Methods and systems for constructing emotional curves for children's autism risk assessment

Abstract

This invention discloses a method and system for constructing emotional curves for children's autism risk assessment, relating to the field of image and video processing technology. The method includes the following steps: acquiring a sequence of facial video images of the test subject watching an emotionally stimulating video; extracting facial action unit (AU) feature values ​​and macroscopic behavioral features from the facial video image sequence; calculating the dynamic weight coefficients of each feature in the multimodal feature set based on predefined effect sizes and significance levels; calculating the emotional intensity score at each time point based on the dynamic weight coefficients and real-time AU feature values, and generating an emotional curve through temporal smoothing; extracting feature indicators from the emotional curve to characterize abnormal emotional responses in autism, conducting autism risk assessment based on these feature indicators, and outputting the results. This achieves refined and dynamic modeling of the test subject's emotional responses, providing quantifiable and interpretable auxiliary diagnostic evidence for early autism screening.

Description

Technical Field

[0001] This invention relates to the field of image and video processing technology, and in particular to a method and system for constructing emotional curves for children's autism risk assessment. Background Technology

[0002] Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that begins in infancy or early childhood. Its core symptoms include impaired social interaction, communication difficulties, and restricted interests and repetitive behaviors. Early detection and intervention are crucial for improving the prognosis of children with autism.

[0003] Currently, the diagnosis of autism mainly relies on subjective assessments by professional doctors based on standardized scales and clinical observations. This approach is not only highly dependent on medical resources and costly, but also easily influenced by the subjective experience of the assessors, limiting its application in primary healthcare institutions and large-scale screening scenarios.

[0004] In recent years, with the development of artificial intelligence technology, especially the application of computer vision in behavioral analysis, new possibilities have been provided for non-invasive and objective auxiliary detection of autism. Some existing studies have attempted to use video analysis of children's behavioral characteristics to assist in diagnosis.

[0005] Existing technology 1: A video analysis method based on general facial expression recognition, proposed at a child behavior research conference in 2022, can identify children's basic emotions such as happiness and sadness to a certain extent. However, it is mainly used for general children's emotion teaching or simple emotional state understanding. Its model relies on overall expression classification and does not deeply analyze the subtle facial action units (AUs) that constitute complex expressions. The emotional expression of children with autism is often atypical, manifested by subtle features such as weak expression activation intensity, delayed response, or mismatch with the situation. These features are difficult to effectively capture through macro-level expression classification, resulting in significant deficiencies in the sensitivity and specificity of this technology in autism detection.

[0006] Existing Technology Two: The article "Behavioral Recognition in Children's Activity Videos" published by Li Na et al. proposes a behavioral recognition model for children's daily activity videos. This model primarily focuses on recognizing children's gross motor behaviors such as running, jumping, and climbing, as well as simple social interaction patterns. This technology emphasizes physical behavior rather than detailed facial expression analysis, and the facial feature extraction method used is relatively coarse, unable to accurately quantify the activation state of specific facial muscle groups (such as the intensity and dynamic changes of AU04 (brow drooping) and AU12 (mouth corner raising). Since emotional abnormalities in autistic children often manifest as transient, low-intensity activation patterns of facial motor units, this method, lacking detailed AU analysis, struggles to effectively reveal the essential differences in their emotional responses, thus limiting its application value in the auxiliary diagnosis of autism.

[0007] In summary, the existing technologies share the following common shortcomings:

[0008] 1. It failed to perform fine-grained, quantitative analysis of facial expressions based on AU, and could not capture atypical micro-expressions related to autism.

[0009] 2. There is a lack of continuous modeling of the dynamic changes in emotional response over time, making it difficult to quantify key temporal characteristics such as reaction delay and volatility.

[0010] 3. The model failed to fully integrate micro-facial movements with macro-behavioral features such as head posture, and thus failed to construct a comprehensive non-verbal behavior analysis model.

[0011] Therefore, it is crucial to develop a program that can automate, refine, and dynamically analyze children's emotional responses in order to effectively support early screening for autism. Summary of the Invention

[0012] The technical problem this invention aims to solve is how to precisely analyze the subtle changes in children's facial expressions when watching videos with specific emotions, in order to assist in the early screening of autism.

[0013] In a first aspect, to address the aforementioned technical problems, this invention provides a method for constructing an emotion curve for children's autism risk assessment, comprising the following steps:

[0014] Collect facial video image sequences of test subjects while they watch emotionally stimulating videos;

[0015] Multiple facial action unit (AU) feature values ​​and macroscopic behavioral features are extracted from the facial video image sequence to form a multimodal feature set;

[0016] Based on predefined effect sizes and significance levels, the dynamic weight coefficients of each feature in the multimodal feature set are calculated; wherein, by introducing a multimodal adjustment factor, the AU feature values ​​and the macro-behavioral features are weighted and fused.

[0017] Based on the dynamic weighting coefficients and real-time AU feature values, the emotion intensity score at each time point is calculated, and an emotion curve is generated through time-series smoothing.

[0018] Feature indicators for characterizing abnormal emotional responses in autism are extracted from the emotion curve, and autism risk assessment is performed based on the feature indicators, with the results output.

[0019] Furthermore, the emotionally stimulating video is a video that contains at least positive, neutral, and negative emotional content.

[0020] Further:

[0021] The AU feature values ​​include AU01, AU04, AU06, AU09, AU12, AU15, AU18, AU20, and AU25; wherein, AU06, AU09, AU12, AU20, and AU25 are positively correlated with positive emotions, and AU01, AU04, and AU15 are negatively correlated with negative emotions.

[0022] The macroscopic behavioral characteristics include the rotation and translation parameters of the head posture.

[0023] Furthermore, the calculation process of the dynamic weighting coefficients includes:

[0024] The effect size d corresponding to the AU feature value is normalized to obtain the normalized effect size;

[0025] The significance weights are determined based on the p-values ​​of the correlation between the AU feature values ​​and the emotion dimension.

[0026] Based on the AU feature values ​​and the macroscopic behavioral features, calculate the multimodal adjustment factor;

[0027] The normalized effect size, the significance weight, and the multimodal adjustment factor are multiplied together to obtain the dynamic weight coefficient.

[0028] Furthermore, the formula for calculating the multimodal adjustment factor is as follows:

[0029]

[0030] In the formula, This represents the multimodal adjustment factor; Indicates the adjustment coefficient; Indicates the intensity of action units at the micro-expression level; This represents a quantitative value indicating emotional / behavioral characteristics at the macro level.

[0031] Furthermore, the mapping rule for determining the significance weight based on the significance p-value is as follows:

[0032]

[0033] in, Indicates the significance weight.

[0034] Furthermore, the formula for calculating the emotion intensity score is as follows:

[0035]

[0036] In the formula, This indicates the intensity score of the emotion; Indicates the first Dynamic weighting coefficients for each AU; Indicates the first The and the first The interaction weight of each AU; Indicates the first Normalized intensity values ​​of each AU; Indicates the first Normalized intensity values ​​of each AU; This represents the total number of AUs.

[0037] Furthermore, the time-series smoothing process employs a sliding window averaging method, which satisfies the expression:

[0038]

[0039] In the formula, The emotional intensity score represents a discrete-time index, indicating each point in time within a sliding window. Indicates window size; Indicates the current frame.

[0040] Furthermore, the characteristic indicators include at least one of the following:

[0041] The amplitude of the emotional response is the extreme value of the emotional curve within a specific stimulus segment.

[0042] Emotional response delay is the time difference between the onset of a stimulus and the occurrence of a significant change in the emotional curve.

[0043] Emotional volatility, which is the standard deviation or variance of the aforementioned emotion curve;

[0044] Atypical response pattern, which is the number of times a preset abnormal waveform appears in the emotion curve;

[0045] The complexity of emotional fluctuations, obtained by calculating the information entropy of the emotional curve, satisfies the following formula:

[0046]

[0047] In the formula, Indicates that the emotion value is within the range The probability of; This indicates the number of intervals for the emotion value.

[0048] Furthermore, the output results include:

[0049] Sentiment curve chart;

[0050] The evaluation index quantification table includes the maximum and minimum values ​​of the emotional response amplitude, the average value of the emotional response delay, the standard deviation of the emotional volatility, the frequency of occurrence of the atypical response pattern, and the entropy value.

[0051] The risk assessment conclusion is based on the comparison results between the aforementioned characteristic indicators and the normal children's emotional database. The similarity is calculated using Euclidean distance, and the autism risk level is marked as low risk, medium risk, or high risk.

[0052] A second aspect of the present invention provides a system for constructing a child's emotion curve for autism risk assessment using the method described above, comprising:

[0053] The data acquisition module is used to acquire facial video data of test subjects when they watch emotionally stimulating videos;

[0054] The feature extraction module is used to extract facial action unit (AU) intensity features and macroscopic behavioral features from the facial video data to form a multimodal feature set.

[0055] The weight calculation module is used to calculate the dynamic weight coefficients of each feature in the multimodal feature set based on the effect size and significance level.

[0056] The emotion intensity calculation module is used to calculate the emotion intensity score based on the dynamic weighting coefficient and the real-time AU feature value, and generate an emotion curve.

[0057] The feature analysis module is used to extract feature indicators that characterize abnormal emotional responses in autism from the emotion curve;

[0058] The risk assessment module is used to conduct autism risk assessment based on the aforementioned characteristic indicators and output the results.

[0059] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0060] This invention integrates facial action unit (AU) features with macroscopic behavioral features and constructs a dynamic weighting calculation mechanism based on effect size and significance level. This enables refined and dynamic modeling of the emotional responses of test subjects, objectively capturing common atypical emotional response patterns in children with autism. It effectively overcomes the shortcomings of traditional assessment methods, such as strong human subjectivity and coarse feature extraction, and provides quantifiable and interpretable auxiliary diagnostic evidence for early screening of autism. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is the overall flowchart disclosed in this invention;

[0063] Figure 2 This is a schematic diagram of the principle of the deep learning-based facial AU feature extraction and emotion calculation model disclosed in this invention.

[0064] Figure 3 This is a schematic diagram of an emotion curve disclosed in an embodiment of the present invention. Detailed Implementation

[0065] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] This invention aims to provide a method and system for constructing emotional curves for children's autism risk assessment. Through multimodal feature fusion and dynamic weight calculation, it achieves precise quantification and visualization of the emotional response of the test subjects, providing an objective basis for early screening of autism.

[0067] Please see Figure 1 This paper showcases the core model architecture for facial action unit (AU) feature extraction and emotion intensity calculation in this solution. Its design aims to achieve refined and dynamic modeling from raw facial videos to emotion curves. Specifically:

[0068] 1. Local Gated Conv module

[0069] This module focuses on subtle features in local facial areas, such as muscle activity in key areas like the corners of the mouth and eyebrows. Through gated convolution, the model can adaptively emphasize emotion-related local features (such as cheek lifting in AU06 and upturned corners of the mouth in AU12) while suppressing irrelevant background information, thereby accurately capturing the low-intensity, transient, atypical micro-expressions commonly seen in children with autism.

[0070] 2. Global Multi-Head Mamba Module

[0071] Mamba, as a next-generation state-space model, excels at handling temporal dependencies in long-sequence data. Combining a multi-head mechanism, this module integrates global facial contextual information (such as spatial relationships between different AUs) and macroscopic behavioral features (such as head pose). This design ensures that the model can understand the dynamic wholeness of emotional expression, rather than analyzing individual AUs in isolation.

[0072] 3. Temporal Mixer Module

[0073] Since emotions are dynamic, temporal modeling is crucial. The Residual Temporal Mixer receives the fused features and is specifically responsible for capturing and mixing information along the temporal dimension. This module is dedicated to temporal modeling and ensures the training stability of deep networks through residual connections. Its core function is to capture the dynamic changing patterns of emotion intensity (e.g., the complete process of emotion from induction to peak and then to decay), thereby generating a smooth and continuous emotion curve and avoiding fluctuations caused by frame-level noise.

[0074] 4. Multi-path Decoder Prediction Weights

[0075] This section corresponds to the "Dynamic Weight Coefficient Calculation" process described below (i.e., step S3). The Decoder dynamically generates the weight coefficients for each AU based on its effect size (d-value), significance (p-value), and multimodal modulatory factor, ultimately obtaining the emotion intensity score through weighted fusion. This mechanism enables adaptive adjustment of the contribution of different AUs, enhancing the model's sensitivity to atypical emotional responses in autism.

[0076] Please see Figure 2 The method for constructing a child's emotion curve for autism risk assessment provided by this invention mainly includes the following steps:

[0077] S1. Collect facial video image sequences of test subjects while they watch emotionally stimulating videos.

[0078] Specifically, a camera device is used to record a sequence of frontal facial images of the test subjects while they watch emotionally stimulating videos (such as animated clips). The video frame rate is 30 frames per second, the duration is usually 1-3 minutes, and the shooting distance is 0.8-1.2 meters.

[0079] In this plan, the emotionally stimulating videos are those that contain at least positive, neutral, and negative emotional content to ensure the diversity of emotional triggers.

[0080] In a specific example:

[0081] Positive emotional content consists of 20-30 seconds of cheerful animation, featuring vibrant colors and pleasant sound effects.

[0082] Neutral content consists of 20-30 second videos of natural scenery, such as slowly flowing rivers or static trees.

[0083] The negative emotion content consists of 15-20 seconds of gentle, sad animation, such as a scene of raindrops falling, without strong negative stimuli.

[0084] This scheme aims to ensure the consistency and comparability of emotion induction through standardized stimuli.

[0085] S2. Extract multiple facial action unit (AU) feature values ​​and macroscopic behavioral features from the facial video image sequence to form a multimodal feature set.

[0086] Specifically, automated facial analysis tools (such as OpenFace) are used to process each frame of the captured video, extract facial regions, and identify relevant AU intensity values ​​and macroscopic behavioral features.

[0087] In this scheme, the extracted AU feature values ​​include, but are not limited to, AU01 (inner eyebrow raised), AU04 (eyebrow lowered), AU06 (cheek lifted), AU09 (nose wrinkled), AU12 (corner of mouth raised), AU15 (corner of mouth lowered), AU18 (lip tightened), AU20 (corner of mouth stretched), and AU25 (lips open). Among them, AU06, AU09, AU12, AU20, and AU25 are positively correlated with positive emotions, while AU01, AU04, and AU15 are negatively correlated with negative emotions.

[0088] The extracted AU feature values ​​are normalized, and they satisfy the formula:

[0089]

[0090] In the formula, Represents the minimum intensity value of a single AU across all acquisition frames; max(AU) is the maximum intensity value of a single AU across all acquisition frames; after normalization The value range is [0,1].

[0091] In this scheme, macroscopic behavioral characteristics include rotation and translation parameters of head posture, which are used to analyze an individual's attention allocation, focus of interest, and level of engagement in real time.

[0092] The combination of macro-level behavioral characteristics and micro-level facial action (AU) data forms a complete nonverbal behavior map.

[0093] S3. Based on predefined effect sizes and significance levels, calculate the dynamic weight coefficients of each feature in the multimodal feature set; among them, by introducing a multimodal adjustment factor, the AU feature values ​​and macro-behavioral features are weighted and fused.

[0094] In this scheme, the calculation process of the dynamic weighting coefficient includes the following steps:

[0095] S31. Normalize the effect size d value corresponding to the AU characteristic value to obtain the normalized effect size.

[0096] The normalized effect size satisfies the following formula:

[0097]

[0098] In the formula, The effect size of the current facial action unit (AU) is usually calculated by statistical methods (such as independent samples t-test) and is used to measure the strength of the effect of the AU in distinguishing different emotional states. This represents the set of effect sizes for all AUs, that is, a list consisting of the absolute values ​​of the effect sizes of all AUs to be evaluated.

[0099] S32. Determine the significance weight based on the p-value of the correlation between AU feature values ​​and the emotion dimension.

[0100] The significance weight is determined by the significance p-value. The mapping rules are as follows:

[0101]

[0102] S33. Calculate the multimodal adjustment factor based on AU eigenvalues ​​and macroscopic behavioral characteristics.

[0103] The formula for calculating the multimodal adjustment factor is as follows:

[0104]

[0105] In the formula, Indicates the multimodal adjustment factor; This represents the adjustment coefficient, with a default value of 0.1. This represents the micro-expression intensity features derived from facial action unit (AU) analysis, used to describe the degree of fine-grained facial muscle activity; High-level macro-emotional features representing images, voice, actions, or cross-modal emotion models provide a coarse-grained characterization of the overall emotional state.

[0106] S34. Multiply the normalized effect size, significance weight, and multimodal adjustment factor to obtain the dynamic weight coefficient.

[0107] The mechanism for determining the dynamic weighting coefficients of emotions in this scheme is based on the pleasure dimension in emotion psychology, and combines the significance p-value and effect size d-value provided in the paper "Sensing Emotional Valence and Arousal Dynamics through Automated Facial Action Unit Analysis" to construct a weighted coefficient generation algorithm. This algorithm first maps the significance p-value to the significance weights. Then, combined with the normalized value of the effect size Perform multiplication to determine the linear regression coefficients of each AU feature value in the sentiment score. .

[0108] The dynamic weighting coefficients satisfy the following expression:

[0109]

[0110] In the formula, Indicates the dynamic weighting coefficient; This represents the normalized effect size; Indicates significance weight; This represents the multimodal regulation factor.

[0111] In a specific example, dynamic weighting coefficients As shown in Table 1 below:

[0112] Table 1

[0113]

[0114] S4. Calculate the emotion intensity score at each time point based on the dynamic weight coefficient and the real-time AU feature value, and generate an emotion curve through time-series smoothing.

[0115] The formula for calculating the emotional intensity score is as follows:

[0116]

[0117] In the formula, This indicates the score for the intensity of emotion; Indicates the first Dynamic weighting coefficients for each AU; Indicates the first The and the first The interaction weight of each AU; Indicates the first Normalized intensity values ​​of each AU; Indicates the first Normalized intensity values ​​of each AU; This represents the total number of AUs participating in the analysis.

[0118] In a further proposed approach, a sliding window averaging method is used to smooth the emotion intensity score over time, satisfying the expression:

[0119]

[0120] In the formula, This indicates the intensity score of the emotion; It is a discrete-time index representing each point in time within the sliding window; Indicates window size; This indicates the current frame. In a specific example, a window of 30 frames is selected, with a step size of 5 frames.

[0121] S5. Extract characteristic indicators from the emotion curve to characterize abnormal emotional responses in autism, conduct autism risk assessment based on the characteristic indicators, and output the results.

[0122] Specifically, the characteristic indicators used to characterize abnormal emotional responses in autistic individuals include, but are not limited to, the following:

[0123] a. Emotional response amplitude, which is the extreme value of the emotional curve within a specific stimulus segment.

[0124] b. Emotional response delay, which is the time difference from the onset of the stimulus to the appearance of a significant change in the emotional curve.

[0125] c. Emotional volatility, which is the standard deviation or variance of the emotion curve.

[0126] d. Atypical response pattern, which is the number of times a pre-defined abnormal waveform appears in the emotion curve.

[0127] e. The complexity of emotional fluctuations, which is obtained by calculating the information entropy of the emotion curve, and satisfies the formula:

[0128]

[0129] In the formula, Indicates that the emotion value is within the range The probability of; This indicates the number of intervals for the emotion value.

[0130] The further proposed solution includes outputs such as a sentiment curve, a quantitative table of assessment indicators, and risk assessment conclusions. Specifically:

[0131] The horizontal axis of the emotion curve is the frame number, and the vertical axis is the emotion intensity value. The time intervals of videos with positive, neutral, and negative emotions are also marked.

[0132] The quantitative evaluation index table includes the maximum and minimum values ​​of emotional response amplitude, the average value of emotional response delay, the standard deviation of emotional fluctuation, the frequency of occurrence of atypical response patterns, and the entropy value.

[0133] The risk assessment conclusions are based on the comparison results of characteristic indicators with a database of normal children's emotions, and the similarity is calculated using Euclidean distance. The risk level of autism is marked as low, medium or high.

[0134] This invention also provides a system for constructing emotional curves for children's autism risk assessment using the above-mentioned method, mainly including a data acquisition module, a feature extraction module, a weight calculation module, an emotion intensity calculation module, a feature analysis module, and a risk assessment module, specifically:

[0135] The data acquisition module is used to obtain facial video data of test subjects when they watch emotionally stimulating videos.

[0136] The feature extraction module is used to extract facial action unit (AU) intensity features and macroscopic behavioral features from facial video data to form a multimodal feature set.

[0137] The weight calculation module is used to calculate the dynamic weight coefficients of each feature in the multimodal feature set based on the effect size and significance level.

[0138] The emotion intensity calculation module is used to calculate the emotion intensity score based on the dynamic weight coefficient and the real-time AU feature value, and generate an emotion curve.

[0139] The feature analysis module is used to extract feature indicators that characterize abnormal emotional responses in autistic individuals from the emotion curve.

[0140] The risk assessment module is used to conduct autism risk assessment based on characteristic indicators and output the results.

[0141] Example

[0142] Facial images of 5-year-old children watching an emotionally stimulating video for 3 minutes were collected. AU features were extracted using OpenFace, and dynamic weights were calculated to generate emotion curves, such as... Figure 3 As shown in the figure. The horizontal axis represents the frame number, and the vertical axis represents the intensity of positive emotions. A positive emotion intensity greater than 0 indicates a positive emotion, and a positive emotion intensity less than 0 indicates a negative emotion.

[0143] The following feature indicators were extracted from the generated sentiment curve:

[0144] Emotional response magnitude: Maximum response value for positive segments is 0.85 (normal range 0.7-1.0).

[0145] Emotional reaction delay: 2.3 seconds (normal <2.0 seconds).

[0146] Emotional fluctuation entropy value: 1.2 seconds (normal > 1.5 seconds).

[0147] The overall similarity score is 0.45, indicating a medium risk for autism. Further clinical evaluation is recommended.

[0148] This invention enables an objective assessment of autism risk by precisely quantifying children's emotional responses, providing an effective tool for early autism screening.

[0149] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a child's emotion curve for autism risk assessment, characterized in that, Includes the following steps: Collect facial video image sequences of test subjects while they watch emotionally stimulating videos; Multiple facial action unit (AU) feature values ​​and macroscopic behavioral features are extracted from the facial video image sequence to form a multimodal feature set; Based on predefined effect sizes and significance levels, the dynamic weight coefficients of each feature in the multimodal feature set are calculated; wherein, by introducing a multimodal adjustment factor, the AU feature values ​​and the macro-behavioral features are weighted and fused; wherein, the calculation process of the dynamic weight coefficients includes: The effect size d corresponding to the AU feature value is normalized to obtain the normalized effect size; The significance weights are determined based on the p-values ​​of the correlation between the AU feature values ​​and the emotion dimension. Based on the AU feature values ​​and the macroscopic behavioral features, calculate the multimodal adjustment factor; The normalized effect size, the significance weight, and the multimodal adjustment factor are multiplied together to obtain the dynamic weight coefficient. Based on the dynamic weighting coefficients and real-time AU feature values, the emotion intensity score at each time point is calculated, and an emotion curve is generated through time-series smoothing. Feature indicators for characterizing abnormal emotional responses in autism are extracted from the emotion curve, and autism risk assessment is performed based on the feature indicators, with the results output.

2. The method for constructing a child's emotional curve for autism risk assessment according to claim 1, characterized in that, The emotionally stimulating video is a video that contains at least positive, neutral, and negative emotional content.

3. The method for constructing a child's emotional curve for autism risk assessment according to claim 1, characterized in that: The AU feature values ​​include AU01, AU04, AU06, AU09, AU12, AU15, AU18, AU20, and AU25; wherein, AU06, AU09, AU12, AU20, and AU25 are positively correlated with positive emotions, and AU01, AU04, and AU15 are negatively correlated with negative emotions. The macroscopic behavioral characteristics include the rotation and translation parameters of the head posture.

4. The method for constructing a child's emotion curve for autism risk assessment according to claim 1, characterized in that, The formula for calculating the multimodal adjustment factor is as follows: In the formula, This represents the multimodal adjustment factor; Indicates the adjustment coefficient; Indicates the intensity of action units at the micro-expression level; This represents a quantitative value indicating emotional / behavioral characteristics at the macro level.

5. The method for constructing a child's emotional curve for autism risk assessment according to claim 1, characterized in that, The mapping rule for determining the significance weight based on the significance p-value is as follows: in, Indicates the significance weight.

6. The method for constructing a child's emotion curve for autism risk assessment according to claim 1, characterized in that, The formula for calculating the emotional intensity score is as follows: In the formula, This indicates the intensity score of the emotion. Indicates the first Dynamic weighting coefficients for each AU; Indicates the first Normalized intensity values ​​of each AU; Indicates the first The and the first The interaction weight of each AU; Indicates the first Normalized intensity values ​​of each AU; , They represent the first , One AU; This represents the total number of AUs.

7. The method for constructing a child's emotional curve for autism risk assessment according to claim 1, characterized in that, The time-series smoothing process employs a sliding window averaging method, which satisfies the expression: In the formula, This indicates the score for the intensity of emotion; It is a discrete-time index representing each point in time within the sliding window; Indicates window size; Indicates the current frame.

8. The method for constructing a child's emotional curve for autism risk assessment according to claim 7, characterized in that, The feature indicators include at least one of the following: The amplitude of the emotional response is the extreme value of the emotional curve within a specific stimulus segment. Emotional response delay is the time difference between the onset of a stimulus and the occurrence of a significant change in the emotional curve. Emotional volatility, which is the standard deviation or variance of the aforementioned emotion curve; Atypical response pattern, which is the number of times a preset abnormal waveform appears in the emotion curve; The complexity of emotional fluctuations, obtained by calculating the information entropy of the emotional curve, satisfies the following formula: In the formula, Indicates that the emotion value is within the range The probability of; This indicates the number of intervals for the emotion value.

9. The method for constructing a child's emotion curve for autism risk assessment according to claim 8, characterized in that, The output results include: Sentiment curve chart; The evaluation index quantification table includes the maximum and minimum values ​​of the emotional response amplitude, the average value of the emotional response delay, the standard deviation or variance of the emotional curve, and the number of times a preset abnormal waveform appears in the emotional curve. The risk assessment conclusion is based on the comparison results between the aforementioned characteristic indicators and the normal children's emotional database. The similarity is calculated using Euclidean distance, and the autism risk level is marked as low risk, medium risk, or high risk.

10. A system for constructing emotional curves for children's autism risk assessment using the method described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire facial video data of test subjects when they watch emotionally stimulating videos; The feature extraction module is used to extract facial action unit (AU) intensity features and macroscopic behavioral features from the facial video data to form a multimodal feature set. The weight calculation module is used to calculate the dynamic weight coefficients of each feature in the multimodal feature set based on the effect size and significance level. The emotion intensity calculation module is used to calculate the emotion intensity score based on the dynamic weighting coefficient and the real-time AU feature value, and generate an emotion curve; The feature analysis module is used to extract feature indicators that characterize abnormal emotional responses in autism from the emotion curve; The risk assessment module is used to conduct autism risk assessment based on the aforementioned characteristic indicators and output the results.

Images