An adaptive high beam control method and device
By using feature-level fusion of cameras and millimeter-wave radar and dynamic scene map matching, the perception and control problems of existing adaptive high beam technology in harsh environments have been solved, achieving highly robust ADB control and improving system stability and driving comfort.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FORYOU GENERAL ELECTRONICS
- Filing Date
- 2025-10-22
- Publication Date
- 2026-07-31
AI Technical Summary
Existing adaptive high beam technology suffers from decreased perception performance in adverse weather and complex lighting conditions, insufficient handling of confidence conflicts between sensors, a single control strategy, difficulty in coping with complex scenarios, and weak anti-interference capability against non-vehicle light sources.
By using feature-level fusion of cameras and millimeter-wave radar, a fused target object containing geometric state variables and semantic attributes is generated. A dynamic scene map is constructed and matched with the optical shape strategy mapping library to achieve feedforward-feedback composite control, thereby improving the reliability of system decision-making and optical shape adaptability.
With cross-validation by multiple sensors, intelligent ADB control with high robustness under all operating conditions is achieved, ensuring system stability and driving comfort in harsh environments, avoiding glare and blind spots, and providing precise anti-glare while maximizing useful vision.
Smart Images
Figure CN121608673B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle control technology, and in particular to an adaptive high beam control method and device. Background Technology
[0002] Adaptive High Beam (ADB) technology is one of the key technologies for improving nighttime driving safety. However, existing ADB solutions still have significant limitations. First, mainstream solutions rely excessively on a single vision sensor, whose perception performance drops sharply in adverse weather conditions such as rain, snow, fog, and haze, as well as complex lighting environments such as strong light and tunnel entrances / exits, leading to functional failures or malfunctions. Second, while some solutions mention the fusion of radar and cameras, the fusion level is superficial, often involving simple comparisons of target lists, failing to effectively handle confidence conflicts between sensors, resulting in insufficient system decision reliability. Third, control strategies are generally singular and fixed, based on fixed brightness attenuation or masking templates, making it difficult to provide refined and predictive responses to complex scenarios such as curves, slopes, and following other vehicles, easily causing glare or blind spots. In addition, existing solutions have weak anti-interference capabilities against non-vehicle light sources such as streetlights and reflective signs, easily leading to unnecessary frequent switching off or flashing of high beams, affecting the driving experience.
[0003] Therefore, there is an urgent need in this field for an intelligent ADB control scheme that can achieve full operating conditions, high robustness, and scene adaptability. Summary of the Invention
[0004] This invention provides an adaptive high beam control method and device, aiming to overcome the deficiencies in the prior art and realize intelligent and scene-adaptive control of high beams.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] This invention provides an adaptive high beam control method, comprising:
[0007] Step 1: Perform feature-level fusion on the perception data from the camera and millimeter-wave radar to generate a fused target object containing geometric state variables and semantic attributes;
[0008] Step 2: Construct a dynamic scene map based on the fusion target object, match the dynamic scene map with the pre-stored light shape strategy mapping library, and determine the current dominant scene and the corresponding light shape control strategy;
[0009] Step 3: Execute feedforward control based on the light shape control strategy and execute feedback control based on the motion state changes of the fusion target object to realize feedforward-feedback composite control of matrix ADB headlights.
[0010] Specifically, step 1 includes:
[0011] Step 101: Acquire visual perception data from the camera, radar perception data from the millimeter-wave radar, and vehicle status data.
[0012] Step 102: Unify the visual perception data and radar perception data into a world coordinate system with the vehicle as the origin;
[0013] Step 103: Perform correlation matching on the targets detected by the camera and radar;
[0014] Step 104: Generate a fused target object for each successfully associated target, wherein:
[0015] The geometric state quantity is obtained by inputting the correlated sensor observation data into the state tracking filter;
[0016] The semantic attributes include target type, visual confidence C_c, and radar confidence C_r.
[0017] Specifically, in step 104, the determination of the target type is based primarily on visual recognition type, and the fused type confidence P_fused is calculated based on the visual confidence C_c and radar confidence C_r.
[0018] Specifically, step 2 includes:
[0019] Step 201: Construct a dynamic scene graph containing nodes, edges, and weights, where nodes represent fusion target objects, edges represent spatial relationships between targets, and weights are calculated based on the target's threat level T, type coefficient K_type, and overall confidence C.
[0020] Step 202: Perform similarity matching between the feature vector of the current scene map and the scene template in the light-form strategy mapping library to determine the current dominant scene;
[0021] Step 203: Invoke the light shape control strategy corresponding to the current dominant scene.
[0022] Specifically, the formula for calculating the weight W is:
[0023] W = T×K_type×C,
[0024] Wherein, the threat level T = α × f_distance(D) + β × f_speed(V_rel) + γ × f_lateral_move(A_lat), the type coefficient K_type is determined by a lookup table, and the overall confidence level C = δ × C_c + ε × C_r.
[0025] Another aspect of the present invention provides an adaptive high beam control device, comprising:
[0026] The perception and fusion module is configured to perform feature-level fusion on perception data from cameras and millimeter-wave radar to generate a fused target object containing geometric state variables and semantic attributes.
[0027] The intelligent decision-making module is configured to construct a dynamic scene map based on the fusion target object, match the dynamic scene map with a pre-stored light shape strategy mapping library, and determine the current dominant scene and the corresponding light shape control strategy.
[0028] The light pattern control module is configured to perform feedforward control based on the light pattern control strategy and feedback control based on the motion state changes of the fusion target object, thereby realizing feedforward-feedback composite control of the matrix ADB headlights.
[0029] Specifically, the sensing and fusion module includes:
[0030] The data acquisition submodule is configured to acquire visual perception data, radar perception data, and vehicle status data.
[0031] The coordinate unification submodule is configured to unify the sensing data to the world coordinate system.
[0032] The target association submodule is configured to perform association matching of targets detected by cameras and radar.
[0033] The object generation submodule is configured to generate the target object for fusion.
[0034] Specifically, the object generation submodule includes:
[0035] The state filtering unit is configured to calculate geometric state quantities through a state tracking filter;
[0036] The semantic attribute generation unit is configured to determine the target type and calculate visual confidence and radar confidence.
[0037] Specifically, the intelligent decision-making module includes:
[0038] The graph construction submodule is configured to build a dynamic scene graph that includes nodes, edges, and weights.
[0039] The scene matching submodule is configured to match the current scene map with the light shape strategy mapping library;
[0040] The strategy invocation submodule is configured to invoke the corresponding light pattern control strategy.
[0041] Specifically, the light shape control module includes:
[0042] The feedforward control submodule is configured to generate basic control commands based on the light pattern strategy.
[0043] The feedback control submodule is configured to fine-tune the control commands in a closed loop using a PID controller.
[0044] The device also includes a central processing unit, which is connected to the sensing and fusion module, the intelligent decision-making module, and the light shape control module to coordinate their operation.
[0045] The beneficial effects of this invention are as follows: By deeply fusing vision and radar at the feature layer and managing confidence, this invention introduces a dynamic scene map and light pattern strategy mapping library for mid-level decision-making, realizing the understanding of traffic scene semantics, and then generating a non-fixed light pattern that best matches the current scene, thereby maximizing the useful field of view while accurately preventing glare; by adopting a feedforward-feedback composite control mechanism, the timeliness and smoothness of light pattern changes are ensured, avoiding abrupt switching between bright and dark, and significantly improving driving comfort; finally, multi-sensor cross-validation and the system's adaptive fault tolerance effectively filter out static interference sources, ensuring the stability of the system's core functions when the performance of some sensors degrades, thus providing an intelligent, scene-adaptive, high-performance ADB solution as a whole. Attached Figure Description
[0046] Figure 1 This is a flowchart illustrating the adaptive high beam control method of the present invention;
[0047] Figure 2 This is a schematic diagram of the adaptive high beam control device of the present invention;
[0048] Figure 3 This is a schematic diagram of the structure of the sensing and fusion module of the present invention;
[0049] Figure 4 This is a schematic diagram of the intelligent decision-making module of the present invention;
[0050] Figure 5 This is a schematic diagram of the optical pattern control module of the present invention. Detailed Implementation
[0051] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. The drawings are for reference and illustration only and do not constitute a limitation on the scope of protection of the present invention.
[0052] In the process described in the specification, claims, or drawings of this invention, each step is numbered (e.g., step 10, 20, etc.). These numbers are used only to distinguish the steps and do not represent any execution order. It should be noted that the terms "first," "second," etc., used herein are only for distinguishing the objects being described and do not represent a chronological order, nor do they indicate that "first," "second," etc., are different types.
[0053] Example 1
[0054] like Figure 1 As shown, this embodiment provides an adaptive high beam control method, including:
[0055] Step 1: Pre-feature fusion of multi-source heterogeneous data.
[0056] This step aims to address the issue of reliance on a single sensor and achieve reliable sensing across all operating conditions. Specifically, it includes:
[0057] Step 101: Acquire multi-source sensing data.
[0058] Specifically, images are captured by a forward-facing camera, and target recognition is performed using deep learning models (such as the YOLO series models). The output includes information such as target type, two-dimensional pixel coordinates, visual confidence C_c, and visual features (such as feature points of vehicle taillights / headlights). Simultaneously, the echo signal is processed by a forward-facing millimeter-wave radar. After clustering and tracking algorithms, a target list is output, including target distance, relative speed, azimuth angle, target ID, RCS value, and radar confidence C_r. Furthermore, vehicle status information, including vehicle speed, steering angle, yaw rate, GPS location, and navigation map data containing the curvature of the road ahead, is obtained through the vehicle's CAN bus interface.
[0059] Step 102: Unify the spatiotemporal coordinate system.
[0060] The sensor data from cameras and radar are transformed and unified into a world coordinate system with the vehicle as the origin, laying the foundation for subsequent data association and fusion.
[0061] Step 103: Perform target association.
[0062] Based on nearest neighbor algorithms or probabilistic data association methods, the targets detected by radar and the targets identified by cameras are matched to determine whether the observation data from different sensors originate from the same physical target.
[0063] Step 104: Generate the fusion target object.
[0064] For each successfully associated target, a unified fusion target object is prepared, which contains two main categories of information: geometric state variables and semantic attributes.
[0065] Preferably, the geometric state quantities (such as the fused 3D position and fused velocity) are obtained by inputting the correlated radar and visual observation data and their confidence scores (C_r, C_c) as measurements into a state tracking filter (e.g., a Kalman filter) for optimal estimation.
[0066] The semantic attributes include:
[0067] Target type: Primarily based on visual recognition type. To improve reliability, the confidence level of this type can be adjusted by combining radar confidence level. For example, the fused type confidence level can be calculated using the formula P_fused=(W_c*C_c+W_r*C_r) (where W_c and W_r are adjustable weights based on sensor state), or the joint confidence level formula P_fused =C_c * C_r can be used to calculate the fused type confidence level.
[0068] The visual confidence score C_c is determined by the probability score output by the target detection neural network combined with the overall brightness and contrast of the current image, while the radar confidence score C_r is determined by the signal-to-noise ratio (SNR) of the target point and the smoothness of its tracking trajectory.
[0069] Step 2: Mid-level decision-making based on dynamic scene graph.
[0070] This step aims to address the issue of rigid control strategies and achieve scenario-adaptive intelligent decision-making, specifically including:
[0071] Step 201: Construct a dynamic scene graph.
[0072] Based on the fusion target object list generated in step 1, and combined with the vehicle's status, the central processing unit constructs a weighted dynamic scene graph in real time. This graph includes:
[0073] Node: Represents each target object to be merged.
[0074] Edges: Represent the spatial relationship between targets (such as the relationship between front and rear vehicles, or the relationship between vehicles in opposite lanes). By fusing the vehicle's GPS location and navigation map data to obtain a lane line model, the relationship between targets is determined. For example, an edge is established between two targets located in the same lane model and close to each other to establish a "front and rear vehicle relationship".
[0075] Weights are dynamically assigned based on the target's threat level T, type coefficient K_type, and overall confidence level C. The specific calculation process is as follows:
[0076] Calculate the base threat level T:
[0077] T = α × f_distance(D) + β × f_speed(V_rel) + γ × f_lateral_move(A_lat)
[0078] Where D is the relative distance, V_rel is the relative velocity, and A_lat is the lateral acceleration; α, β, γ are weighting coefficients and satisfy α + β + γ = 1; f_distance(D) is the distance threat function, preferably 1 / (1 + k_d × D), where k_d is a distance attenuation coefficient that can be configured according to the scene; f_speed(V_rel) is the speed threat function, preferably min(1, |V_rel| / V_max), where V_max is the set maximum threat relative velocity; f_lateral_move(A_lat) is the lateral movement threat function, preferably min(1, |A_lat| / A_max), where A_max is the preset maximum lateral acceleration threshold (e.g., 2~4 m / s²).
[0079] Determine the target type coefficient K_type: Assign a basic threat coefficient to different target types using a lookup table method. For example: oncoming vehicles are 1.0, pedestrians / motorcycles are 0.9, vehicles traveling in the same direction are 0.7, static obstacles are 0.3, and unknown targets are 0.5.
[0080] Calculate the overall confidence level C:
[0081] C = δ × C_c + ε × C_r, where δ and ε are weighting coefficients that can be dynamically adjusted according to the sensor status.
[0082] Calculate the final node weight W: W = T × K_type × C.
[0083] Step 202: Scene recognition and policy matching.
[0084] The feature vectors of the current scene graph constructed in step 201 (e.g., statistical features such as the weight distribution of each node and the type and number of edges) are compared with the feature vectors of scene templates in the preset "light shape strategy mapping library" to calculate similarity (e.g., calculate Euclidean distance). The scene corresponding to the scene template with the highest similarity (i.e., the smallest Euclidean distance) is selected as the current dominant scene (e.g., "highway following", "meeting on a curve", "urban mixed traffic"). The light shape strategy mapping library pre-stores a variety of typical scenes and their optimal light shape control strategies determined by simulation and expert annotation.
[0085] Step 203: Invoke the light pattern control strategy.
[0086] Based on the identified dominant scene, the corresponding light pattern control strategy in the mapping library is invoked. For example:
[0087] In the case of "high-speed following", the strategy is to create a "dark tunnel" behind and to the sides of the vehicle in front, while retaining side lighting to observe the road shoulder.
[0088] In the case of a "meeting on a curve" scenario, the strategy is to make the light-shaped shading area pre-aiming and distorting its shape according to the road curvature and the vehicle's speed, so as to ensure that the shading area continuously and accurately covers the trajectory of the oncoming vehicle.
[0089] In the case of a "pedestrian approaching" scenario, the strategy is to reduce the brightness in the area where the pedestrian is located to a level that is not dazzling, but not to completely block the pedestrian's movements, so as to continuously monitor the pedestrian's movements.
[0090] Step 3: Feedforward-feedback composite optical shape control.
[0091] This step aims to address the issues of control response delay and smoothness, achieving precise and comfortable light pattern adjustment, specifically including:
[0092] Step 301: Feedforward Control. Based on the light pattern control strategy determined in Step 2, basic light pattern control commands (such as the position, shape, and grayscale gradient of the shading area) are directly output to the matrix ADB headlight assembly.
[0093] Step 302: Feedback Control. Continuously monitor the motion state of the target object (especially changes in position and velocity). Use a PID controller to perform closed-loop fine-tuning and smoothing of the feedforward control command.
[0094] Specifically, the feedback controller calculates the fine-tuning value Δu(t) of the control quantity according to the PID control law based on the deviation e(t) between the expected value of the fused target state and the state value corresponding to the current optical command. The fine-tuning value Δu(t) of the control quantity is calculated according to the following proportional-integral-derivative (PID) control law:
[0095]
[0096] in,
[0097] Δu(t) is the control fine-tuning vector at time t, for example, the adjustment amount [Δx, Δy] at the center of the shading area. T ;
[0098] e(t) is the deviation vector at time t, such as the deviation between the target position and the center of the shielding area;
[0099] K p K i K d These are the proportional, integral, and derivative gain coefficients of the controller, respectively.
[0100] This is the integral term of the deviation vector, used to eliminate steady-state error;
[0101] The differential term of the deviation vector is used to predict the trend of error changes and improve system stability.
[0102] For example, in a following vehicle scenario, the deviation between the target position of the rear of the preceding vehicle and the geometric center of the current occlusion area is used as e(t), and the adjustment amount [Δx, Δy] of the occlusion area center is calculated using a PID control algorithm. T It performs real-time, smooth corrections to feedforward instructions.
[0103] Furthermore, the system possesses adaptive capabilities. For example, when it predicts that the vehicle in front will decelerate, the feedback controller will increase the area of the obstruction zone in advance; when it detects that the camera confidence C_c has decreased due to weather conditions, it automatically increases the weight of the radar data in the fusion formula (i.e., ε), ensuring that the system can still operate stably when sensor performance changes.
[0104] Example 2
[0105] like Figure 2 As shown, this embodiment provides an adaptive high beam control device, including: a perception and fusion module, an intelligent decision-making module, a light pattern control module, and an execution module connected in sequence, and also includes a central processing unit, which is connected to the perception and fusion module, the intelligent decision-making module, the light pattern control module, and the execution module;
[0106] The sensing and fusion module is used to realize the acquisition and feature-level fusion of multi-source heterogeneous data;
[0107] The intelligent decision-making module is used to make light-shape strategy decisions based on dynamic scene maps;
[0108] The light pattern control module is used to perform feedforward-feedback composite control;
[0109] The execution module includes an addressable matrix ADB headlight assembly, configured to receive light pattern control commands, realize independent control of at least 100 pixels, and complete complex light pattern distortion and grayscale gradient adjustment.
[0110] The central processing unit is used to connect the various modules and is configured to coordinate the operation of the system.
[0111] In this embodiment, the central processing unit has an NPU computing power of at least 2 TOPS to support real-time inference of deep learning models.
[0112] like Figure 3 As shown, in this embodiment, the perception and fusion module includes a data acquisition submodule, a coordinate unification submodule, a target association submodule, and an object generation submodule connected in sequence.
[0113] The data acquisition submodule is configured as follows:
[0114] Images are captured by the front-facing camera, and the built-in deep learning model is used to identify targets, outputting target type, pixel coordinates, visual confidence C_c and visual features;
[0115] The echo signal is processed by forward millimeter-wave radar, and after being processed by clustering tracking algorithm, the target range, velocity, azimuth angle, RCS value and radar confidence C_r are output.
[0116] The vehicle speed, steering angle, yaw rate, GPS location, and navigation map data are obtained through the vehicle's CAN bus interface.
[0117] The coordinate unification submodule is configured to uniformly convert the perception data from the camera and radar to a world coordinate system with the vehicle as the origin.
[0118] The target association submodule is configured to perform association matching on the same target detected by radar and camera based on nearest neighbor or probabilistic data association algorithms;
[0119] The object generation submodule is configured to generate a fused target object for each successfully associated target, including:
[0120] The state filtering unit is configured to input the correlated sensor observation data and confidence level into the state tracking filter to estimate the fused 3D position and fusion speed.
[0121] The semantic attribute generation unit is configured as follows:
[0122] The target type is determined primarily by visual recognition type, and the confidence level of the type is corrected using radar confidence. For example, the confidence level of the fused type can be calculated using the formula P_fused=(W_c*C_c+W_r*C_r) (where W_c and W_r are adjustable weights based on sensor state), or the joint confidence level formula P_fused = C_c * C_r can also be used to calculate the confidence level of the fused type.
[0123] Record the visual confidence score C_c and the radar confidence score C_r, where C_c is determined by the target detection probability score combined with the image quality, and C_r is determined by the target point signal-to-noise ratio and the smoothness of the tracking trajectory.
[0124] like Figure 4 As shown, in this embodiment, the intelligent decision-making module includes a graph construction submodule, a scene matching submodule, and a strategy invocation submodule connected in sequence;
[0125] The graph construction submodule is configured to construct a weighted dynamic scene graph in real time based on the fusion target object list and the vehicle status. This graph includes:
[0126] Nodes represent the various target objects for fusion;
[0127] The edge represents the spatial relationship between targets determined based on the lane line model;
[0128] The weights are calculated based on the target's threat level T, type coefficient K_type, and overall confidence level C, where:
[0129] The threat level T is calculated using the formula T = α × f_distance(D) + β × f_speed(V_rel) + γ × f_lateral_move(A_lat);
[0130] The type coefficient K_type is determined by a lookup table.
[0131] The overall confidence level C is calculated using the formula C = δ × C_c + ε × C_r;
[0132] The final weight W = T × K_type × C.
[0133] The scene matching submodule is configured to perform similarity matching between the feature vector of the current scene map and the pre-stored light shape strategy mapping library, and identify the current dominant scene based on the principle of minimum Euclidean distance.
[0134] The strategy invocation submodule is configured to invoke the corresponding light pattern control strategy based on the identified dominant scene, including but not limited to:
[0135] "Dark tunnel" strategy in high-speed following scenarios;
[0136] Pre-aiming deflection and shape distortion strategies in curve passing scenarios;
[0137] Local dimming strategy when pedestrians are approaching.
[0138] like Figure 5 As shown, in this embodiment, the light pattern control module includes a feedforward control submodule and a feedback control submodule that are interconnected.
[0139] The feedforward control submodule is configured to generate basic light pattern control commands based on the light pattern strategy output by the decision module and send them to the ADB headlight actuator.
[0140] The feedback control submodule is configured as follows:
[0141] Continuously monitor changes in the motion state of the target object being fused;
[0142] The feedforward control command is fine-tuned in a closed loop using a PID controller. For example, the deviation e(t) between the position of the preceding vehicle and the center of the shielded area is used as input to calculate the adjustment amount [Δx, Δy]. T ;
[0143] It has adaptive adjustment capabilities, dynamically adjusting the fusion weights based on changes in sensor confidence.
[0144] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.
Claims
1. A method of adaptive high beam control, characterized in that include: Step 1: Perform feature-level fusion on the perception data from the camera and millimeter-wave radar to generate a fused target object containing geometric state variables and semantic attributes; Step 2: Construct a dynamic scene map based on the fusion target object, match the dynamic scene map with the pre-stored light shape strategy mapping library, and determine the current dominant scene and the corresponding light shape control strategy; Step 3: Execute feedforward control based on the light pattern control strategy and execute feedback control based on the motion state changes of the fusion target object to realize feedforward-feedback composite control of matrix ADB headlights; Step 2 includes: Step 201: Construct a dynamic scene graph containing nodes, edges, and weights, where nodes represent fusion target objects, edges represent spatial relationships between targets, and weights are calculated based on the target's threat level T, type coefficient K_type, and overall confidence C. Step 202: Perform similarity matching between the feature vector of the current scene map and the scene template in the light-form strategy mapping library to determine the current dominant scene; Step 203: Invoke the light shape control strategy corresponding to the current dominant scene; The formula for calculating the weight W is: W = T×K_type×C, Wherein, the threat level T = α × f_distance(D) + β × f_speed(V_rel) + γ × f_lateral_move(A_lat), where D is the relative distance, V_rel is the relative velocity, A_lat is the lateral acceleration, f_distance(D) is the distance threat function, f_speed(V_rel) is the velocity threat function, and f_lateral_move(A_lat) is the lateral movement threat function; α, β, γ are weighting coefficients and satisfy α + β + γ = 1; the type coefficient K_type is determined by a lookup table method, and the overall confidence level C = δ × C_c + ε × C_r, where C_c represents the visual confidence level, C_r represents the radar confidence level, and δ and ε are weighting coefficients that can be dynamically adjusted according to the sensor status.
2. The adaptive high beam control method according to claim 1, characterized in that, Step 1 includes: Step 101: Acquire visual perception data from the camera, radar perception data from the millimeter-wave radar, and vehicle status data. Step 102: Unify the visual perception data and radar perception data into a world coordinate system with the vehicle as the origin; Step 103: Perform correlation matching on the targets detected by the camera and radar; Step 104: Generate a fused target object for each successfully associated target, wherein: The geometric state quantity is obtained by inputting the correlated sensor observation data into the state tracking filter; The semantic attributes include target type, visual confidence C_c, and radar confidence C_r.
3. The method according to claim 2, characterized in that, In step 104, the determination of the target type is based primarily on visual recognition type, and the fused type confidence P_fused is calculated based on the visual confidence C_c and radar confidence C_r.
4. An adaptive high beam control device, characterized in that, include: The perception and fusion module is configured to perform feature-level fusion on perception data from cameras and millimeter-wave radar to generate a fused target object containing geometric state variables and semantic attributes. The intelligent decision-making module is configured to construct a dynamic scene map based on the fusion target object, match the dynamic scene map with a pre-stored light shape strategy mapping library, and determine the current dominant scene and the corresponding light shape control strategy. The light pattern control module is configured to perform feedforward control based on the light pattern control strategy and to perform feedback control based on the motion state changes of the fusion target object, thereby realizing feedforward-feedback composite control of the matrix ADB headlights. The intelligent decision-making module determines the current dominant scenario and the corresponding light pattern control strategy according to the following method: A dynamic scene graph containing nodes, edges, and weights is constructed, where nodes represent fusion target objects, edges represent spatial relationships between targets, and weights are calculated based on the target's threat level T, type coefficient K_type, and comprehensive confidence C. The feature vector of the current scene map is matched with the scene template in the light shape strategy mapping library to determine the current dominant scene; Invoke the light pattern control strategy corresponding to the current dominant scene; The formula for calculating the weight W is: W = T×K_type×C, Wherein, the threat level T = α × f_distance(D) + β × f_speed(V_rel) + γ × f_lateral_move(A_lat), where D is the relative distance, V_rel is the relative velocity, A_lat is the lateral acceleration, f_distance(D) is the distance threat function, f_speed(V_rel) is the velocity threat function, and f_lateral_move(A_lat) is the lateral movement threat function; α, β, γ are weighting coefficients and satisfy α + β + γ = 1; the type coefficient K_type is determined by a lookup table method, and the overall confidence level C = δ × C_c + ε × C_r, where C_c represents the visual confidence level, C_r represents the radar confidence level, and δ and ε are weighting coefficients that can be dynamically adjusted according to the sensor status.
5. The apparatus according to claim 4, characterized in that, The sensing and fusion module includes: The data acquisition submodule is configured to acquire visual perception data, radar perception data, and vehicle status data. The coordinate unification submodule is configured to unify the sensing data to the world coordinate system. The target association submodule is configured to perform association matching of targets detected by cameras and radar. The object generation submodule is configured to generate the target object for fusion.
6. The apparatus according to claim 5, characterized in that, The object generation submodule includes: The state filtering unit is configured to calculate geometric state quantities through a state tracking filter; The semantic attribute generation unit is configured to determine the target type and calculate visual confidence and radar confidence.
7. The apparatus according to claim 4, characterized in that, The intelligent decision-making module includes: The graph construction submodule is configured to build a dynamic scene graph that includes nodes, edges, and weights. The scene matching submodule is configured to match the current scene map with the light shape strategy mapping library; The strategy invocation submodule is configured to invoke the corresponding light pattern control strategy.
8. The apparatus according to claim 4, characterized in that, The light pattern control module includes: The feedforward control submodule is configured to generate basic control commands based on the light pattern strategy. The feedback control submodule is configured to fine-tune the control commands in a closed loop using a PID controller. The device also includes a central processing unit, which is connected to the sensing and fusion module, the intelligent decision-making module, and the light shape control module to coordinate their operation.