Multi-mode Al-sensing self-adaptive brightness-adjusting low-blue-light head-up display optical system

Abstract

The invention discloses a multi-modal sensing Al self-adaptive brightness adjustment low-blue-light head-up display optical system, which is characterized in that a multi-modal data acquisition layer is arranged to fuse environment, vehicle and driver state information, and precise scene recognition is realized through a multi-modal data preprocessing layer and a core AI processing layer; the blue light adjustment decision layer generates an adjustment instruction giving consideration to eye protection, visibility and comfort, and finally the HUD brightness and the blue light content are dynamically adjusted and controlled through the multi-mode execution feedback layer. According to the method, effective suppression of harmful blue light and intelligent adaptation of display brightness in different driving scenes are achieved, the retina injury risk and the human eye adjustment burden caused by long-time use of the HUD are remarkably reduced, and the visual comfort and safety in the driving process are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of head-up display, and particularly relates to a multi-modal perception Al adaptive brightness and low-blue-light head-up display optical system. BACKGROUND

[0002] With the development of intelligent vehicles, head-up display (HUD) as an important vehicle-mounted human-computer interaction device is widely used in navigation, driving assistance information prompting and other scenarios. However, the existing HUD systems generally have the problems of excessively high eye-entrance brightness and insufficient blue light radiation control. The traditional HUD usually adopts a fixed or semi-automatic brightness adjustment strategy, and the short-wave blue light in the 415 nm-455 nm band in the backlight source is relatively strong. The blue light in this band has strong penetrating power, and long-term irradiation can easily cause the risk of retinal macular degeneration. At the same time, in the complex and changeable driving environment (such as tunnel entry and exit, night meeting, rainy and foggy weather), if the HUD display brightness is too different from the ambient light, it will cause the driver's pupil to adjust frequently, aggravate visual fatigue, and affect driving safety.

[0003] Although the HUD products on the market have been optimized in display clarity and virtual image distance, most of them do not dynamically regulate the harmful blue light content, and lack adaptive adjustment capability based on multi-modal perception. When the environmental light changes dramatically or the physiological state of the driver fluctuates, the system cannot respond and optimize the display parameters in real time, resulting in poor eye protection performance, poor human-computer interaction experience, and difficulty in balancing visual health and driving safety. SUMMARY

[0004] The purpose of the present application is to provide a multi-modal perception Al adaptive brightness and low-blue-light head-up display optical system, which realizes effective inhibition of harmful blue light and intelligent adaptation of display brightness in different driving scenarios, significantly reduces the risk of retinal damage and the burden of human eye adjustment caused by long-term use of HUD, and solves the problems raised in the background.

[0005] To achieve the above purpose, the present application adopts the following technical scheme: a multi-modal perception Al adaptive brightness and low-blue-light head-up display optical system, comprising:

[0006] A multi-modal data acquisition layer is used to acquire real-time data of the environment, the vehicle and the driver, and the multi-modal data acquisition layer comprises a visual road condition module, a voice interaction module, a tactile feedback module, a physiological state module and a traditional upgrade module;

[0007] A multi-modal data preprocessing layer is connected to the multi-modal data acquisition layer and is used to perform time series alignment, missing data completion, modal purification and feature standardization processing on the acquired original data, and output a standardized feature matrix of uniform dimension;

[0008] a core AI processing layer connected to the multi-modal data preprocessing layer, configured to perform multi-modal feature fusion on the standardized feature matrix, and identify a current driving scene type based on a fusion result, and output a benchmark adjustment strategy;

[0009] a blue light adjustment decision layer connected to the core AI processing layer, configured to generate final HUD brightness and blue light content adjustment instructions based on the benchmark adjustment strategy, under the constraints of eye protection, visibility and comfort;

[0010] a multi-modal execution feedback layer connected to the blue light adjustment decision layer, configured to execute the adjustment instructions and synchronously collect execution effect feedback data, forming an execution-verification closed loop;

[0011] a cloud-local closed loop iteration layer connected to the multi-modal execution feedback layer and the core AI processing layer, configured to perform local fine-tuning and cloud model retraining based on the feedback data, and realize continuous optimization of the algorithm.

[0012] Preferably, the visual road condition module includes a front-view high-definition camera and a HUD projection area camera.

[0013] Preferably, the visual road condition module includes a front-view high-definition camera and a HUD projection area camera.

[0014] Preferably, the speech interaction module includes a microphone array and a speech processing chip.

[0015] Preferably, the speech interaction module includes a microphone array and a speech processing chip.

[0016] Preferably, the physiological state module includes a non-contact heart rate sensor and an eye state detector.

[0017] Preferably, the physiological state module includes a non-contact heart rate sensor and an eye state detector.

[0018] Preferably, the traditional upgrade module includes a spectral subdivision detector and a CAN bus.

[0019] Preferably, the traditional upgrade module includes a spectral subdivision detector and a CAN bus.

[0020] Preferably, the core AI processing layer adopts a lightweight cross-modal Transformer model for multi-modal feature fusion, realizes deep fusion of vision, voice, physiology, touch and vehicle signals through modal embedding and cross-modal attention calculation, and combines a rule engine and a supervised learning model to complete recognition of 15 types of compound driving scenes.

[0021] Preferably, the blue light adjustment decision layer adopts a multi-objective particle swarm optimization algorithm to generate an optimal combination of adjustment parameters considering eye protection, visibility and comfort under the constraint conditions of a blue light hazard coefficient H_B < 0.08 mW.h / sr.m2, a visibility coefficient V > 0.85 and a comfort coefficient C > 0.7.

[0022] Preferably, the multi-modal execution feedback layer adjusts the proportion of harmful blue light by controlling the flip frequency of 415 nm / 455 nm / 480 nm three-waveband blue light micro-mirrors in the DLP chip, and maintains color balance by compensating the flip frequency of red and green light micro-mirrors, while collecting spectrometer, AR chip and driver feedback data to realize closed-loop verification.

[0023] Preferably, the cloud-local closed-loop iteration layer comprises a local fine-tuning module and a cloud iteration module.

[0024] The local fine-tuning module fine-tunes the initial constraint threshold of MOPSO based on 7-day feedback data using the gradient descent method, and the cloud iteration module pushes the AI model retrained by the cloud to the vehicle end through OTA.

[0025] Preferably, it also comprises a fault reduction processing mechanism, when any sensor fails, the system automatically switches to a backup sensing path and executes a bottom guarantee strategy.

[0026] The technical effects and advantages of the present application are as follows:

[0027] The present application fuses environmental, vehicle and driver state information through the multi-modal data acquisition layer, and realizes accurate scene recognition through the multi-modal data preprocessing layer and the core AI processing layer, and then generates adjustment instructions considering eye protection, visibility and comfort through the blue light adjustment decision layer, and finally dynamically adjusts and controls the brightness and blue light content of the HUD through the multi-modal execution feedback layer. The method realizes effective suppression of harmful blue light and intelligent adaptation of display brightness in different driving scenes, significantly reduces the risk of retinal damage and the burden of human eye adjustment caused by long-term use of the HUD, and improves the visual comfort and safety during driving. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1This is a block diagram of the multimodal sensing AI adaptive brightness adjustment low blue light head-up display optical system of the present invention;

[0029] Figure 2 This is a block diagram of the multimodal data acquisition layer of the present invention;

[0030] Figure 3 This is a layout diagram of the multimodal data acquisition layer of the present invention;

[0031] Figure 4 This is a schematic diagram of the optical mechanism for achieving a white light source through RGB three-color mixing, as described in this invention.

[0032] Figure 5 This is a schematic diagram of the MicroLED or OLED imaging unit of the present invention. Detailed Implementation

[0033] 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 embodiments of the present invention, and not all embodiments. The specific embodiments described herein are merely used to explain the present invention and are not intended to limit 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.

[0034] This invention provides a multimodal sensing AI adaptive brightness adjustment low blue light head-up display optical system, such as... Figure 1 As shown, the system works by reflecting light emitted from the image display unit onto the windshield via mirrors of varying sizes. The light then forms an image on the windshield's mirror surface, allowing the human eye to see a virtual image. The system employs a multi-modal data acquisition layer to simultaneously collect real-time data from the environment, vehicle, and driver. The second step involves multi-modal data preprocessing to address data heterogeneity, temporal discrepancies, and noise interference, outputting a standardized feature matrix. The third step involves a core AI processing layer that fuses multi-modal features, accurately identifies the scene, and outputs a baseline strategy to resolve modal conflicts. The fourth step involves blue light adjustment decision-making, seeking a balance between eye protection, visibility, and comfort to generate the final adjustment command. The fifth step involves multi-modal execution feedback to implement the adjustment command, providing synchronous feedback on the effect, forming an execution-verification closed loop. The sixth step involves cloud-local closed-loop iteration, continuously optimizing and adding fault redundancy, improving algorithm adaptability through data iteration, and resolving issues such as lower-than-normal performance and modal failure.

[0035] For example, a multimodal data acquisition layer is used to collect real-time data on the environment, vehicle, and driver, such as... Figure 2 and Figure 3As shown, the multi-modal data acquisition layer includes a visual road condition module, a voice interaction module, a tactile feedback module, a physiological state module, and a traditional upgrade module. Through special sensors and synchronous triggering, ambient light, road conditions, driver state, and vehicle signals are converted into quantifiable electrical signals, solving the problems of data comprehensiveness and real-time performance in dynamic automotive scenarios.

[0036] The visual road condition module includes a front-view high-definition camera and a HUD projection area camera. The front-view high-definition camera uses a global shutter CMOS imaging technology and has a 120dB wide dynamic range and a MIPI-CSI2 interface. The HUD projection area camera has a low-light enhancement technology and a region-of-interest collection function.

[0037] The voice interaction module includes a microphone array and a voice processing chip. The microphone array is arranged in an equilateral triangle and uses beamforming technology. The voice processing chip has echo cancellation and voice endpoint detection functions.

[0038] The physiological state module includes a non-contact heart rate sensor and an eye state detector. The non-contact heart rate sensor uses a photoplethysmography method and has a built-in 3-axis acceleration sensor for motion artifact suppression. The eye state detector collects blink frequency and eyelid opening degree information through a binocular high-definition camera.

[0039] Photoplethysmography (PPG): 940nm infrared light is emitted (penetrates the skin depth of 0.5-1mm), hemoglobin in the blood absorbs this band light, and the change of blood flow caused by heart beating leads to the fluctuation of reflected light intensity. The receiving end converts the light intensity change into a current signal (μA level) through a photodiode. Motion artifact suppression: a built-in 3-axis acceleration sensor detects the sensor displacement caused by steering wheel vibration (such as vehicle jolt), and through "adaptive filtering", it eliminates vibration interference, so that the heart rate detection error is <2bpm (when walking / turning).

[0040] The traditional upgrade module includes a spectrum subdivision detector and a CAN bus. The spectrum subdivision detector is built-in with 415 nm / 455 nm / 480 nm three-band optical filters to distinguish between harmful blue light (415-455 nm, weight coefficient K=1) and beneficial blue light (455-480 nm, K=0.1). The CAN bus adopts a flexible data rate mode and has a fault tolerance mechanism. The ambient light brightness is synchronously collected. The ambient light intensity is converted into a 0-3.3V analog voltage through a "full-band photosensitive resistor" (response range 380-780 nm), and is converted into a digital value (0.1-100000 lux) through a 12-bit ADC at a sampling frequency of 10 Hz. The "flexible data rate (FD)" mode is adopted, the arbitration section rate is 250 kbps (to ensure that the bus conflict priority transmits safety signals), the data section rate is 2 Mbps (to adapt to high-frequency update signals such as vehicle speed and windshield wiper), and the single frame data length can reach 64 bytes (only 8 bytes for a traditional CAN); the bus error detection logic (bit error, CRC error) is built-in. When an error is detected, an error frame is automatically sent, and the system switches to a "passive mode" to avoid affecting other nodes on the bus. The fault recovery time is less than 100 ms, which meets the ISO11898-2 vehicle standard.

[0041] The tactile feedback module includes steering wheel touch keys and seat pressure sensors. The steering wheel touch keys are mainly used to receive the driver's steering wheel tactile instructions, such as short press touch keys to increase blue light, long press to increase blue light, etc. The seat pressure sensors are used to sense the driver's sitting posture, such as the driver's mental focus and forward leaning posture, to automatically adjust the HUD brightness and blue light content.

[0042] Exemplarily, the multi-modal data preprocessing layer is connected to the multi-modal data collection layer and is used for time sequence alignment, missing data completion, modal purification and feature standardization processing on the collected original data, and outputs a standardized feature matrix of uniform dimension, which is specifically as follows:

[0043] 1) Global timestamp synchronization principle:

[0044] Hardware triggering: the main control SoC outputs a 1 MHz synchronization clock signal, and all sensors (camera, microphone, heart rate sensor) are triggered to collect data based on this clock, ensuring that "the same physical signal at the same time" corresponds to "the same timestamp";

[0045] Timestamp calibration: through the fusion of "vehicle-mounted GPS (precision 100 ns) + RTC real-time clock (precision 1 ms)", an "year-month-day-hour-minute-second-millisecond" timestamp is added to each frame of collected data, such as a visual frame t=100.001 ms and a voice frame t=100.002 ms, with a deviation controlled to be less than 1 ms.

[0046] Missing data completion principle:

[0047] Linear interpolation algorithm: for low-frequency sensors (such as physiological sensors 10 Hz, camera 30 fps), the missing value in the middle is calculated by the adjacent two effective data points, the formula is:

[0048]

[0049] For the completed target data value, For the measured data value at time t, For the measured data value at time t, t is the "target time" that needs to be completed, is the nearest "effective sampling time" before the target time t, is the nearest "effective sampling time" after the target time t.

[0050] Example: =100ms heart rate 70bpm, =110ms heart rate 72bpm, complete t=105ms heart rate 71bpm, ensure continuous timing data without breakpoints.

[0051] Modal purification: eliminate noise interference specific to the automotive scene;

[0052] Visual noise reduction (multi-scale Retinex dehazing + strong light suppression); Multi-scale Retinex dehazing principle: decompose the foggy image into "reflective component (clear road conditions)" and "illumination component (fog noise)", the formula is:

[0053] ;

[0054] I(x,y) is the pixel gray value of the image (0-255), the original fog / rain image collected by the front-view camera, containing "clear road condition information" and "fog / light interference", is the input data for dehazing processing; (x,y) represents the pixel coordinates of the image, corresponding to the two-dimensional position of the camera imaging plane;

[0055] (x,y) is the two-dimensional spatial coordinate of a single pixel in the image, x represents the horizontal pixel index (from left to right), y represents the vertical pixel index (from top to bottom), used to locate the position of each pixel;

[0056] R(x,y) is the reflective component of the image, i.e. the "true texture information" of the object itself, such as lane lines, tunnel entrance contour, and oncoming vehicle body, which is the core data to be preserved after dehazing processing and is directly used for scene recognition of HUD (such as judging whether to enter a tunnel);

[0057] L(x, y) is the illumination component of the image, containing interference information such as "fog, backlight, uneven light", which is the main cause of image blur and needs to be estimated and separated by algorithm.

[0058] logI(x, y) is the logarithmic gray value of the original image, and the logarithmic operation is performed on each pixel of the original fog image I(x, y), which weakens the gray value of the strong light area (such as the highlight reflection area in fog) and enhances the dark details (such as the shadow of the vehicle in fog);

[0059] logL(x, y) is the logarithmic estimate value of the illumination component, which is smoothed by "multi-scale Gaussian filtering" (σ=15 / 30 / 60, corresponding to different blur degrees) on logI(x, y), and the logarithmic form of the illumination component is estimated - the larger the Gaussian filter kernel, the closer the estimated L(x, y) to the global illumination trend (such as the overall distribution of fog);

[0060] logR(x, y) is the logarithmic gray value of the reflection component. After separating the illumination interference, the logarithmic form of the "clear texture information" is obtained, and then it is restored to the original gray range (0-255) through exponential transformation (R(x, y)=elogR(x, y)), that is, the final dehazing image.

[0061] Through Gaussian filtering (multi-scale σ=15 / 30 / 60) to estimate the illumination component L(x, y), and then reconstruct the reflection component R(x, y), the rain image dehazing is realized (the contrast after dehazing is improved by 50%); adaptive strong light suppression: for the oncoming high beam area (brightness >5000lux), use "piecewise linear compression" to compress the pixel value from 0-255 to 0-51, and keep the brightness of other areas, to avoid the loss of road conditions caused by overexposure in strong light area.

[0062] 2) Voice purification:

[0063] Noise suppression: based on the "deep noise suppression (DNS)" lightweight model (parameters <500,000), input noisy speech MFCC features, output clean speech features;

[0064] Endpoint detection (VAD): calculate the energy of the speech frame, when the energy > threshold (-30dB) and duration > 20ms, it is determined as valid speech, otherwise it is rejected;

[0065] 3) Physiological noise reduction (Kalman filter + wavelet transform):

[0066] Kalman filter to suppress heart rate vibration noise: through "prediction-update" cycle to eliminate the heart rate jump caused by seat vibration, the core formula is:

[0067] Prediction: (Previous heart rate prediction of the current heart rate);

[0068] Update: (Modified predicted value combined with the current measured heart rate Z_K);

[0069] Prediction stage:

[0070] Real number (physiological parameter estimate value, such as heart rate: bpm), prior estimate value at time k: physiological parameter (such as heart rate) estimated by the last time's correction value without combining the current measured data, "-" represents "prior" (unmodified);

[0071] k is an integer (time step index), discrete time step: corresponding to the sampling period of the physiological sensor (10Hz in the system, i.e. 100ms for one time step, k=1 corresponds to 100ms;

[0072] A is a real number (state transition matrix, usually 1) state transition coefficient: describes the influence degree of "last time physiological state" on "current time state". Because the physiological parameters such as heart rate and HRV change gently, A=1 in the system (assuming that there is no sudden change in heart rate at adjacent time).

[0073] B is a real number (control input matrix, usually 0) control input coefficient: describes the influence of "external active control signal" on physiological state. The physiological parameters (heart rate, HRV) in the system are passive physiological signals collected, without active control input, so B=0;

[0074] Uk is a real number (control input signal) external control input: signal actively adjusting physiological state (such as drugs, exercise, etc.), there is no such input in the system, so Uk=0;

[0075] is a real number (physiological parameter correction value, such as heart rate: bpm) k-1 posterior estimate value: the "true physiological parameter" after the last time's "prediction-update" closed-loop correction (which has eliminated the vibration interference of the last time), which is the basis for the current prediction.

[0076] Update stage:

[0077] : real number (physiological parameter correction value, such as heart rate: bpm) k posterior estimate value: the "final true physiological parameter" after the Kalman gain weighted correction combined with the current measured data and the prior estimate value, which is directly used for AI fatigue judgment (such as HRV<50ms to determine fatigue).

[0078] Kk: non-negative real number (Kalman gain), weight coefficient: dynamically adjusts the "prior estimate value Deviation from actual test;

[0079] Contribution ratio of each feature;

[0080] Sensor raw reading at time k: raw data collected by non-contact heart rate sensor, which may contain vibration interference (e.g. when a vehicle passes through a speed bump, the measured heart rate jumps from 70 bpm to 90 bpm);

[0081] H: real number (observation matrix, usually 1) observation coefficient: describes the mapping relationship between "true physiological state" and "sensor measured value". In the system, the sensor directly collects heart rate and HRV, and the measured value is linearly corresponding to the true value;

[0082] Residual error: the difference between the current measured value and the prior estimate, reflecting the "reliability of measured data" - large deviation (e.g. 90-70=20 bpm) usually means that the measured value is disturbed by vibration, and small deviation (e.g. 72-70=2 bpm) means that the measured value is reliable.

[0083] Where the gain K_k is dynamically adjusted to ensure that the heart rate error caused by vibration (e.g. from 70 bpm to 90 bpm) is corrected to a reasonable value;

[0084] Wavelet transform eliminates EOG power frequency interference: db4 wavelet basis is used to decompose EOG signal into 6 layers, and high frequency coefficients corresponding to 50 Hz power frequency interference are removed, and then the signal is reconstructed, so that the blink recognition accuracy is improved from 85% to 95%.

[0085] Feature standardization: unify the dimension and magnitude of multi-modal data

[0086] Numerical feature normalization (Min-Max):

[0087] Principle: map continuous values such as brightness, heart rate, and speed to the [0, 1] interval to eliminate "magnitude difference caused by algorithm weight deviation" (e.g. speed 100 km / h and heart rate 70 bpm have different magnitudes, direct input will cause the algorithm to deviate to speed), the formula is:

[0088]

[0089] Example: ambient light brightness =0.1 lux, =100000 lux, when x=1000 lux, =(1000-0.1) / (100000-0.1)≈0.01, after standardization, the magnitude is unified with other features.

[0090] Image / speech feature extraction and dimension unification:

[0091] Image feature: extract 256-dimensional feature vector of the HUD projection area ROI (256x256) through MobileNetV2's "depth separable convolution" (split 3x3 convolution into 1x1 point convolution + 3x3 depth convolution), and then perform L2 normalization (make the length of the feature vector = 1) to eliminate brightness effects;

[0092] Speech feature: frame the speech signal with a frame length of 20ms and a step length of 10ms, extract 13-dimensional MFCC (Mel Frequency Cepstrum Coefficient), expand to 128 dimensions through "Mel spectrum mapping", and unify the image feature dimension, finally form a three-dimensional matrix of "30-step length x 5 modality x 256 features".

[0093] Exemplarily, the core AI processing layer is connected to the multi-modal data preprocessing layer, and is used for multi-modal feature fusion of the standardized feature matrix, and identification of the current driving scene type based on the fusion result, and output of a benchmark adjustment strategy; the core AI processing layer adopts a lightweight cross-modal Transformer model for multi-modal feature fusion, realizes deep fusion of vision, speech, physiology, touch and vehicle signals through modal embedding and cross-modal attention calculation, and completes identification of 15 types of complex driving scenes in combination with a rule engine and a supervised learning model. Specifically as follows:

[0094] Multi-modal feature fusion (lightweight cross-modal Transformer)

[0095] 1) Modal embedding principle:

[0096] Assign a combination structure of "one-hot encoding + modal feature vector" to each type of modality, such as:

[0097] Visual modality:

[0001] + 256-dimensional image feature;

[0098] Speech modality:

[0010] + 256-dimensional MFCC feature;

[0099] Purpose: to let the model clearly distinguish "different types of features" (such as avoiding confusion between "brightness features" of images and "frequency features" of speech), and the feature dimension after embedding is still 256-dimensional (4 bits for encoding and 252 bits for features, and the dimension is kept unchanged through compression algorithm).

[0100] Cross-modal attention calculation principle: calculate the correlation weight of different modal features based on "scaled dot-product attention", and the formula is

[0101] ;

[0102] Attention(Q, K, V): matrix multi-modal fusion feature, dimension: [modal number d_k], cross-modal attention output: the "fusion feature matrix" obtained by calculating the weight through Q-K similarity and weighting V, directly input into the subsequent feature aggregation layer for scene recognition (such as "fatigue + tunnel" scene classification);

[0103] Q: matrix (query modal feature, dimension, [m, d_k]; query modal feature matrix: "target modal feature" that needs to be focused on (such as judging whether it is a "tunnel scene", Q is the tunnel outline feature of the visual modal), the number of features of the query modal m is the number of features of the query modal (m=1 in the system, single modal single feature vector), dk is the feature dimension (dk=256 in the system, consistent with the dimension of the multi-modal feature matrix);

[0104] K: matrix (key modal feature, dimension, key modal feature matrix: [n, d_k]; used to calculate "association similarity" with Q (such as vehicle CAN speed feature, physiological HRV feature, voice instruction feature), the number of keys (n=4 in the system, 4 types of modal except Q), dk=256 (consistent with the dimension of Q, to ensure that the matrix can be multiplied).

[0105] V: matrix (value modal feature, dimension [n, d_k]); value modal feature matrix: "modal original feature" corresponding to K (K is used to calculate similarity, V is used to output original data), dimension consistent with K ([n, d_k]), V and K in the system come from the same modal (such as K and V of vehicle modal are speed feature vectors).

[0106] QKT: matrix (similarity matrix, dimension: [m, n]); similarity calculation between Q ([m, d_k]) and the transpose of K KT, dimension: [d_k, n]), the result is a similarity matrix of [m, n], each element represents the "association strength" between a feature of Q and a feature of K (the larger the value, the stronger the association);

[0107] : real number (scaling factor, DK=256 in the system); dimension normalization factor: used to alleviate the problem of "DK feature dimension being too large causing similarity value overflow" - when DK increases, the element value of QKT will increase sharply, causing the weight after softmax to tend to 0 or 1, and the gradient to disappear;

[0108] Softmax(): weight normalization function: QKT / similarity matrix, normalize each row to sum up to 1, output "attention weight matrix of each modality" (dimension: [m, n]), ensure that the weight can be directly used for weighted summation.

[0109] Q (query) is the current modality feature, K (key) is the other modality feature, d_k=256 (feature dimension), and the weight is normalized by softmax;

[0110] Vehicle scale modality weight example: "visual tunnel entrance" and "CAN vehicle speed 60km / h" weight = 0.8 (strong association), and "voice no instruction" weight = 0.2 (weak association), ensuring that key modalities play a dominant role in decision-making.

[0111] 3) Feature aggregation principle

[0112] Use "2-head attention + residual connection": divide 256-dimensional features into 2 groups (128-dimensional each) to calculate attention, and then concatenate them into 256-dimensional; At the same time, add residual connection (aggregated features + original embedding features) to avoid deep network gradient vanishing;

[0113] Computing power adaptation: through "model quantization (32-bit floating point-8-bit integer)" and "layer pruning (remove redundant convolution layers)", the model parameters are reduced from 1.2M to less than 3 million, and the inference time on NPU (1.5TOPS) is <30ms, meeting the real-time requirements of cars.

[0114] Scene recognition decision (rule engine + supervised learning):

[0115] 1) Modality priority determination principle:

[0116] Based on "automobile functional safety priority", rules are formulated, and the core logic is:

[0117] Safety level (highest): visually recognize "collision warning", CAN detect "emergency brake" - pause blue light adjustment, and prefer to increase HUD brightness to 1000cd / ㎡ (to ensure that the driver focuses on the road conditions);

[0118] Intention level (high): driver's voice "lower blue light", touch long press instruction - combined with ambient light data fine-tuning (such as reducing blue light from 30% to 20% in strong light, instead of 10%, to avoid visibility decline);

[0119] State level (medium): physiological HRV <50ms (fatigue), visually recognize "tunnel" - actively adjust and voice confirmation (execute within 3s without objection);

[0120] Environmental level (low): slight changes in ambient light, seat pressure fluctuations - only as a reference, do not actively trigger adjustment.

[0121] 2) 15 categories of scene classification principle:

[0122] Supervised learning training: based on 500,000+ car scene samples (covering rainy days + voice, fatigue + tunnel, etc. Compound scene), input the "global feature vector" into the fully connected layer, and output the probability distribution of 15 categories of scenes (such as "fatigue + tunnel" probability = 0.92, "ordinary night" probability = 0.05);

[0123] Classification accuracy guarantee: Use "cross-entropy loss function" to optimize the model, measure the gap between the scene probability distribution predicted by the AI model and the true scene label, and guide the model iteration optimization (such as improving the classification accuracy of "fatigue + tunnel", "voice + rainy day" and other compound scenes) The formula is:

[0124]

[0125] Among them is the true scene label (0 / 1) of sample i, is the model prediction probability, and the training scene classification accuracy is greater than 98%;

[0126] LOSS: Model overall loss: the average loss value of all training samples, reflecting the "overall deviation degree" of the model prediction result and the true scene - the smaller the loss, the more accurate the model classification (system training goal Loss <0.05, corresponding to scene classification accuracy >98%)

[0127] N: Total number of training samples: The total amount of "multi-modal scene samples" used for model training, N = 500,000 in the system (covering all complex scenes such as rainy days, tunnels, fatigue, and voice instructions, each category of scene sample ≥30,000, ensuring data balance)

[0128] : Sample average operation, take the average of the loss values of all N samples, eliminate the influence of "sample number difference" on the loss value (such as the total loss of 100 samples and the total loss of 1000 samples cannot be directly compared, and after averaging, they can be compared horizontally).

[0129] : Sample dimension summation: traverse all training samples from the first (i = 1) to the Nth (i = N), calculate the loss of each sample and accumulate it to get the "total loss";

[0130] i: The number of a single training sample: used to locate a specific multi-modal scene sample such as "rainy day + voice instruction 'lower blue light' + vehicle speed 50km / h", which contains complete multi-modal data of vision / voice / physiology / vehicle);

[0131] C: Total number of scene categories: Total number of "multimodal composite scenes" that the model needs to distinguish, C = 15 in the system (such as "fatigue + tunnel", "voice + rainy day", "tactile adjustment + night driving", etc. 15 core scenes).

[0132] : Sum of category dimensions: For each sample, from the 1st (C = 1) to the Cth (C = 15) of all scene categories, the "local loss" of the sample in each scene category is calculated and added up to get the "loss of a single sample";

[0133] c: Number of a single scene category: used to locate a specific composite scene (such as c = 3 for "fatigue + tunnel" scene, c = 8 for "voice command + rainy day" scene, the number corresponds to the "scene-strategy mapping table")

[0134] : True class label of sample i: "one-hot encoding" is used (i.e. each sample corresponds to only one true class, =1 other classes yic=0), such as sample i is "fatigue + tunnel" (c = 3), then =1, =, =....=, =0;

[0135] : Model predicts the probability that sample i belongs to category c: the probability value output by the AI model (such as lightweight Transformer), reflecting the model's confidence that "sample i is category c". For example, sample i is "fatigue + tunnel", the model predicts =0.92 high confidence), =0.03 (low confidence);

[0136] Log(.): Log penalty operation, taking the logarithm of the predicted probability pic, the core function is to "amplify the penalty of low probability" - when approaches 0, the prediction is wrong), log( ) approaches infinity, the loss value increases sharply; when approaches 01, log( ) approaches 0, the loss value is small;

[0137] -: loss value conversion to positive, because belongs to [0, 1], log( ) ≤ 0, after taking the negative sign, makes log( ) ≥ 0 to ensure that the overall LOSS is non-negative (for ease of understanding and optimization, the loss value cannot be negative).

[0138] 3) Reference strategy output principle:

[0139] Pre-stored "scene-strategy mapping table" to vehicle-mounted Flash (partition storage, commonly used scene cache), such as:

[0140] Scene "fatigue+tunnel" - blue light ratio 8%, HUD brightness 8000 nits, AR contrast 70%;

[0141] Scene "voice+rainy day" - blue light ratio 15%, HUD brightness 1500 nits, AR contrast 75%;

[0142] Strategy making basis: reference IEC62471 blue light safety standard (H_B<0.08 mW·h / sr·m²) and human eye visual comfort experimental data (1000 drivers test, 90% approved parameter interval).

[0143] Exemplarily, the blue light adjustment decision layer is connected to the core AI processing layer, and is used to generate final HUD brightness and blue light content adjustment instructions based on the reference adjustment strategy, under the constraints of eye protection, visibility and comfort three targets. The blue light adjustment decision layer adopts a multi-objective particle swarm optimization algorithm to generate an optimal combination of final adjustment parameters that take into account eye protection, visibility and comfort under the constraints of blue light hazard coefficient H_B<0.08 mW·h / sr·m², visibility coefficient V>0.85, and comfort coefficient C>0.7. Specifically as follows:

[0144] 1) Three-target optimization (multi-objective particle swarm optimization MOPSO):

[0145] 1. Blue light hazard coefficient (HB) calculation: based on IEC62471 standard, calculate the hazard of HUD blue light to retina;

[0146] Formula: KB(λ) t; Φ(λ) is the blue light flux, KB(λ) is the hazard weight, t is the gaze duration, and HB is constrained to be less than 0.08 mWh / srm²;

[0147] Imaging unit, blue light source band avoids harmful blue light band of 415-455nm, shifts to long wave band above 460nm, meets the requirements of Rhine eye protection:

[0148]

[0149] The ratio of harmful blue light of 415-455nm to blue light band of 400-500nm is less than 50%;

[0150] 2. Visibility Coefficient (V): HUD brightness, AR contrast, color restoration;

[0151] Formula: V = 0.5LHUD + 0.3CAR + 0.2Rcolor, constraint V>0.85 (ensure navigation clear);

[0152] 3. Comfort Coefficient (C):

[0153] Fusion of driver physiological feedback (ΔHRV: heart rate variability change), voice confirmation (S_voice: 1=satisfied, 0=unsatisfied), tactile operation (S_touch: 1=no correction, 0=manual correction), formula: C=0.4ΔHRV+0.3Svoice+0.3Stouch, constraint C>0.7 (ensure driver comfort);

[0154] C: Comprehensive comfort score: quantitative result of physiological, voice, and tactile multi-modal feedback-C is closer to 1, the more comfortable the driver is with the current HUD adjustment (blue light ratio, brightness, AR parameters); System constraint requires C>0.7 (otherwise, the adjustment parameters need to be re-optimized);

[0155] 0.4: physiological feedback weight, ΔHRV (heart rate variability change) weight ratio, is the highest among the three factors-physiological signals such as HRV are "unconscious objective feedback", which can better reflect comfort than voice / tactile "active feedback" (e.g., the driver may not realize it, but the physiology is already showing fatigue);

[0156] ΔHRV: Physiological comfort indicator: the difference between the current driver's heart rate variability (HRV) and the "comfort baseline HRV", the normalized result-ΔHRV is larger, the more relaxed the physiological state is (the higher the comfort is).・Calculation logic;

[0157] 0.3: voice / tactile feedback weight: lower but higher than single feedback-because voice / tactile is "subjective demand actively expressed by the driver", it needs to be complementary to physiological objective feedback (e.g., the driver feels comfortable but HRV is slightly reduced, follow the active feedback);

[0158] Svoice: voice feedback identifier: "voice confirmation result" of the driver on the HUD adjustment effect, used to directly obtain subjective evaluation-Svoice=1 the driver actively says "comfortable" "can" and other positive feedback, or confirms the adjustment (e.g., the system asks "whether to reduce blue light", and answers "yes"); Svoice=0 the driver says "too dark" "yellowish" and other negative feedback, or refuses to adjust (e.g., answers "no");・When there is no voice: default Svoice=0.5 neutral, not biased towards positive / negative;

[0159] Stouch touch operation identifier: indirectly judge subjective comfort through the driver's touch operation on the steering wheel / center control-Stouch=1: After adjustment, the driver does not manually correct the HUD parameters (such as not pressing the "increase blue light" "brighten" button), or long press "confirm" key to indicate satisfaction; Stouch=0: After adjustment, the driver manually corrects the parameters within 30 seconds (such as adjusting the blue light from 15% to 20%), or short press "cancel" key to indicate dissatisfaction;

[0160] Example: ΔHRV=0.2 (HRV increases by 20%, indicating comfort), S_voice=1, S_touch=1, then C=0.4×0.2+0.3×1+0.3×1=0.78;

[0161] 4. MOPSO optimization principle: initialize 100 "particles" (each particle represents a set of adjustment parameters: blue light ratio B, brightness L_hud, AR contrast C_ar); Particle update: guide particle movement through "individual optimal (Pbest)" and "global optimal (Gbest)", formula:

[0162] ;

[0163] = + ;

[0164] i: the number of individual optimization particles: each particle represents a set of HUD adjustment parameter combinations (blue light ratio, HUD brightness, AR contrast), the total number of particles in the system is usually set to 100 (balance computing power and optimization accuracy);

[0165] d: the dimension of the HUD adjustment parameter: d=3 in the system corresponds to 3 core adjustment parameters-d=1: blue light ratio (5%-30%); d=2, HUD brightness (5000-20000 nits); d=3: AR contrast (50%-80%).

[0166] t: iteration step of MOPSO algorithm: each iteration corresponds to a parameter optimization attempt, the number of iterations in the system is set to 50 (to ensure that the optimization is completed within 100ms, meeting the real-time requirements of vehicle-mounted);

[0167] t+1: next iteration step: based on the speed and position of the current iteration (t), calculate the parameter adjustment direction and final value of the next iteration (t+1);

[0168] : adjustment amplitude of particle i in dimension d at iteration t: represents the adjustment amount of "dimension d parameter" of particle i in the next iteration (positive number = increase parameter, negative number = decrease parameter);

[0169] The adjustment magnitude of particle i in dimension d at iteration t+1: based on the current position (xid(t)) and the updated velocity;

[0170] The new parameter values ​​obtained serve as the starting point for the next iteration of optimization.

[0171] The best historical parameter value of particle i in dimension d at iteration t (i.e., Pbest): Among all parameters that satisfy the "three objective thresholds" from the beginning of iteration to iteration t, the dimension parameter value with the best overall performance (e.g., minimum, maximum);

[0172] The globally optimal parameter values ​​(Gbest) of all particles in dimension d at iteration t: The parameter values ​​in dimension d corresponding to the "Pareto optimal solution" selected from all particles' Pbest solutions (in multi-objective optimization, there is no unique global optimum; Gbest is the best solution that is comprehensively selected from the Pareto optima).

[0173] w: Velocity inertia coefficient: controls the degree to which a particle "retains its current velocity trend" - the larger w is, the more the particle tends to maintain its current optimization direction (strong stability); the smaller w is, the easier it is for the particle to be guided by Pbest / Gbest (strong exploratory nature);

[0174] Pbest guiding weight: controls the strength of the particle's "learning from its own historical best (Pbest)". The larger the c1, the more the particle relies on its own experience and avoids deviating from the effective optimization direction.

[0175] Gbest's guiding weights: control the strength of particle "learning towards the swarm's global optimum (Gbest)". The larger the particle size, the more it relies on collective experience, thus accelerating the overall convergence speed;

[0176] Pbest-guided random perturbation: Introduces randomness to prevent particles from getting trapped in "local optima" (e.g., a certain combination of parameters seems to meet the requirements, but there are better solutions that have not been explored);

[0177] Gbest-guided random perturbation: same effect Further enhance the algorithm's global search capability to adapt to the parameter diversity of HUD scenarios (such as different drivers having different sensitivities to blue light, requiring multiple sets of optimal solutions).

[0178] Where w = 0.8 (inertia weight), c1 = c2 = 2 (acceleration coefficient). / A random number in the range [0,1].

[0179] Final output: Among the particles that meet the H_B, V, C constraints, select the parameter combination with the highest "overall score" (e.g. B=12%, L_hud=12000nits, C_ar=72%).

[0180] 5. Personalized preference learning: multi-modal reinforcement learning MQL

[0181] 1. Reinforcement learning modeling:

[0182] State (S): Scene type + driver historical interaction (e.g. "3 times manual blue light reduction when meeting at night");

[0183] Action (A): Blue light proportion adjustment amplitude (±1%-±5%), feedback mode (voice / tactile);

[0184] Reward (R): Voice "satisfaction" R=+3, no feedback R=+2, manual correction R=-2;

[0185] 2. Model training: Run "Q-Learning" algorithm on vehicle MCU (STM32L4), update Q table every 10 interactions, after 100 iterations, personalized matching degree > 98%;

[0186] 3. Preference storage: Store driver preferences (e.g. "night meeting blue light 5%") to vehicle EEPROM (no loss after power off), directly call next time in the same scene.

[0187] Q-Learning update formula:

[0188] Q(S,A)=Q(S,A)+α[R+βmaxQ(S’,A’)-Q(S,A)],α=0.1 learning rate, β=0.9 discount factor.

[0189] Exemplarily, a multi-modal execution feedback layer is connected to the blue light adjustment decision layer for executing the adjustment instructions and synchronously collecting execution effect feedback data to form an execution-verification closed loop; the multi-modal execution feedback layer adjusts the proportion of harmful blue light by controlling the flipping frequency of the 415nm / 455nm / 480nm three-waveband blue light micro-mirrors in the DLP chip, and maintains color balance by compensating the flipping frequency of the red and green light micro-mirrors, while collecting spectrometer, AR chip and driver feedback data to realize closed loop verification, multi-modal execution (precise control of hardware actions) is as follows:

[0190] Low blue light adjustment (DLP chip multi-waveband driving):

[0191] DLP5532-Q1 chip working principle: through the 1.3 million pieces of micro-mirror (10.8 μm each) of "digital micro-mirror device (DMD)" flip (± 12°) control light reflection, wherein "blue light micro-mirror group" is divided into 415 nm / 455 nm / 480 nm three groups according to wavelength, and the flip frequency (1000-2000 Hz) is independently controlled;

[0192] Adjustment logic: if the decision instruction is "blue light ratio 12%", control the flip ratio of 415-455 nm micro-mirror from 30% to 8%, 455-480 nm micro-mirror remains 4%, while compensating the flip frequency of red / green light micro-mirror (red light +5%, green light +3%), to avoid yellowish picture.

[0193] The HUD imaging unit composed of LCD, DLP, LOCS, MicroLED or OLED light machine controls the content of blue light by adjusting the current size of blue light source. Figure 4 For RGB three-color mixed light to realize white light source light machine, Figure 5 For MicroLED or OLED imaging unit, only the current size of blue chip needs to be adjusted to control the content of blue light. The brightness adjustment of HUD is the same by adjusting the current size.

[0194] Multi-dimensional feedback (data back and effect verification):

[0195] Hardware feedback principle:

[0196] Spectrometer: real-time acquisition of HUD output light blue light ratio (accuracy ± 1%), if the deviation is > 5% (such as instruction 12%, actual measurement 18%), trigger hardware calibration (re-adjust DMD micro-mirror flip frequency);

[0197] AR chip feedback: chip output AR navigation contrast (error < 5%), if V < 0.85 (such as rain contrast 60%), automatically improve AR brightness ( ).

[0198] Driver feedback principle:

[0199] Voice feedback: through the loudspeaker to broadcast the adjustment result (such as "blue light from 25% to 12%"), collect the driver's voice evaluation (such as "too dark" - trigger secondary adjustment);

[0200] Physiological feedback: heart rate sensor monitors the change of HRV after adjustment, if HRV decreases > 10% (indicates discomfort), automatically recall blue light ratio (+ 2%).

[0201] Exemplarily, the cloud-local closed-loop iteration layer connects the multi-modal execution feedback layer and the core AI processing layer, and is used for local fine-tuning and cloud model retraining based on feedback data, so as to realize continuous optimization of the algorithm.

[0202] Specifically, the cloud-local closed-loop iteration layer includes a local fine-tuning module and a cloud iteration module; wherein the local fine-tuning module fine-tunes the initial constraint threshold of MOPSO by using the gradient descent method based on 7-day feedback data, and the cloud iteration module pushes the AI model retrained by the cloud to the vehicle end through the OTA mode. Specifically as follows:

[0203] 1) Local iteration, vehicle end optimization:

[0204] 1. Weekly parameter fine-tuning: based on 7-day feedback data (such as “after adjusting the rain scene, 50% of the drivers manually reduce the blue light from 15% to 12%”), the “initial constraint threshold” of MOPSO is fine-tuned by using the “gradient descent method” (such as reducing the rain night blue light benchmark from 15% to 12%);

[0205] 2. Fault reduction processing:

[0206] Modal failure detection: through the “sensor self-diagnosis signal” (such as no output of the spectrum sensor - determine failure), trigger ISO14229 UDS protocol fault code storage (such as P0101: ambient light sensor failure);

[0207] Reduction strategy: if the spectrum sensor fails, automatically switch to “visual camera + CAN headlight signal” joint judgment of ambient light (such as headlight on - determine night - blue light 10%), the bottom strategy meets the safety standard of “night 10% / day 25%”.

[0208] 2) Cloud iteration, global optimization of vehicle end

[0209] 1. Mass data collection:

[0210] Anonymous upload: upload “scene data + adjustment effect + feedback record” to the vehicle cloud through the vehicle Ethernet (100BASE-T1) (encrypted transmission, in line with ISO21434 network security);

[0211] Data screening: filter invalid data (such as sensor fault data) in the cloud, keep valid samples (such as southern rain, northern ice and snow, and other regional scene data), and update 100,000+ new samples every month.

[0212] 2. Model retraining and OTA pushing principle

[0213] Cloud training: retrain MobileCMT model using GPU cluster (e.g. NVIDIA A100), add new scenarios (e.g. "ice and snow reflection scenario" - blue light ratio 18%, brightness 10000nts);

[0214] OTA push: push optimized model (compressed to 500KB) to vehicles in batches through ISO15765 protocol, update time <5min (choose to update when the vehicle is off to avoid affecting driving).

[0215] It also includes fault reduction processing mechanism, when any sensor fails, the system automatically switches to backup sensing path and executes the bottom line strategy. Cloud fault analysis: through "fault code big data analysis" (e.g. 10% of a batch of spectral sensors fail - determine the hardware batch problem), generate repair strategy (e.g. firmware update compensation algorithm); Remote repair: for non-hardware damage faults (e.g. algorithm parameter drift), repair firmware through OTA push, repair success rate >95%; for hardware failure, push "nearby 4S shop maintenance reminder".

[0216] In summary, the application sets up a multi-modal data acquisition layer to fuse environmental, vehicle and driver state information, and realizes accurate scene recognition through multi-modal data preprocessing layer and core AI processing layer, and then generates adjustment instructions that take into account eye protection, visibility and comfort through the blue light adjustment decision layer, and finally dynamically adjusts and controls the HUD brightness and blue light content through the multi-modal execution feedback layer.

[0217] The method realizes effective suppression of harmful blue light and intelligent adaptation of display brightness in different driving scenarios, significantly reduces the risk of retinal damage and the burden of human eye adjustment caused by long-term use of HUD, and improves the visual comfort and safety during driving.

[0218] Finally, it should be noted that: the above only for the preferred embodiments of the application, and not for limiting the application, although the application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application, should be included in the protection scope of the application.

Claims

1. A multi-modal sensing Al adaptive brightness low blue light head-up display optical system, characterized in that, include: A multimodal data acquisition layer is used to collect real-time data on the environment, vehicles, and drivers. The multimodal data acquisition layer includes a visual road condition module, a voice interaction module, a tactile feedback module, a physiological state module, and a traditional upgrade module. The multimodal data preprocessing layer, connected to the multimodal data acquisition layer, is used to perform temporal alignment, missing data completion, modality cleanup and feature standardization on the acquired raw data, and output a standardized feature matrix with a unified dimension. The core AI processing layer, connected to the multimodal data preprocessing layer, is used to perform multimodal feature fusion on the standardized feature matrix, identify the current driving scenario type based on the fusion result, and output a baseline adjustment strategy. The blue light adjustment decision layer, connected to the core AI processing layer, is used to generate the final HUD brightness and blue light content adjustment instructions based on the benchmark adjustment strategy, while satisfying the three objectives of eye protection, visibility and comfort. A multimodal execution feedback layer, connected to the blue light adjustment decision layer, is used to execute the adjustment instructions and synchronously collect execution effect feedback data to form an execution-verification closed loop; The cloud-local closed-loop iteration layer connects the multimodal execution feedback layer and the core AI processing layer, and is used to perform local fine-tuning and cloud model retraining based on feedback data to achieve continuous algorithm optimization. 2.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, The visual road condition module includes a forward-facing high-definition camera and a HUD projection area camera; The front-view high-definition camera uses global shutter CMOS imaging technology and has a 120dB wide dynamic range and MIPI-CSI2 interface. The HUD projection area camera has low-light enhancement technology and region of interest acquisition function. 3.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, The voice interaction module includes a microphone array and a voice processing chip; The microphone array is arranged in an equilateral triangle and employs beamforming technology, and the voice processing chip has echo cancellation and voice endpoint detection functions. 4.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, The physiological state module includes a non-contact heart rate sensor and an eye state detector; The non-contact heart rate sensor uses photoplethysmography and incorporates a 3-axis accelerometer for motion artifact suppression. The eye state detector uses a binocular high-definition camera to collect information on blink frequency and eyelid opening and closing.

5. The multi-modal sensing Al adaptive brightness low-blue head-up display optical system according to claim 1, wherein, The traditional upgrade module includes a spectral subdivision detector and a CAN bus; The spectral subdivision detector has a built-in 415nm / 455nm / 480nm three-band filter to distinguish between harmful and beneficial blue light, and the CAN bus adopts a flexible data rate mode and has a fault tolerance mechanism. 6.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, The core AI processing layer uses a lightweight cross-modal Transformer model to fuse multimodal features. It achieves deep fusion of visual, speech, physiological, tactile and vehicle signals through modal embedding and cross-modal attention calculation, and combines a rule engine and supervised learning model to complete the recognition of 15 types of composite driving scenarios. 7.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, The blue light adjustment decision layer adopts a multi-objective particle swarm optimization algorithm, and under the constraint conditions of a blue light hazard coefficient H_B < 0.08 mW.h / sr.m2, a visibility coefficient V > 0.85 and a comfort coefficient C > 0.7, the final adjustment parameter combination is optimized to give consideration to eye protection, visibility and comfort. 8.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, The multi-modal execution feedback layer adjusts the proportion of harmful blue light by controlling the flipping frequency of the 415 nm / 455 nm / 480 nm three-waveband blue light micro-mirror in the DLP chip, maintains color balance by compensating the flipping frequency of the red and green light micro-mirror, and realizes closed-loop verification by collecting the spectrometer, AR chip and driver feedback data. 9.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, The cloud-end closed-loop iteration layer includes a local fine-tuning module and a cloud-end iteration module. The local fine-tuning module fine-tunes the initial constraint threshold of MOPSO based on 7-day feedback data using the gradient descent method, and the cloud-end iteration module pushes the AI model retrained in the cloud to the vehicle end through OTA. 10.The multi-modal sensing Al adaptive brightness low-blue light head-up display optical system according to claim 1, wherein, It also includes a fault order reduction processing mechanism. When any sensor fails, the system automatically switches to the backup sensing path and executes the bottom protection strategy.

Images