Early autism assessment method based on man-machine touch interaction

Abstract

The invention provides an autism early evaluation method based on man-machine tactile interaction. The autism early evaluation method is used for an autism early evaluation system based on man-machine tactile interaction. The system comprises an accompanying robot which is used as an interaction object of a testee and issues a touch instruction to the testee; the tactile acquisition module comprises array type tactile sensors arranged on a plurality of body parts of the accompanying robot and is used for acquiring tactile interaction data when the testee interacts with the accompanying robot; the tactile sensing module is used for performing action recognition and position detection on the tactile interaction data acquired by the tactile acquisition module through a designed tactile gesture recognition model to obtain a data sensing result comprising a tactile gesture type and a touch position; the autism evaluation module is used for outputting a tactile behavior evaluation result of the testee according to the tactile interaction data through a designed autism tactile analysis model; and an autism early evaluation result is output according to the tactile interaction data through a designed autism tactile evaluation model.

Description

Technical Field

[0001] The present invention relates to the fields of machine touch, human-computer interaction and pattern recognition, and in particular to an early assessment method for autism based on human-computer touch interaction. Background Art

[0002] Autism is one of the common brain neurodevelopmental disorders. The number of children suffering from autism is huge, and the incidence rate is increasing year by year. Autism not only causes language and behavioral damage to children, but also causes serious social interaction and emotional cognitive disorders, affecting children's social development and quality of life. Therefore, early screening and intervention treatment of autism has become an urgent problem to be solved in rehabilitation training.

[0003] Traditional autism diagnosis methods mainly rely on behavioral observation and developmental history assessment. The doctor will conduct detailed interviews with the child and his or her parents to understand the child's development in language, social interaction, and repetitive behaviors. Commonly used diagnostic tools include the Autism Diagnostic Interview-Revised (ADI-R) and the Autism Diagnostic Observation Scale (ADOS). ADI-R is a structured interview that mainly assesses a child's communication, social interaction, and behavior patterns by asking parents or caregivers. ADOS directly observes a child's social behavior, communication skills, and play activities through interaction with the child. However, the above methods require professional physicians to conduct tests in the clinic, and the testing process is relatively complicated and not easy to promote. In addition, the above subjective factors affect the test results.

[0004] At present, autism assisted diagnosis and treatment technology based on accompanying robots has been widely studied. Robots can monitor children's gaze, behavior, facial expressions and other information, provide quantitative indicators for the assessment of autism, and can also intervene and treat through gamification training. For example, Liu Honghai et al. from Harbin Institute of Technology [Liu j, Wang Z, Xu K, et al. Early screening of autism in toddlers via response-to-instructions protocol [J]. IEEE Transactions on Cybernetics, 2022, 52(5): 3914-3924.] developed an early screening system for autism based on expert consensus, using a multi-sensor fusion network to obtain behavioral characteristics during the interaction between children and doctors, and designed an automatic line of sight estimation method based on multi-sensor data fusion and probability model, which improved the accuracy and universality of autism assisted diagnosis. At present, autism assisted diagnosis and treatment systems are mainly based on visual technology. This method is mostly carried out under laboratory conditions. In real environments, it is difficult to accurately capture children's gaze and behavioral movements due to occlusion and perspective issues. In addition, autistic children are often taciturn and lack facial expressions, and it is difficult to obtain comprehensive behavioral and emotional characteristics by relying solely on visual and auditory technology, which brings certain limitations to the intelligent diagnosis and treatment of autism.

[0005] Touch is essential for children to establish social relationships. Especially in the early stages when children's language systems are not yet mature, touch can promote children's emotional development, enhance their sense of security and trust, and thus increase their willingness to communicate with others. However, children with autism have difficulties in perceiving and responding to tactile stimuli, and often find it difficult to make normal tactile social movements. Studies have shown that up to 77% of children with autism have tactile dysfunction. Therefore, using tactile sensing technology to analyze children's behavioral movements can not only broaden the observation perspective of the autism diagnosis and treatment system, but also provide objective and quantitative evaluation indicators to improve diagnostic accuracy.

[0006] In order to solve the above problems, people have been seeking an ideal technical solution. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the prior art and provide an early assessment method for autism based on human-computer tactile interaction. By using a tactile interaction device to collect the interactive movements of autistic children and using tactile interaction detection technology to identify the abnormal movements of the children, an objective judgment basis is provided for the early screening of autism.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is: In a first aspect, the present invention provides an early assessment system for autism based on human-computer tactile interaction, comprising: The companion robot, as the subject's interaction object, issues touch commands to the subject; A tactile collection module, including array-type tactile sensors arranged at multiple body parts of the companion robot, for collecting tactile interaction data when the subject interacts with the companion robot; The tactile perception module uses the designed tactile gesture recognition model to perform action recognition and position detection on the tactile interaction data collected by the tactile acquisition module, and obtains data perception results including tactile gesture type and touch position; The autism assessment module uses the designed autism tactile analysis model to output the subject's tactile behavior assessment results based on the tactile interaction data; The designed autism tactile assessment model also outputs early autism assessment results based on tactile interaction data; Among them, the autism tactile analysis model is used to count the number of correct responses of the subjects based on touch instructions and data perception results, and evaluate the tactile behavior of autistic children; The tactile gesture recognition model is a neural network model trained using a tactile interaction dataset; The autism tactile assessment model is a neural network model trained using the autism tactile assessment dataset.

[0009] Based on the above, the construction of the tactile interaction dataset includes the following steps: (1) Recruiting a group of normally developing children as participants, and showing them five common gestures, including hitting, stroking, patting, poking, and knocking, through videos; (2) During the data collection process, one participant was selected each time to interact with the companion robot; (3) During the interaction between each participant and the companion robot, one gesture was randomly shown to the participant. Each gesture was collected 10 times at the back, arm, and head. After each gesture was collected, there was a 5-minute break until all five gestures were collected. (4) After collecting the gesture data of all participants, the total sample is obtained; (5) Assign labels to the total recorded samples, label the gesture corresponding to each sample, and obtain the gesture label and position label.

[0010] Based on the above, the tactile gesture recognition model includes: The first 3D convolutional layer is used to transform the original data X i Perform 3D convolution operation to obtain feature map f 1; The first pooling layer is used to pool the feature map f1 Use the filter to filter the feature map f 1 Perform maximum pooling, and then use the RELU activation function for nonlinear activation operation to obtain the feature map ; The second 3D convolutional layer is used to map the feature map Perform 3D convolution operation to obtain feature map f 2; The second pooling layer is used to pool the feature maps f 2 Use filters to filter the feature map f 2 Perform maximum pooling, and then use the RELU activation function for nonlinear activation operation to obtain the feature map ; The third 3D convolution layer is used to map the feature map Perform 3D convolution operation to obtain feature map f 3; Global average pooling layer, used to pool feature maps f 3. Perform global average pooling to obtain 1D features; The gesture classifier and the position classifier are used to classify the 1D features and output the gesture type and the touch position respectively; wherein the gesture classifier and the position classifier are both two-layer linear layers.

[0011] Based on the above, the construction of the autism tactile assessment dataset includes the following steps: (1) Recruiting a group of children with autism as the patient group and a group of children with normal development as the control group; (2) During the data collection process of each round of testing, one child from the patient group and the control group was randomly selected as a participant to interact with the companion robot; The companion robot issued a touch command to the participant, and the therapist observed whether the participant would respond and recorded the participant group, gesture category, and touch position as a sample after each touch command was issued; For incorrectly performed or unresponsive behaviors, the therapist guided the participant by demonstrating the action, and then the companion robot issued the touch command again; (3) Each round of testing for each participant issues four touch commands, with a certain time interval between each two touch commands; after each round of testing, the companion robot takes a break before the next round of testing. The companion robot issues a touch command to the participant that is different from the previous round of testing, and repeats the steps of the previous round of testing until each participant has repeated three rounds of testing; (4) After collecting data from all participants, the total sample is obtained; (5) Label the recorded samples and number the gestures corresponding to each sample to obtain participant labels, gesture labels, and position labels; (6) Data cleaning and data set construction; Detect the data and remove noise and outliers in the data; Divide the total cleaned samples into training set and test set: The data of several autistic children were randomly selected from the children with autism group, and the data of the same number of normally developed children were randomly selected from the control group, and put into the training set for the training and adjustment of the autism tactile assessment model; The remaining data of autistic children and normal children were put into the test set to evaluate the effectiveness and generalization performance of the autism tactile assessment model.

[0012] Based on the above, the autism tactile assessment model includes a basic network, a gesture classification network, a position detection network, and an autism recognition network; The basic network is used to extract time domain and space domain features from touch data; The gesture classification network uses a machine learning method to determine whether the subject can make a correct action in response to the touch command issued by the companion robot; The position detection network uses a machine learning method to determine whether the touch command issued by the subject to the companion robot can be executed to the correct part of the companion robot; The autism identification network uses a machine learning method to analyze the differences in movement characteristics between autistic children and normally developing children, and assess the autism risk of the subjects.

[0013] Based on the above, a multi-task learning method is used in the autism tactile assessment model to train the gesture classification network, position detection network and autism recognition network; In each training iteration, the entropy of each network output probability vector is calculated as the task uncertainty, and the inverse of the uncertainty is used as the weight. For tasks with higher uncertainty, the weight of its loss function is lower, thereby achieving adaptive adjustment of the weight of each task.

[0014] Based on the above, the array structure of the array tactile sensor is 4×4, and the sampling frame rate is 50fps.

[0015] In a second aspect, the present invention provides an early assessment method for autism based on human-computer tactile interaction, which is used in the early assessment system for autism based on human-computer tactile interaction, and comprises the following steps: Step 1: After the subject is ready, the test begins. The companion robot randomly issues a touch command to the subject, and then collects the subject's interactive actions of executing the touch command through the tactile acquisition module and transmits it to the tactile perception module in real time; Step 2: The companion robot chooses to repeat step 1 or conduct the next round of testing through the output of the perception module; Step 3: The companion robot outputs the perception module and repeats steps 1 and 2 for multiple rounds of testing; The time interval between each two rounds of testing shall not be less than 5 minutes; After multiple rounds of testing are completed, the autism assessment module outputs tactile behavior assessment results and autism early assessment results; The method to obtain the tactile behavior assessment results is: The autism tactile analysis model compares the gesture type and touch position output by the tactile perception module with the instructions issued by the companion robot. Complete consistency is scored as 0 points; correct behavior after guidance is scored as 1 point; incorrect behavior after guidance is scored as 2 points; and no response after guidance is scored as 3 points. Accumulating the scores of multiple rounds of tests to obtain a tactile behavior assessment score as a tactile behavior assessment result; Methods for obtaining early assessment results of autism are: The autism tactile assessment model is used to evaluate the tactile interaction data of each round of testing to obtain multiple groups of autism assessment results; the multiple groups of autism assessment results obtained are fused using an average weighted method to output the type of the subject, i.e., normal development or autism.

[0016] The present invention has the following beneficial effects and advantages: (1) The present invention does not rely on parent interviews and clinical tests. By processing the tactile data during the human-machine interaction between the subject and the companion robot, the subject's behavior in the screening is qualitatively judged and the subject's tactile function is evaluated, providing an objective and quantitative evaluation standard for the early screening of children with autism.

[0017] (2) The present invention uses a companion robot as an interaction object for autistic children, which can collect natural interaction data and improve the fun of screening and the involvement of children.

[0018] (3) By collecting physical interaction information, the present invention avoids the problem of difficulty in obtaining data due to occlusion and viewing angle in visual detection, and further broadens the observation perspective of the autism diagnosis and treatment system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a structural block diagram of an early autism assessment system based on human-computer tactile interaction.

[0020] Figure 2 Diagram of the tactile gesture recognition model.

[0021] Figure 3 Diagram of a tactile assessment model for autism.

[0022] Figure 4 This is a flowchart of the early autism assessment method based on human-computer tactile interaction. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0024] Example 1 like Figure 1 As shown, this embodiment provides an early assessment system for autism based on human-computer tactile interaction, including: The companion robot, as the subject's interaction object, issues touch commands to the subject; A tactile collection module, including array-type tactile sensors arranged at multiple body parts of the companion robot, for collecting tactile interaction data when the subject interacts with the companion robot; The tactile perception module uses the designed tactile gesture recognition model to perform action recognition and position detection on the tactile interaction data collected by the tactile acquisition module, and obtains data perception results including tactile gesture type and touch position; The autism assessment module uses the designed autism tactile analysis model to output the subject's tactile behavior assessment results based on the tactile interaction data; The designed autism tactile assessment model also outputs early autism assessment results based on tactile interaction data; Among them, the autism tactile analysis model is used to count the number of correct responses of the subjects based on touch instructions and data perception results, and evaluate the tactile behavior of autistic children; The tactile gesture recognition model is a neural network model trained using a tactile interaction dataset; The autism tactile assessment model is a neural network model trained using the autism tactile assessment dataset.

[0025] Example 2 The difference between this embodiment and Embodiment 1 is that: a specific autism early assessment system based on human-computer tactile interaction is provided.

[0026] Companion Robot The Nao robot used is from the French company Aldebaran Robotics. The Nao robot has visual, auditory detection and language conversation capabilities and can interact with people.

[0027] Tactile acquisition module Array tactile sensors are set on the back, arms and head of the companion robot to collect tactile interaction data (actions) when the subject interacts with the companion robot. The array tactile sensor uses the matrix flexible film pressure sensor M0404 of Viccos, which has a 4-row × 4-column pattern, a total of 16 sensing points, a sensor spatial resolution of 20mm, and a sampling frequency of 16Hz; the sensor can collect pressure information for analyzing the subject's interactive actions.

[0028] Tactile sensing module Building a tactile interaction dataset includes the following steps: (1) Recruit a group of normally developing (non-autistic) children as participants, 20 of whom will be selected, and show the five commonly used gestures, including hitting, stroking, patting, poking, and knocking, to each participant through a video; (2) During the data collection process, one participant was selected each time to interact with the companion robot; (3) During the interaction between each participant and the companion robot, one gesture was randomly shown to the participant. Each gesture was collected 10 times at the back, arm, and head. After each gesture was collected, there was a 5-minute break until all five gestures were collected. (4) After collecting the gesture data of 20 participants, a total of 3000 samples (20×5×3×10) were obtained; the data format of each sample is , where 4×4 represents the spatial distribution of the flexible film pressure sensors, and 16 is the number of touch gesture sequence frames; (5) Assign labels to the total number of recorded samples, label the gestures corresponding to each sample, and obtain the gesture labels l g ∈ {1,2,3,4,5} and position labels l a ∈ {1,2,3}.

[0029] like Figure 2 As shown, the tactile gesture recognition model includes: The first 3D convolution layer includes 32 3×3×3 3D convolution kernels, which are used to transform the original data X i Perform 3D convolution operation to obtain feature map f 1; The first pooling layer is used to pool the feature map f 1 Use a 2×2 filter with a step size of 2 to filter the feature map f 1 Perform maximum pooling, and then use the RELU activation function for nonlinear activation operation to obtain the feature map ; The second 3D convolution layer includes 128 3×3×3 3D convolution kernels for feature maps. Perform 3D convolution operation to obtain feature map f 2; The second pooling layer is used to pool the feature maps f 2 Use a 2×2 filter and filter the feature map with a step size of 2 f 2 Perform maximum pooling, and then use the RELU activation function for nonlinear activation operation to obtain the feature map ; The third 3D convolution layer includes 256 3×3×3 3D convolution kernels, which are used to transform the feature map Perform 3D convolution operation to obtain feature map f 3; Global average pooling layer, used to pool feature maps f 3. Perform global average pooling to obtain 256-dimensional 1D features; The gesture classifier and position classifier are used to classify the 1D features and output the gesture type and touch position respectively. The gesture classifier and position classifier are both two-layer linear layers, and the neuron mapping of the two linear layers is 256→64 and 64→ y (Gesture Classifier y =5, location classifier y =3).

[0030] Autism Assessment Module The construction of the autism tactile assessment dataset includes the following steps: (1) Recruit a group of children with autism as the patient group (40 children are planned to be selected) and a group of children with normal development (non-autistic) as the control group (40 children are planned to be selected); (2) During the data collection process of each round of testing, one child from the patient group and the control group was randomly selected as a participant to interact with the companion robot; The companion robot issues a touch command to the participant (e.g., “Please pat my back gently,” “Please poke my arm,” or “Please touch my head”). The therapist observes whether the participant responds and records the participant group (whether the child is autistic), gesture category, and touch position as a sample after each touch command is issued; For incorrectly performed or unresponsive behaviors, the therapist guided the participant by demonstrating the action, and then the companion robot issued the touch command again; (3) Each round of testing for each participant issues four touch commands, with a certain time interval between each two touch commands; after each round of testing, the companion robot takes a break before the next round of testing. The companion robot issues a touch command to the participant that is different from the previous round of testing, and repeats the steps of the previous round of testing until each participant has repeated three rounds of testing; (4) After collecting data from all participants, the total sample is obtained; (5) Label the recorded samples and number the gestures corresponding to each sample to obtain the participant labels. l s ∈ {1,2} (1 for normally developed children, 2 for autistic children), gesture label l g ∈ {1,2,3,4,5} (representing hit, stroke, slap, poke, and knock, respectively) and position labels l a ∈ {1,2,3} (representing back, arms, and head respectively); (6) Data cleaning and data set construction; Since noise is inevitably mixed in during data collection, the data must be cleaned first after it is obtained. The data of subjects who do not cooperate or cannot complete the experiment must be eliminated. At the same time, the data must be tested to remove noise and outliers in the data.

[0031] Divide the total cleaned samples into training set and test set: The data of 30 autistic children were randomly selected from the autism group, and the data of 30 normally developing children were randomly selected from the control group, and put into the training set for the training and adjustment of the autism tactile assessment model; The data of the remaining 20 autistic children and normal children were put into the test set to evaluate the effectiveness and generalization performance of the autism tactile assessment model.

[0032] 1.1 Results of tactile behavior assessment The constructed autism tactile analysis model is: Focusing on the participation, accuracy and repeatability of tactile interaction behaviors, extract the tactile characteristics of interactive actions, including mechanics, time domain and kinematics; A test of homogeneity of variance was performed for each tactile feature to determine whether the variances of the two groups were equal; Under the premise of homogeneity of variance, test whether the data conforms to normal distribution; Two-way ANOVA was used to parse out the differences between different tactile features in children with autism and typically developing children.

[0033] 2. Results of early autism assessment The autism tactile assessment model includes a base network, a gesture classification network, a position detection network, and an autism recognition network.

[0034] The basic network uses a three-layer 3D convolutional neural network to extract the temporal and spatial domain features of touch data. The convolution kernel size is set to 3×3×3, and the number of convolution kernels in each layer is 32, 128, and 256; Gesture classification network,designs two linear layers to determine whether the subject can make correct,actions in response to the touch commands issued by the companion robot.,The number of linear layers in each layer is 64,5, respectively, and the gesture type is output; Position detection network,designs two linear layers to determine whether the touch command issued by the subject to the companion robot can be executed to the correct part of the companion robot.,The number of linear layers in each layer is 64 and 3 respectively, and the touch position is output; Autism recognition network,designs two linear layers to analyze the differences in movement characteristics between autistic children and normally developing children,,assisting in identifying the autism risk of the subjects.,The number of linear layers in each layer is 64 and 2 respectively,,and the subject type is output.

[0035] The autism tactile assessment model uses multi-task learning to train gesture classification networks, position detection networks, and autism identification networks. In multi-task learning, the magnitude of loss function values ​​and loss gradients of different tasks are often different. How to assign appropriate weights to different tasks to obtain the best model performance is an issue that needs to be focused on. The autism tactile assessment model intends to use an uncertainty weighting method to dynamically adjust the weight of its loss function according to the uncertainty of each task. Specifically, in each training iteration, the entropy of the probability vector output by each network is calculated as the task uncertainty, and the inverse of the uncertainty is used as the weight. For tasks with higher uncertainty, the weight of its loss function is lower, so that the weight of each task can be adaptively adjusted.

[0036] Example 3 This embodiment provides a method for early assessment of autism based on human-computer tactile interaction, which is used in the early assessment of autism based on human-computer tactile interaction described in Example 2. Figure 4 As shown, the following steps are included: Step 1: After the subject is ready, the test begins. The companion robot randomly issues a touch command to the subject, and then collects the subject's interactive actions of executing the touch command through the tactile acquisition module and transmits it to the tactile perception module in real time; Step 2: The companion robot chooses to repeat step 1 or conduct the next round of testing through the output of the perception module; Step 3: The companion robot outputs the perception module and repeats steps 1 and 2 for multiple rounds of testing; The time interval between each two rounds of testing shall not be less than 5 minutes; After multiple rounds of testing are completed, the autism assessment module outputs tactile behavior assessment results and autism early assessment results; The method to obtain the tactile behavior assessment results is: The autism tactile analysis model compares the gesture type and touch position output by the tactile perception module with the instructions issued by the companion robot. Complete consistency is scored as 0 points; correct behavior after guidance is scored as 1 point; incorrect behavior after guidance is scored as 2 points; and no response after guidance is scored as 3 points. Accumulating the scores of multiple rounds of tests to obtain a tactile behavior assessment score as a tactile behavior assessment result; Methods for obtaining early assessment results of autism are: The autism tactile assessment model is used to evaluate the tactile interaction data of each round of testing to obtain multiple groups of autism assessment results; the multiple groups of autism assessment results obtained are fused using an average weighted method to output the type of the subject, i.e., normal development or autism.

[0037] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. An early assessment system for autism based on human-computer tactile interaction, characterized in that: include: The companion robot, as the subject's interaction object, issues touch commands to the subject; A tactile collection module, including array-type tactile sensors arranged at multiple body parts of the companion robot, for collecting tactile interaction data when the subject interacts with the companion robot; The tactile perception module uses the designed tactile gesture recognition model to perform action recognition and position detection on the tactile interaction data collected by the tactile acquisition module, and obtains data perception results including tactile gesture type and touch position; The autism assessment module uses the designed autism tactile analysis model to output the subject's tactile behavior assessment results based on the tactile interaction data; The designed autism tactile assessment model also outputs early autism assessment results based on tactile interaction data; Among them, the autism tactile analysis model is used to count the number of correct responses of the subjects based on touch instructions and data perception results, and evaluate the tactile behavior of autistic children; The tactile gesture recognition model is a neural network model trained using a tactile interaction dataset; The autism tactile assessment model is a neural network model trained using the autism tactile assessment dataset.

2. The autism early assessment system based on human-computer tactile interaction according to claim 1 is characterized in that: The construction of the tactile interaction dataset includes the following steps: (1) Recruiting a group of normally developing children as participants, and showing them five common gestures, including hitting, stroking, patting, poking, and knocking, through videos; (2) During the data collection process, one participant was selected each time to interact with the companion robot; (3) During the interaction between each participant and the companion robot, one gesture was randomly shown to the participant. Each gesture was collected 10 times at the back, arm, and head. After each gesture was collected, there was a 5-minute break until all five gestures were collected. (4) After collecting the gesture data of all participants, the total sample is obtained; (5) Assign labels to the total recorded samples, label the gesture corresponding to each sample, and obtain the gesture label and position label.

3. The autism early assessment system based on human-computer tactile interaction according to claim 1, characterized in that: The tactile gesture recognition model includes: The first 3D convolutional layer is used to transform the original data X i Perform 3D convolution operation to obtain feature map f 1; The first pooling layer is used to pool the feature map f 1 Use the filter to filter the feature map f 1 Perform maximum pooling, and then use the RELU activation function for nonlinear activation operation to obtain the feature map ; The second 3D convolutional layer is used to map the feature map Perform 3D convolution operation to obtain feature map f 2; The second pooling layer is used to pool the feature maps f 2 Use filters to filter the feature map f 2 Perform maximum pooling, and then use the RELU activation function for nonlinear activation operation to obtain the feature map ; The third 3D convolution layer is used to map the feature map Perform 3D convolution operation to obtain feature map f 3; Global average pooling layer, used to pool feature maps f 3. Perform global average pooling to obtain 1D features; The gesture classifier and the position classifier are used to classify the 1D features and output the gesture type and the touch position respectively; wherein the gesture classifier and the position classifier are both two-layer linear layers.

4. The autism early assessment system based on human-computer tactile interaction according to claim 1, characterized in that: The construction of the autism tactile assessment dataset includes the following steps: (1) Recruiting a group of children with autism as the patient group and a group of children with normal development as the control group; (2) During the data collection process of each round of testing, one child from the patient group and the control group was randomly selected as a participant to interact with the companion robot; The companion robot issued a touch command to the participant, and the therapist observed whether the participant would respond and recorded the participant group, gesture category, and touch position as a sample after each touch command was issued; For incorrectly performed or unresponsive behaviors, the therapist guided the participant by demonstrating the action, and then the companion robot issued the touch command again; (3) Each round of testing for each participant issues four touch commands, with a certain time interval between each two touch commands; after each round of testing, the companion robot takes a break before the next round of testing. The companion robot issues a touch command to the participant that is different from the previous round of testing, and repeats the steps of the previous round of testing until each participant has repeated three rounds of testing; (4) After collecting data from all participants, the total sample is obtained; (5) Label the recorded samples and number the gestures corresponding to each sample to obtain participant labels, gesture labels, and position labels; (6) Data cleaning and data set construction; Detect the data and remove noise and outliers in the data; Divide the total cleaned samples into training set and test set: The data of several autistic children were randomly selected from the children with autism group, and the data of the same number of normally developed children were randomly selected from the control group, and put into the training set for the training and adjustment of the autism tactile assessment model; The remaining data of autistic children and normal children were put into the test set to evaluate the effectiveness and generalization performance of the autism tactile assessment model.

5. The autism early assessment system based on human-computer tactile interaction according to claim 1, characterized in that: The autism tactile assessment model includes a basic network, a gesture classification network, a position detection network, and an autism recognition network; The basic network is used to extract time domain and space domain features from touch data; The gesture classification network uses a machine learning method to determine whether the subject can make a correct action in response to the touch command issued by the companion robot; The position detection network uses a machine learning method to determine whether the touch command issued by the subject to the companion robot can be executed to the correct part of the companion robot; The autism identification network uses a machine learning method to analyze the differences in movement characteristics between autistic children and normally developing children, and assess the autism risk of the subjects.

6. The autism early assessment system based on human-computer tactile interaction according to claim 5, characterized in that: In the autism tactile assessment model, multi-task learning is used to train the gesture classification network, position detection network and autism recognition network; In each training iteration, the entropy of each network output probability vector is calculated as the task uncertainty, and the inverse of the uncertainty is used as the weight. For tasks with higher uncertainty, the weight of its loss function is lower, thereby achieving adaptive adjustment of the weight of each task.

7. The early autism assessment system based on human-computer tactile interaction according to any one of claims 1 to 6, characterized in that: The array structure of the array tactile sensor is 4×4, and the sampling frame rate is 50fps.

8. A method for early assessment of autism based on human-computer tactile interaction, used in the system for early assessment of autism based on human-computer tactile interaction as claimed in any one of claims 1 to 7, characterized in that: The following steps are involved: Step 1: After the subject is ready, the test begins. The companion robot randomly issues a touch command to the subject, and then collects the subject's interactive actions of executing the touch command through the tactile acquisition module and transmits it to the tactile perception module in real time; Step 2: The companion robot chooses to repeat step 1 or conduct the next round of testing through the output of the perception module; Step 3: The companion robot outputs the perception module and repeats steps 1 and 2 for multiple rounds of testing; The time interval between each two rounds of testing shall not be less than 5 minutes; After multiple rounds of testing are completed, the autism assessment module outputs tactile behavior assessment results and autism early assessment results; The method to obtain the tactile behavior assessment results is: The autism tactile analysis model compares the gesture type and touch position output by the tactile perception module with the instructions issued by the companion robot. Complete consistency is scored as 0 points; correct behavior after guidance is scored as 1 point; incorrect behavior after guidance is scored as 2 points; and no response after guidance is scored as 3 points. Accumulating the scores of multiple rounds of tests to obtain a tactile behavior assessment score as a tactile behavior assessment result; Methods for obtaining early assessment results of autism are: The autism tactile assessment model is used to evaluate the tactile interaction data of each round of testing to obtain multiple groups of autism assessment results; the multiple groups of autism assessment results obtained are fused using an average weighted method to output the type of the subject, i.e., normal development or autism.

Images