LED vehicle lamp module dynamic light type adjusting system based on adaptive environment perception

Abstract

The present application relates to the field of electric lighting technology, in particular to a LED vehicle lamp module dynamic light type adjusting system based on adaptive environment sensing. The system comprises a front camera, a millimeter wave radar, a main control chip and a light emitting diode constant current driving circuit. The space-time feature extraction network inside the main control chip aligns the spatial coordinates and fuses the time dimension of the light spot pixel features in the environment image and the distance and radial velocity features in the point cloud data, and outputs a three-dimensional tensor containing the spatial position prediction coordinates of the oncoming vehicle in the future preset period. The light type calculation module calculates the target pixel coordinate set and generates a pulse width modulation control signal according to the projection mapping relationship between the three-dimensional tensor and the light emitting diode matrix. The light emitting diode constant current driving circuit receives the control signal and completes the state update of the corresponding pixel node in advance. The present application eliminates the light type switching lag caused by the physical time consumption of the signal processing link and avoids transient glare.

Description

Technical Field

[0001] This invention relates to the field of electric lighting technology, and more specifically to a dynamic light pattern adjustment system for LED vehicle lamp modules based on adaptive environmental perception. Background Technology

[0002] Existing adaptive headlight systems typically rely on a front-facing camera to capture images of the environment in front of the vehicle. The main control chip processes these images frame by frame, identifying the headlight beams or vehicle body outline pixels of oncoming vehicles. After determining the position of the oncoming vehicle in the image coordinate system, the main control chip calculates the pixel areas in the LED matrix that need to be turned off or have their brightness reduced based on a preset projection mapping relationship, and sends control commands to the LED driver circuit. Upon receiving the commands, the LED driver circuit changes the power state of the corresponding pixel nodes, creating dark areas in the original lighting pattern to prevent direct light from hitting oncoming vehicles.

[0003] Based on the specific implementation of the aforementioned existing technologies, a core technical problem exists: when vehicles meet at night, they are all in high-speed motion. From the camera acquiring images, the main control chip performing image recognition and pixel coordinate calculation, to the LED driving circuit completing the state switching, the entire signal processing and transmission link has inherent physical time consumption. This physical time consumption causes the actual switching action of the LED matrix's light pattern to lag behind the actual spatial position change of the oncoming vehicle. When the oncoming vehicle travels to the light cutoff line area, the LED matrix has not yet completed the extinguishing or dimming state update of the corresponding pixel nodes, causing the oncoming driver to experience transient glare. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic light pattern adjustment system for LED vehicle lighting modules based on adaptive environmental perception, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The LED vehicle headlight module dynamic light pattern adjustment system based on adaptive environment perception includes a front camera, millimeter-wave radar, main control chip and LED constant current drive circuit. The front camera acquires images of the environment in front, and the millimeter-wave radar outputs point cloud data of the target in front.

[0007] The spatiotemporal feature extraction network inside the main control chip performs spatial coordinate alignment and temporal dimension fusion of the spot pixel features in the environmental image with the distance features and radial velocity features in the point cloud data, and outputs a three-dimensional tensor containing the predicted coordinates of the spatial positions of oncoming vehicles within a preset future time period.

[0008] The light pattern calculation module inside the main control chip calculates the set of target pixel coordinates in the LED matrix that need to be turned off or have their brightness reduced based on the projection mapping relationship between the three-dimensional tensor and the LED matrix, and generates a pixel-level PWM control signal. The LED constant current drive circuit receives the pixel-level PWM control signal and completes the update of the off or dimming status of the LED pixel nodes in the corresponding set of target pixel coordinates before the oncoming vehicle enters the preset illumination cutoff line area.

[0009] Preferably, the spatial coordinate alignment and temporal dimension fusion includes projecting radar coordinate points in the point cloud data onto the pixel coordinate system of the environmental image to generate a radar projection map, and extracting the fusion feature vector of the region in the radar projection map that overlaps with the spot pixel features;

[0010] The fused feature vector is input into a temporal sliding window set inside the spatiotemporal feature extraction network. The temporal sliding window extracts fused feature vectors from multiple consecutive frames according to a preset time step and inputs them into a three-dimensional convolutional layer. The three-dimensional convolutional layer performs convolution operations on the fused feature vectors from multiple consecutive frames along the time dimension.

[0011] Preferably, the output includes a three-dimensional tensor of the predicted spatial coordinates of oncoming vehicles within a future preset time period, which includes inputting the high-dimensional spatiotemporal features extracted by the spatiotemporal feature extraction network into a long short-term memory network layer.

[0012] The Long Short-Term Memory (LSTM) network layer extracts the temporal state variables from the spatiotemporal high-dimensional features, inputs the temporal state variables into the fully connected layer, and decodes them in conjunction with the vehicle kinematic constraint equations. The vehicle kinematic constraint equations construct the vehicle kinematic model based on the vehicle's yaw rate and longitudinal velocity, and construct the road boundary constraint equations in conjunction with the road curvature of the high-precision map. The predicted coordinates that exceed the road range are truncated and corrected so that the predicted position of oncoming vehicles remains within the drivable area of ​​the road. The three-dimensional tensor is generated by multiplying the state transition matrix with the output vector of the fully connected layer.

[0013] Preferably, the projection mapping relationship is constructed by obtaining the intrinsic and extrinsic parameter matrices of the front-facing camera and the nonlinear refractive coefficient of the LED matrix lens.

[0014] A mapping equation is established from the pixel coordinate system of the environment image to the physical coordinate system of the LED matrix based on the intrinsic parameter matrix, the extrinsic parameter matrix and the nonlinear refractive coefficient. The mapping equation is discretized into a mapping lookup table containing the correspondence between the physical coordinates of each LED pixel node in the LED matrix and the coordinates of each pixel in the environment image. The mapping lookup table is stored in the flash memory chip connected to the main control chip.

[0015] Preferably, generating the pixel-level PWM control signal includes fitting the edge of the anti-glare cutoff line in the LED matrix based on the target pixel coordinate set, and extracting the set of transition zone pixel coordinates adjacent to the edge of the anti-glare cutoff line;

[0016] Calculate the Euclidean distance from each pixel node in the transition zone pixel coordinate set to the edge of the anti-glare cutoff line. Calculate the PWM duty cycle value of each pixel node in the transition zone pixel coordinate set according to the Euclidean distance and a preset interpolation rule. Write the PWM duty cycle value into the buffer of the direct memory access controller and generate the pixel-level PWM control signal based on the buffer.

[0017] Preferably, the step of updating the off or dimming state of the LED pixel node in the target pixel coordinate set includes the shift register set inside the LED constant current drive circuit receiving the pixel-level PWM control signal and performing data shifting and writing within the current display frame period;

[0018] After the shift register is written, it outputs a latch enable signal to the latch circuit. The latch circuit latches the data in the shift register to the LED driver output register synchronously according to the rising edge of the latch enable signal. The LED driver output register outputs the updated driving current at the beginning of the next display frame cycle.

[0019] Preferably, the step of projecting the radar coordinate points in the point cloud data onto the pixel coordinate system of the environmental image to generate a radar projection map includes converting the point cloud data from a spherical coordinate system to a Cartesian coordinate system.

[0020] Based on the extrinsic matrix of the front-facing camera, the radar coordinates in the Cartesian coordinate system are transformed to the camera coordinate system. The radar coordinates in the camera coordinate system are projected onto the pixel plane of the environmental image using a pinhole imaging model. The projected radar coordinates are then processed by polar coordinate rasterization to generate a depth map. The depth map is used as the radar projection map, and invalid pixels with a confidence level lower than a preset threshold are removed from the radar projection map.

[0021] Preferably, the vehicle kinematic constraint equation is based on the vehicle's yaw rate and longitudinal velocity to construct the vehicle's kinematic model, and combined with the road curvature of the high-precision map to construct the road boundary constraint equation. The predicted coordinates that exceed the road range are truncated and corrected so that the predicted position of oncoming vehicles is kept within the road drivable area. This includes acquiring the yaw rate data from the vehicle's yaw rate sensor and the longitudinal velocity data from the vehicle's longitudinal velocity sensor.

[0022] A kinematic model of the vehicle is constructed based on the yaw rate data and the longitudinal velocity data. A road boundary constraint equation is constructed by combining the road curvature data in the high-precision map. The kinematic model of the vehicle and the road boundary constraint equation are combined to generate the state transition matrix. The state transition matrix is ​​used to truncate the predicted coordinates in the three-dimensional tensor that exceed the road boundary constraint equation.

[0023] Preferably, obtaining the nonlinear refractive coefficient of the LED matrix lens includes attaching a thin-film temperature sensor to the surface of the LED matrix lens, the thin-film temperature sensor collecting surface temperature distribution data of the LED matrix lens at a preset sampling frequency, and retrieving a pre-stored table of thermal expansion coefficients and refractive index temperature gradients of lens materials.

[0024] Based on the surface temperature distribution data, the lens material thermal expansion coefficient table and the refractive index temperature gradient table are consulted to calculate the lens surface deformation compensation and refractive index offset at the current temperature. The nonlinear refractive index is then corrected in real time based on the lens surface deformation compensation and refractive index offset.

[0025] Preferably, the step of calculating the PWM duty cycle value of each pixel node in the transition zone pixel coordinate set according to the Euclidean distance and a preset interpolation rule includes obtaining the current ambient light intensity value collected by the ambient light sensor, matching the corresponding reference duty cycle slope in a preset mapping table based on the current ambient light intensity value, calculating the maximum and minimum allowable brightness values ​​of each pixel node in the transition zone pixel coordinate set, constructing a piecewise linear interpolation function based on the reference duty cycle slope, the maximum allowable brightness value, the minimum allowable brightness value, and the Euclidean distance, and inputting the Euclidean distance into the piecewise linear interpolation function to calculate the PWM duty cycle value.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. By introducing point cloud data from millimeter-wave radar, the distance and radial velocity features in the point cloud data are spatially aligned and fused with the light spot pixel features in the environmental image, outputting a three-dimensional tensor containing predicted spatial coordinates for a future preset time period. Based on this three-dimensional tensor, the target pixel coordinate set is calculated, and a pulse width modulation control signal is generated. This allows the LED constant current drive circuit to preemptively update the extinguishing or dimming state of the corresponding pixel node before an oncoming vehicle enters the preset illumination cutoff area. This technique transforms light pattern adjustment from a real-time, lagging response dependent on a single frame image to a forward-looking response based on spatiotemporal prediction, reducing the lag in light pattern switching caused by the physical time consumption of the signal processing link and avoiding transient glare.

[0028] 2. Point cloud data is projected onto the pixel coordinate system of the environmental image to generate a radar projection map. This projection map is then input into a 3D convolutional layer via a time-series sliding window for convolution operations, ensuring the continuity of features across multiple data sources in the time dimension. A long short-term memory network layer is used to extract temporal state variables, and a state transition matrix is ​​constructed by combining it with vehicle kinematic constraint equations to decode and generate a 3D tensor. The introduction of physical motion constraints avoids divergence in predicted coordinates. When generating the pulse width modulation (PWM) control signal, the edge of the anti-glare cutoff line is fitted, and the set of pixel coordinates in the transition zone is extracted. The PWM duty cycle value is calculated based on the Euclidean distance from each pixel node in the transition zone pixel coordinate set to the edge of the anti-glare cutoff line, avoiding brightness jumps at the boundary between the anti-glare area and the illumination area and maintaining a smooth transition of the light pattern. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the overall workflow of the LED vehicle headlight module dynamic light pattern adjustment system based on adaptive environment perception, as described in this invention.

[0030] Figure 2 This is a flowchart of the multi-source sensor data spatial coordinate alignment and temporal dimension fusion process of the present invention;

[0031] Figure 3 This is a flowchart of the three-dimensional tensor generation process for predicting the spatial position of opposing vehicles according to the present invention.

[0032] Figure 4 This is a flowchart illustrating the LED matrix projection mapping relationship construction and temperature correction process of the present invention.

[0033] Figure 5 This is a flowchart of the pixel-level PWM control signal generation and transition zone dimming calculation of the present invention;

[0034] Figure 6 This is a flowchart illustrating the pixel node state update process of the LED constant current driving circuit of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please refer to Figure 1This embodiment provides a dynamic light pattern adjustment system for LED vehicle headlight modules based on adaptive environmental perception, including a front-facing camera, millimeter-wave radar, a main control chip, and an LED constant current drive circuit. The front-facing camera is fixedly installed in the upper-middle part of the inner side of the vehicle's windshield, with its optical axis parallel to the vehicle's longitudinal axis. The front-facing camera continuously acquires environmental images of the driving environment in front of the vehicle at a preset acquisition frame rate. The pixel coordinate system of the environmental image has the upper left corner of the image as the origin, with the positive u-axis pointing horizontally to the right and the positive v-axis pointing vertically downwards. The environmental images acquired by the front-facing camera are transmitted to the image signal processing unit of the main control chip via a MIPI interface. The image signal processing unit sequentially performs distortion correction, white balance adjustment, and Gaussian filtering noise reduction on the environmental image to obtain a pre-processed environmental image. The feature extraction unit inside the main control chip extracts spot pixel features from the preprocessed environmental image. Specifically, the feature extraction unit converts the environmental image from the RGB color space to the YCbCr color space, extracts the brightness component data of the Y channel, performs binarization segmentation on the brightness component data of the Y channel through a preset brightness threshold, marks the pixels with brightness values ​​higher than the brightness threshold as candidate spot pixels, performs connected component analysis on the candidate spot pixels, removes invalid connected components with an area smaller than a preset area threshold or an aspect ratio exceeding a preset ratio range, and retains the set of pixels corresponding to the valid connected components as spot pixel features. The spot pixel features include the centroid coordinates, contour boundary coordinates, circumscribed rectangle size, and average brightness value of the valid connected components.

[0037] The millimeter-wave radar is fixedly mounted in the center of the vehicle's front bumper, with its detection surface perpendicular to the vehicle's longitudinal axis. The millimeter-wave radar outputs point cloud data of the target ahead at the same frame rate as the front-facing camera. Each data point in the point cloud data includes range characteristics, radial velocity characteristics, azimuth characteristics, elevation characteristics, and signal-to-noise ratio (SNR) characteristics. The raw point cloud data output by the millimeter-wave radar is transmitted via a CANFD bus to the radar data processing unit of the main control chip. The radar data processing unit performs constant false alarm rate (CFAR) detection on the raw point cloud data, filtering out static clutter and invalid data points with an SNR lower than a preset confidence threshold to obtain valid point cloud data. The valid point cloud data retains the range, radial velocity, azimuth, and elevation characteristics.

[0038] The main control chip incorporates a spatiotemporal feature extraction network, deployed within its neural network processing unit. This network performs spatial coordinate alignment and temporal fusion on the pixel features of light spots in the environmental image and the distance and radial velocity features in the point cloud data. Specifically, spatial coordinate alignment establishes a one-to-one mapping between the radar coordinate system and the pixel coordinate system of the environmental image, transforming the point cloud data from the radar coordinate system to the pixel coordinate system, ensuring that the coordinate positions of the same spatial target coincide in both data sources. Temporal fusion correlates the aligned feature data from multiple consecutive frames along a temporal dimension, extracting the target's motion temporal features. After completing spatial coordinate alignment and temporal dimension fusion, the spatiotemporal feature extraction network outputs a three-dimensional tensor containing the predicted spatial coordinates of oncoming vehicles within a future preset time period. The dimensions of the three-dimensional tensor are T×H×W, where T is the number of prediction time steps within the future preset time period, H is the height dimension of the predicted coordinates, and W is the width dimension of the predicted coordinates. The duration of the future preset time period is greater than or equal to the signal processing and execution time of the entire system link. In this embodiment, the future preset time period is set to 200ms, the number of prediction time steps is set to 10, and the duration of each time step is 20ms.

[0039] The main control chip also includes a light pattern calculation module, which runs within the chip's general-purpose processor core. This module receives a 3D tensor output from the spatiotemporal feature extraction network and, based on the projection mapping between the 3D tensor and the LED matrix, calculates the set of target pixel coordinates in the LED matrix that need to be turned off or have their brightness reduced. The projection mapping establishes the correspondence between the pixel coordinate system of the environmental image and the physical coordinate system of the LED matrix. The physical coordinate system of the LED matrix has its origin at the top-left pixel node, with the positive x-axis pointing horizontally to the right and the positive y-axis pointing vertically downwards. Each LED pixel node corresponds to a unique coordinate in the physical coordinate system. Based on the predicted coordinates of oncoming vehicles in the 3D tensor and the projection mapping, the light pattern calculation module determines the LED matrix projection area corresponding to when an oncoming vehicle enters the preset illumination cutoff line area within a preset time period in the future, and includes the coordinates of the LED pixel nodes within this projection area into the target pixel coordinate set.

[0040] After the light pattern calculation module generates the target pixel coordinate set, it generates a pixel-level PWM control signal based on the target pixel coordinate set. The pixel-level PWM control signal contains the PWM duty cycle value corresponding to each LED pixel node in the LED matrix. The PWM duty cycle value has a linear relationship with the output brightness of the LED pixel node. The main control chip's direct memory access controller writes the pixel-level PWM control signal into the SPI bus transmit buffer between the main control chip and the LED constant current drive circuit, and transmits the pixel-level PWM control signal to the LED constant current drive circuit through the SPI bus.

[0041] The LED constant current drive circuit is electrically connected to the LED matrix, which consists of M rows and N columns of LED pixel nodes. Each LED pixel node corresponds to an independent drive channel in the LED constant current drive circuit. The LED constant current drive circuit receives pixel-level PWM control signals and updates the off or dimming status of the LED pixel nodes in the corresponding target pixel coordinate set before an oncoming vehicle enters the preset illumination cutoff line area. The preset illumination cutoff line area is a region set according to road traffic safety regulations where vehicle high beams are prohibited from directly shining into the eyes of oncoming drivers. This area corresponds to a fixed coordinate range in the pixel coordinate system of the environmental image and a fixed projection area in the physical coordinate system of the LED matrix.

[0042] In this embodiment, in order to clarify the input and output parameters and physical meaning of each core module of the system, the following parameter comparison table is constructed to standardize the core parameters during system operation.

[0043] Table 1 Comparison of Core Input and Output Parameters of Dynamic Light Pattern Adjustment System

[0044] Image input The environmental image acquired at time t, where u and v are pixel coordinates. 1920×1080×3 30Hz Point cloud input Radar point cloud data acquired at time t, where r is the distance. Radial velocity, θ is the azimuth angle, and φ is the pitch angle. 30Hz Light spot characteristics The spot pixel features extracted at time t , Let C be the centroid coordinates, C be the contour coordinate set, and A be the area of ​​the connected region. 30Hz Fusion features The fused feature vector after spatial alignment at time t, where u and v are pixel coordinates and r is the distance. Radial velocity 1920×1080×4 30Hz Prediction Tensor future Three-dimensional tensor for predicting the position of oncoming vehicles at any given time. For predicting time steps 10×1920×1080 30Hz PWM control signal LED matrix PWM duty cycle value corresponding to the coordinate pixel node 30Hz Drive output LED matrix Drive current value corresponding to the coordinate pixel node 30Hz

[0045] In the above table, The number of valid points in a single frame of point cloud data. M represents the number of effective connected components extracted from a single frame of environmental image, M represents the number of rows in the LED matrix, and N represents the number of columns in the LED matrix. This reference table standardizes the data parameters across the entire system, ensuring that data interaction between modules has a unified coordinate reference and physical meaning, and avoiding spatial alignment errors and data transmission errors caused by inconsistent parameter definitions.

[0046] In this embodiment, the total signal processing and execution time of the system includes image acquisition time, point cloud acquisition time, feature extraction and fusion time, prediction calculation time, light pattern calculation time, signal transmission time, and drive circuit state update time. The total time of the entire link does not exceed 150ms, and the preset time period is set to 200ms to ensure that the LED constant current drive circuit can complete the state update of the corresponding LED pixel node before the oncoming vehicle enters the preset illumination cutoff line area, eliminating the lag in light pattern switching caused by the physical time consumption of the signal processing link. This embodiment realizes the fusion processing of multi-source sensor data and the forward prediction of the oncoming vehicle position. Based on the prediction results, the dynamic light pattern adjustment of the LED headlight module is completed, and the light pattern state update is completed before the oncoming vehicle enters the illumination cutoff line area, avoiding the generation of transient glare.

[0047] In a preferred embodiment, spatial coordinate alignment and temporal fusion include projecting radar coordinate points from point cloud data onto the pixel coordinate system of an environmental image to generate a radar projection map, extracting fusion feature vectors from regions overlapping with spot pixel features in the radar projection map, inputting the fusion feature vectors into a temporal sliding window set within a spatiotemporal feature extraction network, the temporal sliding window extracting multiple consecutive frames of fusion feature vectors according to a preset time step and inputting them into a three-dimensional convolutional layer, and the three-dimensional convolutional layer performing convolution operations on the multiple consecutive frames of fusion feature vectors along the temporal dimension.

[0048] Specifically, refer to Figure 2 To generate a radar projection map, the radar coordinates in the point cloud data are projected onto the pixel coordinate system of the environmental image. First, the point cloud data is converted from a spherical coordinate system to a Cartesian coordinate system. The origin of the radar coordinate system is the antenna phase center of the millimeter-wave radar. The x-axis is positive forward along the vehicle's longitudinal axis, the y-axis is positive left along the vehicle's lateral axis, and the z-axis is positive upward along the vehicle's vertical axis. The conversion formula from spherical to Cartesian coordinates is as follows:

[0049]

[0050] In the formula, , , These are the three axis components of the Cartesian coordinates in the radar coordinate system. For the distance features of radar coordinate points in point cloud data, The azimuth features of radar coordinate points in point cloud data. This represents the elevation angle characteristics of radar coordinate points in point cloud data.

[0051] After completing the transformation from spherical coordinates to Cartesian coordinates, the radar coordinates in the Cartesian coordinates are transformed to the camera coordinates based on the extrinsic parameter matrix of the front-facing camera. The origin of the camera coordinates is the optical center of the front-facing camera. The X-axis is forward along the camera's optical axis with a positive direction, the Y-axis is horizontal to the left along the camera's imaging plane with a positive direction, and the Z-axis is vertically upward along the camera's imaging plane with a positive direction. The transformation formula is as follows:

[0052]

[0053] In the formula, , , These are the three axis components in the camera coordinate system. The extrinsic rotation matrix for the front-facing camera has a dimension of 3×3. The translation vector is the extrinsic parameter of the front-facing camera, with a dimension of 3×1. The rotation matrix and translation vector are obtained through joint calibration of the front-facing camera and the millimeter-wave radar. The joint calibration process is completed by the PnP algorithm based on the multi-modal calibration board and the joint optimization method of 3D point cloud registration. During the calibration process, it is ensured that the origin of the radar coordinate system and the camera coordinate system are in the same vehicle coordinate system.

[0054] After completing the transformation from the radar coordinate system to the camera coordinate system, the radar coordinate points in the camera coordinate system are projected onto the pixel plane of the environmental image using the pinhole imaging model, thus obtaining the coordinates of the radar coordinate points in the pixel coordinate system. The transformation formula is as follows:

[0055]

[0056] In the formula, , These are the coordinates in the pixel coordinate system. , This represents the equivalent focal length along the x and y axes in the intrinsic parameter matrix of the front-facing camera. , These are the pixel coordinates of the principal point of the imaging plane in the intrinsic parameter matrix of the front-facing camera. The intrinsic parameter matrix is ​​obtained through the factory calibration of the front-facing camera.

[0057] After projecting the radar coordinates to the pixel coordinate system, the projected radar coordinates are then subjected to polar coordinate rasterization to generate a depth map. Specifically, centered on the principal point coordinates of the pixel coordinate system, the pixel plane is divided into multiple equally spaced sector grids and equally spaced annular grids. The pixel value within each grid is assigned the average distance feature of all projected radar coordinates within that grid. Grids without projected radar coordinates are assigned a value of 0. The generated depth map has the same size as the environment image and is used as the radar projection map. After generating the radar projection map, invalid pixels with a confidence level below a preset threshold are removed. The confidence level is calculated using the signal-to-noise ratio (SNR) feature of the projected radar coordinates. The confidence level of a single grid is the average SNR of all projected radar coordinates within that grid. Pixels corresponding to grids with confidence levels below the preset threshold are removed, and their pixel values ​​are set to 0, resulting in the final radar projection map.

[0058] After generating the radar projection map, the fusion feature vector of the region overlapping with the spot pixel features in the radar projection map is extracted. Specifically, the connected region corresponding to the spot pixel features in the environmental image is taken as the region of interest. All pixels in the region of interest in the radar projection map are extracted. The pixel coordinates, distance features, and radial velocity features corresponding to each pixel are concatenated with the contour coordinates, centroid coordinates, and average brightness value of the spot pixel features to generate a fusion feature vector with a dimension of 1×D, where D is the feature dimension. In this embodiment, D is set to 8.

[0059] The fused feature vector is input into a temporal sliding window set within the spatiotemporal feature extraction network. The temporal sliding window extracts multiple consecutive frames of fused feature vectors according to a preset time step. The preset time step corresponds to the acquisition frame rate of the front-facing camera and millimeter-wave radar. In this embodiment, the acquisition frame rate is 30Hz, the single frame time step is 33.3ms, and the window length of the temporal sliding window is set to 5, meaning it extracts 5 consecutive frames of fused feature vectors, generating a dimension of... The temporal feature tensor is denoted by H, where H is the height of the environment image in pixels, W is the width of the environment image in pixels, and D is the dimension of the fused feature vector.

[0060] The temporal feature tensor is input into a 3D convolutional layer. The 3D convolutional layer performs convolution operations along the temporal dimension on the fused feature vectors from multiple consecutive frames. The kernel size of the 3D convolutional layer is [size missing]. Where T is the kernel size in the time dimension and k is the kernel size in the spatial dimension, in this embodiment, T is set to 3, k is set to 3, the stride of the convolution operation is 1, and the padding method is the same to ensure that the size of the output feature map is consistent with the size of the input feature map. The formula for the 3D convolution operation is as follows:

[0061]

[0062] In the formula, The output feature map of the 3D convolutional layer in the time dimension t and spatial dimension eigenvalues ​​at that location The input temporal feature tensor contains the feature values ​​at the corresponding positions. The weight parameters of the 3D convolution kernel. These are the bias parameters for the convolutional layer. is the convolution kernel index in the time dimension, and m and n are the convolution kernel indices in the spatial dimension.

[0063] The feature map output from the 3D convolutional layer is input to the pooling layer for downsampling. The pooling layer adopts max pooling with a kernel size of 2×2 and a stride of 2 to obtain high-dimensional spatiotemporal features. These high-dimensional spatiotemporal features are used for subsequent oncoming vehicle position prediction.

[0064] In this embodiment, to clarify the configuration of the operation parameters for the temporal sliding window and the three-dimensional convolutional layer, the following parameter configuration table is constructed to standardize the core operation parameters in the feature fusion process.

[0065] Table 2 Configuration of Core Operational Parameters for Temporal Feature Fusion

[0066] Timing Sliding Window Window length 5 frames Number of frames from which continuous fused feature vectors are extracted Timing Sliding Window sliding step 1 frame Number of frames per window slide 3D convolutional layer Time dimension convolution kernel size 3 Size of the convolution kernel in the time dimension 3D convolutional layer Spatial dimension convolution kernel size 3×3 Size of convolution kernel in spatial dimension 3D convolutional layer convolution stride 1 The stride of convolution operations in the time and space dimensions 3D convolutional layer Fill method Same fill Boundary padding methods for convolution operations 3D convolutional layer Number of output channels 32 Number of channels in the output feature map of a convolutional layer Max pooling layer Pooling core size 2×2 Spatial dimension and kernel size of pooling operations Max pooling layer Pooling step size 2 Spatial dimension stride of pooling operations Max pooling layer Fill method Effective filling Boundary padding methods for pooling operations

[0067] The parameter configurations in the table above can be adaptively adjusted according to different vehicle driving scenarios. In high-speed driving scenarios, the window length of the temporal sliding window is increased to improve the extraction range of temporal features; in urban road driving scenarios, the size of the spatial dimension convolution kernel is reduced to improve the extraction accuracy of spatial features.

[0068] In this embodiment, a radar projection map is generated by projecting radar point cloud data onto the pixel coordinate system of the environmental image, thereby achieving precise spatial alignment between image data and radar data. By extracting fused feature vectors from multiple consecutive frames through a time-series sliding window and performing convolution operations through a three-dimensional convolutional layer, the feature continuity extraction of multi-source data in the time dimension is achieved, avoiding feature loss and false target detection caused by single-frame image processing, and improving the stability and accuracy of oncoming vehicle feature extraction.

[0069] In a preferred embodiment, the output of a three-dimensional tensor containing the predicted coordinates of the spatial positions of oncoming vehicles within a preset future time period includes inputting high-dimensional spatiotemporal features extracted by a spatiotemporal feature extraction network into a long short-term memory network layer. The long short-term memory network layer extracts temporal state variables from the high-dimensional spatiotemporal features. The temporal state variables are input into a fully connected layer and decoded in conjunction with vehicle kinematic constraint equations. The vehicle kinematic constraint equations construct a vehicle kinematic model based on the vehicle's yaw rate and longitudinal velocity, and construct road boundary constraint equations in conjunction with the road curvature of a high-precision map. The predicted coordinates that exceed the road range are truncated and corrected so that the predicted position of oncoming vehicles remains within the drivable area of ​​the road. The three-dimensional tensor is generated by multiplying the state transition matrix with the output vector of the fully connected layer.

[0070] Specifically, refer to Figure 3 The spatiotemporal high-dimensional features output by the spatiotemporal feature extraction network are input into a Long Short-Term Memory (LSTM) network layer. The LSTM network layer contains multiple cascaded LSTM units. Each LSTM unit contains a forget gate, an input gate, a cell state gate, and an output gate. The state update formula for the LSTM unit is as follows:

[0071]

[0072] In the formula, The output of the forget gate at time t. The input gate outputs at time t. Let t represent the candidate cell state at time t. Let t represent the cell state at time t. The output of the output gate at time t. Let be the hidden state output of the LSTM cell at time t. The spatiotemporal high-dimensional features input at time t, The hidden state output of the LSTM cell at time t-1. The cell state at time t-1. , , , This is the weight matrix corresponding to the gating structure. , , , This is the bias vector corresponding to the gating structure. It is the sigmoid activation function. It is the hyperbolic tangent activation function.

[0073] The hidden state output of the Long Short-Term Memory (LSTM) network layer is the extracted temporal state variable. The temporal state variable contains temporal features such as the historical motion trajectory, speed change trend, and acceleration change trend of the oncoming vehicle. The temporal state variable is input into the fully connected layer, which performs linear transformation and nonlinear activation on the temporal state variable and outputs a feature vector with a dimension of 1×4. The four components of the feature vector are the longitudinal position, lateral position, longitudinal velocity, and lateral velocity of the oncoming vehicle at the current moment.

[0074] The output vector of the fully connected layer is decoded by combining it with the vehicle kinematic constraint equations. The vehicle kinematic constraint equations are constructed using a bicycle model, which equates the two front wheels of the vehicle to one front wheel and the two rear wheels to one rear wheel. The vertical and pitch motions of the vehicle are ignored, and only the longitudinal and lateral motions of the vehicle in the horizontal plane are considered. The vehicle kinematic constraint equations are as follows:

[0075]

[0076] In the formula, , This refers to the vehicle's longitudinal and lateral positions in the world coordinate system. The yaw angle of the vehicle. Let be the longitudinal speed of the vehicle. The steering angle of the vehicle's front wheels. This refers to the vehicle's wheelbase. For the longitudinal acceleration of the vehicle, , , , This is the first derivative of the corresponding state variable.

[0077] A kinematic model of the vehicle is constructed based on its yaw rate and longitudinal velocity. Road boundary constraint equations are then established using the road curvature from a high-precision map. Predicted coordinates outside the road boundary are truncated and corrected to ensure that the predicted position of oncoming vehicles remains within the drivable road area. The steering wheel angle and front wheel angle have a fixed transmission ratio; the front wheel angle can be calculated from the steering wheel angle. The relationship between lateral acceleration and longitudinal and yaw rates is as follows: The vehicle kinematic constraint equations are discretized to obtain the state transition matrix in the discrete time domain. The discretized sampling time step is... The formula for constructing the state transition matrix is ​​as follows:

[0078]

[0079] In the formula, Here is the state transition matrix, and the state vector is... By multiplying the state transition matrix with the output vector of the fully connected layer, the predicted coordinates of the oncoming vehicle positions at each time step within a preset future time period are calculated. The predicted coordinates of all predicted time steps are then concatenated to generate a dimension... The three-dimensional tensor is given by T, where T is the number of prediction time steps, and H and W are the pixel dimensions of the environment image.

[0080] Furthermore, yaw rate data from the vehicle's yaw rate sensor and longitudinal velocity data from the vehicle's longitudinal velocity sensor are acquired. Based on the yaw rate and longitudinal velocity data, a kinematic model of the vehicle is constructed. The kinematic model of the vehicle is also constructed using a bicycle model. Road boundary constraint equations are constructed by combining road curvature data from a high-precision map. The road boundary constraint equations define the allowable range of lateral position of oncoming vehicles in the road plane. The kinematic model of the vehicle and the road boundary constraint equations are combined to generate a state transition matrix. The state transition matrix is ​​used to truncate the predicted coordinates in the three-dimensional tensor that exceed the road boundary constraint equations. The predicted coordinates that exceed the road boundary are corrected to the coordinate values ​​corresponding to the road boundary, avoiding the light pattern adjustment error caused by the predicted coordinates exceeding the road range.

[0081] In this embodiment, the projection mapping relationship is constructed in the following way: the intrinsic and extrinsic parameter matrices of the front camera are obtained, the nonlinear refractive coefficient of the LED matrix lens is obtained, and a mapping equation from the pixel coordinate system of the environmental image to the physical coordinate system of the LED matrix is ​​established based on the intrinsic and extrinsic parameter matrices and the nonlinear refractive coefficient. The mapping equation is discretized into a mapping lookup table containing the correspondence between the physical coordinates of each LED pixel node in the LED matrix and the coordinates of each pixel in the environmental image. The mapping lookup table is stored in the flash memory chip connected to the main control chip.

[0082] Specifically, refer to Figure 4The process of establishing the mapping equation is as follows: First, the transformation relationship between the world coordinate system, camera coordinate system, pixel coordinate system, LED matrix physical coordinate system, and lens refraction coordinate system is established. The light emitted by the LED matrix is ​​refracted by the lens and projected onto the road plane in front. The projection point on the road plane corresponds to the pixel coordinates in the environment image after being imaged by the front camera. The mapping equation is the inverse process of this transformation relationship. The corresponding pixel node coordinates in the LED matrix are derived from the pixel coordinates of the environment image. The mapping equation is as follows:

[0083]

[0084] In the formula, , The coordinates are in the physical coordinate system of the LED matrix. , These are the coordinates in the pixel coordinate system of the environment image. This is the inverse of the intrinsic parameter matrix of the front-facing camera. This is the inverse of the transformation matrix from the camera coordinate system to the LED matrix coordinate system. Let be the rotation matrix of the road plane. Let be the translation vector of the road plane. Let be the refractive transformation matrix corresponding to the nonlinear refractive index of the lens. is the nonlinear refractive index of the lens.

[0085] The mapping equation is discretized according to the pixel coordinate interval of the environmental image and the pixel node interval of the LED matrix to generate a mapping lookup table. Each entry in the mapping lookup table contains the pixel coordinates of the environmental image and the corresponding physical coordinates of the LED matrix. The light pattern calculation module can directly obtain the corresponding relationship by looking up the table during operation, without performing matrix operations in real time, thus improving the calculation efficiency.

[0086] Furthermore, obtaining the nonlinear refractive index of the LED matrix lens includes attaching a thin-film temperature sensor to the surface of the LED matrix lens. The thin-film temperature sensor collects surface temperature distribution data of the LED matrix lens at a preset sampling frequency. It then retrieves pre-stored tables of the thermal expansion coefficients and refractive index temperature gradients of lens materials. Based on the surface temperature distribution data, it calculates the lens surface deformation compensation and refractive index offset at the current temperature. The nonlinear refractive index is then corrected in real time based on the lens surface deformation compensation and refractive index offset. The temperature correction formula for the lens refractive index is as follows:

[0087]

[0088] In the formula, Let T be the refractive index of the lens at the current temperature T. Standard temperature The reference refractive index of the lens is below. The refractive index temperature gradient of the lens material is obtained by looking up the refractive index temperature gradient table. The lens surface deformation compensation is calculated by looking up the linear expansion coefficient and the temperature difference from the thermal expansion coefficient table. The surface equation of the lens is corrected based on the surface deformation compensation. The nonlinear refractive coefficient is updated in real time by combining the corrected refractive index. The updated nonlinear refractive coefficient is substituted into the mapping equation to complete the real-time correction of the mapping lookup table.

[0089] In this embodiment, to clarify the temperature correction parameters for the nonlinear refractive index of the lens, the following parameter table is constructed to standardize the thermal and optical parameters of the lens material.

[0090] Table 3 Thermal and Optical Parameters of LED Matrix Lens

[0091] Reference refractive index 1.491 none Reference refractive index of lens material at standard temperature of 25°C Refractive index temperature gradient 1 / ℃ Gradient of refractive index of lens material as a function of temperature coefficient of linear expansion 1 / ℃ The coefficient of thermal expansion of lens materials Radius of curvature of reference surface 25.4 mm Reference radius of curvature of lens surface at standard temperature of 25℃ Lens center thickness 3.2 mm Reference thickness at the optical center of the lens Operating temperature range -40 85 ℃ Normal operating temperature range of the lens Temperature sampling frequency 10 Hz Temperature sampling frequency of thin-film temperature sensor

[0092] The parameters in the table above can be adjusted according to the material type and structural design of the lens. During the temperature correction process, when the maximum temperature difference of the lens surface temperature distribution exceeds the preset threshold, the refractive index correction and surface deformation compensation are performed in different regions of the lens to improve the calculation accuracy of the mapping equation.

[0093] In this embodiment, temporal state variables are extracted through a long short-term memory network layer and combined with vehicle kinematic constraint equations to construct a state transition matrix and decode to generate a three-dimensional tensor. By introducing physical motion law constraints, the divergence of predicted coordinates is avoided, improving the accuracy of oncoming vehicle position prediction. By establishing a mapping equation that includes the nonlinear refractive coefficient of the lens and performing real-time temperature correction, the accuracy of the mapping relationship between the pixel coordinate system and the physical coordinate system of the LED matrix is ​​ensured, avoiding light pattern projection errors caused by lens temperature changes.

[0094] In a preferred embodiment, generating a pixel-level PWM control signal includes fitting the edge of the anti-glare cutoff line in the LED matrix based on the target pixel coordinate set, extracting the transition zone pixel coordinate set adjacent to the edge of the anti-glare cutoff line, calculating the Euclidean distance from each pixel node in the transition zone pixel coordinate set to the edge of the anti-glare cutoff line, calculating the PWM duty cycle value of each pixel node in the transition zone pixel coordinate set according to the Euclidean distance and a preset interpolation rule, writing the PWM duty cycle value into the buffer of the direct memory access controller, and generating a pixel-level PWM control signal based on the buffer.

[0095] Specifically, refer to Figure 5The anti-glare cutoff line edge is fitted in the LED matrix based on the target pixel coordinate set. The target pixel coordinate set includes the pixel coordinates of the LED matrix projection area when an oncoming vehicle enters the illumination cutoff line area within a preset time period in the future. The upper boundary coordinate point of this area is used as the fitting sample point. The least squares method is used to perform polynomial fitting on the fitting sample point to generate the anti-glare cutoff line edge. The order of the fitting polynomial is set to 3 to ensure that the anti-glare cutoff line edge can adapt to the contour changes of oncoming vehicles and the road curvature changes.

[0096] After fitting the edge of the anti-glare cutoff line, the set of pixel coordinates of the transition area adjacent to the edge of the anti-glare cutoff line is extracted. The transition area is divided into an upper transition area and a lower transition area. The upper transition area is a region with a preset width range on the side of the anti-glare cutoff line edge facing the lighting area, and the lower transition area is a region with a preset width range on the side of the anti-glare cutoff line edge facing the anti-glare dark area. In this embodiment, the total width of the transition area is set to 8 LED pixel nodes, and the width of the upper transition area and the lower transition area is 4 LED pixel nodes each. The coordinates of all LED pixel nodes within the transition area are included in the set of pixel coordinates of the transition area.

[0097] Calculate the Euclidean distance from each pixel node in the transition zone pixel coordinate set to the edge of the anti-glare cutoff line. The formula for calculating the Euclidean distance is as follows:

[0098]

[0099] In the formula, LED pixel node coordinates The Euclidean distance to the edge of the anti-glare cutoff line, the piecewise linearized expression for the edge of the anti-glare cutoff line is: a, b, and c are the coefficients of the piecewise linear equation. For the curved segment at the edge of the anti-glare cutoff line, it is discretized into multiple continuous straight line segments. The distance from the pixel node to each straight line segment is calculated, and the minimum value is taken as the final Euclidean distance.

[0100] The PWM duty cycle of each pixel node in the transition zone pixel coordinate set is calculated according to the Euclidean distance and the preset interpolation rule. The preset interpolation rule adopts a piecewise linear interpolation method. The PWM duty cycle of the LED pixel node in the anti-glare dark zone is set to 0, that is, the off state; the PWM duty cycle of the LED pixel node in the lighting area is set to the reference duty cycle, that is, the fully lit state; the PWM duty cycle of the LED pixel node in the transition zone is linearly interpolated between 0 and the reference duty cycle according to the Euclidean distance.

[0101] Further, the current ambient light intensity value is acquired from an ambient light sensor installed inside the vehicle's windshield. The sensor collects ambient light intensity values ​​in front of the vehicle at a preset sampling frequency. Based on the current ambient light intensity value, a corresponding reference duty cycle slope is matched in a preset mapping table. The maximum and minimum allowable brightness values ​​of each pixel node in the transition zone pixel coordinate set are calculated. The maximum allowable brightness value corresponds to the reference duty cycle of the illuminated area, and the minimum allowable brightness value corresponds to the 0 duty cycle of the anti-glare dark area. A piecewise linear interpolation function is constructed based on the reference duty cycle slope, the maximum allowable brightness value, the minimum allowable brightness value, and the Euclidean distance. The Euclidean distance is input into the piecewise linear interpolation function to calculate the PWM duty cycle value. The piecewise linear interpolation function is as follows:

[0102]

[0103] In the formula, This represents the PWM duty cycle value corresponding to the Euclidean distance d. The duty cycle corresponding to the minimum allowable brightness value. The duty cycle corresponding to the maximum permissible brightness value. The boundary distance of the lower transition zone. This represents the boundary distance of the upper transition zone. As the baseline duty cycle slope, .

[0104] After calculating the PWM duty cycle values ​​corresponding to all LED pixel nodes in the LED matrix, the PWM duty cycle values ​​are written into the buffer of the direct memory access controller. The buffer adopts a dual buffer structure, including a front buffer and a back buffer. The PWM duty cycle values ​​in the current frame period are written into the back buffer. After writing is completed, the direct memory access controller switches the back buffer and the front buffer during the blanking period of the current frame period. Based on the switched front buffer, a pixel-level PWM control signal is generated and transmitted to the LED constant current drive circuit through the SPI bus.

[0105] In this embodiment, the LED constant current driving circuit completes the update of the off or dimming state of the LED pixel nodes in the corresponding target pixel coordinate set. This includes the shift register set inside the LED constant current driving circuit receiving the pixel-level PWM control signal and performing data shifting and writing within the current display frame period. After the shift register is written, it outputs a latch enable signal to the latch circuit. The latch circuit synchronously latches the data in the shift register to the LED driver output register according to the rising edge of the latch enable signal. The LED driver output register outputs the updated driving current at the beginning of the next display frame period.

[0106] Specifically, refer to Figure 6The shift register inside the LED constant current drive circuit has a bit width consistent with the total number of pixels in the LED matrix. Each bit in the shift register corresponds to the PWM duty cycle value of one LED pixel node. The pixel-level PWM control signal is input to the shift register in the form of a serial data stream. Driven by the serial clock of the SPI bus, the shift register shifts the serial data stream into the register one by one, completing the shifting and writing of all data within the current display frame cycle. After the shift register is written, a latch enable signal is generated. The rising edge of the latch enable signal is aligned with the end of the current display frame cycle. The latch circuit is triggered on the rising edge of the latch enable signal, synchronously latching all data in the shift register to the LED driver output register, avoiding light pattern flicker caused by premature updates of some LED pixel node states during the data writing process. After the LED driver output register latches the data, at the beginning of the next display frame cycle, it outputs the corresponding drive current to each pixel node of the LED matrix according to the latched PWM duty cycle value, completing the update of the LED pixel node's off or dimming state.

[0107] In this embodiment, to clarify the interpolation parameter configuration of the PWM duty cycle of the transition zone pixel node, the following parameter table is constructed to standardize the core parameters in the interpolation process.

[0108] Table 4. Transition Zone PWM Duty Cycle Interpolation Parameter Configuration Table

[0109] Total width of transition zone 8 Pixels Total pixel width of the transition area on both sides of the anti-glare cutoff line edge Lower transition zone width 4 Pixels Pixel width of the transition zone on the dark side of the anti-glare area Upper transition zone width 4 Pixels Pixel width of the transition area on one side of the illumination area Minimum duty cycle 0 % Reference PWM duty cycle for anti-glare dark areas Maximum duty cycle 100 % Reference PWM duty cycle for the lighting area interpolation polynomial order 1 none Polynomial order of duty cycle interpolation Ambient light illuminance sampling frequency 5 Hz Illuminance sampling frequency of ambient light sensor PWM signal frequency 20 kHz carrier frequency of pixel-level PWM control signal Duty cycle adjustment resolution 12 bit PWM duty cycle numerical adjustment resolution

[0110] The parameters in the table above can be adaptively adjusted according to the vehicle driving scenario and ambient lighting conditions. In the highway driving scenario, the width of the transition zone is increased to improve the smoothness of the light pattern transition. In the scenario with low ambient light, the baseline duty cycle slope is reduced to decrease the brightness change gradient of the transition zone and avoid visual discomfort caused by brightness jumps.

[0111] In this embodiment, by fitting the edge of the anti-glare cutoff line and extracting the set of pixel coordinates in the transition area, the PWM duty cycle value is calculated based on the Euclidean distance from the pixel node in the transition area to the edge of the anti-glare cutoff line. This avoids brightness jumps at the boundary between the anti-glare area and the illumination area, maintaining a smooth transition of the light pattern. Through the double buffer structure and synchronous latching mechanism, it is ensured that all pixel nodes of the LED matrix complete the state update at the same time, avoiding glare problems caused by asynchronous light pattern flickering and state updates.

Claims

1. A dynamic light pattern adjustment system for LED vehicle headlight modules based on adaptive environmental perception, characterized in that, It includes a front-facing camera, a millimeter-wave radar, a main control chip, and an LED constant current drive circuit. The front-facing camera captures images of the environment in front, and the millimeter-wave radar outputs point cloud data of the target in front. The spatiotemporal feature extraction network inside the main control chip performs spatial coordinate alignment and temporal dimension fusion of the spot pixel features in the environmental image with the distance features and radial velocity features in the point cloud data, and outputs a three-dimensional tensor containing the predicted coordinates of the spatial positions of oncoming vehicles within a preset future time period. The light pattern calculation module inside the main control chip calculates the set of target pixel coordinates in the LED matrix that need to be turned off or have their brightness reduced based on the three-dimensional tensor and the correspondence between the projected light patterns constructed from the environmental image pixel coordinate system to the LED matrix physical coordinate system based on primary and secondary light distribution. It generates a pixel-level PWM control signal. The LED constant current drive circuit receives the pixel-level PWM control signal and completes the update of the off or dimming state of the LED pixel nodes in the corresponding target pixel coordinate set before the oncoming vehicle enters the preset illumination cutoff line area.

2. The LED vehicle headlight module dynamic light pattern adjustment system based on adaptive environment perception according to claim 1, characterized in that, The spatial coordinate alignment and temporal dimension fusion includes projecting radar coordinate points in the point cloud data onto the pixel coordinate system of the environmental image to generate a radar projection map, and extracting the fusion feature vector of the region in the radar projection map that overlaps with the spot pixel features. The fused feature vector is input into a temporal sliding window set inside the spatiotemporal feature extraction network. The temporal sliding window extracts fused feature vectors from multiple consecutive frames according to a preset time step and inputs them into a three-dimensional convolutional layer. The three-dimensional convolutional layer performs convolution operations on the fused feature vectors from multiple consecutive frames along the time dimension.

3. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 1, characterized in that, The output includes a three-dimensional tensor containing the predicted coordinates of the spatial positions of oncoming vehicles within a future preset time period, comprising: inputting the spatiotemporal high-dimensional features extracted by the spatiotemporal feature extraction network into a long short-term memory network layer, and extracting temporal state variables by the long short-term memory network layer; The time-series state variables are input into a fully connected layer to perform a linear transformation, and the vehicle motion characteristics are output. Based on the vehicle-bicycle kinematic constraint equations, a state transition matrix is ​​generated by combining the longitudinal velocity, lateral acceleration, and steering wheel angle parameters of the oncoming vehicle. By combining the vehicle's kinematic model with the road boundary constraint equations of the high-precision map, the initial predicted coordinates are obtained through matrix operations, and the predicted coordinates that exceed the road boundary are truncated. By stitching together the truncated and corrected predicted coordinates at each time step, we obtain the three-dimensional tensor for predicting the spatial position of the oncoming vehicle.

4. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 1, characterized in that, The projection mapping relationship is constructed by obtaining the intrinsic and extrinsic parameter matrices of the front-facing camera and the nonlinear refractive coefficient of the LED matrix lens. A mapping equation is established from the pixel coordinate system of the environment image to the physical coordinate system of the LED matrix based on the intrinsic parameter matrix, the extrinsic parameter matrix and the nonlinear refractive coefficient. The mapping equation is discretized into a mapping lookup table containing the correspondence between the physical coordinates of each LED pixel node in the LED matrix and the coordinates of each pixel in the environment image. The mapping lookup table is stored in the flash memory chip connected to the main control chip.

5. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 1, characterized in that, The generation of pixel-level PWM control signal includes fitting the edge of the anti-glare cutoff line in the LED matrix based on the target pixel coordinate set, and extracting the set of pixel coordinates of the transition zone adjacent to the edge of the anti-glare cutoff line. Calculate the Euclidean distance from each pixel node in the transition zone pixel coordinate set to the edge of the anti-glare cutoff line. Calculate the PWM duty cycle value of each pixel node in the transition zone pixel coordinate set according to the Euclidean distance and a preset interpolation rule. Write the PWM duty cycle value into the buffer of the direct memory access controller and generate the pixel-level PWM control signal based on the buffer.

6. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 1, characterized in that, The step of updating the off or dimming state of the LED pixel node in the target pixel coordinate set includes the shift register set inside the LED constant current drive circuit receiving the pixel-level PWM control signal and performing data shifting and writing within the current display frame period; After the shift register is written, it outputs a latch enable signal to the latch circuit. The latch circuit latches the data in the shift register to the LED driver output register synchronously according to the rising edge of the latch enable signal. The LED driver output register outputs the updated driving current at the beginning of the next display frame cycle.

7. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 2, characterized in that, The step of projecting radar coordinate points from the point cloud data into the pixel coordinate system of the environmental image to generate a radar projection map includes converting the point cloud data from a spherical coordinate system to a Cartesian coordinate system. Based on the extrinsic matrix of the front-facing camera, the radar coordinates in the Cartesian coordinate system are transformed to the camera coordinate system. The radar coordinates in the camera coordinate system are projected onto the pixel plane of the environmental image using a pinhole imaging model. The projected radar coordinates are then processed by polar coordinate rasterization to generate a depth map. The depth map is used as the radar projection map, and invalid pixels with a confidence level lower than a preset threshold are removed from the radar projection map.

8. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 3, characterized in that, The generation of the three-dimensional tensor also includes: A kinematic model of the vehicle is constructed based on its yaw rate and longitudinal velocity; a road spatial constraint equation is constructed by combining the road curvature and road boundary information of the high-precision map; using the road spatial constraint equation, the predicted coordinates of oncoming vehicles that exceed the road boundary in the three-dimensional tensor are truncated and corrected; so that the predicted spatial position of oncoming vehicles is kept within the actual drivable area of ​​the road.

9. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 4, characterized in that, The nonlinear refractive coefficient of the LED matrix lens is obtained by attaching a thin-film temperature sensor to the surface of the LED matrix lens, the thin-film temperature sensor collecting surface temperature distribution data of the LED matrix lens at a preset sampling frequency, and retrieving a pre-stored table of thermal expansion coefficients and refractive index temperature gradients of lens materials. Based on the surface temperature distribution data, the lens material thermal expansion coefficient table and the refractive index temperature gradient table are consulted to calculate the lens surface deformation compensation and refractive index offset at the current temperature. The nonlinear refractive index is then corrected in real time based on the lens surface deformation compensation and refractive index offset.

10. The LED vehicle lamp module dynamic light pattern adjustment system based on adaptive environment perception according to claim 5, characterized in that, The step of calculating the PWM duty cycle value of each pixel node in the transition zone pixel coordinate set according to the Euclidean distance and a preset interpolation rule includes obtaining the current ambient light intensity value collected by the ambient light sensor, matching the corresponding reference duty cycle slope in a preset mapping table based on the current ambient light intensity value, calculating the maximum and minimum allowable brightness values ​​of each pixel node in the transition zone pixel coordinate set, constructing a piecewise linear interpolation function based on the reference duty cycle slope, the maximum allowable brightness value, the minimum allowable brightness value, and the Euclidean distance, and inputting the Euclidean distance into the piecewise linear interpolation function to calculate the PWM duty cycle value.

Images