Vehicle atmosphere lamp interaction method, device and equipment and storage medium

Abstract

The invention provides a vehicle atmosphere lamp interaction method and device, equipment and a storage medium. The method comprises the steps that micro-expression information of a driver is acquired, and driving information and in-vehicle language information are acquired; inputting the micro-expression information, the driving information and the in-vehicle language information into a preset language model to obtain a first event set and a corresponding first event relationship; generating a state mapping curve of the driver according to the first event set and the first event relation through a preset graph neural network; determining a state fluctuation interval of the driver according to the state mapping curve and a preset fluctuation judgment rule so as to determine a second event set and a second event relation; inputting the second event set and the second event relationship into a preset language model to obtain a state switching event; and matching in a preset atmosphere lamp regulation and control rule base according to the state switching event to obtain an atmosphere lamp regulation and control strategy, and controlling an atmosphere lamp of the vehicle according to the atmosphere lamp regulation and control strategy.

Description

Technical Field

[0001] The present application relates to the technical field of lighting control, and particularly to a vehicle ambient light interaction method, device, equipment and storage medium. Background Art

[0002] Currently, during driving, the emotional state of the driver directly affects driving safety. The conversations among passengers in the vehicle and the external environment information are two main factors affecting the driver's emotions. When passengers in the vehicle have intense discussions or arguments, it will distract the driver's attention and trigger negative emotions; while environmental factors such as traffic congestion, bad weather, and emergencies outside the vehicle will also cause the driver to have adverse emotions such as anxiety and irritability. These emotional fluctuations will reduce the driver's judgment and reaction ability, increasing the risk of traffic accidents. Some vehicles use technologies such as speech recognition and facial expression recognition to monitor the driver's state. However, these technologies are often single-dimensional detections and cannot comprehensively capture the driver's emotional changes. At the same time, the existing in-vehicle ambient light systems are mostly preset fixed modes and lack the ability to interact with the driver's emotional state in real time. Summary of the Invention

[0003] The present application provides a vehicle ambient light interaction method, device, equipment and storage medium, which are used to analyze the driver's driving state by combining external environmental factors and in-vehicle language factors, and adjust the vehicle ambient light to control the vehicle ambient light to regulate the driver's adverse emotions.

[0004] In a first aspect, an embodiment of the present application provides a vehicle ambient light interaction method, which includes: Obtain the driver's micro-expression information, driving information, and in-vehicle language information; Input the micro-expression information, the driving information, and the in-vehicle language information into a preset language model to obtain a first event set and a corresponding first event relationship; Generate a state mapping curve of the driver through a preset graph neural network according to the first event set and the first event relationship; According to the state mapping curve and a preset fluctuation determination rule, determine the state fluctuation interval of the driver, divide the events corresponding to the state fluctuation interval in the first event set into a second event set, and set the event relationship corresponding to the second event set as a second event relationship; Input the second event set and the second event relationship into the preset language model to obtain state concern events; Match the state concern events in a preset ambient light regulation rule library to obtain an ambient light regulation strategy, and control the vehicle's ambient light according to the ambient light regulation strategy.

[0005] Second aspect, an embodiment of the present application provides a vehicle atmosphere light interaction device, which includes: An information acquisition module, configured to acquire the micro-expression information of the driver, the driving information, and the in-vehicle language information; An event determination module, configured to input the micro-expression information, the driving information, and the in-vehicle language information into a preset language model to obtain a first event set and a corresponding first event relationship; A curve generation module, configured to generate a state mapping curve of the driver according to the first event set and the first event relationship through a preset graph neural network; An event division module, configured to determine the state fluctuation interval of the driver according to the state mapping curve and a preset fluctuation determination rule, divide the events corresponding to the state fluctuation interval in the first event set into a second event set, and set the event relationship corresponding to the second event set as a second event relationship; A state analysis module, configured to input the second event set and the second event relationship into the preset language model to obtain state concern events; An atmosphere regulation module, configured to match according to the state concern events in a preset atmosphere light regulation rule library to obtain an atmosphere light regulation strategy, and control the atmosphere light of the vehicle according to the atmosphere light regulation strategy.

[0006] Third aspect, an embodiment of the present application provides a vehicle, which includes a memory and a processor; The memory is used to store a computer program; The processor is configured to execute the computer program and implement the vehicle atmosphere light interaction method according to any one of the embodiments of the present application when executing the computer program.

[0007] Fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the processor is caused to implement the vehicle atmosphere light interaction method according to any one of the embodiments of the present application.

[0008] An embodiment of the present application provides a vehicle ambient light interaction method, which includes: obtaining the micro-expression information of the driver, obtaining driving information and in-vehicle language information; inputting the micro-expression information, driving information and in-vehicle language information into a preset language model to obtain a first event set and corresponding first event relationships; through a preset graph neural network, generating a state mapping curve of the driver according to the first event set and the first event relationships; according to the state mapping curve and a preset fluctuation determination rule, determining the state fluctuation interval of the driver, dividing the events corresponding to the state fluctuation interval in the first event set into a second event set, and setting the event relationships corresponding to the second event set as second event relationships; inputting the second event set and the second event relationships into the preset language model to obtain state concern events; matching the state concern events in a preset ambient light control rule library to obtain an ambient light control strategy, and controlling the ambient light of the vehicle according to the ambient light control strategy. Through the above method, the language model is used to process multi-modal information, generate an event set and its relationships, transform discrete state information into a structured event description, use a graph neural network to generate a state mapping curve, effectively establish the temporal correlation of the driver's state, be able to track and predict the trend of emotional changes, improve the real-time performance of state judgment, determine the state fluctuation interval through the fluctuation determination rule, and screen key events based on this to form a second event set, reducing the interference of irrelevant information. Based on the matching mechanism between the state concern events and the preset rule library, the automatic generation of the ambient light control strategy is realized, enabling the ambient light to make corresponding adjustments according to the real-time state of the driver, enhancing the intelligence of the human-vehicle interaction, and thus improving the driving safety. Description of the Drawings

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0010] Figure 1 It is a schematic flowchart of a vehicle ambient light interaction method provided by an embodiment of the present application; Figure 2 It is a schematic block diagram of a vehicle ambient light interaction device provided by an embodiment of the present application. Detailed Embodiments

[0011] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0012] The flowcharts shown in the accompanying drawings are only illustrative examples, and do not necessarily include all contents and operations / steps, nor are they necessarily executed in the described order. For example, some operations / steps can be decomposed, combined, or partially merged, so the actual execution order may change according to the actual situation.

[0013] It should also be understood that the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0014] It should be further understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0015] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of a vehicle atmosphere light interaction method provided by an embodiment of this application. As Figure 1 shown, the specific steps of the vehicle atmosphere light interaction method include: S101 - S106.

[0016] S101. Obtain the micro-expression information of the driver, obtain the driving information and the in-vehicle language information.

[0017] Exemplarily, capture the real-time image of the driver, and obtain the micro-expression information of the real-time image through a pre-set image processing model. For example, the YOLO series model. The above process can be quickly realized through existing lightweight image models. The micro-expression information includes key features such as facial expressions, eye activities, and head postures. At the same time, collect the driving information, such as driving behavior data such as vehicle speed, acceleration, steering angle, and braking force, which can reflect the driving state of the driver. The in-vehicle language information is collected through an in-vehicle microphone, including the voice commands of the driver, the conversation content with passengers, and acoustic features such as tone and intonation. The comprehensive collection of these three types of information provides a rich data basis for subsequent analysis.

[0018] It should be noted that the collection of the above information needs to comply with existing laws and regulations and be carried out on the basis of user consent to prevent the infringement of personal privacy.

[0019] S102. Input the micro-expression information, the driving information, and the in-vehicle language information into a pre-set language model to obtain a first event set and the corresponding first event relationship.

[0020] Exemplarily, the pre - set language model first extracts features and performs unified encoding on information of different modalities, converts micro - expressions, driving data, and voice information into processable feature vectors, and then through a multi - modal fusion mechanism, integrates and analyzes these features to identify key events, i.e., the first event set. These events may include the driver's emotional changes, abnormal driving behaviors, important conversation contents, etc. The model also analyzes the temporal, causal, and other relationships between these events to obtain the first event relationship.

[0021] S103. Through the pre - set graph neural network, generate the state mapping curve of the driver according to the first event set and the first event relationship.

[0022] Exemplarily, using the graph neural network, the events in the first event set are taken as nodes, and the event relationships are taken as edges to construct an event association graph. Through operations such as graph convolution, the network can capture the deep associations between events. On this basis, temporal analysis of the events is carried out to connect discrete event points into a continuous state curve. This curve can intuitively show the changing trend of the driver's state over time, including dimensions such as emotional fluctuations and attention changes.

[0023] S104. According to the state mapping curve and the preset fluctuation determination rules, determine the state fluctuation interval of the driver, divide the events corresponding to the state fluctuation interval in the first event set into the second event set, and set the event relationships corresponding to the second event set as the second event relationship.

[0024] Exemplarily, based on the preset fluctuation determination rules, analyze the state mapping curve to identify the abnormal fluctuation intervals in the curve. These rules include fluctuation amplitude thresholds, duration thresholds, change rate thresholds, etc. Once it is found that the state change within a certain time period exceeds the preset threshold, the system will mark this interval as the state fluctuation interval. Then, the system searches for the original events corresponding to these intervals, classifies them into the second event set, and at the same time retains the relationship structure between these events.

[0025] S105. Input the second event set and the second event relationship into the pre - set language model to obtain the state - concerned events.

[0026] Exemplarily, when processing the second event set, the language model adopts a multi-level analysis framework. The model conducts a preliminary classification of events, dividing them into emotional events (such as anxiety, irritability), behavioral events (such as frequent lane changes, hard braking), and environmental events (such as complex road conditions, bad weather). Through deep semantic understanding technology, the causal chain between events is analyzed. For example, an event sequence like "the vehicle in front suddenly brakes → the vehicle brakes emergently → the driver is nervous" is identified. On this basis, the urgency and duration of each event are also evaluated to provide a basis for subsequent intervention strategies. For complex multi-event combinations, an event impact graph is constructed to analyze the interactions and cumulative effects between events. For example, when a combination of "long driving + road congestion + hot weather" is detected, it is identified as a potentially high-risk state. In addition, the model also considers personalized factors. Based on the historical data of different drivers, personalized criteria for determining state-concerned events are established. The model assigns a weight coefficient to each state-concerned event for precise control of subsequent ambient light adjustment strategies.

[0027] S106. Match according to the state-concerned event in the preset ambient light regulation rule library to obtain an ambient light regulation strategy, and control the ambient light of the vehicle according to the ambient light regulation strategy.

[0028] Exemplarily, in the rule library, corresponding basic dimming schemes are designed according to different types of state-concerned events. For example, for fatigue driving, the system will adopt a gradual blue dimming sequence to enhance the driver's alertness; for emotional agitation, soft warm tones will be used to create a calm atmosphere. The rule library also contains detailed parameter configurations, such as precise control of light intensity (usually adjustable between 50 - 500 lumens), color temperature range (adjustable between 2700K - 6500K), beam angle (adjustable between 15 - 120 degrees), etc. The system will fine-tune these parameters according to the driver's personal preferences and physiological characteristics. In addition, the rule library also designs a dynamic dimming mode that can achieve special effects such as gradual change, flicker, and pulsation. These effects have been verified by human factors engineering to ensure that they will not interfere with driving. When executing the regulation strategy, the system will consider external factors such as ambient light, vehicle speed, and road conditions to dynamically adjust the light parameters. For example, the light brightness will be reduced during night driving, and the light change frequency will be reduced during high-speed driving. The system will also record the driver's reactions to different dimming schemes and continuously optimize the regulation strategy through machine learning to achieve personalized and precise intervention. To ensure safety, a special dimming mode in an emergency state is also set up to alert the driver to potential dangers.

[0029] An embodiment of the present application provides a method for vehicle ambient light interaction, which includes: obtaining the micro-expression information of the driver, obtaining driving information and in-vehicle language information; inputting the micro-expression information, driving information and in-vehicle language information into a preset language model to obtain a first event set and corresponding first event relationships; through a preset graph neural network, generating a state mapping curve of the driver according to the first event set and the first event relationships; according to the state mapping curve and a preset fluctuation determination rule, determining the state fluctuation interval of the driver, dividing the events corresponding to the state fluctuation interval in the first event set into a second event set, and setting the event relationships corresponding to the second event set as second event relationships; inputting the second event set and the second event relationships into the preset language model to obtain state concern events; matching the state concern events in a preset ambient light control rule library to obtain an ambient light control strategy, and controlling the ambient light of the vehicle according to the ambient light control strategy. Through the above method, the language model is used to process multi-modal information, generate event sets and their relationships, transform discrete state information into structured event descriptions, use a graph neural network to generate a state mapping curve, effectively establish the temporal correlation of the driver's state, be able to track and predict the trend of emotional changes, improve the real-time performance of state judgment, determine the state fluctuation interval through the fluctuation determination rule, and filter key events based on this to form a second event set, reducing the interference of irrelevant information. Based on the matching mechanism between the state concern events and the preset rule library, the automatic generation of the ambient light control strategy is realized, enabling the ambient light to make corresponding adjustments according to the real-time state of the driver, enhancing the intelligence of the human-vehicle interaction, and thus improving the driving safety.

[0030] To more clearly introduce the technical solution of the present application, the technical solution of the present application will also be introduced through specific embodiments below. It should be noted that the specific embodiments are used to expand the description of the technical solution of the present application, rather than limiting the present application.

[0031] In some embodiments, the preset language model includes: a multi-modal encoder and a hierarchical decoder. The multi-modal encoder includes: a feature extraction layer, a dynamic attention layer, a multi-head fusion layer, and a projection transformation layer. Inputting the micro-expression information, driving information and in-vehicle language information into the preset language model to obtain an associated event set and event association relationships, including: S1021 - S1028.

[0032] S1021. Based on the feature extraction layer of the multi-modal encoder, perform multi-scale feature extraction on the micro-expression information to obtain a multi-scale feature tensor.

[0033] Exemplarily, a BERT pre-trained language model is used to encode the text. To fully extract the multi-level semantic information in the text, a multi-scale feature extraction mechanism is designed: at the word level, the basic semantics of a single micro-expression action are captured through word vector encoding, such as atomic action features like "frowning" and "blinking"; at the phrase level, the sequential features of micro-expression combinations are extracted through a bidirectional LSTM network, such as temporal combination features like "frowning first and then smiling"; at the sentence level, the semantic features of the overall expression state are extracted through a self-attention mechanism, such as high-level semantics like "nervous and tired state". These features at different scales are fused through residual connections, and a feature selection gating mechanism is introduced to dynamically adjust the importance weights of each scale feature according to different scenarios. Thus, the output multi-scale feature tensor is a three-dimensional matrix, which contains word-level features (the first dimension), phrase-level features (the second dimension), and sentence-level features (the third dimension), obtaining the multi-scale feature tensor, providing a rich feature representation for subsequent dynamic attention processing.

[0034] S1022. Input the multi-scale feature tensor into the dynamic attention layer, and use an adaptive weight mechanism to perform feature recombination to obtain an enhanced feature tensor.

[0035] Exemplarily, an attention weight matrix is established, and the matrix dimension of the attention weight matrix matches the three dimensions (word level, phrase level, sentence level) of the multi-scale feature tensor. For the features of each dimension, the correlation score with the context semantics is calculated, and the scaled dot-product attention mechanism is used for score calculation. For example, when a keyword like "frowning eyebrows" appears in the text description, the system will increase the attention weight of the word-level features; when a temporal description like "frowning first and then relaxing" appears, the weight of the phrase-level features will be increased; when a description of the overall emotional state appears, the weight of the sentence-level features will be increased. Through this dynamic adjustment mechanism, the system can adaptively focus on the most relevant feature levels according to different micro-expression description scenarios. The weights are normalized through the softmax function to ensure that the sum of all attention weights is 1, and the weights are weighted-summed with the original feature tensor to obtain an enhanced feature representation. This process ensures that the system can highlight the most informative feature representation according to the specific content and context of the text description.

[0036] S1023. Through the multi-head fusion layer, multiple parallel attention heads are set, and multi-head interaction calculations are performed on the enhanced feature tensor and a preset context vector to obtain a fused feature tensor.

[0037] Exemplarily, the multi-head fusion layer is provided with 8 parallel attention heads, and each attention head independently learns the text semantic feature combination patterns of different dimensions. Each attention head contains three transformation matrices: a query matrix (Q), a key matrix (K), and a value matrix (V). The enhanced feature tensor output by S1022 (including word-level, phrase-level, and sentence-level information) is transformed through these three matrix transformations to obtain query vectors, key vectors, and value vectors. At the same time, a preset context vector (including various standard micro-expression description patterns, such as "fatigue expression", "tension expression", "distraction expression", etc.) also undergoes the same transformation, then the dot product of the query vector and the key vector is calculated, and divided by the scaling factor (the square root of the attention head dimension), and the attention scores are obtained through the softmax function. The attention scores are multiplied by the value vectors to obtain the output of each attention head. The outputs of the 8 attention heads are concatenated and linearly transformed to form a fusion feature tensor, which contains the micro-expression semantic information interpreted from multiple perspectives.

[0038] S1024. Through the projection transformation layer, the fusion feature tensor is mapped to a preset semantic space through a non-linear transformation and normalized to obtain a first semantic matrix.

[0039] Exemplarily, the projection transformation layer is used to define a high-dimensional semantic space, including multiple interpretation dimensions such as expression intensity, emotion type, mental state, etc. Through a non-linear transformation function (such as a multi-layer perceptron), the fusion feature tensor is mapped into this semantic space. The transformation process includes: using the ReLU activation function for non-linear transformation to retain the positive features described in the text; adjusting the feature dimensions through a fully connected layer to unify the micro-expression description features at different levels into the same semantic space; then using LayerNorm for feature normalization to ensure the consistency of the feature distributions in different dimensions; through L2 normalization processing, the modulus lengths of all feature vectors are made 1, which is convenient for subsequent semantic similarity calculation. In the output first semantic matrix, each row represents the micro-expression semantic feature vector at a time point, which contains the normalized expression feature representation extracted from the text description.

[0040] S1025. Semantically encode the driving information and in-vehicle language information to obtain a fusion information vector, and map the fusion information vector to a preset semantic space to obtain a second semantic matrix.

[0041] Exemplarily, driving information and in-vehicle language information are also encoded by the BERT model, and a bidirectional LSTM network is used to capture the temporal features of the behavior description. The hidden layer state of the network filters important features through the attention mechanism, focusing on those behavior descriptions related to the emotional state and cognitive load. These features are mapped to the same semantic space as the micro-expression features through the fully connected layer to ensure that features of different modalities can be directly compared. The behavior description features and the language description features are fused by weighted summation to form a comprehensive behavior feature vector, which is then dimensionally adjusted and normalized to obtain the second semantic matrix.

[0042] S1026. Perform contrastive consistency learning on the first semantic matrix and the second semantic matrix to obtain the third semantic matrix.

[0043] Exemplarily, positive sample pairs (time-aligned micro-expression descriptions and behavior descriptions) and negative sample pairs (descriptions at different time points or unrelated description pairs) are constructed based on the first semantic matrix and the second semantic matrix. The improved InfoNCE loss function is used as the objective of contrastive learning. By maximizing the semantic similarity of positive sample pairs and minimizing the semantic similarity of negative sample pairs, the consistent representation of features is learned. The momentum encoder is used to incrementally update the feature representation to ensure that the model can stably learn the semantic association between micro-expressions and other behaviors, and the third semantic matrix is obtained.

[0044] S1027. Calculate the semantic similarity score between the third semantic matrix and the second semantic matrix according to the contrastive loss function, and determine the third semantic matrix with a similarity score higher than the preset score threshold as the target semantic matrix.

[0045] Exemplarily, the matching degree between the micro-expression action description and the standard action description in the template is calculated at the word level; the temporal similarity between the micro-expression sequence and the standard state sequence is calculated at the phrase level; the semantic similarity between the overall state description and the standard state category is calculated at the sentence level. The similarities at these three levels are weighted and fused, and the weights are dynamically adjusted according to the specific content of the current description. Hierarchical similarity thresholds are set. The threshold at the lower level (such as the word level) is set higher (such as 0.9), and the threshold at the higher level (such as the sentence level) is relatively lower (such as 0.7). Matching is performed based on these multi-level similarity thresholds, and the feature vectors that meet the threshold conditions are retained according to the matching results to form the target semantic matrix.

[0046] S1028. Generate a multi-modal feature vector according to the target semantic matrix, and input the multi-modal feature vector into the hierarchical decoder to obtain an associated event set and event association relationships.

[0047] Exemplarily, the decoder is based on an improved Transformer structure and includes three decoding levels: the first-level decoder is responsible for decoding word-level features into specific micro-expression action descriptions; the second-level decoder integrates phrase-level features to generate a description of the temporal changes of micro-expressions; the third-level decoder outputs an overall state assessment of the driver based on sentence-level features. During the decoding process, temporal information is added to the feature sequence through a positional encoder, and a multi-layer self-attention mechanism is used to capture the internal dependencies of the feature sequence. At the same time, a logical association is established between different-level descriptions through a graph attention network to construct a multi-level evaluation structure including micro-expression-behavior-state, and an evaluation report containing three levels is output: a micro-expression action list, a description of temporal behavior changes, and an overall state assessment conclusion, and the relationship between these evaluation results is displayed through an associated graph structure to obtain an associated event set and event association relationships.

[0048] In some embodiments, according to the contrast loss function, the semantic similarity score between the third semantic matrix and the second semantic matrix is calculated, including: S271 - S278.

[0049] S271. Calculate the conditional probability distribution of the third semantic matrix and the second semantic matrix to obtain a conditional probability distribution sample set; Exemplarily, the input layer receives the feature vectors (dimension 128) of the third semantic matrix and the second semantic matrix. The first hidden layer contains 256 nodes, the second hidden layer 128 nodes, and the third hidden layer 64 nodes. A fully connected manner is adopted between each layer, and the LeakyReLU activation function (slope parameter 0.2) is used for non-linear transformation. Batch normalization is performed on the data of each batch to maintain the stability of the data distribution. Through this network structure, the conditional probability distribution sample set P(Y|X) is calculated, where X represents the features of the third semantic matrix and Y represents the features of the second semantic matrix, and the obtained conditional probability distribution sample set contains the conditional dependence relationship between the two matrices.

[0050] S272. According to the conditional probability distribution sample set, calculate the sample mean and variance, and substitute the sample mean and variance as parameters into a preset variational distribution function to obtain a parameterized variational function; Exemplarily, the sample mean μ and variance σ² are calculated for the conditional probability distribution sample set. For example, for a sample sequence of a certain feature dimension [0.1, 0.2, 0.3, 0.4, 0.5], the calculated mean μ = 0.3 and variance σ² = 0.025. Then, these statistics are substituted as parameters into the predefined Gaussian variational distribution function q(z|x), specifically: q(z|x)=N(z|μ(x), σ²(x)). This parameterized variational function is used to approximate the true posterior distribution, where μ(x) and σ²(x) are the mean and variance functions output by the neural network. Through this parameterization method, subsequent methods such as stochastic gradient descent can be used to optimize the parameters of the variational distribution.

[0051] S273. According to the parameterized variational function, perform random sampling on the third semantic matrix and the second semantic matrix to obtain a sampled feature matrix; Exemplarily, based on the parameterized variational function q(z|x), the reparameterization trick is used for random sampling. Specifically, ε is sampled from the standard normal distribution N(0, 1), and then the sample conforming to the target distribution is obtained through the transformation z = μ + σ×ε. For example, if μ = 0.3, σ = 0.158 for a certain dimension, and the sampled ε = 1.2, then the sampled value for this dimension is 0.3 + 0.158×1.2 = 0.49. This sampling process is repeated multiple times (such as 100 times) for each feature dimension of the two semantic matrices to obtain a sampled feature matrix, and these samples will be used for subsequent probability density estimation.

[0052] S274. Perform kernel density calculation on the sampled feature matrix to obtain the joint probability distribution value; Exemplarily, the process of kernel density calculation requires the use of a Gaussian kernel function. The bandwidth parameter h of the appropriate Gaussian kernel function is selected (such as automatically determined using the Silverman criterion), and then for each sample point x, the specific calculation formula is: ; where K is the Gaussian kernel function , x is the target point of the probability density (i.e., a certain position in the feature space where the probability density is to be calculated), xi is the sampled sample point, n is the total number of samples, h is the bandwidth parameter, and K is the kernel function (here the Gaussian kernel is used).

[0053] For a certain two-dimensional feature point, the distribution density of the surrounding sample points is statistically calculated to obtain the joint probability density estimation value of this point. In this way, a continuous probability distribution representation of the two matrix feature spaces can be obtained.

[0054] S275. Substitute the joint probability distribution value and the conditional probability distribution sample set into the preset upper bound formula of variational mutual information, and perform gradient descent iteration to obtain the upper bound estimation value of mutual information; Exemplarily, substitute the calculated joint probability distribution value p(x,y) and conditional probability distribution q(y|x) into the upper bound formula of variational mutual information CLUB(X,Y) = Ep(x,y)[logq(y|x)] - Ep(x)p(y)[logq(y|x)]. Use the Adam optimizer for gradient descent optimization, set the learning rate to 0.001, β1 = 0.9, and β2 = 0.999. In each iteration, calculate the gradient of the loss function with respect to the network parameters and update the parameters to minimize the upper bound of mutual information. For example, if the estimated value of the upper bound of mutual information in a certain iteration is 2.5, adjust the network parameters through backpropagation so that the estimated value in the next iteration is reduced to 2.3.

[0055] S276. Calculate the cosine distance between the third semantic matrix and the second semantic matrix according to the estimated value of the upper bound of mutual information to obtain a feature distance matrix; Exemplarily, construct an improved cosine distance calculation formula based on the estimated value of the upper bound of mutual information. For any two feature vectors a and b in the third semantic matrix A and the second semantic matrix B, calculate their cosine distance d(a,b) = 1 - cos(a,b) = 1 - (a·b) / (||a||·||b||). For example, the calculation process of the cosine distance between vector a = [0.3, 0.4, 0.5] and b = [0.4, 0.3, 0.6] is as follows: dot product a·b = 0.12 + 0.12 + 0.3 = 0.54, norm ||a|| = 0.707, ||b|| = 0.781, and the final distance d = 1 - 0.54 / (0.707×0.781) = 0.023. Repeat this calculation process for all feature vector pairs to obtain a feature distance matrix.

[0056] S277. Perform a softmax normalization operation on the feature distance matrix to obtain a normalized distance matrix; Exemplarily, apply the Softmax function to the feature distance matrix D for normalization processing, converting each distance value into a probability form. For each element dij in the matrix, calculate its normalized value through the formula softmax(dij) = exp(dij) / Σk = 1~nexp(dik). For example, if the distance values in a certain row are [0.2, 0.3, 0.4], then first calculate the exponential values [exp(0.2), exp(0.3), exp(0.4)] = [1.22, 1.35, 1.49], and then divide by the sum 4.06 to obtain the normalized values [0.30, 0.33, 0.37]. This can obtain a standardized distance distribution, which is convenient for subsequent similarity calculation.

[0057] S278. Perform a weighted average operation on the elements in the normalized distance matrix to obtain a semantic similarity score.

[0058] Exemplarily, weighted average calculation is performed on the elements of the normalized distance matrix. First, weight coefficients are designed. For example, weights can be assigned based on the importance of features. The weights of important features are higher (such as 0.4), and the weights of secondary features are lower (such as 0.1). Then, these weights are multiplied by the corresponding normalized distance values and summed to obtain the semantic similarity score. For instance, if the normalized distance values of three features are [0.3, 0.33, 0.37], and the corresponding weights are [0.4, 0.35, 0.25], then the similarity score is 0.3×0.4 + 0.33×0.35 + 0.37×0.25 = 0.329. This score reflects the overall similarity degree of the two semantic matrices.

[0059] In some embodiments, the pre - configured graph neural network includes an input layer, a convolutional layer, a cross - attention layer, a multi - layer perceptron, and an output layer. Through the pre - configured graph neural network, according to the first event set and the first event relationship, a state mapping curve of the driver is generated, including: S1031 - S1035.

[0060] S1031. Based on the input layer, event nodes are generated according to each event node in the first event set, connection edges of the event nodes are generated according to the first event relationship, and an event - associated hypergraph is generated.

[0061] Exemplarily, each event node in the first event set is represented by a 256 - dimensional vector, which contains feature information such as the time, type, and duration of the event. For example, a speeding event can be represented as a vector composed of a timestamp [1, 0, 0, 0] (indicating it occurs in the 1st hour), an event type [0, 1, 0, 0] (indicating it is a speeding type), a duration [0.5] (indicating it lasts for 30 minutes), etc. Connection edges are constructed according to the first event relationship. If there is a temporal relationship, spatial relationship, or semantic relationship between two events, a weighted edge is added between the corresponding nodes. The weight value is determined according to the strength of the relationship. For example, the closer the time of two events, the greater the weight of the edge. An event - associated hypergraph structure is formed according to the above process, where each node may have multiple types of connection relationships with multiple other nodes. This structure can effectively capture the complex interaction patterns between events.

[0062] S1032. Based on the convolutional layer, convolutional processing is performed on the event - associated hypergraph to extract the hidden state representation of each event node. The hidden state representation includes: semantic information and context information of the event node.

[0063] Exemplarily, a three-layer GCN structure is defined. The first layer maps the original features of the nodes (256 dimensions) to a 512-dimensional hidden space. The second layer compresses the features to 256 dimensions, and the third layer obtains a 128-dimensional hidden state representation. In each layer of convolutional operation, the representation of the node aggregates the information of its adjacent nodes. For example, for node i, the formula for calculating its representation in the (l + 1)-th layer is: h(l+1) i =σ(Σj∈N(i)(1 / c ij )W(l)h(l) j ), where N(i) is the set of neighbors of node i, c ij is the normalization coefficient, W(l) is the weight matrix of the l-th layer, and σ is the ReLU activation function. In this way, the hidden state of each node contains both its own semantic information and the context information from adjacent nodes.

[0064] S1033. Based on the cross-attention layer, perform a concatenation operation on the hidden state representations of the event pairs to generate pairwise event semantic vectors.

[0065] Exemplarily, for each pair of events (i, j), their hidden state representations hi and hj are input into 8 attention heads. Each attention head independently calculates the attention weights, then calculates the weighted feature representations, and then concatenates the outputs of all attention heads and passes them through a linear transformation to obtain a 256-dimensional pairwise event semantic vector. This process can capture the fine-grained interaction features between event pairs. For example, if event i is "hard braking" and event j is "lane departure", the cross-attention mechanism can learn the correlation strength between these two events.

[0066] S1034. Based on the multi-layer perceptron, classify the pairwise event semantic vectors to obtain the causal relationship probability distribution of each event pair.

[0067] Exemplarily, the multi-layer perceptron uses a three-layer structure to classify the pairwise event semantic vectors. The first layer contains 512 neurons, the second layer 256 neurons, and the third output layer corresponds to different categories of causal relationships (such as "causal relationship", "correlation relationship", "no relationship", etc.). The LeakyReLU activation function (slope parameter 0.2) and Dropout (probability 0.3) are used between each layer to prevent overfitting. For the input 256-dimensional pairwise event semantic vector, the probability distribution of each category is calculated through forward propagation, so as to obtain the causal relationship probability distribution. For example, for the pair of events "hard braking" and "lane departure", a probability distribution [0.7, 0.2, 0.1] may be obtained, indicating a 70% probability of a causal relationship, a 20% probability of a correlation relationship, and a 10% probability of no relationship. The cross-entropy loss function is used during training, and the Adam optimizer is used for parameter update.

[0068] S1035. Based on the output layer, generate the state mapping curve of the driver according to the causal relationship probability distribution, the paired event semantic vectors, and the preset mapping rules.

[0069] Exemplarily, define the dimensions of the state space (such as fatigue level, distraction level, emotion value, etc.), and the value range of each dimension is [0, 1]. Design a mapping function to map the causal relationship strength and semantic features of the event pair to the change amount in the state space. For example, if it is detected that there is a strong causal relationship (probability 0.8) between the event of "yawning continuously multiple times" and the event of "lane departure", then increase the value by 0.3 in the fatigue level dimension. For each time point, accumulate the impacts of all relevant event pairs, and use a smoothing function (such as exponential moving average) to process to obtain a continuous state mapping curve. The output state mapping curve reflects the dynamic change process of the driver's state over time.

[0070] In some embodiments, according to the state mapping curve and the preset fluctuation determination rules, determine the state fluctuation interval of the driver, including: S1041 - S1049.

[0071] S1041. Sample the state mapping curve through a preset sampling time window to obtain a sequence of state values.

[0072] S1042. Input the sequence of state values into a preset peak detection algorithm, detect the state peaks and state valleys in the sequence of state values, and perform state fluctuation amplitude grading according to the state peaks, state valleys, and a preset amplitude threshold to obtain the state fluctuation amplitude feature.

[0073] S1043. Perform a fast Fourier transform on the sequence of state values to obtain a frequency domain feature sequence, and perform statistical grading on the frequency domain feature sequence to obtain the state fluctuation frequency feature.

[0074] S1044. Perform sliding statistics through a preset time window to obtain the change rate of the state value within each time window. Perform grading on the change rate according to a preset duration threshold to obtain the state fluctuation duration feature.

[0075] S1045. Multiply the state fluctuation amplitude feature, the state fluctuation frequency feature, and the state fluctuation duration feature by the corresponding feature weight matrices respectively to obtain a weighted feature vector.

[0076] Exemplarily, the weight matrix dimension of the state fluctuation amplitude feature is 50×32, with larger weights assigned to abnormal amplitude fluctuations; the weight matrix dimension of the state fluctuation frequency feature is 40×32, focusing on high-frequency fluctuation components; the weight matrix dimension of the state fluctuation duration feature is 30×32, assigning higher weights to fluctuations with longer durations. For example, if a fluctuation with an amplitude exceeding 0.8 is detected, the weight value will be set to 0.4; the weight of a fluctuation with a frequency above 0.5 Hz is 0.3; the weight of a fluctuation with a duration exceeding 10 minutes is 0.3. Through matrix multiplication operations, the three features are multiplied by their corresponding weight matrices respectively to obtain a 96-dimensional weighted feature vector.

[0077] S1046. Perform non-linear activation processing on the weighted feature vector and perform feature selection through an adaptive gating mechanism to obtain a fused feature vector.

[0078] Exemplarily, the weighted feature vector is first non-linearly transformed using the LeakyReLU activation function, with the slope parameter set to 0.2 to retain negative value information. Feature selection is performed through an adaptive gating mechanism. The gating network consists of two fully connected layers, with an input dimension of 96, a hidden layer of 64, and an output dimension of 96. The gating network learns to generate weight values between 0 and 1 to dynamically adjust the importance of different features, obtaining a 96-dimensional fused feature vector, where important features are enhanced and irrelevant features are suppressed. For example, when it is detected that the driver is in a fatigued state, the weights of fatigue-related features are automatically increased and the influence of other features is reduced.

[0079] S1047. Add the weighted feature vector and the fused feature vector through residual connection to obtain a comprehensive fluctuation feature vector.

[0080] Exemplarily, an important feature value in the weighted feature vector is 0.8, while the value after gating processing is 0.3. The importance of the original feature can be retained through residual connection. The 96-dimensional comprehensive fluctuation feature vector obtained after residual connection not only retains the information of the original feature but also contains the optimized information after non-linear transformation and feature selection.

[0081] S1048. Cluster the comprehensive fluctuation feature vector to obtain multiple cluster centers, calculate the Mahalanobis distance between the cluster centers, and determine the number of categories according to the distance threshold.

[0082] Exemplarily, the elbow method is used to determine that the range of the initial number of cluster centers is 3 - 7. Then, clustering is performed with multiple random initializations, and the result with the optimal silhouette coefficient is selected. The Mahalanobis distance between each pair of the obtained cluster centers is calculated, taking into account the covariance structure of the features. For example, if the Mahalanobis distance between two cluster centers is less than the preset threshold of 1.5, they are merged into one category. Through this adaptive distance threshold adjustment method, the appropriate number of categories is determined to avoid overly fragmented or general classifications.

[0083] S1049. Map the number of categories to the preset interval division rule to determine the state fluctuation interval of the driver.

[0084] Exemplarily, if it is determined to be 3 categories through the above process, the state fluctuation interval of the driver is mapped to a "stable interval" (fluctuation amplitude less than 0.2), a "fluctuation interval" (fluctuation amplitude 0.2 - 0.5), and a "severe fluctuation interval" (fluctuation amplitude greater than 0.5). Each interval has detailed descriptive features. For example, the stable interval indicates that the driver's state remains stable, the fluctuation interval indicates obvious state changes, and the severe fluctuation interval indicates possible abnormal states.

[0085] In some embodiments, the second event set and the second event relationship are input into a pre - set language model to obtain state - concerned events, including: S1051 - S1056.

[0086] S1051. Perform semantic encoding on each event in the second event set to obtain an event semantic vector sequence.

[0087] S1052. Perform structured processing on the second event relationship to obtain a relationship adjacency matrix, and perform graph attention operation on the event semantic vector sequence according to the relationship adjacency matrix to obtain a context - enhanced vector.

[0088] S1053. Calculate the correlation score between events according to the context - enhanced vector and perform normalization processing on the correlation score to obtain an event association weight.

[0089] S1054. Perform weighted summation operation on the event semantic vector sequence according to the event association weight to obtain a fused semantic representation.

[0090] S1055. Perform multi - layer perceptron transformation on the fused semantic representation to obtain an event importance score, and sort the events in the second event set according to the event importance score to obtain a sorted event sequence.

[0091] S1056. Use the events at the preset ordinal positions in the sorted event sequence as state - concerned events.

[0092] In the above process, through semantic encoding and graph attention operations, in order to deeply understand the inherent meaning of events and the correlation relationships between events, not only the features of individual events are captured, but also the structured dependencies between events are modeled. By adopting the methods of relevance calculation and weight normalization, the features of important events are prominently strengthened in the fusion process, avoiding the loss of key information. Through the multi-layer perceptron and sorting mechanism, the system can adaptively identify the concern events that are most indicative of the driver's state, providing an intelligent and quantifiable event screening solution. This multi-step processing flow ensures that the identification of state concern events takes into account both the importance of the events themselves and the associated impacts between events, so as to more accurately reflect the actual state changes of the driver.

[0093] In some embodiments, match the state concern events in a preset atmosphere light control rule library to obtain an atmosphere light control strategy, including: Retrieve the state concern events in a preset atmosphere light control rule library to obtain a target control rule set, where the target control rule set includes: atmosphere light brightness parameters, color parameters, and flashing parameters.

[0094] Generate atmosphere light configuration parameters according to the target control rule set.

[0095] Match the atmosphere light configuration parameters in a preset atmosphere light control instruction library to obtain an atmosphere light control instruction sequence.

[0096] Obtain an atmosphere light control strategy according to the atmosphere light control instruction sequence.

[0097] Please refer to Figure 2 , Figure 2 which is a schematic block diagram of a vehicle atmosphere light interaction device provided by an embodiment of the present application. The vehicle atmosphere light interaction device 200 is used to execute the aforementioned vehicle atmosphere light interaction method. Among them, the vehicle atmosphere light interaction device 200 can be configured in a server.

[0098] Among them, the server can be an independent server, a server cluster, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.

[0099] As Figure 2 shown, the vehicle atmosphere light interaction device 200 includes: an information acquisition module 201, an event determination module 202, a curve generation module 203, an event division module 204, a state analysis module 205, and an atmosphere regulation module 206.

[0100] An information acquisition module 201 is configured to acquire the micro-expression information of the driver, the driving information, and the in-vehicle language information.

[0101] An event determination module 202 is configured to input the micro-expression information, the driving information, and the in-vehicle language information into a preset language model to obtain a first event set and corresponding first event relationships.

[0102] A curve generation module 203 is configured to generate a state mapping curve of the driver according to the first event set and the first event relationships through a preset graph neural network.

[0103] An event division module 204 is configured to determine the state fluctuation interval of the driver according to the state mapping curve and a preset fluctuation determination rule, divide the events corresponding to the state fluctuation interval in the first event set into a second event set, and set the event relationships corresponding to the second event set as second event relationships.

[0104] A state analysis module 205 is configured to input the second event set and the second event relationships into the preset language model to obtain state concern events.

[0105] An atmosphere regulation module 206 is configured to match according to the state concern events in a preset atmosphere lamp regulation rule library to obtain an atmosphere lamp regulation strategy, and control the atmosphere lamp of the vehicle according to the atmosphere lamp regulation strategy.

[0106] An embodiment of the present application provides a vehicle, which includes a memory and a processor; the memory is used to store a computer program; the processor is configured to execute the computer program and implement the vehicle atmosphere lamp interaction method according to any one of the embodiments of the present application when executing the computer program.

[0107] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the processor is enabled to implement the vehicle atmosphere lamp interaction method according to any one of the embodiments of the present application.

[0108] As mentioned above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of various equivalent modifications or replacements, and these modifications or replacements should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A vehicle atmosphere lamp interaction method, characterized in that, The method includes: Obtaining the micro-expression information of the driver, obtaining driving information and in-vehicle language information; Inputting the micro-expression information, the driving information and the in-vehicle language information into a preset language model to obtain a first event set and corresponding first event relationships; Generating a state mapping curve of the driver according to the first event set and the first event relationships through a preset graph neural network; Determining the state fluctuation interval of the driver according to the state mapping curve and a preset fluctuation determination rule, dividing the events corresponding to the state fluctuation interval in the first event set into a second event set, and setting the event relationships corresponding to the second event set as second event relationships; Inputting the second event set and the second event relationships into the preset language model to obtain state concern events; Performing matching according to the state concern events in a preset atmosphere light control rule library to obtain an atmosphere light control strategy, and controlling the atmosphere light of the vehicle according to the atmosphere light control strategy.

2. The vehicle atmosphere light interaction method according to claim 1, characterized in that, The preset language model includes: a multi-modal encoder and a hierarchical decoder. The multi-modal encoder includes: a feature extraction layer, a dynamic attention layer, a multi-head fusion layer and a projection transformation layer. The inputting the micro-expression information, the driving information and the in-vehicle language information into the preset language model to obtain an associated event set and event association relationships includes: Performing multi-scale feature extraction on the micro-expression information based on the feature extraction layer of the multi-modal encoder to obtain a multi-scale feature tensor; Inputting the multi-scale feature tensor into the dynamic attention layer and performing feature recombination using an adaptive weight mechanism to obtain an enhanced feature tensor; Through the multi-head fusion layer, setting multiple parallel attention heads to perform multi-head interaction calculations on the enhanced feature tensor and a preset context vector to obtain a fused feature tensor; Through the projection transformation layer, mapping the fused feature tensor to a preset semantic space through a non-linear transformation and performing normalization processing to obtain a first semantic matrix; Performing semantic encoding on the driving information and the in-vehicle language information to obtain a fused information vector, and mapping the fused information vector to the preset semantic space to obtain a second semantic matrix; Performing contrastive consistency learning on the first semantic matrix and the second semantic matrix to obtain a third semantic matrix; Calculating the semantic similarity score between the third semantic matrix and the second semantic matrix according to a contrastive loss function, and determining the third semantic matrix with a similarity score higher than a preset score threshold as the target semantic matrix; Generating a multi-modal feature vector according to the target semantic matrix and inputting the multi-modal feature vector into the hierarchical decoder to obtain the associated event set and event association relationships.

3. The vehicle atmosphere light interaction method according to claim 2, characterized in that, The calculating the semantic similarity score between the third semantic matrix and the second semantic matrix according to the contrastive loss function includes: Performing conditional probability distribution calculation on the third semantic matrix and the second semantic matrix to obtain a conditional probability distribution sample set; According to the conditional probability distribution sample set, calculate the sample mean and variance, and substitute the sample mean and the variance as parameters into a preset variational distribution function to obtain a parameterized variational function; According to the parameterized variational function, perform random sampling on the third semantic matrix and the second semantic matrix to obtain a sampled feature matrix; Perform kernel density calculation on the sampled feature matrix to obtain a joint probability distribution value; Substitute the joint probability distribution value and the conditional probability distribution sample set into a preset upper bound formula of variational mutual information, and perform gradient descent iteration to obtain an estimated value of the upper bound of mutual information; According to the estimated value of the upper bound of mutual information, calculate the cosine distance between the third semantic matrix and the second semantic matrix to obtain a feature distance matrix; Perform softmax normalization operation on the feature distance matrix to obtain a normalized distance matrix; Perform weighted average operation on the elements in the normalized distance matrix to obtain the semantic similarity score.

4. The vehicle atmosphere light interaction method according to claim 1, wherein The preset graph neural network includes an input layer, a convolutional layer, a cross-attention layer, a multi-layer perceptron, and an output layer. Generating a state mapping curve of the driver according to the first event set and the first event relationship through the preset graph neural network includes: Based on the input layer, generate event nodes according to each event node in the first event set, and generate connection edges of the event nodes according to the first event relationship to generate an event-associated hypergraph; Based on the convolutional layer, perform convolutional processing on the event-associated hypergraph to extract the hidden state representation of each event node, where the hidden state representation includes: semantic information and context information of the event node; Based on the cross-attention layer, perform a splicing operation on the hidden state representations of event pairs to generate paired event semantic vectors; Based on the multi-layer perceptron, classify the paired event semantic vectors to obtain the causal relationship probability distribution of each event pair; Based on the output layer, generate a state mapping curve of the driver according to the causal relationship probability distribution, the paired event semantic vectors, and a preset mapping rule.

5. The vehicle atmosphere light interaction method according to claim 1, characterized in that, Determining the state fluctuation interval of the driver according to the state mapping curve and a preset fluctuation determination rule includes: Sample the state mapping curve through a preset sampling time window to obtain a state numerical sequence; Input the state numerical sequence into a preset peak detection algorithm to detect state peaks and state valleys in the state numerical sequence, and perform state fluctuation amplitude grading according to the state peaks, the state valleys, and a preset amplitude threshold to obtain state fluctuation amplitude characteristics; Perform fast Fourier transform on the state numerical sequence to obtain a frequency domain feature sequence, and perform statistical grading on the frequency domain feature sequence to obtain state fluctuation frequency characteristics; Perform sliding statistics through a preset time window to obtain the change rate of the state value within each time window; perform grading on the change rate according to a preset duration threshold to obtain state fluctuation duration characteristics; Multiply the state fluctuation amplitude feature, the state fluctuation frequency feature, and the state fluctuation duration feature by their corresponding feature weight matrices respectively to obtain weighted feature vectors; Perform non-linear activation processing on the weighted feature vectors and perform feature selection through an adaptive gating mechanism to obtain fused feature vectors; Add the weighted feature vectors and the fused feature vectors through residual connection to obtain a comprehensive fluctuation feature vector; Cluster the comprehensive fluctuation feature vector to obtain multiple cluster centers, calculate the Mahalanobis distance between the cluster centers, and determine the number of categories according to the distance threshold; Map the number of categories into a preset interval partitioning rule to determine the state fluctuation interval of the driver.

6. The vehicle atmosphere light interaction method according to claim 1, wherein The step of inputting the second event set and the second event relationship into the preset language model to obtain state concern events includes: Perform semantic encoding on each event in the second event set to obtain an event semantic vector sequence; Perform structural processing on the second event relationship to obtain a relationship adjacency matrix, and perform graph attention operation on the event semantic vector sequence according to the relationship adjacency matrix to obtain context-enhanced vectors; Calculate the correlation score between events according to the context-enhanced vectors and perform normalization processing on the correlation score to obtain event association weights; Perform weighted summation operation on the event semantic vector sequence according to the event association weights to obtain a fused semantic representation; Perform multi-layer perceptron transformation on the fused semantic representation to obtain event importance scores, and sort the events in the second event set according to the event importance scores to obtain a sorted event sequence; Use the event at a preset order position in the sorted event sequence as the state concern event.

7. The vehicle atmosphere lamp interaction method according to claim 1, wherein Match according to the state concern event in a preset atmosphere light regulation rule library to obtain an atmosphere light regulation strategy, including: Retrieve in a preset atmosphere light regulation rule library according to the state concern event to obtain a target regulation rule set, and the target regulation rule set includes: atmosphere light brightness parameters, color parameters, and flashing parameters; Generate atmosphere light configuration parameters according to the target regulation rule set; Match in a preset atmosphere light control instruction library according to the atmosphere light configuration parameters to obtain an atmosphere light control instruction sequence; Obtain the atmosphere light regulation strategy according to the atmosphere light control instruction sequence.

8. An in-vehicle atmosphere lamp interaction device, characterized in that, The vehicle atmosphere light interaction device includes: An information acquisition module, configured to acquire the micro-expression information of the driver, acquire driving information and in-vehicle language information; An event determination module, configured to input the micro-expression information, the driving information, and the in-vehicle language information into a preset language model to obtain a first event set and a corresponding first event relationship; A curve generation module, configured to generate a state mapping curve of the driver according to the first event set and the first event relationship through a preset graph neural network; An event division module, configured to determine a state fluctuation interval of a driver according to the state mapping curve and a preset fluctuation determination rule, divide events corresponding to the state fluctuation interval in the first event set into a second event set, and set an event relationship corresponding to the second event set as a second event relationship; A state analysis module, configured to input the second event set and the second event relationship into the preset language model to obtain state concern events; An atmosphere regulation module, configured to match according to the state concern events in a preset atmosphere lamp regulation rule library to obtain an atmosphere lamp regulation strategy, and control the atmosphere lamp of the vehicle according to the atmosphere lamp regulation strategy.

9. A vehicle, characterized in that, The vehicle includes a memory and a processor; The memory is used for storing a computer program; The processor is configured to execute the computer program and, when executing the computer program, implement the vehicle atmosphere lamp interaction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the processor is caused to implement the vehicle atmosphere lamp interaction method according to any one of claims 1 to 7.

Images