Adaptive multi-scene weight distribution for vehicle ambient light intelligent control method

Abstract

This application provides an intelligent control method for in-vehicle ambient lighting with adaptive multi-scene weight allocation, comprising: acquiring normalized feature vectors of multiple sensing signals; determining the scene type based on the current state and obtaining the corresponding weight vector from a preset weight lookup table; weighting and fusing the weight vector and the normalized feature vector to obtain a fusion index; determining the target color light parameter based on the fusion index and the scene type; applying a smooth transition filter to the target color light parameter to obtain the output color light parameter, and driving the ambient lighting. This method orthogonally separates the scene weight layer and the signal feature layer, utilizes a weighted fusion engine to achieve dynamic fusion control of multiple signals, and eliminates parameter step abrupt changes during scene switching through a smooth transition filter, achieving a continuous and smooth transition of color light parameters.

Description

Technical Field

[0001] This application relates to the field of vehicle-mounted intelligent lighting control technology, and in particular to an intelligent control method for vehicle-mounted ambient lighting that adapts to multi-scene weight allocation. Background Technology

[0002] Existing in-vehicle ambient lighting control systems typically employ a scene preset mode, where fixed color parameters are pre-set for each driving scenario. When a scene changes, the corresponding preset parameters are directly applied to drive the light-emitting elements. However, this mode has the following shortcomings: Firstly, the fixed color parameters cannot be dynamically adjusted based on multiple real-time sensing signals within the vehicle, resulting in a disconnect between the ambient lighting effect and the actual state of the occupants. Secondly, the color parameters undergo abrupt changes during scene switching, creating visual abruptness and discontinuity. Therefore, how to maintain the flexibility of scene expansion while achieving multi-channel signal fusion and visual continuity during scene switching has become a pressing technical problem to be solved in the field of in-vehicle ambient lighting control. Summary of the Invention

[0003] To address the shortcomings of existing technologies where in-vehicle ambient lighting uses scene preset modes, resulting in the inability to dynamically adjust color parameters based on the fusion of multiple real-time sensing signals and the occurrence of abrupt parameter changes during scene switching, this application provides an intelligent control method for in-vehicle ambient lighting with adaptive multi-scene weight allocation. This method orthogonally separates the scene decision layer and the signal sensing layer, uses scene-related weight vectors and normalized feature vectors for weighted fusion to output color parameters, and eliminates abrupt scene switching changes through smooth transition filtering. This enables dynamic fusion control of multiple signals and continuous smooth transition of color parameters during scene switching. Specifically, this application provides the following technical solutions:

[0004] This application provides an intelligent control method for in-vehicle ambient lighting with adaptive multi-scene weight allocation, including:

[0005] Obtain the normalized feature vector of the multi-channel sensing signal;

[0006] The scene type is determined based on the current state, and the corresponding weight vector is obtained from a preset weight lookup table based on the scene type, wherein the sum of the components of the weight vector is a preset value.

[0007] The weight vector and the normalized feature vector are weighted and fused to obtain a fusion index. The target color light parameters are determined based on the fusion index and the scene type.

[0008] A smooth transition filter is applied to the target color light parameters to obtain the output color light parameters, and the ambient light is driven according to the output color light parameters.

[0009] Optionally, the multi-channel sensing signals include emotion signals, audio signals, vehicle condition signals, and ambient light signals.

[0010] Optionally, before obtaining the normalized feature vector, the method further includes: performing zero-point calibration on the sensor and performing gain calibration on each light-emitting channel.

[0011] Optionally, determining the scene type based on the current state includes: performing a coarse classification of the current state using the vehicle's motion state as a primary criterion, and performing a fine classification of the coarse classification result using the amplitude distribution of the normalized feature vector as a secondary criterion to obtain the scene type.

[0012] Optionally, it further includes: determining the emotional fit based on the occupant's facial features, generating a weight correction amount based on the deviation between the emotional fit and the target fit, and superimposing the weight correction amount onto the weight vector to obtain a corrected weight vector.

[0013] Optionally, generating the weight correction amount based on the deviation includes: generating a scalar correction intensity based on the deviation through a closed-loop controller, and distributing the scalar correction intensity to each dimension according to the current proportion of each component in the weight vector to obtain the weight correction amount.

[0014] Optionally, the smooth transition filtering includes applying a time-domain low-pass filter independently to each component of the target color light parameter.

[0015] Optionally, the time-domain low-pass filter determines the current output color light parameter based on the output color light parameter of the previous moment and the target color light parameter of the current moment; when the scene type changes, the output color light parameter of the previous moment is maintained, and the time-domain low-pass filter is continued to be applied to the target color light parameter after the switch based on the output color light parameter of the previous moment.

[0016] Optionally, when driving the ambient light according to the output color light parameters, the output color light parameters are converted into driving signals for each light-emitting channel, and the driving signals are corrected according to the pre-calibrated gain of each light-emitting channel.

[0017] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0018] The beneficial effects of this application are as follows: by structurally orthogonally separating the scene weight layer and the signal feature layer, the scene decision only outputs the weight vector and the signal perception only outputs the normalized feature vector. The two are superimposed by a weighted fusion engine to output the color light parameters, thereby eliminating the rigid binding relationship between the scene and the color light parameters. This allows the addition of a new scene to only require defining the corresponding weight vector without modifying the fusion logic. At the same time, by applying a smooth transition filter to the target color light parameters, the parameter step during scene switching is transformed into a continuous gradual output, eliminating the visual abrupt change at the moment of switching. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the intelligent control method for vehicle ambient lighting provided in the embodiments of this application.

[0020] Figure 2 This is a schematic diagram of the multi-source signal acquisition and normalization sub-process provided in an embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the weighted fusion and color parameter calculation sub-process provided in the embodiments of this application.

[0022] Figure 4 This is a schematic diagram of the smooth transition and LED output sub-process provided in the embodiments of this application.

[0023] Figure 5 A schematic diagram of the hue and brightness timing response curves for the smooth transition effect of scene switching provided in the embodiments of this application.

[0024] Figure 6 This is a schematic diagram of the convergence curve of the micro-expression closed-loop weight correction provided in the embodiments of this application. Detailed Implementation

[0025] The intelligent control method for in-vehicle ambient lighting involved in this application embodiment can be applied to various vehicles such as passenger cars, commercial vehicles, new energy vehicles, and intelligent cockpit platforms. The ambient lighting can be deployed in positions such as below the dashboard, door interior panels, sides of the center console, bottom of the seats, and edges of the roof to provide occupants with dynamic lighting effects that are adapted to the driving scenario and occupant status. This solves the technical problem that existing in-vehicle ambient lighting systems only support fixed color switching and cannot adaptively adjust color light parameters based on multi-dimensional perception signals.

[0026] For example, in a welcoming scenario, the ambient lighting uses a warm orange soft light to greet occupants as they take their seats; in a music-themed scenario, the color and flicker of the ambient lighting dynamically change with the rhythm of the music; in a soothing scenario, the ambient lighting uses a low-brightness, constant blue light to create a relaxing atmosphere; and in a high-intensity driving scenario, the ambient lighting uses a high-brightness, flickering red light to enhance the immersive driving experience. The ambient lighting typically uses RGB tri-color LEDs as light-emitting elements, and the output of any color and brightness can be achieved by adjusting the PWM duty cycle of the red, green, and blue channels. The aforementioned vehicles can be equipped with perception and control hardware including, but not limited to, onboard cameras, microphone arrays, ambient light sensors, and body control domain controllers. The onboard camera is used to capture facial images of occupants for emotion recognition and micro-expression analysis, and a resolution of at least 640 is recommended. The sensor has 480 pixels and a frame rate of at least 15 fps. The microphone array is used to collect in-vehicle audio signals to extract music rhythm features, with a sampling rate of at least 16 kHz. The ambient light sensor is used to detect changes in the color temperature and illuminance of the in-vehicle ambient light and includes three light-sensing channels: red, green, and blue.

[0027] The execution entity of the method described in this application embodiment is an in-vehicle controller. The in-vehicle controller includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the method steps described in this application embodiment. The in-vehicle controller can be an independent ambient lighting control module or integrated into a body domain controller or a smart cockpit domain controller. The processor's main frequency is not less than 100 MHz to meet the full-process calculation requirements of completing four-channel signal acquisition, feature extraction, scene classification, fusion calculation, EWMA filtering, and PWM output update within each 100 ms sampling cycle. The memory contains at least 128 KB of RAM for storing runtime variables, calibration parameters, and weight lookup tables, and at least 256 KB of Flash for storing program code and factory default parameters.

[0028] like Figure 1 As shown, the intelligent control method for vehicle ambient lighting provided in this application includes the following steps. The overall processing flow of the method is divided into four stages: the first stage is signal normalization (S100), which maps the four heterogeneous sensing signals to a unified map. The first stage involves forming comparable feature vectors within a given interval. The second stage is scene recognition and weight generation (S200), which determines the current scene and retrieves the corresponding weight vector based on the vehicle state and feature distribution. The third stage is weighted fusion and parameter mapping (S300), which fuses the weight vector and feature vector to obtain a fusion index and maps it to the target color light parameter. The fourth stage is smoothing output and driving (S500), which applies EWMA filtering to the target color light parameter and converts it into a PWM drive signal. Between the third and fourth stages, a micro-expression closed-loop feedback correction (S400) is included, which generates a weight correction amount based on the occupant's facial micro-expressions and superimposes it onto the baseline weight vector to achieve adaptive optimization. The design motivation for this closed-loop correction mechanism is that the baseline weight vector is determined only based on the scene type and cannot reflect the individual differences and dynamic changes in the occupant's emotions in the same scene. By introducing a micro-expression feedback closed loop, the weight vector can be finely adjusted according to the occupant's real-time emotional state, improving the matching degree between the ambient lighting effect and the occupant's subjective feeling. The four stages are executed sequentially in each sampling period, forming a complete control loop of "perception → decision → fusion → output". The typical value of the sampling period is 100 ms, meaning the system cyclically executes the above four stages at a frequency of 10 Hz. Within each sampling period, the stages are executed sequentially in a fixed order: S100, S200, S300, S400, S500.

[0029] Before executing the real-time control process, the vehicle controller performs zero-point calibration on the sensors and gain calibration on each light-emitting channel during the system power-on phase. During the zero-point calibration, the vehicle controller continuously acquires 32 dark-state samples from the ambient light sensor under shaded or off-light conditions, and uses the average value as the zero-point bias. The purpose of acquiring data 32 times is to reduce the standard error of the zero-point estimation to approximately one-fifth of the standard deviation of the ADC noise, meeting engineering accuracy requirements. The 32 dark-state samples are acquired continuously within 200 ms. During acquisition, the vehicle controller shuts down all LED light-emitting channels to ensure the sensor receives only residual ambient light. The vehicle controller calibrates its gain coefficient using a reference light source of known brightness. This ensures that subsequent ambient light readings meet the requirements. The vehicle controller collects microphone audio data in a silent state inside the vehicle and calculates the short-time average energy as the noise floor threshold. .

[0030] During the gain calibration of each light-emitting channel, the vehicle controller sequentially applies five levels of stepped drive signals to the R, G, and B light-emitting channels, with duty cycles of 0, 0.25, 0.5, 0.75, and 1.0, respectively. After each level stabilizes for 200 ms, eight brightness samples are continuously collected using an ambient light sensor, and the average value is taken to reduce measurement noise. The gain of each channel is then determined by linear fitting using the least squares method. and residual bias This ensures that the brightness output of each channel meets the requirements. After calibration, the vehicle controller checks the consistency of the three-channel gain. If the deviation between the gain of any channel and the average of the three channels exceeds 3%, the deviation value is recorded for subsequent color compensation. The vehicle controller then uses the calibration parameters and weight lookup table... The data is loaded from non-volatile memory to the RAM working area, and after calibration, the flag is set to enter real-time control mode. The entire initialization process is executed automatically after the system is powered on, typically taking about 2 to 5 seconds. During calibration, the ambient light remains off.

[0031] S100, obtain the normalized feature vector of the multi-channel sensing signal.

[0032] like Figure 2 As shown, the multi-channel sensing signals include emotion signals, audio signals, vehicle condition signals, and ambient light signals. The vehicle controller acquires, calibrates, compensates for, and extracts features from the above four sensing signals respectively, and maps the features of each signal to a normalized value. The heterogeneity of the four signals is reflected in the fact that each signal has different physical dimensions, dynamic range, and sampling rate: the emotion signal is a discrete category label, the audio signal is a statistical quantity of a time-domain waveform, the vehicle condition signal is a normalized combination of vehicle motion parameters, and the ambient light signal is a normalized value of color temperature. After the normalization process maps these heterogeneous signals to the same numerical range, the contributions of each feature to the fusion index become comparable and additive. The sampling period... Typically, the time ranges from 50 ms to 200 ms; this embodiment uses... The control frequency is set at 10 Hz, which is sufficient to capture the dynamic changes in occupant emotions and music rhythm without placing an excessive computational burden on the onboard processor, thus obtaining a four-dimensional normalized feature vector. The vehicle controller is... A complete signal acquisition and feature extraction process is executed once within a fixed sampling period of milliseconds (ms). The four signals are acquired sequentially and their feature vectors are updated synchronously within the same sampling period, ensuring that subsequent fusion operations use feature snapshots from the same moment. The sampling period is... A ms update frequency corresponds to a 10 Hz update frequency, which can meet the visual response requirements of ambient lighting (the human eye's perception time for changes in light is about 100-200 ms) without wasting computing resources due to excessively fast updates.

[0033] S110 collects emotional signals and extracts emotional feature values.

[0034] The vehicle controller acquires the current frame image from the vehicle camera, identifies the occupant's facial region through a face detection front-end, and outputs an emotion label. The face detection front-end employs a multi-scale sliding window strategy, performing traversal detection at multiple scaling levels of the input image. Through multi-level filtering by a cascaded classifier, non-facial regions are quickly eliminated, and finally, the pixel coordinates and detection confidence of the facial region are output.

[0035] When the detection confidence level is lower than a preset threshold of 0.7, the vehicle controller considers that no valid face has been detected in the current frame, and maintains the emotion feature value of the previous frame unchanged to avoid abnormal fluctuations in emotion features due to false detections. The face detection front end performs localization on the valid face region in the input image, extracts the texture and geometric features of the face region, and inputs the features into a pre-trained emotion classification model. The classification model can be a support vector machine classifier or a lightweight convolutional neural network, outputting the probability distribution of each emotion category, and taking the category with the highest probability as the emotion label of the current frame. The emotion labels include four discrete categories: pleasure, calmness, fatigue, and irritability. These four categories cover the most common passenger emotional states in driving scenarios, where pleasure and calmness are positive or neutral emotions, and fatigue and irritability are negative emotions. The four categories are arranged in an ordered order from high to low in terms of emotion valence. The vehicle controller converts the emotion labels into continuous values ​​according to a preset emotion-value mapping relationship. .

[0036] In the mapping relationship, pleasure corresponds to calm response Fatigue Irritability The selection of the mapping values ​​is based on the theory of emotional valence: the valence of pleasure is set to a maximum of 0.8 instead of 1.0 to reserve space for stronger positive emotions in the fusion calculation; the valence of irritability is set to a minimum of 0.2 instead of 0 to avoid the emotional channel contributing nothing to the fusion index when experiencing negative emotions, and to prevent facial expressions from being detected when no facial features are detected. The mapping values ​​are set based on the emotional valence dimension, with positive emotions mapped to a high value range and negative emotions mapped to a low value range. Specifically, pleasure is set to 0.8 instead of 1.0 because extreme positive emotions are less common in daily driving, leaving room for future expansion of more emotion categories; irritability is set to 0.2 instead of 0 because 0 is reserved for extreme negative emotions (such as anger); a neutral value of 0.5 is used when no face is detected to prevent the emotional components from skewing the fusion index. For example, when the face detection output label is calm, The vehicle controller uses this value as an emotion feature component.

[0037] S120, acquires audio signals and extracts rhythm intensity feature values.

[0038] The vehicle controller acquires the PCM audio stream from the vehicle microphone and takes... Each sampling point constitutes one frame. Typically, the value is in the range of 512 to 2048; in this embodiment, it is taken as... The frame length of 1024 is chosen to balance temporal resolution and rhythm perception requirements: at a sampling rate of 16 kHz, this corresponds to 1024 / 16000. A time window of 0.064 seconds captures short-term energy changes in the music beat without causing excessive energy estimation variance due to excessively short frame lengths. The vehicle controller calculates the short-term energy of this frame. and zero crossing rate :

[0039]

[0040]

[0041] The on-board controller obtains the original value of rhythm intensity by performing noise compensation and weighted aggregation on short-time energy and zero-crossing rate. .

[0042] in The background noise energy threshold, collected during system initialization, is used to subtract ambient background noise from energy calculations, ensuring that the audio feature value is close to 0 when no music is playing, thus avoiding interference from background noise. This background noise energy threshold is continuously updated during runtime using an exponential moving average: when the energy of the current frame is detected to be lower than... When the time is right, it is determined to be a state without music. Update the noise floor estimate to adapt to slow changes in ambient noise (such as increased wind noise at high speeds). Original audio features. The calculation formula is as follows:

[0043]

[0044] in The normalized maximum energy reference value, in physical terms, represents the upper limit of short-time energy that an audio signal can achieve under the current sampling configuration, used to account for energy differences. Normalization to The interval. The value depends on the microphone sensitivity, ADC range, and frame length. Combination: Under the 16-bit ADC range, the full-scale value of a single sampling point is Frame length The theoretical maximum short-time energy is However, in a car environment, the peak factor of music signals is typically 6 dB to 12 dB, and the actual peak energy is about one-quarter to one-sixteenth of the theoretical maximum value. Usually in to Within the scope, this embodiment takes This corresponds to approximately 25% of the full amplitude (i.e., a peak factor of approximately 6 dB), a value that results in the normalized energy distribution of most music signals being... Within the range, leave a margin above to avoid frequent saturation. This is the normalized maximum zero-crossing rate reference value, which physically represents the upper engineering limit of the zero-crossing rate of an audio signal, used to adjust the zero-crossing rate. Normalization to The theoretical zero-crossing rate of pure white noise is 0.5 (there is a 50% probability of zero crossing between any two adjacent sampling points). In-vehicle audio signals contain music and speech components, and their zero-crossing rate is usually lower than that of white noise.

[0045] The Typically, it is in the range of 0.3 to 0.5; in this embodiment, it is taken as... The theoretical zero-crossing rate of white noise is taken as the upper limit of normalization, so that the normalized zero-crossing rate of any actual audio signal does not exceed 1. and This is a weighted average of energy and zero-crossing rate. Typically, the value is in the range of 0.5 to 0.9; in this embodiment, it is taken as... The Typically, it is in the range of 0.1 to 0.5; in this embodiment, it is taken as... Energy is a stronger indicator of rhythm intensity than zero-crossing rate because short-time energy directly reflects volume, while zero-crossing rate reflects spectral distribution characteristics. Combining both can distinguish between rhythmically strong rock music and rhythmically weak, calm background music. Audio characteristics are obtained through clamping processing. For example, when the short-time energy of a PCM frame... Noise floor energy Zero crossing rate hour, , This reflects the low tempo intensity of the gentle background music. In an alternative implementation, Mel frequency cepstral coefficients can be used instead of the combination of short-time energy and zero-crossing rate to extract audio features.

[0046] S130 collects vehicle condition signals and extracts characteristic values ​​of driving intensity.

[0047] The vehicle controller reads the vehicle speed from the CAN bus. angular velocity of steering and longitudinal acceleration The speed of the vehicle The absolute speed of a vehicle is the most stable indicator of the intensity of driving; the steering angular velocity... This reflects the dynamic changes in the vehicle's left and right directions; the value increases during sharp turns. The longitudinal acceleration... This value reflects the vehicle's acceleration and deceleration characteristics in the forward and backward directions, increasing during rapid acceleration and emergency braking. The onboard controller obtains the raw value of driving intensity through weighted summation. :

[0048]

[0049] in As a reference upper limit for vehicle speed normalization, the Typically, speeds of 150 km / h or 200 km / h are used; in this embodiment, the speed is taken as... It covers the maximum speed of most passenger vehicles; As the upper limit of the normalized reference for steering angular velocity, this embodiment takes... For emergency bending conditions; As the upper limit of the normalized longitudinal acceleration reference, this embodiment takes Approximately 0.8g, covering scenarios involving rapid acceleration and deceleration. Weight , and The value of follows these principles: Within the range of 0.3 to 0.5, this embodiment takes... Vehicle speed is given a high weight as the most stable indicator of driving intensity because the vehicle speed signal has small fluctuations and high reliability. Within the range of 0.2 to 0.4, this embodiment takes... ; Within the range of 0.2 to 0.4, this embodiment takes... Steering angular velocity and acceleration serve as auxiliary indicators reflecting the instantaneous dynamic characteristics of driving. This is obtained after caliper processing. For example, when the vehicle speed angular velocity of steering longitudinal acceleration hour, This reflects a moderately intense driving state.

[0050] In another embodiment, the driving intensity feature value can also be obtained by a state estimation method based on a vehicle dynamics model. The independent normalized combination of steering angular velocity and acceleration is replaced by vehicle lateral stability indexes. For example, yaw rate gain and sideslip angle are used as inputs, and the current driving intensity comprehensive score is estimated by a vehicle stability dynamic model.

[0051] S140 collects ambient light signals and extracts environmental feature values.

[0052] The vehicle controller reads the ADC value from the ambient light sensor and uses the gain coefficient calibrated during the system initialization phase. and zero-point offset The physical illuminance is obtained by performing calibration compensation on the ADC value. The ambient light sensor includes three photosensitive channels: R, G, and B. The vehicle controller processes the calibrated response values ​​of the three channels. Transformed into CIE 1931 chromaticity coordinates using the standard chromaticity matrix. The transformation matrix is The constant matrix, whose elements are determined by the spectral response curves provided by the sensor manufacturer, is then used by the vehicle controller to estimate the current ambient light color temperature using the McCamy approximation formula. The McCamy approximation formula is a widely used color temperature estimation method, which has the advantages of low computational cost and accuracy that meets the requirements of automotive applications. The McCamy approximation formula first calculates the chromaticity shift. Then substitute into the third-order polynomial This approximation is in The deviation from the precise color temperature within the K range is less than 2%. The environmental characteristic values ​​are:

[0053]

[0054] in As the lower limit of color temperature normalization, this embodiment takes K; As the upper limit of color temperature normalization, this embodiment takes K. For example, when the ambient light color temperature At time K, .

[0055] S150, merge the four feature values ​​to obtain the normalized feature vector.

[0056] The vehicle controller merges the four feature values ​​into a four-dimensional normalized feature vector. Each component is in Within the interval. When any sensing signal cannot be acquired due to hardware failure or communication timeout, the vehicle controller sets the feature value of that signal to a neutral default value of 0.5 to prevent it from having a biased impact on the fusion index, and records the fault information in the internal log. All four signals use Min-Max normalization, with upper and lower limits determined by engineering calibration experiments and stored in non-volatile memory, ensuring that the contributions of each feature to the fusion index are comparable under different scenarios. For example, when... , , , At that time, the normalized feature vector .

[0057] In another embodiment, the normalization process can also use the Z-score normalization method instead of Min-Max normalization. Specifically, the normalization process is first performed on each feature. ,in and The statistical mean and standard deviation of this signal are obtained during the system initialization phase by collecting signal data from the first 1000 sampling periods (i.e., the first 100 seconds), and continuously updated during runtime using an exponential moving average, then compressed using a hyperbolic tangent function. interval, i.e. This maps each feature to an output with a mean of 0.5 and a monotonically increasing value.

[0058] S200, determine the scene type based on the current state, and obtain the corresponding weight vector from a preset weight lookup table based on the scene type, wherein the sum of each component of the weight vector is a preset value.

[0059] The scene recognition and weight vector generation structurally orthogonally separate scene decision-making from signal perception. Scene recognition is only responsible for outputting weight vectors representing scene characteristics, while signal acquisition is only responsible for outputting normalized feature vectors; the two do not interfere with each other. This orthogonal separation structure allows adding a new scene type only to append the corresponding weight vector row to the weight lookup table, without modifying the logic of signal acquisition and fusion operations, thus reducing system coupling. Specifically, to add a "conference" scene, only one row needs to be appended to the weight lookup table. By adding a set of hue, brightness, and flicker anchor points for the conference scene to the mapping anchor point parameters, scene expansion can be completed without modifying the code of the signal acquisition module and the fusion calculation module. The vehicle controller determines the current scene type based on the vehicle's current operating state and the amplitude distribution of the normalized feature vector using hierarchical criteria, and retrieves the corresponding baseline weight vector from a preset weight lookup table using the scene type as an index. The sum of all components of the benchmark weight vector is 1.

[0060] S210 uses the vehicle's motion state as the primary criterion to perform a coarse classification of the current state.

[0061] The vehicle controller reads the vehicle speed from the CAN bus. Gear signal and door status The macroscopic motion state of the vehicle is used as the primary criterion for coarse scenario classification. This primary criterion includes a joint determination of four conditions: when the vehicle speed... Furthermore, the gear is in park, the door is open, and the audio characteristics... When all four conditions are met simultaneously, the vehicle controller will determine the current scene as a welcoming scene. The aforementioned Typically within the range of 0 to 10 km / h, this embodiment takes... This threshold distinguishes between a stationary or idling vehicle and a normal driving state. Typically, the value is in the range of 0.1 to 0.4; in this embodiment, it is taken as... This threshold ensures that when an occupant actively plays music while the vehicle is parked, it does not mistakenly enter the welcome scene, but instead enters the music scene in response to the occupant's entertainment intention. If any of the above four conditions are not met, the vehicle controller determines whether the vehicle is currently in a driving state or a non-welcome state, and proceeds to the secondary fine classification. In the secondary fine classification, when the audio features... Vehicle condition characteristics When the thresholds are exceeded, music scenarios take precedence over passionate driving scenarios. This priority design is based on the fact that the occupant's intention to actively play music is clearer than the passive response of the vehicle's motion state.

[0062] S220, using the amplitude distribution of the normalized feature vector as a secondary criterion, perform fine classification on the coarse classification result to obtain the scene type.

[0063] The vehicle controller reads the normalized feature vector. Secondary fine classification is performed according to the following rules: when audio features At that time, it was determined to be a music scene. ;when And vehicle condition characteristics At that time, it was determined to be a scene of passionate driving. ;when and At that time, it was determined to be a relaxing scene. The aforementioned Typically, the value is in the range of 0.2 to 0.6; in this embodiment, it is taken as... This corresponds to a medium to high volume where the current audio energy exceeds 40% of full scale; Typically, the value is in the range of 0.3 to 0.7; in this embodiment, it is taken as... This corresponds to high-speed driving or combinations of frequent sharp turns and rapid acceleration. This classification design ensures that the decision paths for the four scenarios are mutually exclusive and complete: the primary criterion separates the welcoming scenario, and the secondary criterion fully covers the remaining three scenarios through a two-layer progressive judgment using audio thresholds and vehicle condition thresholds. and In this scenario, music-related scenarios take priority because the occupants' intention to play music is clear. In another embodiment, the scenario classification can also be achieved using a multi-class classifier based on a statistical model, utilizing support vector machines or random forests to classify the normalized feature vectors, thus transforming the classification decision from rule-driven to data-driven.

[0064] S230, retrieve the baseline weight vector from the weight lookup table using the scenario type as the index.

[0065] The vehicle controller uses the scene identifier to look up a weighted lookup table pre-stored in the RAM workspace. Retrieve the corresponding benchmark weight vector The baseline weight vectors corresponding to each scenario in the weight lookup table are as follows: Welcoming Scenario Ambient light signals are given the highest weight to create a warm and welcoming lighting effect; music scenes The audio signal has the highest weight to make the ambient lighting change with the rhythm of the music; for a soothing scene. The weights of emotional signals and ambient light signals are evenly distributed to achieve a balance between occupant emotional perception and environmental adaptation; in passionate driving scenarios. The vehicle condition signal has the highest weight to ensure the ambient lighting responds to driving intensity. The sum of the four components of each of the above baseline weight vectors is 1; this constraint guarantees the fusion index. As a mathematical property of weighted averages, its range is strictly limited to... Within the interval, a stable input range is provided for subsequent piecewise linear mapping. The weight allocation of each scene in the weight lookup table follows the design principle that the dominant signal accounts for 60% to 75% of the weight. The basis for this design principle is that when the weight of the dominant signal is less than 60%, the influence of this signal on the fusion index is diluted by the other three signals, and the correlation between the ambient lighting effect and the core semantics of the scene weakens; when the weight of the dominant signal is greater than 75%, the contribution of the other signals is too low, and the fusion mechanism degenerates into single-signal driving, losing the advantage of multi-signal collaborative perception. The selection of the dominant signal follows the principle of scene core semantics: the welcoming scene focuses on ambient light adaptation, the music scene focuses on audio rhythm, the soothing scene balances the two dimensions of emotion and environment, and the passionate driving scene focuses on vehicle condition response.

[0066] The weight lookup table is stored as a four-row, four-column floating-point matrix, with each row corresponding to the weight vector of a scene, occupying a total of 64 bytes (4 scenes). 4 components The weight lookup table (4-byte single-precision floating-point) consumes negligible amounts of RAM resources. During system initialization, the weight lookup table is loaded from non-volatile storage into RAM. When a new scene type is added, only the corresponding weight vector row needs to be appended to the weight lookup table; the fusion operation logic requires no modification. In another embodiment, the baseline weight vector can also be obtained through a prediction method based on a machine learning model. The trained regression model directly outputs the weight vector based on the current feature vector, transforming weight allocation from discrete table retrieval to continuous prediction. This allows for capturing the transition states between predefined scene types and provides a smoother weight transition in the boundary regions of scene switching.

[0067] For example, the normalized feature vector output by S100 is used. Assuming the current vehicle speed The gear is in Drive (D) and the doors are closed. (Level 1 coarse classification) If the welcoming criteria are not met, the item will be moved to the secondary refined category. (In the secondary refined category...) Non-musical scenes The scenario was not one of spirited driving and was therefore classified as a relaxed driving scenario. Retrieved from the weighted lookup table The scene classifier also maintains the current scene identifier. Previous scene identifier ,when Scene switching is triggered at any time.

[0068] S300, the weight vector and the normalized feature vector are weighted and fused to obtain a fusion index, and the target color light parameters are determined according to the fusion index and the scene type.

[0069] like Figure 3 As shown, the vehicle controller uses the weight vector With the normalized feature vector Perform a weighted inner product operation to obtain a fusion index that reflects the current state's influence on the total intensity of the ambient light color drive. Then, the fusion index is mapped to the target color light parameters through a set of scene-related mapping functions. The fusion index The physical meaning is that it compresses four heterogeneous sensing signals into a single scalar value, which comprehensively reflects the degree of stimulation of the ambient light by the current environment and the occupant's state. The closer to 1, the stronger the external stimulus, and the more vibrant, brighter, and more dynamic the ambient light should output. A value closer to 0 indicates a weaker external stimulus, and the ambient light should output a softer, dimmer light effect. This design simplifies the complex interaction of multi-dimensional signals into a single-variable control problem, effectively reducing the design complexity of color parameter mapping. In alternative solutions involving independent mapping of four signals, a more complex design is required. The group mapping function maps four signals to three parameters: hue, brightness, and flicker. Through the intermediate abstraction layer of the fusion index, only the design... Group mapping function.

[0070] S310, the fusion index is obtained by weighting and summing the components of the weight vector with the corresponding components of the normalized feature vector.

[0071] The vehicle controller calculates the fusion index according to the following formula. :

[0072]

[0073] in , , , These are weighted components representing emotion, audio, vehicle condition, and ambient light signals, respectively. Since each component of the weight vector is positive and sums to 1, each component of the normalized feature vector is within... Within the interval, the fusion index can be determined based on the properties of the weighted average. The range of values ​​is Specifically, when all eigencomponents are 0 This indicates that all sensory channels are in the lowest energy state, at which point the color parameters are mapped to the low-end keypoint values ​​of each scene; when all feature components are 1... This indicates that all sensory channels are in their highest energy state, and the color parameters are mapped to high-end keypoint values. The fusion index... As a weighted average, it is inherently bounded, ensuring it falls within a valid range without additional pruning. However, to prevent abnormal input values ​​due to sensor malfunctions or communication errors, the vehicle controller performs range clamping on the input data. If any component of the normalized feature vector exceeds the range due to abnormal input... At that time, clamp the component to Ensure that the fusion index does not exceed the limit.

[0074] In another embodiment, the weighted fusion can also be implemented using a geometric weighted average, i.e. This causes any signal to be at an extremely low value, suppressing the fusion index. For example, using the aforementioned soothing scenario parameters, the geometrically weighted average fusion index... , with linear weighted sum The value is close to but slightly lower, reflecting the suppression effect of geometric averaging on low-value components. The geometrically weighted averaging method is suitable for applications requiring suppression capabilities from any signal path. For example, when the vehicle condition signal approaches zero (the vehicle is stationary), even if the other three signals are relatively high, the fusion index obtained by geometrically weighted averaging is still suppressed, thus avoiding overly intense lighting effects when the vehicle is stationary. In contrast, the linear weighted sum method independently accumulates the contributions of each signal, without the rejection effect of a single signal. It is suitable for conventional scenarios where equal participation of all signals in fusion is desired, and on an automotive embedded processor, it only requires four multiplications and three additions, making its computational efficiency superior to the exponentiation operations required by geometrically weighted averaging.

[0075] For example, the soothing scene output by S200 is used. and eigenvectors The fusion index .

[0076] S320, the target color parameters are determined by a mapping function based on the fusion index and the scene type.

[0077] The vehicle controller calculates the target hue using the scene type as an index. Target brightness and target flicker frequency The mapping function adopts a piecewise linear form, with the segmentation points... Will The fusion index interval is divided into two equal segments, and linear interpolation is performed within each segment. The design basis for choosing a segmentation point value of 0.5 is: 0.5 is... At the midpoint of the interval, the fusion index is evenly divided into a low-energy segment and a high-energy segment, allowing the parameter change rate within each segment to be set independently. In the low-energy segment, the change is slow to avoid visual jumps caused by minor sub-threshold changes, while in the high-energy segment, a rapid response is used to reflect the dynamic nature of the ambient lighting.

[0078]

[0079] in Represents hue, brightness, or flicker frequency. , , Scenes The low, medium, and high fusion indices correspond to Parameter anchor value. The hue. The range of values ​​is The brightness The range of values ​​is (Percentage scale), the flicker frequency The range of values ​​is The flicker frequency This indicates the always-on mode.

[0080] The mapping anchor point parameters for each scene are as follows, and the hue range selection for each scene is based on the correspondence between color psychology and scene semantics. (Welcoming Scene) Hue Anchor Point , , Locked in the warm orange range, warm colors are associated with warmth and friendliness in environmental psychology, making them suitable for creating an atmosphere for welcoming scenes; brightness anchor point , , Brightness is maintained at medium to high to convey a sense of solemnity; flicker anchor points are all 0, with flicker-free constant-on output to avoid visual interference during welcoming. Music Scene Hue Anchor Point , , Covering a wide color gamut from blue to magenta, the rich color variations match the dynamics of the music; brightness anchor point , , Covering the full range from soft to bright; strobe anchors , , The flashing lights follow the rhythm of the music. A soothing scene. Hue Anchor Point , , Located in the cyan range, cyan-blue tones are associated with tranquility and relaxation in color psychology, and the narrow 40° hue range avoids color abruptness in soothing scenes; brightness anchor point , , Maintain low to medium brightness to avoid glare; flicker anchor point , , It produces a very faint breathing light effect only at extremely high fusion indices. (For high-performance driving scenarios) Hue Anchor Point , , Located in the red zone, red is associated with excitement and alertness in color psychology, matching the psychological state of aggressive driving; brightness anchor point , , High brightness enhances visual impact; flicker anchor points , , The flashing light is synchronized with the intensity of driving.

[0081] In the piecewise linear mapping, the intermediate anchor point The range of the fusion index is divided into two equal segments: the first segment corresponds to the gradual transition from low to medium excitation, and the second segment corresponds to the gradual transition from medium to high excitation. This equal division design ensures that the change in the fusion index near the medium excitation level produces a symmetrical response sensitivity to the color parameters. The piecewise linear mapping ensures that the output changes continuously with the fusion index, and the slopes of the two segments at the middle anchor point can be different, allowing the same scene to have different response sensitivities in the low and high excitation ranges. In another embodiment, the mapping function can also be implemented using cubic spline interpolation, making the change curve of the color parameters smoother.

[0082] For example, the soothing scene is used. Hue mapping: , Luminosity mapping: Strobe mapping: The target color light parameter triplet is: This demonstrates a cyan-blue, medium-to-low brightness, and flicker-free light effect in a soothing scene. In contrast, if the current scene were a welcoming scene... Assuming the fusion index , , , , It embodies the warm orange, high-brightness, and flicker-free color light effect in a welcoming scene.

[0083] S400, determine the emotion fit degree based on the occupant's facial features, generate a weight correction amount based on the deviation between the emotion fit degree and the target fit degree, and superimpose the weight correction amount onto the weight vector to obtain the corrected weight vector.

[0084] The onboard controller assesses the compatibility between the current ambient lighting effect and the occupants' emotional state by collecting facial micro-expressions. Based on the deviation signal, it drives the closed-loop controller to generate a weight correction amount, which is then superimposed on the baseline weight vector to achieve adaptive optimization. This micro-expression closed-loop feedback correction operates independently of the feedforward color light parameter calculation path, indirectly affecting the color light output only through the correction weight vector. This independence ensures that even if micro-expression detection fails or the camera is obstructed, the system can still continue to operate normally based on the baseline weight vector, only losing its adaptive optimization capability without causing system failure. The closed-loop correction period and the feedforward sampling period... The same applies, meaning that the deviation calculation and weight update are performed every 100 ms.

[0085] S410, extracts facial action unit intensity values ​​from facial images.

[0086] The vehicle controller acquires the current frame image from the vehicle camera, detects the facial region, and locates the coordinates of 68 facial key points. The vehicle controller calculates the intensity values ​​of three facial action units closely related to emotional valence: AU1 is inner eyebrow elevation, characterized by the change in vertical distance between key points 21 and 22 and the bridge of the nose key point 27; AU6 is cheek lifting, characterized by the change in distance between lower eyelid key points 46 and 47 and cheekbone key points 35 and 36; AU12 is corner of the mouth elevation, characterized by the change in horizontal distance between corner of the mouth key points 48 and 54 and the facial midline key point 33. Intensity values ​​of each action unit are shown below. Calculate using the following formula:

[0087]

[0088] in For the current frame and the first The Euclidean distance between keypoint pairs associated with each action unit This serves as the baseline distance for a neutral facial expression. The neutral baseline distance is collected during the system initialization phase, specifically by continuously capturing 16 frames of images in a static facial expression state after the occupant is seated, and calculating the median distance between each key point as the baseline. The median, rather than the mean, is used to suppress the impact of occasional abnormal frames. This is a scaling factor, which amplifies the relative rate of change of keypoint displacement to a larger scale. The interval. For example, it can be 3.0, 5.0, 8.0, etc., and this embodiment uses... A 20% change in keypoint displacement is mapped to an intensity value of 1, a value that balances the sensitivity and robustness of expression detection. In an alternative implementation, a pre-trained facial action unit recognition convolutional neural network can be used instead of intensity estimation based on geometric keypoints.

[0089] S420 calculates the emotional fit score based on the intensity value of facial action units.

[0090] The vehicle controller weights and combines the intensity values ​​of the three action units to obtain an emotion fit score. :

[0091]

[0092] in , , The contribution weights of each action unit are assigned. The combined occurrence of AU6 and AU12 is a hallmark of a genuine smile (Duchenne smile), and is most indicative of emotional valence in emotion recognition research; therefore, each is assigned a relatively high weight of 0.4. AU1, with its raised inner eyebrows, is often seen in negative emotions such as surprise or sadness, and is assigned a lower weight of 0.2 as a negative indicator. The selection of these three action units is based on the facial muscle movement regions most relevant to emotional valence in the facial action coding system. AU6 and AU12 correspond to the movement of the orbicularis oculi and zygomaticus major muscles, respectively, and their combined activation is a reliable marker of a genuine smile. The sum of the three weights is 1, a constraint that... The range of values ​​is defined and controllable, ensuring that... The range of values ​​is Internal: When only AU1 reaches maximum strength 1 When both AU6 and AU12 reach their maximum strength of 1 . Indicates positive emotions. It indicates negative emotions. For example, when , , hour, .

[0093] S430, the closed-loop controller generates a scalar correction intensity based on the deviation, and distributes the scalar correction intensity to each dimension according to the current proportion of each component in the weight vector to obtain the weight correction amount.

[0094] The onboard controller calculates the emotion adaptation deviation. ,in For target fit, the Typically, the value is in the range of 0 to 0.8; in this embodiment, it is taken as... This target value is slightly higher than the neutral value of 0, indicating that the system expects occupants to maintain a slightly positive emotional state. When the actual emotional fit is lower than the target value, the closed-loop controller increases the weight of the emotional signal to make the ambient lighting respond more actively to the occupants' emotions; when the actual emotional fit is higher than the target value, the closed-loop controller decreases the weight of the emotional signal to stabilize the ambient lighting effect and avoid over-response. The closed-loop controller uses a proportional-integral controller, calculating the scalar correction intensity based on the deviation. :

[0095]

[0096] in For proportional gain, the Typically, the value is in the range of 0.01 to 0.15; in this embodiment, it is taken as... This value makes the maximum deviation ( The contribution of the proportional term is 0.025, approximately... This ensures both effective response and avoids weight jitter. For integral gain, the Typically, the value is in the range of 0.001 to 0.05; in this embodiment, it is taken as... The integral gain is one-fifth of the proportional gain, so that the integral term accumulates slowly to eliminate steady-state deviation without causing overshoot; The upper limit of a single correction magnitude, the Typically, the value is in the range of 0.01 to 0.10; in this embodiment, it is taken as... To ensure that a single weight adjustment does not exceed one-eighth of each component of the baseline weight, perceptible abrupt changes in light output are avoided. When the integral term... Exceed Accumulation stops when the time comes, meaning that no new deviation value will be added to the integral term in subsequent cycles until the integral term falls back to a safe range. This prevents integral saturation from causing the correction amount to be clamped at the extreme value for a long time and unable to respond to changes in the direction of deviation in a timely manner.

[0097] The onboard controller distributes the scalar correction intensity to four dimensions according to the current proportions of each component in the baseline weight vector: , , , The sum of all components of the four maintenance positive quantities is 0, satisfying the conservation of the total weight before and after the correction, i.e. The physical meaning of this zero-sum constraint is that the closed-loop correction only changes the relative importance distribution of each signal, without changing the overall amplitude range of the fusion index, thus ensuring that the corrected color parameters still vary within the preset mapping range and will not exceed the safe output range. Each component of the corrected weights is non-negatively clamped, with a minimum clamp value of 0.01 to prevent any signal weight from returning to zero and causing the channel to completely fail. Then, L1 normalization is performed to restore the sum of each component to 1, resulting in the corrected weight vector. The weighted proportional distribution design allows components with larger weight values ​​to bear more reduction. The proportional integral controller resets the accumulated integral to zero during scene switching, i.e., when the current scene identifier output by the scene classifier is zero... Scene identifiers from the previous period When they are different, the vehicle controller will integrate the terms. Reset to 0 to avoid the cumulative deviation of the previous scenario affecting the weight correction of the new scenario.

[0098] For example, the soothing scene is used. benchmark weights ,exist time ,deviation Scalar correction strength The sum of non-emotional components Four-dimensional allocation , , , After correction .exist At that moment, assuming the adjusted weights slightly improved the ambient lighting effect, , Integral term , After correction .exist time , , , As the deviation gradually decreases, the weight correction also converges, demonstrating the steady-state tracking characteristics of the proportional-integral controller. When the deviation converges to... The proportional term contributes less than 0.001, while the integral term dominates the correction, achieving zero steady-state error tracking. The step response rise time of the PI controller is approximately... approximately 1 sampling period (i.e., approximately (seconds) This response speed allows the weight adjustment to reflect changes in passenger emotions promptly without causing frequent light flickering due to overly rapid adjustments. In actual operation, since changes in passenger facial micro-expressions are usually gradual rather than abrupt, the weight adjustment exhibits a smoother convergence process.

[0099] S500, apply a smooth transition filter to the target color light parameters to obtain the output color light parameters, and drive the ambient light according to the output color light parameters.

[0100] like Figure 4 As shown, the vehicle controller independently applies an exponentially weighted moving average filter to each component of the target color light parameter to eliminate parameter abrupt changes during scene switching, and converts the smoothed color light parameters into drive signals for each light emission channel, which are then output to the ambient light. The core idea of ​​the EWMA filter is to perform a weighted mixing between the old and new values, giving the new value a smaller weight so that it gradually integrates, thereby achieving a smooth transition without abrupt changes. This filtering method only needs to save the output value of the previous frame, resulting in minimal storage overhead, making it suitable for the real-time computational constraints of embedded systems. The carrier frequency of the PWM signal is typically set to the range of 1 kHz to 20 kHz. In this embodiment, the carrier frequency is set to 10 kHz, which is much higher than the flicker threshold perceptible to the human eye (approximately 60 Hz), ensuring that occupants will not perceive flicker caused by PWM dimming under any duty cycle. The vehicle controller generates the PWM signal through a hardware timer, where the timer's counting period corresponds to the carrier frequency, and the duty cycle corresponds to the drive value of each channel.

[0101] S510, apply a smooth transition filter to the target color parameters to obtain the output color parameters.

[0102] The vehicle controller targets the hue. Target brightness and target flicker frequency Apply EWMA filtering independently:

[0103]

[0104] in represent , or any one of them, This represents the smoothing coefficient for the corresponding component. The relationship between the time constant and the smoothing coefficient is as follows: ,in ms is the sampling period. Typically, the value is in the range of 0.05 to 0.20; in this embodiment, it is taken as... Corresponding time constant The transition time is approximately 1 second. The hue filter coefficient is set to a minimum because hue changes are the most stimulating to the human eye. Color psychology research shows that hue jumps are the primary factor causing visual fatigue. A slow hue transition allows occupants to adapt to a new color environment unconsciously. Typically, the value is in the range of 0.10 to 0.30; in this embodiment, it is taken as... Corresponding time constant The visual sensitivity to changes in brightness is lower than that to changes in hue, therefore a slightly faster response can be used; Typically, the value is in the range of 0.10 to 0.30; in this embodiment, it is taken as... Corresponding time constant The flicker parameter needs to change quickly to keep up with the rhythm of the music, so the maximum value is taken.

[0105] Due to hue It has a ring topology. and Representing the same hue, the vehicle controller calculates the difference before performing hue filtering. ,when The difference will be corrected to Then perform the EWMA operation to avoid incorrect hue jumps at the ring boundary. For example... Figure 5 As shown, after the scene switches from soothing to music, the hue gradually changes from 197.1° to the target value of 284°, and the brightness smoothly increases from 41.4, presenting a visually continuous transition without abrupt changes. In another embodiment, the smooth transition filter can also be implemented using a first-order low-pass filter, whose continuous domain transfer function is... After discretization with a zero-order hold, the following is obtained: ,in The sampling period is As a time constant, when It is equivalent to the EWMA form.

[0106] For example, Time-soothing scene steady-state output , The scene switches to music, and the target EWMA filtering , , . hour , , . hour , , . hour , , It can be observed that the hue gradually approaches the target 284° from 197.1° in an exponential decay manner, with each increment decreasing. The brightness and flicker also show a smooth approach curve, achieving scene switching without visual abrupt changes.

[0107] S520, the output color light parameters are converted into driving signals for each light-emitting channel and used to drive the ambient light.

[0108] The vehicle controller will smooth the color parameters. Convert to RGB three-channel color values ​​using the standard HSL to RGB conversion formula. , of which saturation Fixed at 1.0. During the HSL to RGB conversion process, the brightness is first determined... and saturation Calculate intermediate variables (Saturation) and (Minimum value offset), then based on hue The 60° partition (out of six partitions) determines the raw values ​​of the R, G, and B channels. For example, when (When dividing by purple and blue) , , This indicates that the blue channel dominates within this zone. The vehicle controller determines the gain of each light-emitting channel based on the values ​​calibrated during system initialization. Calculate normalized channel gain The RGB color values ​​are adjusted for gain to obtain the PWM duty cycle for each channel. The HSL to RGB conversion uses the CIE standard color space conversion algorithm, determining the primary and secondary color channels based on the 60-degree sector of the hue angle, and then calculating the normalized values ​​for each channel based on brightness and saturation. The gain adjustment utilizes the channel gain obtained during the calibration phase. and bias RGB values ​​are compensated to ensure the actual output color matches the target color. Specifically, the PWM duty cycle for each channel is calculated.

[0109]

[0110] The normalized channel gain The calculation aims to scale the gain of each channel to a relative proportion, ensuring that the channel with the highest gain has a normalized value of 1, while the normalized values ​​of the remaining channels are less than 1. This gain correction compensates for the differences in luminous efficiency of different colored LEDs by dividing the RGB color value by the normalized gain of the corresponding channel, ensuring that the actual output color matches the target color. When the smoothed flicker frequency... At that time, the vehicle controller sets the strobe timer period. In the first half of each timer cycle ( The output has a normal PWM duty cycle, and the duty cycle is reduced to 20% of the original value in the second half of the cycle, creating a visual rhythm of alternating light and dark. In constant-on mode, the strobe timer is not activated. The duty cycle attenuation coefficient of the strobe timer is fixed at 0.2, ensuring that the brightness in the dark state is one-fifth of that in the bright state, producing a noticeable flicker without completely extinguishing the light and causing gaps in darkness. The vehicle controller will... The ambient lighting is driven by writing data to the PWM hardware register. In an example where the scene switches from soothing to music, the hue gradually changes from a cyan-blue area to a blue-purple area. Passengers experience a continuous color transition rather than an abrupt color jump, enhancing the comfort of the driving experience. The previous frame output state of the EWMA filter is not reset during scene transitions, utilizing the filter's natural attenuation characteristics to achieve a smooth transition. Figure 6 As shown, the weight vector converges to a stable value after about 5 to 10 iterations during the micro-expression closed-loop feedback correction process.

[0111] To more fully illustrate the end-to-end computation process of this method, the following example uses a music scene to demonstrate the complete data flow from signal acquisition to light driving. Assume an occupant is playing rock music inside the vehicle, the vehicle is stationary, and the occupant's expression is one of pleasure. In step S110, the face detection front-end outputs a "pleasant" label. In step S120, the short-time energy of the rock music frame. Noise floor energy Zero crossing rate , , In step S130, the vehicle is stationary. , , , In step S140, the interior lighting color temperature is 5500 K. Normalized eigenvectors In step S200, the scene classifier determines... km / h and the doors are closed, but The audio activity is high, but it does not meet the requirements for welcoming guests. It will be moved to the secondary refined category. It was determined to be a music scene. In step S230, the weight lookup table is retrieved... In step S310, the weighted inner product... In step S320, Hue mapping Brightness mapping strobe mapping The target color light parameter triplet is: It features dynamic lighting effects with blue-purple hues, high brightness, and high-frequency flashing, reflecting the musical scene and matching the high-energy rhythmic characteristics of rock music.

[0112] The following example, using a passionate driving scenario, illustrates another typical end-to-end computation process. Assume the driver is driving at high speed, without music playing, and with a calm expression. In step S110, In step S120, the short-time energy of ambient noise With background noise near, , In step S130, , , , , In step S140, the ambient light color temperature for nighttime driving is 3500 K. Normalized eigenvectors In step S200, If the conditions for welcoming guests are not met, the item will be classified into a secondary category. Non-musical scenes; It was determined to be a scene of reckless driving. In step S230, the retrieval... In step S310, In step S320, , , , .

[0113] The target color light parameter triplet is The dynamic lighting effects—red, high-brightness, and medium-flicker—performed in a high-energy driving scenario, contrasting sharply with the blue-purple lighting effects in a music scenario. This demonstrates the differentiated guidance of weight allocation on color light output across different scenarios. The two end-to-end examples above validate the core design idea of ​​this method: by altering the relative importance of each signal through scene-related weight vectors, the same set of normalized feature vectors generates different fusion indices in different scenarios, thus mapping to differentiated color light parameters. This mechanism enables the ambient light's color light output to reflect both the comprehensive intensity of multi-dimensional perceived signals and the core semantics of the current scene, achieving an adaptive intelligent control effect.

[0114] Based on the same inventive concept, this application also provides an intelligent control system for in-vehicle ambient lighting. The system includes a signal normalization module, a scene weight generation module, a fusion calculation module, and a smoothing output module.

[0115] The signal normalization module is used to acquire normalized feature vectors from multiple sensing signals. The input interface of the signal normalization module connects to four sensing channels: an emotion sensor, a microphone, a CAN bus, and an ambient light sensor. The output interface outputs a four-dimensional normalized feature vector. The signal normalization module is implemented by the processor executing instructions from the memory. The scene weight generation module is used to determine the scene type based on the current state and obtain the corresponding weight vector from a preset weight lookup table. The input interface of the scene weight generation module receives the vehicle CAN bus status and the normalized feature vector, and the output interface outputs the scene identifier. and benchmark weight vector The scene weight generation module is implemented by the processor executing instructions from the memory. The fusion operation module is used to perform weighted fusion of the weight vector and the normalized feature vector to obtain a fusion index, and determine the target color light parameters based on the fusion index and the scene type. The feature input terminal of the fusion operation module receives the normalized feature vector, the weight input terminal receives the reference weight vector or the corrected weight vector, and the output terminal outputs the target color light parameter triplet. The fusion computing module is implemented by the processor executing instructions in the memory.

[0116] The smoothing output module is used to apply a smooth transition filter to the target color light parameters to obtain the output color light parameters, and to drive the ambient light according to the output color light parameters. The input terminal of the smoothing output module receives the target color light parameter triplet, and the output terminal outputs the PWM duty cycle of each channel. The smooth output module, connected to the LED hardware driver, is implemented by the processor executing instructions from memory.

[0117] The signal normalization module and the fusion calculation module transmit a four-dimensional floating-point vector via an internal data bus. The scene weight generation module and the fusion calculation module transmit scene identifiers and four-dimensional weight vectors via an internal data bus. The output of the signal normalization module is connected to the feature input terminal of the fusion calculation module, the output of the scene weight generation module is connected to the weight input terminal of the fusion calculation module, and the output of the fusion calculation module is connected to the input terminal of the smoothing output module.

[0118] The system also includes a calibration module for calibration initialization. This module automatically performs sensor zero-point calibration and LED channel gain calibration after the system is powered on, writes the calibration parameters into the RAM working area, and triggers a real-time control mode. The system also includes a micro-expression feedback module. The input interface of this module receives images from the vehicle-mounted camera, and its output interface feeds back the weight correction amount to the weight input of the fusion computing module, achieving closed-loop correction of the weight vector.

[0119] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, such as replacing piecewise linear mapping with spline curve mapping, replacing the PI controller with a PID controller or fuzzy logic controller, replacing EWMA filtering with Kalman filtering, etc. These modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the appended technical solutions.

[0120] Based on the same inventive concept, this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the above-described intelligent control method for in-vehicle ambient lighting. The computer-readable storage medium can be a read-only memory, an erasable programmable read-only memory, flash memory, a hard disk, a solid-state drive, an optical disk, or other forms of non-volatile storage media. The computer program includes a first program module for implementing the signal normalization step, a second program module for implementing the scene recognition and weight generation step, a third program module for implementing the weighted fusion and parameter mapping step, a fourth program module for implementing the micro-expression closed-loop feedback correction step, and a fifth program module for implementing the smooth output and driving step.

[0121] Based on the same inventive concept, this application also provides an electronic device, including a processor and a memory communicatively connected to the processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-described intelligent control method for in-vehicle ambient lighting. The processor can be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field-programmable gate array, or other programmable logic devices. The memory can include volatile memory and non-volatile memory. The electronic device also includes an input / output interface communicatively connected to the processor. The input / output interface is used to connect external sensing devices such as an in-vehicle camera, microphone, ambient light sensor, and CAN bus controller, as well as a PWM output port for connecting an LED driver. The electronic device can be implemented as an independent ambient lighting electronic control unit, or it can be integrated and deployed as a functional submodule of a body domain controller or a smart cockpit domain controller.

Claims

1. A method for intelligent control of in-vehicle ambient lighting with adaptive multi-scene weight allocation, characterized in that, include: Obtain the normalized feature vector of the multi-channel sensing signal; The scene type is determined based on the current state, and the corresponding weight vector is obtained from a preset weight lookup table based on the scene type, wherein the sum of the components of the weight vector is a preset value. The weight vector and the normalized feature vector are weighted and fused to obtain a fusion index. The target color light parameters are determined based on the fusion index and the scene type. A smooth transition filter is applied to the target color light parameters to obtain the output color light parameters, and the ambient light is driven according to the output color light parameters.

2. The method according to claim 1, characterized in that, The multi-channel sensing signals include emotion signals, audio signals, vehicle condition signals, and ambient light signals.

3. The method according to claim 1, characterized in that, Before obtaining the normalized feature vector, the process also includes: Zero-point calibration is performed on the sensor, and gain calibration is performed on each light-emitting channel.

4. The method according to claim 1, characterized in that, Determining the scenario type based on the current state includes: The current state is coarsely classified using the vehicle's motion state as the primary criterion, and the coarse classification result is finely classified using the amplitude distribution of the normalized feature vector as the secondary criterion to obtain the scene type.

5. The method according to claim 1, characterized in that, Also includes: Emotional fit is determined based on the occupant's facial features. A weight correction amount is generated based on the deviation between the emotional fit and the target fit. The weight correction amount is then superimposed on the weight vector to obtain the corrected weight vector.

6. The method according to claim 5, characterized in that, The step of generating a weight correction amount based on the deviation includes: The closed-loop controller generates a scalar correction intensity based on the deviation, and distributes the scalar correction intensity to each dimension according to the current proportion of each component in the weight vector to obtain the weight correction amount.

7. The method according to claim 1, characterized in that, The smooth transition filter includes applying a time-domain low-pass filter independently to each component of the target color light parameter.

8. The method according to claim 1, characterized in that, When driving the ambient light according to the output color light parameters, the output color light parameters are converted into driving signals for each light-emitting channel, and the driving signals are corrected according to the pre-calibrated gain of each light-emitting channel.

9. The method according to claim 7, characterized in that, The time-domain low-pass filter determines the current output color light parameter based on the output color light parameter of the previous moment and the target color light parameter of the current moment. When the scene type changes, the output color light parameter of the previous moment is maintained, and the time-domain low-pass filter is applied to the target color light parameter after the switch based on the output color light parameter of the previous moment.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images