehicle, its contour boundaries, and the location of key areas of the cockpit in the image coordinate system, including: The pre-analyzed image is input into a pre-trained vehicle detection and segmentation neural network model, and the model outputs an initial parsing result containing the target vehicle semantic segmentation mask and the vehicle type probability vector. Based on semantic segmentation mask, the precise contour boundary of the target vehicle is calculated; Select the category with the highest probability value from the vehicle type probability vector as the initial type of the target vehicle; For a target vehicle whose initial type is a passenger car, visual saliency analysis based on gradient magnitude and direction statistics is performed in the windshield area corresponding to the semantic segmentation mask to locate the center point of the lower edge of the windshield, and the coordinates of the center point of the lower edge of the windshield are used as the location of the key area of ​​the cockpit. For target vehicles whose initial type is freight vehicle or passenger vehicle, the three-dimensional size of the vehicle is estimated based on semantic segmentation mask. Combined with the pre-set prior knowledge of the geometric relationship of vehicle key points, the estimated area of ​​the cockpit in the image is inferred through perspective projection geometry calculation, and the center point coordinates of the estimated area are used as the location of the key area of ​​the cockpit. By integrating precise contour boundaries, preliminary types, and key cockpit area locations, the final output of scene semantic parsing is generated.

[0006] In some embodiments, the three-dimensional dimensions of the vehicle are estimated based on a semantic segmentation mask, and combined with pre-defined prior knowledge of the geometric relationships of key vehicle points, the estimated area of ​​the cockpit in the image is inferred through perspective projection geometry calculations, including: Perform minimum bounding rectangle fitting on the semantic segmentation mask to obtain the two-dimensional bounding box of the target vehicle on the image plane and its orientation angle; Based on the preliminary type, retrieve the typical three-dimensional dimensions of the corresponding vehicle from the prior knowledge base of vehicle key point geometric relationships. The typical three-dimensional dimensions include at least the total length, total width, total height of the vehicle, and the longitudinal position of the front surface of the driver's cab relative to the front reference point of the vehicle. Based on the geometric relationship between the pinhole camera model and the two-dimensional bounding box, the approximate three-dimensional position and attitude of the target vehicle in the camera coordinate system are calculated by reverse calculation. Using the rough 3D position and attitude, typical 3D dimensions, and longitudinal position of the front surface of the cockpit as input, the projected quadrilateral of the front surface of the cockpit on the image plane is calculated through a 3D to 2D perspective projection transformation. The projected quadrilateral is shrunk inward by a preset safety boundary distance to form the estimated area of ​​the cockpit in the image.

[0007] In some embodiments, vehicle point cloud trajectory data and the location of key areas in the cockpit are fused, and a spatiotemporal trajectory prediction algorithm is used to generate a predicted three-dimensional motion trajectory of the target vehicle's cockpit within a future time window, including: Establish the coordinate transformation relationship between the image coordinate system where the pre-analyzed image is located and the world coordinate system where the vehicle point cloud trajectory data is located; Based on the coordinate transformation relationship, the location of the key areas of the cockpit is mapped from the image coordinate system to the world coordinate system to obtain the initial three-dimensional coordinates of the key points of the cockpit; The initial three-dimensional coordinates of the key points in the cockpit are spatiotemporally aligned and correlated with the vehicle point cloud trajectory data to form an extended trajectory sequence with the key points in the cockpit as the core. The extended trajectory sequence is input into an attention-based sequence prediction model, which learns the motion patterns and contextual dependencies in the trajectory sequence. Forward computation is performed using a sequence prediction model based on an attention mechanism, and the predicted 3D coordinates of key points in the cockpit at multiple consecutive future moments are output. The predicted three-dimensional coordinates of key points in the cockpit at multiple consecutive future moments are used to construct the predicted three-dimensional motion trajectory of the cockpit.

[0008] In some embodiments, the initial three-dimensional coordinates of key points in the cockpit are spatiotemporally aligned and correlated with vehicle point cloud trajectory data to form an extended trajectory sequence centered on the key points in the cockpit, including: The initial three-dimensional coordinates of the key points in the cockpit are used as the trajectory points at the current moment, and a timestamp and an instantaneous velocity vector from the vehicle point cloud trajectory data are added to them. The system retrieves data points of vehicle point cloud trajectory data within a historical time window. Based on the principles of spatial proximity and motion continuity, it associates historical data points with trajectory points at the current moment and filters out continuous historical trajectory point sets belonging to the same target vehicle. The continuous historical trajectory point set is smoothed and filtered to suppress measurement noise, and the smoothed three-dimensional coordinates and velocity vectors at each historical moment are calculated. The smoothed 3D coordinates and velocity vectors are concatenated with the trajectory point at the current moment and its additional instantaneous velocity vector in chronological order to form an extended trajectory sequence. Each element in the extended trajectory sequence contains 3D coordinates, velocity vectors and corresponding timestamps.

[0009] In some embodiments, the attention-based sequence prediction model is a spatiotemporal graph neural network with an encoder-decoder architecture; The extended trajectory sequence is input into an attention-based sequence prediction model, including: The extended trajectory sequence is constructed as a spatiotemporal graph, where each trajectory point is a graph node. The node features include its three-dimensional coordinates and velocity vector, and the edges between nodes are defined by temporal order and spatial proximity. The encoder part of the spatiotemporal graph neural network in the encoder-decoder architecture aggregates the features of nodes in the spatiotemporal graph through a multi-layer graph attention network and captures the long-range spatiotemporal dependencies between nodes to generate context-enhanced encoded features for each trajectory node. The decoder part of the spatiotemporal graph neural network in the encoder-decoder architecture, with context-enhanced coding features as a condition, gradually generates the predicted distribution of the three-dimensional coordinates of the cockpit key points at each time within the future time window through autoregressive or parallel prediction methods. By sampling or taking the expected value from the predicted distribution, the predicted three-dimensional coordinates of key points in the cockpit at multiple consecutive future moments are obtained.

[0010] In some embodiments, based on the type of the target vehicle and the predicted three-dimensional motion trajectory of the cockpit, a preset glare evaluation physical model is invoked for calculation to generate a set of anti-glare control parameters that meet the license plate recognition illuminance constraints and have the optimal glare evaluation index, including: Based on the type of the target vehicle, the corresponding reference optical parameters are retrieved from the preset parameter mapping table. The reference optical parameters include at least the typical reflectivity of the vehicle license plate and the spatial position offset of the license plate area relative to the vehicle body. By combining the three-dimensional motion trajectory prediction of the cockpit, the known three-dimensional position and orientation parameters of the roadside supplementary lighting unit, and the reference optical parameters, a glare evaluation function and a license plate area illuminance function are constructed, which include the adjustable parameters of the supplementary lighting unit as variables. The first constraint is that the value of the illuminance function of the license plate area is not lower than the preset recognition threshold. The optimization objective is to minimize the value of the glare evaluation function. Multi-objective optimization is performed in the feasible solution space of the adjustable parameters of the supplementary lighting unit. From the results of the multi-objective optimization solution, extract the adjustable parameter combination of the supplementary lighting unit that enables the glare evaluation function to obtain the global or local optimal value under the condition of satisfying the first constraint. The adjustable parameters of the supplementary lighting unit are encoded into a structured data set to generate an anti-glare control parameter set.

[0011] In some embodiments, the glare evaluation function is constructed based on a simplified human eye glare perception model, and includes a glare evaluation function and a license plate area illuminance function that incorporate adjustable parameters of the supplementary lighting unit as variables, including: Based on the prediction of the three-dimensional motion trajectory of the cockpit, the observation vector from the light source center of each supplementary lighting unit to the driver's eye is calculated at each sampling time in the future. Calculate the deviation angle between the observation vector and the driver's line-of-sight vector, where the driver's line-of-sight vector is preset to the vehicle's forward direction or dynamically estimated based on the scene. Based on the deviation angle, the light intensity value in the adjustable parameters of the supplementary lighting unit, and the ambient background brightness, the instantaneous glare value generated by each supplementary lighting unit for the driver at each sampling moment is calculated using a preset glare perception empirical formula. The instantaneous glare values ​​of all supplementary lighting units at all sampling times are weighted by time and accumulated in space, and the accumulated result is used as the output value of the glare evaluation function, wherein the time weighting coefficient decreases according to the prediction time. Based on the predicted three-dimensional motion trajectory of the cockpit and the spatial position offset of the license plate area in the reference optical parameters, the three-dimensional coordinates of the center of the license plate area at each future sampling time are calculated. Based on the three-dimensional coordinates of the license plate area center, the known three-dimensional position and orientation parameters of each supplementary lighting unit, and the light intensity value and beam spatial distribution model in the adjustable parameters of the supplementary lighting unit, the normal illuminance component generated by each supplementary lighting unit illuminating the center of the license plate area is calculated. The normal illuminance components generated by all supplementary lighting units are accumulated and multiplied by the typical reflectance of the vehicle license plate in the reference optical parameters and the preset imaging system gain coefficient to obtain the expected signal intensity value of the license plate area on the image sensor. This value is used as the output value of the illuminance function of the license plate area.

[0012] In some embodiments, the anti-glare control parameter set and the cockpit's three-dimensional motion trajectory prediction are input into a multi-agent cooperative decision-making algorithm. The multi-agent cooperative decision-making algorithm assigns cooperative working modes and parameters to multiple supplementary lighting units deployed on the roadside, and generates pixel-level spot control instructions containing dynamic shadow area location information, including: The baseline value of supplementary light intensity and the definition of dynamic shadow areas in the anti-glare control parameter set, as well as the prediction of the three-dimensional motion trajectory of the cockpit, are used as global state information for collaborative decision-making. Based on global state information and a predefined supplementary lighting unit performance model, a local observation and action space is constructed for each supplementary lighting unit. The action space includes at least the activation flag, working intensity coefficient, and light spot projection angle fine-tuning amount. A distributed reinforcement learning framework based on policy gradient is adopted, enabling the agent corresponding to each lighting unit to output action proposals in parallel based on its local observations; A coordinator module is introduced, which collects action proposals from all agents and arbitrates and corrects conflicting action proposals based on preset conflict resolution rules and global performance evaluation functions, generating consistent cooperative action decisions. The collaborative action decision assigns a final activation flag, working intensity coefficient, and spot projection angle fine-tuning amount to each lighting unit, and calculates the precise pixel coordinate sequence of the dynamic shadow area in the spot pattern to be generated for each activated lighting unit based on the definition of dynamic shadow area and the prediction of the three-dimensional motion trajectory of the cockpit. The final parameters assigned to each enabled fill light unit and its corresponding dynamic shadow region pixel coordinate sequence are encapsulated into an independent control instruction unit; Aggregate all control command units that enable the supplementary lighting unit to generate pixel-level spot control commands.

[0013] In some embodiments, the coordinator module collects action proposals from all agents and, based on preset conflict resolution rules and a global performance evaluation function, arbitrates and corrects conflicting action proposals to generate consistent cooperative action decisions, including: Identify the conflict types in the agent's action proposals. Conflict types include competition for lighting resources in the same road space area, and shading failure conflicts where the superposition of light spots from multiple supplementary lighting units may lead to the accidental illumination of dynamic shadow areas. In response to conflicts over lighting resources, the coordinator module evaluates the contribution of each competing agent to the global performance when it executes its action proposal individually, based on the global performance evaluation function. It prioritizes action proposals with higher contributions and reduces or rejects action proposals with lower contributions. In response to shading failure conflicts, the coordinator module checks the theoretical illuminance distribution of the dynamic shadow area under the superimposed light spot. If there is an area where the illuminance exceeds the preset shadow maintenance threshold, conflict resolution is triggered. Conflict resolution is achieved by iteratively adjusting the working intensity coefficient or the fine-tuning amount of the light spot projection angle in the action proposal of the relevant supplementary lighting unit until the theoretical illuminance of the dynamic shadow area under the superimposed light spot is lower than the shadow maintenance threshold, and the adjusted action proposal is used as the correction result. Integrate all the arbitrated and revised action proposals to form a consistent and collaborative action decision.

[0014] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: The present invention provides an active supplementary lighting glare suppression method. It acquires traffic scene perception data containing pre-analyzed images and vehicle point cloud trajectory data, performs scene semantic parsing on the pre-analyzed images to identify the target vehicle type, contour boundaries, and key cockpit area locations; fuses the vehicle point cloud trajectory data and the key cockpit area locations to generate a predicted three-dimensional motion trajectory of the cockpit; based on the target vehicle type and the predicted three-dimensional motion trajectory of the cockpit, it calls a glare assessment physical model to generate an anti-glare control parameter set; and inputs the anti-glare control parameter set and the predicted three-dimensional motion trajectory of the cockpit into a multi-agent collaborative decision-making algorithm to generate pixel-level spot control commands, thereby driving multiple supplementary lighting units to perform collaborative supplementary lighting operations, so that the spot contains a dynamic shadow area that moves synchronously with the predicted three-dimensional motion trajectory of the cockpit. The present invention achieves effective glare suppression for the driver of the target vehicle while ensuring the illumination required for license plate recognition. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of steps S101 to S106 of the method described in the specific implementation embodiment; Figure 2 This is a schematic diagram of steps S201 to S206 of the method described in the specific implementation embodiment; Figure 3 This is a schematic diagram of steps S301 to S305 of the method described in the specific implementation. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This embodiment provides a method for suppressing glare from an active fill light, including: S101. Obtain traffic scene perception data collected by the roadside perception unit. The traffic scene perception data shall include at least a low-light high dynamic range pre-analyzed image and vehicle point cloud trajectory data. S102. Perform scene semantic parsing on the pre-analyzed image to identify the type of the target vehicle, its outline boundary, and the position of the key areas of the cockpit in the image coordinate system. S103. By integrating vehicle point cloud trajectory data with the key areas of the cockpit, a spatiotemporal trajectory prediction algorithm is used to generate a prediction of the target vehicle's three-dimensional motion trajectory within a future time window. S104. Based on the type of the target vehicle and the prediction of the three-dimensional motion trajectory of the cockpit, the preset glare evaluation physical model is called to perform calculations to generate an anti-glare control parameter set that meets the license plate recognition illuminance constraints and has the best glare evaluation index. The anti-glare control parameter set includes at least the supplementary light intensity benchmark value and the dynamic shadow area definition. S105. Input the anti-glare control parameter set and the cockpit three-dimensional motion trajectory prediction into the multi-agent collaborative decision-making algorithm. The multi-agent collaborative decision-making algorithm assigns collaborative working modes and parameters to multiple supplementary lighting units deployed on the roadside, and generates pixel-level light spot control instructions containing dynamic shadow area location information. S106. Based on the pixel-level light spot control command, drive multiple supplementary lighting units to perform coordinated supplementary lighting operation, wherein at least one supplementary lighting unit projects a light spot containing a dynamic shadow area that moves synchronously with the three-dimensional motion trajectory prediction of the cockpit, so as to suppress glare for the driver of the target vehicle.

[0019] In step S101, a roadside perception unit is deployed on the side of the road to collect traffic scene information. This unit may consist of a visible light camera and a lidar sensor. The traffic scene perception data collected by this unit integrates information from different sensors. Specifically, the low-light, high dynamic range pre-analysis image is obtained by using a camera in low-light environments with multi-exposure fusion or in-sensor extended dynamic range technology to ensure that the image simultaneously preserves details of bright headlights and vehicles in dark areas. Vehicle point cloud trajectory data is generated by lidar scanning, providing a sequence of three-dimensional point sets on the vehicle surface changing over time, including position, speed, and contour information. This step achieves the synchronous acquisition and initial alignment of multi-dimensional traffic scene data.

[0020] In step S102, the scene semantic parsing of the pre-analyzed image can be completed using a deep learning-based image recognition model. This model can output the vehicle category label, pixel-level segmentation mask, and key point coordinates in the image. The identified target vehicle type is, for example, a sedan, truck, or bus; the contour boundary can be extracted by the segmentation mask, defining the precise shape of the vehicle in the image; the location of the key area of ​​the cockpit is determined based on the vehicle type by locating the center of the lower edge of the windshield or estimating the center of the cockpit projection area, and represented by pixel coordinates in the image coordinate system. This step extracts semantic and geometric information directly related to glare suppression from the visual data.

[0021] In step S103, the fusion process first establishes the transformation relationship between the image coordinate system and the world coordinate system, mapping the key areas of the cockpit to three-dimensional space. The spatiotemporal trajectory prediction algorithm receives the mapped three-dimensional position sequence and vehicle point cloud trajectory data, and infers future displacement by analyzing historical motion patterns. The generated three-dimensional motion trajectory prediction of the cockpit is represented as a continuous coordinate sequence of key points in the cockpit in three-dimensional space over a future period. This step combines two-dimensional visual positioning with three-dimensional motion perception, enabling the prediction of the future movement path of the driver's viewpoint area.

[0022] In step S104, the pre-set glare assessment physical model incorporates a simplified relationship describing light propagation, surface reflection, and human visual perception, used to quantify the glare levels generated by different supplementary lighting schemes. This model takes the target vehicle type and the predicted three-dimensional motion trajectory of the cockpit as input, and performs optimization calculations based on the minimum illumination constraints required for license plate recognition, outputting a set of anti-glare control parameters. In this parameter set, the supplementary lighting intensity benchmark defines the reference level of illumination intensity; the dynamic shadow area definition specifies the shape, size, or optical characteristic parameters of the local low-illuminance area that needs to be created within the light spot and moves with the cockpit. This step, through solving the physical model, yields the theoretical control parameters that minimize glare while meeting recognition requirements.

[0023] In step S105, the multi-agent cooperative decision-making algorithm uses the anti-glare control parameter set and the prediction of the cockpit's three-dimensional motion trajectory as global optimization objectives. It assigns appropriate operating modes and specific parameters to the multiple roadside lighting units, such as the on / off state, light intensity coefficient, and beam pointing angle of each unit. By coordinating the actions of each unit, the algorithm generates pixel-level spot control commands. These commands contain the precise pixel position information of the dynamic shadow area in the spot pattern to be projected by each lighting unit at each moment. This step decomposes the global optimization problem into a distributed control task and generates executable, high-precision control signals.

[0024] In step S106, according to pixel-level spot control commands, multiple supplementary lighting units deployed on the roadside are driven to perform coordinated supplementary lighting operations. Each supplementary lighting unit adjusts its optical output according to the commands, so that a dynamic shadow area that moves synchronously with the predicted three-dimensional motion trajectory of the cockpit is formed within the spot projected by at least one unit. That is, the position of this dynamic shadow area in three-dimensional space always corresponds to the predicted position of the cockpit. Through this coordinated projection, while ensuring sufficient illumination of the license plate area, the driver's viewing window area is kept continuously within the shadow of the spot, thereby effectively suppressing glare.

[0025] This embodiment proactively determines the cockpit's movement path by integrating visual semantic parsing and three-dimensional trajectory prediction. It also optimizes control parameters that include dynamic shadow areas using a physical model. Finally, through multi-agent collaborative decision-making, it drives the supplementary lighting unit to generate shadow light spots that move precisely and synchronously in space. This achieves a leap from passive overall dimming to active and precise avoidance, significantly reducing glare interference to the driver while ensuring machine vision recognition performance, and improving the safety and adaptability of the intelligent roadside lighting system.

[0026] In some embodiments, scene semantic parsing is performed on the pre-analyzed image to identify the type of the target vehicle, its contour boundaries, and the location of key areas of the cockpit in the image coordinate system, including: The pre-analyzed image is input into a pre-trained vehicle detection and segmentation neural network model, and the model outputs an initial parsing result containing the target vehicle semantic segmentation mask and the vehicle type probability vector. Based on semantic segmentation mask, the precise contour boundary of the target vehicle is calculated; Select the category with the highest probability value from the vehicle type probability vector as the initial type of the target vehicle; For a target vehicle whose initial type is a passenger car, visual saliency analysis based on gradient magnitude and direction statistics is performed in the windshield area corresponding to the semantic segmentation mask to locate the center point of the lower edge of the windshield, and the coordinates of the center point of the lower edge of the windshield are used as the location of the key area of ​​the cockpit. For target vehicles whose initial type is freight vehicle or passenger vehicle, the three-dimensional size of the vehicle is estimated based on semantic segmentation mask. Combined with the pre-set prior knowledge of the geometric relationship of vehicle key points, the estimated area of ​​the cockpit in the image is inferred through perspective projection geometry calculation, and the center point coordinates of the estimated area are used as the location of the key area of ​​the cockpit. By integrating precise contour boundaries, preliminary types, and key cockpit area locations, the final output of scene semantic parsing is generated.

[0027] In this embodiment, the pre-trained vehicle detection and segmentation neural network model is a computer vision model based on a deep convolutional neural network architecture, such as Mask R-CNN or a similar structure. This model is obtained through supervised training on a large-scale labeled image dataset containing various vehicle types, lighting conditions, and scenes. Each image in the training data is labeled with the vehicle's bounding box, category, and pixel-level segmentation mask, enabling it to simultaneously perform vehicle instance detection, pixel-level segmentation, and coarse type classification. After inputting the pre-analyzed image into the model, the model outputs an initial analytical result containing the target vehicle semantic segmentation mask and the vehicle type probability vector. The semantic segmentation mask is a binary image where the pixel value of the target vehicle is 1, and the background pixel value is 0.

[0028] The precise contour boundary of the target vehicle can be obtained by semantic segmentation mask calculation. This can be achieved by extracting the outer contour pixel chain of the mask image, for example, by applying an edge detection algorithm (such as the Canny operator) to the mask region, or by directly calculating the circumscribed polygon of the mask region. The resulting precise contour boundary is represented by a series of ordered pixel coordinate points, defining the precise shape of the vehicle in the image.

[0029] The vehicle type probability vector is a multi-dimensional vector, with each dimension corresponding to a confidence score of a preset vehicle type (such as passenger car, freight vehicle, and passenger vehicle). The category with the highest probability value is selected as the preliminary type of the target vehicle.

[0030] For target vehicles initially classified as passenger cars, the determination of the key areas of the cockpit focuses on the windshield region. Visual saliency analysis based on gradient magnitude and direction statistics first calculates the gradient magnitude and direction of each pixel within the windshield region using image gradient operators (such as the Sobel operator). Since the lower edge of the windshield typically appears as a horizontal or near-horizontal edge with a large gradient magnitude, the pixel locations where the gradient direction is concentrated in the horizontal direction can be statistically analyzed. Combined with the spatial distribution of gradient magnitude, the horizontal line with the strongest gradient response is located, and the coordinates of the center point of this line are determined as the center point of the lower edge of the windshield, serving as the location of the key areas of the cockpit.

[0031] For target vehicles initially categorized as freight or passenger vehicles, a geometric reasoning-based approach is employed due to the diverse structures of their cockpits and the potential for incompleteness in images caused by perspective distortion. A prior knowledge base of vehicle key point geometric relationships stores standard 3D dimensions (e.g., length, width, height) for different vehicle types, as well as the positional relationship of the cockpit relative to a reference point at the front of the vehicle. Based on the 2D projection information of the vehicle provided by a semantic segmentation mask, the approximate 3D pose and position of the vehicle in the scene can be deduced by establishing a geometric mapping relationship between the image coordinate system and the 3D world coordinate system (e.g., using a pinhole camera model). Combining this with the retrieved prior knowledge of the vehicle's 3D dimensions, the possible projection area of ​​the cockpit's front surface on the image plane can be calculated through 3D-to-2D perspective projection. The coordinates of the center point of this projection area are then used as the location of the key areas of the cockpit.

[0032] By integrating precise contour boundaries, preliminary types, and key cockpit area locations, the final output of scene semantic analysis is generated. This output is a structured data object that provides accurate input for subsequent trajectory prediction and glare assessment.

[0033] This embodiment refines the localization method for key areas of the cockpit, utilizing a pre-trained deep learning model to achieve precise pixel-level recognition of vehicles and components. By differentiating between passenger cars and large vehicles and employing differentiated technical approaches: for passenger cars with regular structures, direct analysis based on image gradient features is used for rapid and accurate localization; for large vehicles with complex structures, reasoning is performed using 3D dimension estimation and geometric prior knowledge, overcoming the limitations of purely visual analysis. This differentiated localization strategy improves the system's adaptability to different vehicle models and the accuracy of cockpit position estimation, laying a reliable perceptual foundation for generating high-precision 3D motion trajectory predictions for the cockpit.

[0034] In some embodiments, the three-dimensional dimensions of the vehicle are estimated based on a semantic segmentation mask, and combined with pre-defined prior knowledge of the geometric relationships of key vehicle points, the estimated area of ​​the cockpit in the image is inferred through perspective projection geometry calculations, including: Perform minimum bounding rectangle fitting on the semantic segmentation mask to obtain the two-dimensional bounding box of the target vehicle on the image plane and its orientation angle; Based on the preliminary type, retrieve the typical three-dimensional dimensions of the corresponding vehicle from the prior knowledge base of vehicle key point geometric relationships. The typical three-dimensional dimensions include at least the total length, total width, total height of the vehicle, and the longitudinal position of the front surface of the driver's cab relative to the front reference point of the vehicle. Based on the geometric relationship between the pinhole camera model and the two-dimensional bounding box, the approximate three-dimensional position and attitude of the target vehicle in the camera coordinate system are calculated by reverse calculation. Using the rough 3D position and attitude, typical 3D dimensions, and longitudinal position of the front surface of the cockpit as input, the projected quadrilateral of the front surface of the cockpit on the image plane is calculated through a 3D to 2D perspective projection transformation. The projected quadrilateral is shrunk inward by a preset safety boundary distance to form the estimated area of ​​the cockpit in the image.

[0035] In this embodiment, a minimum bounding rectangle fitting is performed on the semantic segmentation mask to find a rectangle with the smallest area that can completely enclose all foreground pixels in the mask. This rectangle is the two-dimensional bounding box of the target vehicle on the image plane, and is typically described by the coordinates of the rectangle's center point, width, height, and rotation angle (orientation angle) relative to the horizontal axis of the image. The orientation angle reflects the approximate driving direction of the vehicle in the image.

[0036] The vehicle key point geometric relationship prior knowledge base is a pre-built data storage structure that associates data on different vehicle types (such as heavy trucks and buses) with their typical three-dimensional dimensions and the relative positions of key components. Typical three-dimensional dimensions are retrieved from this knowledge base based on the initial vehicle type, including common overall length, width, and height values ​​for that type of vehicle, as well as the longitudinal distance between the front surface of the cab and the frontmost point of the vehicle (front reference point). This data is derived from statistical summarization of numerous physical measurements of similar vehicles or standard model data.

[0037] The pinhole camera model describes the linear relationship between the projection of a point in 3D space onto a 2D image plane. Its intrinsic parameters (such as focal length and principal point) are obtained in advance through camera calibration. Approximating a vehicle as a cuboid with known typical dimensions (length, width, and height), and assuming its bottom is in contact with the ground, a set of equations can be established based on the geometric information (position, size, and orientation) of its 2D bounding box in the image. These equations can then be used to solve for the cuboid's 3D position (such as center point coordinates) and attitude (such as yaw angle) in the camera coordinate system. This solution process is typically an optimization problem, aiming to achieve the best match between the projection of the 3D cuboid and the 2D bounding box in the image.

[0038] Calculating the projected quadrilateral of the front surface of the cockpit onto the image plane involves: constructing a rectangular plane representing the front surface of the cockpit in three-dimensional space based on the vehicle's three-dimensional attitude and dimensions. The position of this plane is determined by the vehicle's three-dimensional position and the longitudinal position of the front surface of the cockpit. Using a pinhole camera model, the four corner points of this three-dimensional rectangular plane are projected onto the image plane to obtain four corresponding pixel coordinate points. The quadrilateral formed by connecting these four points is the projected quadrilateral of the front surface of the cockpit.

[0039] The safety boundary distance is a preset pixel value or scale value used to compensate for geometric estimation errors, differences in individual vehicle dimensions, and uncertainties in perspective projection. The inward contraction operation is achieved by shifting each side of the quadrilateral inward by a specified distance, resulting in a smaller, more conservative quadrilateral region, which serves as the final image coordinate estimate for the critical areas of the cockpit.

[0040] This embodiment combines the 2D segmentation results from the image with the vehicle's 3D prior knowledge. By establishing a geometric calculation chain from 2D to 3D and then back to 2D, it overcomes the problem of incomplete visual features of the cockpit caused by occlusion, truncation, or perspective distortion in images of large vehicles. This method utilizes a standard pinhole camera model and optimization techniques to robustly estimate the vehicle's 3D pose and position from monocular images, thereby deriving the cockpit's projection area. The robustness of the estimation results is enhanced by introducing a safety boundary, providing reliable technical support for subsequent trajectory prediction and glare suppression control requiring high-precision cockpit position, particularly improving the processing capabilities for large vehicle types such as freight vehicles and passenger vehicles.

[0041] Please see Figure 2 In some embodiments, vehicle point cloud trajectory data and the location of key areas in the cockpit are fused, and a spatiotemporal trajectory prediction algorithm is used to generate a predicted three-dimensional motion trajectory of the target vehicle's cockpit within a future time window, including: S201. Establish the coordinate transformation relationship between the image coordinate system where the pre-analyzed image is located and the world coordinate system where the vehicle point cloud trajectory data is located; S202. Based on the coordinate transformation relationship, the location of the key area of ​​the cockpit is mapped from the image coordinate system to the world coordinate system to obtain the initial three-dimensional coordinates of the key points of the cockpit. S203. The initial three-dimensional coordinates of the key points in the cockpit are spatiotemporally aligned and correlated with the vehicle point cloud trajectory data to form an extended trajectory sequence with the key points in the cockpit as the core. S204. Input the extended trajectory sequence into the sequence prediction model based on the attention mechanism. The model learns the motion patterns and contextual dependencies in the trajectory sequence. S205. Forward calculation is performed using a sequence prediction model based on an attention mechanism to output the predicted three-dimensional coordinates of key points in the cockpit at multiple consecutive future moments. S206. Connect the predicted three-dimensional coordinates of key points in the cockpit at multiple consecutive future moments to form the predicted three-dimensional motion trajectory of the cockpit.

[0042] In step S201, the coordinate transformation relationship between the image coordinate system where the pre-analyzed image is located and the world coordinate system where the vehicle point cloud trajectory data is located is typically obtained through camera calibration and extrinsic parameter calibration. Camera calibration determines the relationship between the image coordinate system and the camera coordinate system, while extrinsic parameter calibration (through joint calibration or calibration objects) determines the rotation and translation relationship between the camera coordinate system and the world coordinate system. Combining both methods yields the projection matrix or transformation function from image pixel coordinates to world 3D coordinates.

[0043] In step S202, based on the coordinate transformation relationship established in step S201, the location of the key area of ​​the cockpit (a single image pixel coordinate point) is mapped to the world coordinate system. The mapping process uses the aforementioned projection matrix, combined with the camera imaging model (such as a pinhole model) and the assumed height of the key point in the cockpit (for example, assuming a typical value based on the vehicle type, or estimating it using point cloud data), to calculate the coordinates of the point in three-dimensional space, i.e., the initial three-dimensional coordinates of the key point in the cockpit.

[0044] In step S203, spatiotemporal alignment refers to unifying both to the same time reference and coordinate system; association aims to determine which data points in the vehicle point cloud trajectory data correspond to the same target vehicle, and to include the cockpit key points as a specific component of the vehicle in the trajectory sequence. Based on spatial proximity and motion continuity, the cockpit key points can be associated with the set of points in the point cloud trajectory representing the main body of the vehicle (such as the roof or rear), thereby forming an extended trajectory sequence centered on the cockpit key points, which includes not only the overall position of the vehicle but also the relative position information of the cockpit. This sequence is a three-dimensional coordinate point list ordered by time, and preferably, it may also include the instantaneous velocity or direction information of each point.

[0045] In step S204, the attention-based sequence prediction model is a deep learning model capable of processing sequence data. Its attention mechanism enables it to focus on important dependencies between different time steps in the sequence, thereby learning dynamic patterns of vehicle motion (such as acceleration, deceleration, and turning) and contextual rules (such as the influence of road direction). The model is trained using a large amount of historical vehicle trajectory data to learn the mapping relationship from historical sequences to future sequences.

[0046] In step S205, forward computation is performed using an attention-based sequence prediction model, that is, the trained model is used to process the input extended trajectory sequence. Based on the learned motion patterns, the model infers and outputs the predicted 3D coordinates of the cockpit key points at various times within a future continuous time period (i.e., the future time window). These predicted values ​​constitute the probability distribution or deterministic estimate of the future positions.

[0047] In step S206, the predicted three-dimensional coordinates of key cockpit points at multiple consecutive future moments are connected and linked together in chronological order to form a continuous spatial curve, i.e., the predicted three-dimensional motion trajectory of the cockpit. This trajectory prediction directly describes the future movement path of the cockpit in three-dimensional space.

[0048] This embodiment constructs a complete computational chain from multi-source sensing data to high-precision trajectory prediction. Through coordinate transformation, it fuses image semantic parsing results with LiDAR point cloud trajectories in three-dimensional space, focusing on the cockpit to construct an extended trajectory sequence. Then, it utilizes an attention-based sequence prediction model to learn the motion patterns. This embodiment elevates the prediction object from the "vehicle as a whole" to the "cockpit," significantly improving the targeting and accuracy of trajectory prediction for glare suppression applications. The generated three-dimensional motion trajectory prediction of the cockpit is a key input for subsequent forward-looking glare assessment and dynamic shadow area control, providing technical support for achieving a leap from reactive to active supplemental lighting control.

[0049] In some embodiments, the initial three-dimensional coordinates of key points in the cockpit are spatiotemporally aligned and correlated with vehicle point cloud trajectory data to form an extended trajectory sequence centered on the key points in the cockpit, including: The initial three-dimensional coordinates of the key points in the cockpit are used as the trajectory points at the current moment, and a timestamp and an instantaneous velocity vector from the vehicle point cloud trajectory data are added to them. The system retrieves data points of vehicle point cloud trajectory data within a historical time window. Based on the principles of spatial proximity and motion continuity, it associates historical data points with trajectory points at the current moment and filters out continuous historical trajectory point sets belonging to the same target vehicle. The continuous historical trajectory point set is smoothed and filtered to suppress measurement noise, and the smoothed three-dimensional coordinates and velocity vectors at each historical moment are calculated. The smoothed 3D coordinates and velocity vectors are concatenated with the trajectory point at the current moment and its additional instantaneous velocity vector in chronological order to form an extended trajectory sequence. Each element in the extended trajectory sequence contains 3D coordinates, velocity vectors and corresponding timestamps.

[0050] In this embodiment, the timestamps attached to the initial three-dimensional coordinates of the cockpit key points are derived from the acquisition time of the pre-analyzed image or a time source synchronized with it. The instantaneous velocity vector is obtained by analyzing the vehicle point cloud trajectory data. Specifically, it can extract the positional changes of consecutive frames representing the main body of the vehicle (such as the center of the roof) from the point cloud data, calculate the quotient of its displacement difference and the time interval, and use it as an estimate of the vehicle's instantaneous velocity. This velocity vector is then assigned to the cockpit key points at the same time.

[0051] Backtracking retrieval refers to searching for all data points in the vehicle point cloud trajectory data within a preset time period (historical time window) prior to the current moment. The spatial proximity criterion requires that the spatial distance between candidate historical data points and the current cockpit key points be less than a threshold; the motion continuity criterion requires that the motion direction of the candidate points be approximately consistent with the velocity vector direction of the current point. Based on these two criteria, a set of continuous historical trajectory points that are most likely to belong to the same vehicle motion sequence as the current cockpit key points can be selected from the historical point cloud data.

[0052] Smoothing filtering of continuous historical trajectory point sets aims to suppress inherent noise in lidar measurements. Kalman filtering, moving average filtering, or spline curve fitting methods can be employed. The process takes the original historical 3D coordinate sequence as input and outputs a smoothed 3D coordinate sequence. Based on the smoothed position changes, a more reliable velocity vector for each historical moment is recalculated.

[0053] The smoothed 3D coordinates and velocity vectors from each historical moment are concatenated with the current cockpit key point trajectory (including its initial 3D coordinates and additional instantaneous velocity vector) in chronological order. This concatenation operation forms a new, temporally coherent sequence, known as the extended trajectory sequence. Each element in this sequence is a data structure containing at least one 3D coordinate, one velocity vector, and one timestamp, fully describing the motion state of the cockpit key point from the past to the present.

[0054] This embodiment ensures the continuity of motion state information in the sequence by attaching instantaneous velocities to key points in the cockpit and associating them with filtered historical trajectories. By smoothing and filtering the historical trajectories, the interference of sensor noise on the prediction model input is effectively reduced. The resulting extended trajectory sequence not only contains accurate position history but also incorporates smoothed velocity information, providing cleaner and more dynamically rich input data for subsequent attention-based sequence prediction models. This lays a solid data foundation for generating accurate and smooth 3D cockpit motion trajectory predictions.

[0055] In some embodiments, the attention-based sequence prediction model is a spatiotemporal graph neural network with an encoder-decoder architecture; The extended trajectory sequence is input into an attention-based sequence prediction model, including: The extended trajectory sequence is constructed as a spatiotemporal graph, where each trajectory point is a graph node. The node features include its three-dimensional coordinates and velocity vector, and the edges between nodes are defined by temporal order and spatial proximity. The encoder part of the spatiotemporal graph neural network in the encoder-decoder architecture aggregates the features of nodes in the spatiotemporal graph through a multi-layer graph attention network and captures the long-range spatiotemporal dependencies between nodes to generate context-enhanced encoded features for each trajectory node. The decoder part of the spatiotemporal graph neural network in the encoder-decoder architecture, with context-enhanced coding features as a condition, gradually generates the predicted distribution of the three-dimensional coordinates of the cockpit key points at each time within the future time window through autoregressive or parallel prediction methods. By sampling or taking the expected value from the predicted distribution, the predicted three-dimensional coordinates of key points in the cockpit at multiple consecutive future moments are obtained.

[0056] In this embodiment, the encoder-decoder architecture spatiotemporal graph neural network is a deep learning model specifically designed for processing spatiotemporal sequence data. The encoder is responsible for compressing and extracting information from the input sequence, while the decoder generates future sequences based on the encoded information.

[0057] Constructing an extended trajectory sequence into a spatiotemporal graph means treating each trajectory point in the sequence as a node in the graph. The feature vector of each node contains the three-dimensional coordinates and velocity vector of that point. The edges between nodes are used to define the relationships between points, which are determined by temporal order (e.g., connecting points at adjacent time steps) and spatial proximity (e.g., connecting points with a spatial distance of less than a threshold), thus embedding both temporal and spatial structures in the graph.

[0058] The encoder operates through a multi-layered graph attention network. Each layer of the graph attention network allows each node to dynamically aggregate the feature information of its neighbors based on its relationship with them (connected by edges). By stacking multiple such networks, the encoder can capture the dependencies between nodes across long-term time and space, i.e., long-range spatiotemporal dependencies. Finally, the encoder outputs a feature representation for each trajectory node that incorporates global contextual information, called a context-enhanced encoded feature.

[0059] The decoder section uses context-enhanced encoded features as initial conditions or global context to generate future trajectories. The autoregressive approach involves the decoder sequentially generating the coordinates for each future moment, using the output of the previous moment as part of the input for the next moment. The parallel prediction approach generates the coordinate distribution for all future moments at once. Regardless of the approach used, the decoder outputs the predicted distribution of the 3D coordinates of key points in the cockpit at each moment within the future time window. This distribution is typically represented by Gaussian distribution parameters (mean and variance) or classification probabilities, reflecting the uncertainty of the prediction.

[0060] Sampling or taking the expected value from the prediction distribution is a step in obtaining deterministic prediction results. Taking the expected value (i.e., the mean) is the most common method, directly obtaining the predicted values ​​of the three-dimensional coordinates of key points in the cockpit at multiple consecutive future times; sampling can be used to assess the diversity of predictions or to perform probabilistic inference.

[0061] This embodiment constructs a spatiotemporal graph of the trajectory sequence and encodes it using a graph attention mechanism. The sequence prediction model can explicitly model the complex spatiotemporal interactions between trajectory points, thereby more accurately learning the group motion patterns and individual motion dynamics of the vehicle. The prediction distribution generated by the decoder further quantifies the uncertainty of the prediction. This embodiment significantly improves the accuracy and robustness of the cockpit's three-dimensional motion trajectory prediction, providing crucial and reliable technical support for subsequent glare assessment and dynamic shadow control that require high-precision forward-looking information.

[0062] Please see Figure 3 In some embodiments, based on the type of the target vehicle and the predicted three-dimensional motion trajectory of the cockpit, a pre-set glare evaluation physical model is invoked for calculation to generate a set of anti-glare control parameters that meet the license plate recognition illuminance constraints and have the optimal glare evaluation index, including: S301. Based on the type of the target vehicle, retrieve the corresponding reference optical parameters from the preset parameter mapping table. The reference optical parameters include at least the typical reflectivity of the vehicle license plate and the spatial position offset of the license plate area relative to the vehicle body. S302. Combining the three-dimensional motion trajectory prediction of the cockpit, the known three-dimensional position and orientation parameters of the roadside supplementary lighting unit, and the reference optical parameters, a glare evaluation function and a license plate area illuminance function are constructed, which include the adjustable parameters of the supplementary lighting unit as variables. S303. Taking the value of the illuminance function of the license plate area not lower than the preset recognition threshold as the first constraint condition, and minimizing the value of the glare evaluation function as the optimization objective, perform multi-objective optimization within the feasible solution space of the adjustable parameters of the supplementary lighting unit. S304. From the results of the multi-objective optimization solution, extract the adjustable parameter combination of the supplementary lighting unit that enables the glare evaluation function to obtain the global or local optimal value under the first constraint condition. S305. Encode the adjustable parameter combination of the supplementary lighting unit into a structured data set to generate an anti-glare control parameter set.

[0063] In step S301, the preset parameter mapping table is a pre-configured data table that establishes a mapping relationship between different vehicle types and their corresponding reference optical parameters. This mapping can be established through measurement statistics or by referencing industry standard data. For example, sampling measurements are performed on license plate materials of different vehicle types, and the average reflectance is statistically analyzed as the typical reflectance. The reference optical parameters include the typical reflectance of the vehicle license plate, which is obtained through measurement and statistical analysis of the reflectivity of various license plate materials; and the spatial offset of the license plate area relative to the vehicle body. This offset defines the fixed three-dimensional offset of the license plate center point relative to the vehicle's geometric center (or a specific reference point at the front of the vehicle) in the vehicle coordinate system, obtained by measuring typical vehicle models. This mapping table is pre-established through experimental measurements and data summarization.

[0064] In step S302, the glare evaluation function takes as input adjustable parameters of the supplementary lighting unit (such as the light intensity and beam angle of each unit) and outputs as a scalar value to comprehensively evaluate the degree of glare generated by the supplementary lighting scheme on the driver. Preferably, the glare evaluation function is constructed based on the principles of light propagation geometry and visual perception. Its calculation process involves determining the estimated position of the driver's eyes based on the predicted three-dimensional motion trajectory of the cockpit, calculating the illumination contribution of each supplementary lighting unit to that position, and evaluating its glare effect based on factors such as light intensity and incident angle. The license plate area illuminance function is used to calculate the expected illumination intensity received by the vehicle's license plate area under given supplementary lighting parameters. Preferably, the license plate area illuminance function is constructed based on an illuminance model, combining the position and beam characteristics of the supplementary lighting unit, the three-dimensional position of the license plate area (calculated from the vehicle trajectory and the spatial offset of the license plate), and the typical reflectivity of the license plate, to calculate the effective luminous flux of light illuminating the license plate surface and reflecting towards the camera.

[0065] In step S303, the preset recognition threshold is an illuminance value pre-set based on the minimum image brightness or signal-to-noise ratio required for reliable operation of the license plate recognition algorithm. The first constraint requires that the calculated result of the illuminance function of the license plate area must be greater than or equal to this recognition threshold. The optimization objective is to find a set of adjustable parameters for the supplementary lighting units, under the premise of satisfying this constraint, so that the value of the glare evaluation function is minimized. The feasible solution space of the adjustable parameters of the supplementary lighting units is limited by the technical specifications of each supplementary lighting unit, such as the intensity adjustment range and the angle adjustment range. The multi-objective optimization solution here is essentially a single-objective (glare minimization) optimization problem with constraints, and optimization algorithms such as gradient descent, genetic algorithms, or sequential quadratic programming can be used to search within the feasible solution space.

[0066] In step S304, the result of the multi-objective optimization solution may contain one or more locally optimal solutions that satisfy the constraints. The extraction process involves selecting the combination of adjustable parameters for the supplementary lighting unit that minimizes the glare evaluation function value from these solutions, as the final optimization result. This combination of adjustable parameters for the supplementary lighting unit includes the specific parameter settings for each supplementary lighting unit.

[0067] In step S305, the extracted adjustable parameters of the supplementary lighting units are structured and encoded, for example, organized according to a predetermined format (such as JSON, XML, or binary protocol) to form an anti-glare control parameter set. This parameter set includes at least a supplementary lighting intensity reference value (such as the light intensity setting value of each unit) and a dynamic shadow area definition (such as the shape and size parameters of the shadow, the specific location of which is dynamically determined by trajectory prediction).

[0068] This embodiment introduces physical properties into the model by establishing a mapping between vehicle type and optical parameters. By constructing two mutually constraining physical functions—glare and illuminance—the subjective goals of "glare suppression" and the objective goal of "recognition assurance" are transformed into a computable optimization problem. Finally, through constrained optimization, the theoretically optimal supplementary lighting parameters that minimize glare while ensuring recognition performance are automatically derived. This embodiment achieves a fundamental shift from empirical, qualitative adjustment to model-based, quantitative optimization, providing a crucial computational framework for generating scientific and precise anti-glare control parameters.

[0069] In some embodiments, the glare evaluation function is constructed based on a simplified human eye glare perception model, and includes a glare evaluation function and a license plate area illuminance function that incorporate adjustable parameters of the supplementary lighting unit as variables, including: Based on the prediction of the three-dimensional motion trajectory of the cockpit, the observation vector from the light source center of each supplementary lighting unit to the driver's eye is calculated at each sampling time in the future. Calculate the deviation angle between the observation vector and the driver's line-of-sight vector, where the driver's line-of-sight vector is preset to the vehicle's forward direction or dynamically estimated based on the scene. Based on the deviation angle, the light intensity value in the adjustable parameters of the supplementary lighting unit, and the ambient background brightness, the instantaneous glare value generated by each supplementary lighting unit for the driver at each sampling moment is calculated using a preset glare perception empirical formula. The instantaneous glare values ​​of all supplementary lighting units at all sampling times are weighted by time and accumulated in space, and the accumulated result is used as the output value of the glare evaluation function, wherein the time weighting coefficient decreases according to the prediction time. Based on the predicted three-dimensional motion trajectory of the cockpit and the spatial position offset of the license plate area in the reference optical parameters, the three-dimensional coordinates of the center of the license plate area at each future sampling time are calculated. Based on the three-dimensional coordinates of the license plate area center, the known three-dimensional position and orientation parameters of each supplementary lighting unit, and the light intensity value and beam spatial distribution model in the adjustable parameters of the supplementary lighting unit, the normal illuminance component generated by each supplementary lighting unit illuminating the center of the license plate area is calculated. The normal illuminance components generated by all supplementary lighting units are accumulated and multiplied by the typical reflectance of the vehicle license plate in the reference optical parameters and the preset imaging system gain coefficient to obtain the expected signal intensity value of the license plate area on the image sensor. This value is used as the output value of the illuminance function of the license plate area.

[0070] In this embodiment, the simplified human eye glare perception model is a mathematical model that maps the physical parameters of the light source to the degree of subjective discomfort. Calculating the observation vector based on the predicted three-dimensional motion trajectory of the cockpit means that for each future sampling moment, based on the predicted three-dimensional coordinates of key points in the cockpit (representing the approximate position of the driver's eyes) and the three-dimensional coordinates of the center of each roadside supplementary lighting unit, a vector pointing from the eye position to the center of the light source is calculated.

[0071] The driver's gaze direction vector is used to characterize the driver's primary gaze direction. In the absence of precise eye tracking, it can be preset to the vehicle's forward direction, which can be obtained from the velocity vector extracted from the vehicle's point cloud trajectory data; or it can be dynamically estimated based on the scene, for example, assuming the driver is looking at the center area of ​​the road ahead.

[0072] The deviation angle is the angle between the observation vector and the driver's line-of-sight vector, reflecting the degree of deviation of the light source relative to the driver's line of sight. The preset glare perception empirical formula is based on mathematical relationships summarized from visual science experiments, such as the glare evaluation method of the International Commission on Illumination (CIE) or similar simplified models. The glare perception empirical formula takes the deviation angle, the luminous intensity of the supplementary lighting unit, and the ambient background brightness as inputs, and outputs a quantified instantaneous glare value. The larger this value, the stronger the glare sensation felt by the light source at that moment.

[0073] The time-weighted coefficient is used to adjust the contribution of instantaneous glare values ​​at different future times to the overall evaluation. It is typically set to decrease as the prediction time point becomes further away, for example, using an exponential decay function to reflect higher confidence in recent predictions. Spatial accumulation refers to summing the contributions generated by all supplementary lighting units. The output value of the glare evaluation function is the total glare evaluation value after time weighting and spatial accumulation.

[0074] The calculation of the three-dimensional coordinates of the license plate area center at each future sampling time is achieved by superimposing the coordinates predicted from the three-dimensional motion trajectory of the cockpit onto the spatial offset of the license plate area relative to the vehicle body in the reference optical parameters. This offset is a fixed three-dimensional vector, and adding it to the cockpit coordinates yields the coordinates of the license plate area center.

[0075] The spatial distribution model of the supplementary lighting unit describes the distribution characteristics of the light intensity emitted by the unit in different directions in space, such as the Lambertian distribution or Gaussian distribution model. Calculating the normal illuminance component generated by each supplementary lighting unit illuminating the center of the license plate area requires using illuminance calculation formulas (such as the inverse square law and the cosine law) based on this spatial distribution model, the light intensity value of the supplementary lighting unit, and the position and orientation of the center of the license plate area relative to the supplementary lighting unit.

[0076] The imaging system gain coefficient is a preset scaling factor used to convert theoretical illuminance values ​​into expected signal intensity values ​​on the image sensor. This coefficient incorporates imaging link factors such as camera sensitivity and lens transmittance. The output value of the license plate area illuminance function, i.e., the expected signal intensity value, reflects the brightness level of the license plate area in the final image under given illumination parameters.

[0077] This embodiment integrates cockpit trajectory, vehicle geometry, supplementary lighting unit characteristics, and environmental parameters into specific physical calculations to construct a quantifiable glare evaluation function and a license plate area illuminance function. The glare evaluation function simulates the dynamic perception process of the human eye towards multiple moving light sources, while the license plate area illuminance function accurately calculates the imaging brightness of the license plate area. This allows the system to automatically optimize through calculation, finding a scientific supplementary lighting solution in a complex multivariate space that both meets recognition requirements and minimizes driver discomfort, thus supporting the realization of intelligent and refined glare suppression.

[0078] In some embodiments, the anti-glare control parameter set and the cockpit's three-dimensional motion trajectory prediction are input into a multi-agent cooperative decision-making algorithm. The multi-agent cooperative decision-making algorithm assigns cooperative working modes and parameters to multiple supplementary lighting units deployed on the roadside, and generates pixel-level spot control instructions containing dynamic shadow area location information, including: The baseline value of supplementary light intensity and the definition of dynamic shadow areas in the anti-glare control parameter set, as well as the prediction of the three-dimensional motion trajectory of the cockpit, are used as global state information for collaborative decision-making. Based on global state information and a predefined supplementary lighting unit performance model, a local observation and action space is constructed for each supplementary lighting unit. The action space includes at least the activation flag, working intensity coefficient, and light spot projection angle fine-tuning amount. A distributed reinforcement learning framework based on policy gradient is adopted, enabling the agent corresponding to each lighting unit to output action proposals in parallel based on its local observations; A coordinator module is introduced, which collects action proposals from all agents and arbitrates and corrects conflicting action proposals based on preset conflict resolution rules and global performance evaluation functions, generating consistent cooperative action decisions. The collaborative action decision assigns a final activation flag, working intensity coefficient, and spot projection angle fine-tuning amount to each lighting unit, and calculates the precise pixel coordinate sequence of the dynamic shadow area in the spot pattern to be generated for each activated lighting unit based on the definition of dynamic shadow area and the prediction of the three-dimensional motion trajectory of the cockpit. The final parameters assigned to each enabled fill light unit and its corresponding dynamic shadow region pixel coordinate sequence are encapsulated into an independent control instruction unit; Aggregate all control command units that enable the supplementary lighting unit to generate pixel-level spot control commands.

[0079] In this embodiment, the global state information for collaborative decision-making integrates the results of upstream optimization calculations (anti-glare control parameter set) and prediction information (cockpit 3D motion trajectory prediction), providing a unified decision-making basis for the collaboration of all lighting units. A predefined lighting unit performance model describes the technical characteristics of each lighting unit, such as its maximum and minimum light intensity, beam angle range, response speed, and spatial resolution (i.e., controllable pixel granularity). This model is pre-established through equipment calibration and specification parameters.

[0080] The local observation constructed for each supplementary lighting unit includes some global state information that the unit can perceive (such as local trajectory predictions related to it) and its own state; the action space defines the range of actions that the agent can perform, the enable flag is a Boolean value, the working intensity coefficient is a real number between 0 and 1, and the spot projection angle fine-tuning amount is a small angle adjustment value.

[0081] The policy gradient-based distributed reinforcement learning framework is a machine learning method in which each supplementary unit corresponds to an agent, and each agent has a policy network. This framework trains each agent to learn to output action proposals beneficial to the global goal based on local observations by continuously trying actions in simulated or historical traffic scenarios, receiving environmental feedback (rewards), and updating its policy network using a policy gradient algorithm. After training, during the deployment phase, each agent outputs action proposals in parallel based on real-time local observations.

[0082] The coordinator module is a central processing unit responsible for integrating action proposals from all distributed agents. Conflict resolution rules are used to determine whether there are contradictions or mutually destructive effects between action proposals from different agents. The global performance evaluation function is used to quantitatively evaluate the overall effect of a set of cooperative actions, with the design goal of minimizing glare and satisfying illuminance constraints. Based on these rules and functions, the coordinator module adjusts or rejects conflicting action proposals, ultimately outputting a cooperative action decision that is consistently executed by all supplementary lighting units.

[0083] Based on the definition of dynamic shadow areas (such as shape and size) and the prediction of the three-dimensional motion trajectory of the cockpit, the precise pixel coordinate sequence of the dynamic shadow areas is calculated. Specifically, the three-dimensional coordinates of the trajectory prediction are mapped onto the image plane corresponding to the light spot projected by the unit through the projection model of the supplementary lighting unit itself (considering its position, orientation and fine-tuned angle), so as to obtain the pixel coordinates of the center point of the shadow area at each moment. Then, the pixel coordinate sequence of the contour of the area is calculated according to the definition of the shadow area.

[0084] The control command unit is a data structure that encapsulates all the information required for a lighting unit to perform coordinated actions, including its activation flag, final working intensity coefficient, fine-tuning amount of the light spot projection angle, and dynamic shadow area pixel coordinate sequence. The pixel-level light spot control command is aggregated from the control command units of all enabled lighting units and is the final command that drives the lighting units to perform coordinated lighting operations.

[0085] This embodiment uses a distributed reinforcement learning framework to enable each lighting unit to make autonomous decisions based on local information. Then, a central coordinator resolves conflicts and optimizes the whole system, forming a highly efficient collaborative mechanism of "distributed proposal and centralized arbitration". This mechanism can flexibly cope with complex road geometry and vehicle motion, and transform the theoretical parameters of global optimization into precise control commands that can be executed by each lighting unit and contain pixel-level dynamic shadow position information. Thus, at the physical level, multiple lighting units work together as a whole to generate and maintain a dynamic shadow area that moves with the cockpit.

[0086] In some embodiments, the coordinator module collects action proposals from all agents and, based on preset conflict resolution rules and a global performance evaluation function, arbitrates and corrects conflicting action proposals to generate consistent cooperative action decisions, including: Identify the conflict types in the agent's action proposals. Conflict types include competition for lighting resources in the same road space area, and shading failure conflicts where the superposition of light spots from multiple supplementary lighting units may lead to the accidental illumination of dynamic shadow areas. In response to conflicts over lighting resources, the coordinator module evaluates the contribution of each competing agent to the global performance when it executes its action proposal individually, based on the global performance evaluation function. It prioritizes action proposals with higher contributions and reduces or rejects action proposals with lower contributions. In response to shading failure conflicts, the coordinator module checks the theoretical illuminance distribution of the dynamic shadow area under the superimposed light spot. If there is an area where the illuminance exceeds the preset shadow maintenance threshold, conflict resolution is triggered. Conflict resolution is achieved by iteratively adjusting the working intensity coefficient or the fine-tuning amount of the light spot projection angle in the action proposal of the relevant supplementary lighting unit until the theoretical illuminance of the dynamic shadow area under the superimposed light spot is lower than the shadow maintenance threshold, and the adjusted action proposal is used as the correction result. Integrate all the arbitrated and revised action proposals to form a consistent and collaborative action decision.

[0087] In this embodiment, the lighting resource competition conflict refers to the situation where the actions of multiple supplementary lighting units cause their light spots to be excessively concentrated in the same area of ​​the road, resulting in excessively high illumination in that area while other areas may be insufficiently illuminated. This not only wastes energy but may also cause secondary glare due to localized overbrightness. The shading failure conflict refers to the situation where, when the light spots of multiple supplementary lighting units are superimposed in space, the dynamic shadow area originally created by one unit may be partially or completely illuminated by the light spots of other units, causing the shadow effect to fail and unable to effectively shield the driver.

[0088] To address conflicts over lighting resources, the coordinator module evaluates contributions based on a global performance assessment function. Specifically, contribution can be quantified by calculating the global performance improvement (such as glare reduction or illuminance compliance) that would result from executing only the action proposal of the competing agent (while other units remain in a default or disabled state). The coordinator module compares the contributions of competing agents, prioritizing action proposals with higher contributions. For action proposals with lower contributions, adjustments are made by reducing their workload coefficient or directly rejecting their activation flag.

[0089] The preset shadow maintenance threshold is a key upper limit of illumination intensity used to define the maximum permissible illuminance that needs to be maintained in dynamic shadow areas. This threshold is preset based on the brightness level required to ensure that the driver does not experience significant glare. Checking the theoretical illuminance distribution of dynamic shadow areas under superimposed light spots refers to calculating the total illuminance generated by the combined illumination of all relevant units at each location point in the dynamic shadow area, based on the action proposals (light intensity, angle) of each relevant supplementary lighting unit and using a light propagation model.

[0090] The iterative adjustment used for conflict resolution is an optimization process. The coordinator module sequentially or simultaneously fine-tunes the working intensity coefficient (reducing light intensity) or the spot projection angle (shifting the spot center away from the shadow area) in the relevant supplementary lighting unit action proposals. After each adjustment, the superimposed illuminance distribution is recalculated until the theoretical illuminance at all locations within the dynamic shadow area is lower than the shadow maintenance threshold. The adjustment process can employ simple optimization strategies such as gradient descent and binary search.

[0091] Integrating all arbitrated and revised action proposals means bringing together the final action settings of all agents that have resolved conflicts and are no longer contradictory, forming a complete scheme that is internally consistent and can be executed collaboratively, i.e., consistent collaborative action decision-making.

[0092] This embodiment addresses two fundamental practical problems inherent in multi-lighting unit collaboration: illumination competition and occlusion interference. It provides a specific and operable conflict detection and resolution mechanism. By quantifying contribution levels for resource competition arbitration, it ensures efficient utilization of limited light energy. Furthermore, it resolves occlusion failures through iterative adjustments based on a physical illuminance model, fundamentally guaranteeing the generation quality of dynamic shadow areas. This embodiment enables the multi-agent collaborative decision-making algorithm not only to generate theoretical collaborative solutions but also to effectively handle physical constraints and conflicts in actual deployments. It ensures the physical feasibility and effectiveness of the final pixel-level spot control commands, achieving precise and robust collaboration among multiple lighting units.

[0093] By adopting the above technical solutions, this invention differs from existing technologies and possesses the following beneficial effects: By integrating multimodal perception data and scene semantic analysis, high-precision positioning of key areas in the target vehicle's cockpit is achieved; further, by combining vehicle point cloud trajectory data and utilizing a spatiotemporal trajectory prediction algorithm, a three-dimensional motion trajectory prediction of the cockpit is generated proactively, providing temporal and spatial basis for active glare suppression. The glare evaluation function and license plate area illuminance function constructed based on a physical model transform the subjective glare suppression and objective license plate recognition illuminance assurance objectives into a computable optimization problem. Through constraint optimization, the theoretically optimal anti-glare control parameter set is automatically generated. Using a multi-agent collaborative decision-making algorithm, the global optimization objective is decomposed and the actions of multiple supplementary lighting units are coordinated, generating pixel-level light spot control commands containing precise location information of dynamic shadow areas. Ultimately, this drives the supplementary lighting units to perform collaborative operations, ensuring that the light spot contains a dynamic shadow area that moves synchronously with the predicted cockpit trajectory. The above-mentioned technical solution realizes a fundamental shift from passive, overall supplementary lighting avoidance to active, precise, and collaborative glare suppression. While ensuring the illumination required by the license plate recognition system, it significantly reduces glare interference to the driver, effectively improving road safety and driving comfort in intelligent traffic monitoring scenarios.

[0094] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0095] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0096] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for suppressing glare from an active fill light, characterized in that, include: Acquire traffic scene perception data collected by the roadside perception unit, wherein the traffic scene perception data includes at least a low-light high dynamic range pre-analyzed image and vehicle point cloud trajectory data. Scene semantic parsing is performed on the pre-analyzed image to identify the type of the target vehicle, its outline boundary, and the position of the key areas of the cockpit in the image coordinate system; By integrating the vehicle point cloud trajectory data with the key area location of the cockpit, a spatiotemporal trajectory prediction algorithm is used to generate a prediction of the target vehicle's three-dimensional motion trajectory within a future time window. Based on the type of the target vehicle and the predicted three-dimensional motion trajectory of the cockpit, a preset glare evaluation physical model is invoked for calculation to generate an anti-glare control parameter set that meets the license plate recognition illuminance constraints and has the optimal glare evaluation index. The anti-glare control parameter set includes at least a supplementary light intensity benchmark value and a dynamic shadow area definition. The anti-glare control parameter set and the cockpit three-dimensional motion trajectory prediction are input into the multi-agent collaborative decision-making algorithm. The multi-agent collaborative decision-making algorithm assigns collaborative working modes and parameters to multiple supplementary lighting units deployed on the roadside, and generates pixel-level light spot control instructions containing dynamic shadow area location information. According to the pixel-level light spot control command, the multiple supplementary lighting units are driven to perform coordinated supplementary lighting operations. Among them, the light spot projected by at least one supplementary lighting unit contains a dynamic shadow area that moves synchronously with the three-dimensional motion trajectory prediction of the cockpit, so as to suppress glare for the driver of the target vehicle.

2. The active fill light glare suppression method according to claim 1, characterized in that, Scene semantic parsing is performed on the pre-analyzed image to identify the type of the target vehicle, its outline boundary, and the location of key areas of the cockpit in the image coordinate system, including: The pre-analyzed image is input into a pre-trained vehicle detection and segmentation neural network model, and the model outputs an initial parsing result containing a semantic segmentation mask for the target vehicle and a probability vector for the vehicle type. Based on the semantic segmentation mask, the precise contour boundary of the target vehicle is calculated; Select the category with the highest probability value from the vehicle type probability vector as the initial type of the target vehicle; For the target vehicle whose initial type is a passenger car, visual saliency analysis based on gradient magnitude and direction statistics is performed in the windshield area corresponding to the semantic segmentation mask to locate the center point of the lower edge of the windshield, and the coordinates of the center point of the lower edge of the windshield are used as the location of the key area of ​​the cockpit. For the target vehicle whose initial type is a freight vehicle or a passenger vehicle, the three-dimensional size of the vehicle is estimated based on the semantic segmentation mask. Combined with the preset prior knowledge of the geometric relationship of the vehicle's key points, the estimated area of ​​the cockpit in the image is inferred through perspective projection geometry calculation, and the center point coordinates of the estimated area are used as the location of the key area of ​​the cockpit. By integrating the precise contour boundaries, the preliminary type, and the locations of key areas in the cockpit, the final output of scene semantic parsing is generated.

3. The active fill light glare suppression method according to claim 2, characterized in that, Based on the semantic segmentation mask, the three-dimensional dimensions of the vehicle are estimated. Combined with pre-defined prior knowledge of the geometric relationships of key vehicle points, the estimated area of ​​the cockpit in the image is inferred through perspective projection geometry calculations, including: Perform minimum bounding rectangle fitting on the semantic segmentation mask to obtain the two-dimensional bounding box of the target vehicle on the image plane and its orientation angle; Based on the preliminary type, typical three-dimensional dimensions of the corresponding vehicle are retrieved from the prior knowledge base of geometric relationships of key vehicle points. The typical three-dimensional dimensions include at least the total length, total width, total height of the vehicle, and the longitudinal position of the front surface of the driver's cab relative to the front reference point of the vehicle. Based on the geometric relationship between the pinhole camera model and the two-dimensional bounding box, the approximate three-dimensional position and attitude of the target vehicle in the camera coordinate system are calculated in reverse. Using the rough three-dimensional position and attitude, the typical three-dimensional dimensions, and the longitudinal position of the front surface of the cockpit as input, the projected quadrilateral of the front surface of the cockpit on the image plane is calculated through a three-dimensional to two-dimensional perspective projection transformation. The projected quadrilateral is shrunk inward by a preset safety boundary distance to form the estimated area of ​​the cockpit in the image.

4. The active fill light glare suppression method according to claim 1, characterized in that, By fusing the vehicle point cloud trajectory data with the key area locations of the cockpit, and using a spatiotemporal trajectory prediction algorithm, a three-dimensional motion trajectory prediction of the target vehicle's cockpit within a future time window is generated, including: Establish the coordinate transformation relationship between the image coordinate system where the pre-analyzed image is located and the world coordinate system where the vehicle point cloud trajectory data is located; Based on the coordinate transformation relationship, the location of the key area of ​​the cockpit is mapped from the image coordinate system to the world coordinate system to obtain the initial three-dimensional coordinates of the key points of the cockpit; The initial three-dimensional coordinates of the key points in the cockpit are spatiotemporally aligned and correlated with the vehicle point cloud trajectory data to form an extended trajectory sequence with the key points in the cockpit as the core. The extended trajectory sequence is input into an attention-based sequence prediction model, which learns the motion patterns and contextual dependencies in the trajectory sequence. The forward calculation is performed by the attention-based sequence prediction model to output the predicted three-dimensional coordinates of key points in the cockpit at multiple consecutive future moments. The predicted three-dimensional coordinates of key points in the cockpit at multiple consecutive future moments are used to construct the predicted three-dimensional motion trajectory of the cockpit.

5. The active fill light glare suppression method according to claim 4, characterized in that, The initial three-dimensional coordinates of the cockpit key points are spatiotemporally aligned and correlated with the vehicle point cloud trajectory data to form an extended trajectory sequence centered on the cockpit key points, including: The initial three-dimensional coordinates of the key points in the cockpit are used as the trajectory points at the current moment, and a timestamp and an instantaneous velocity vector from the vehicle point cloud trajectory data are added to them. The system retrieves data points of the vehicle point cloud trajectory data within a historical time window, and associates the historical data points with the trajectory points at the current moment based on the criteria of spatial proximity and motion continuity, thereby filtering out a set of continuous historical trajectory points belonging to the same target vehicle. The continuous historical trajectory point set is subjected to smoothing filtering to suppress measurement noise, and the smoothed three-dimensional coordinates and velocity vectors at each historical moment are calculated. The smoothed three-dimensional coordinates and velocity vectors are concatenated with the trajectory point at the current moment and its additional instantaneous velocity vector in chronological order to form the extended trajectory sequence. Each element in the extended trajectory sequence contains three-dimensional coordinates, a velocity vector, and a corresponding timestamp.

6. The active fill light glare suppression method according to claim 4, characterized in that, The attention-based sequence prediction model is a spatiotemporal graph neural network with an encoder-decoder architecture; The extended trajectory sequence is input into an attention-based sequence prediction model, including: The extended trajectory sequence is constructed as a spatiotemporal graph, where each trajectory point is a graph node, the node features include its three-dimensional coordinates and velocity vector, and the edges between nodes are defined by temporal order and spatial proximity. The encoder part of the spatiotemporal graph neural network in the encoder-decoder architecture aggregates the features of nodes in the spatiotemporal graph through a multi-layer graph attention network, captures the long-range spatiotemporal dependencies between nodes, and generates context-enhanced encoded features for each trajectory node. The decoder part of the spatiotemporal graph neural network of the encoder-decoder architecture, using the context-enhanced coding features as a condition, gradually generates the predicted distribution of the three-dimensional coordinates of the cockpit key points at each time within the future time window through autoregression or parallel prediction. By sampling or taking the expected value from the predicted distribution, the predicted three-dimensional coordinates of the key points in the cockpit at multiple consecutive future moments are obtained.

7. The active fill light glare suppression method according to claim 1, characterized in that, Based on the type of the target vehicle and the predicted three-dimensional motion trajectory of the cockpit, a pre-set glare evaluation physical model is invoked for calculation to generate a set of anti-glare control parameters that meet the license plate recognition illuminance constraints and have the optimal glare evaluation index, including: Based on the type of the target vehicle, the corresponding reference optical parameters are retrieved from a preset parameter mapping table. The reference optical parameters include at least the typical reflectivity of the vehicle license plate and the spatial position offset of the license plate area relative to the vehicle body. Combining the predicted three-dimensional motion trajectory of the cockpit, the known three-dimensional position and orientation parameters of the roadside supplementary lighting unit, and the reference optical parameters, a glare evaluation function and a license plate area illuminance function are constructed, which include the adjustable parameters of the supplementary lighting unit as variables. The first constraint is that the value of the illuminance function of the license plate area is not lower than the preset recognition threshold. The optimization objective is to minimize the value of the glare evaluation function. Multi-objective optimization is performed in the feasible solution space of the adjustable parameters of the supplementary lighting unit. From the results of the multi-objective optimization solution, extract the adjustable parameter combination of the supplementary lighting unit that enables the glare evaluation function to obtain a global or local optimal value under the first constraint condition; The adjustable parameters of the supplementary lighting unit are combined and encoded into a structured data set to generate the anti-glare control parameter set.

8. The active fill light glare suppression method according to claim 7, characterized in that, The glare evaluation function is constructed based on a simplified human eye glare perception model. It includes a glare evaluation function and a license plate area illuminance function, both containing adjustable parameters of the supplementary lighting unit as variables. Based on the predicted three-dimensional motion trajectory of the cockpit, the observation vector from the light source center of each supplementary lighting unit to the driver's eye is calculated at each sampling time in the future. Calculate the deviation angle between the observation vector and the driver's line-of-sight vector, wherein the driver's line-of-sight vector is preset to the vehicle's forward direction or dynamically estimated based on the scene; Based on the deviation angle, the light intensity value in the adjustable parameters of the supplementary lighting unit, and the ambient background brightness, the instantaneous glare value generated by each supplementary lighting unit for the driver at each sampling moment is calculated using a preset glare perception empirical formula. The instantaneous glare values ​​of all supplementary lighting units at all sampling times are weighted by time and accumulated in space, and the accumulated result is used as the output value of the glare evaluation function, wherein the time weighting coefficient decreases according to the prediction time. Based on the predicted three-dimensional motion trajectory of the cockpit and the spatial position offset of the license plate area in the reference optical parameters, the three-dimensional coordinates of the center of the license plate area at each future sampling time are calculated. Based on the three-dimensional coordinates of the center of the license plate area, the known three-dimensional position and orientation parameters of each supplementary lighting unit, and the light intensity value and beam spatial distribution model in the adjustable parameters of the supplementary lighting unit, the normal illuminance component generated by each supplementary lighting unit illuminating the center of the license plate area is calculated. The normal illuminance components generated by all supplementary lighting units are accumulated and multiplied by the typical reflectance of the vehicle license plate in the reference optical parameters and the preset imaging system gain coefficient to obtain the expected signal intensity value of the license plate area on the image sensor. This value is used as the output value of the illuminance function of the license plate area.

9. The active fill light glare suppression method according to claim 1, characterized in that, The anti-glare control parameter set and the cockpit's three-dimensional motion trajectory prediction are input into a multi-agent collaborative decision-making algorithm. This algorithm assigns collaborative working modes and parameters to multiple roadside supplementary lighting units and generates pixel-level light spot control instructions containing dynamic shadow area location information, including: The supplementary light intensity benchmark value and dynamic shadow area definition in the anti-glare control parameter set, as well as the three-dimensional motion trajectory prediction of the cockpit, are used as global state information for collaborative decision-making. Based on the global state information and the predefined supplementary lighting unit performance model, a local observation and action space is constructed for each supplementary lighting unit. The action space includes at least an activation flag, a working intensity coefficient, and a fine-tuning amount of the light spot projection angle. A distributed reinforcement learning framework based on policy gradient is adopted, enabling the agent corresponding to each lighting unit to output action proposals in parallel based on its local observations; A coordinator module is introduced, which collects action proposals from all agents and arbitrates and corrects conflicting action proposals based on preset conflict resolution rules and global performance evaluation functions to generate consistent cooperative action decisions. The collaborative action decision assigns a final activation flag, working intensity coefficient, and spot projection angle fine-tuning amount to each supplementary lighting unit, and calculates the precise pixel coordinate sequence of the dynamic shadow area in the spot pattern to be generated for each activated supplementary lighting unit based on the definition of the dynamic shadow area and the prediction of the three-dimensional motion trajectory of the cockpit. The final parameters assigned to each enabled fill light unit and its corresponding dynamic shadow region pixel coordinate sequence are encapsulated into an independent control instruction unit; All control command units that enable the supplementary lighting unit are aggregated to generate the pixel-level spot control command.

10. The active fill light glare suppression method according to claim 9, characterized in that, The coordinator module collects action proposals from all agents and, based on preset conflict resolution rules and a global performance evaluation function, arbitrates and corrects conflicting action proposals to generate consistent cooperative action decisions, including: Identify conflict types in the agent's action proposals, including conflicts over lighting resources in the same road space area, and shading failure conflicts where the superposition of light spots from multiple supplementary lighting units may lead to the dynamic shadow area being accidentally illuminated. In response to conflicts over lighting resources, the coordinator module evaluates the contribution of each competing agent to the global performance when it individually executes its action proposal, based on the global performance evaluation function, and prioritizes the adoption of action proposals with higher contributions, while reducing or rejecting action proposals with lower contributions. In response to shading failure conflicts, the coordinator module checks the theoretical illuminance distribution of the dynamic shadow area under superimposed light spots. If there is an area where the illuminance exceeds the preset shadow maintenance threshold, conflict resolution is triggered. The conflict resolution is achieved by iteratively adjusting the working intensity coefficient or the fine-tuning amount of the light spot projection angle in the action proposal of the relevant supplementary lighting unit until the theoretical illuminance of the dynamic shadow area under the superimposed light spot is lower than the shadow maintenance threshold, and the adjusted action proposal is used as the correction result. Integrate all the arbitrated and revised action proposals to form the aforementioned consistent collaborative action decision.

Images