A community pension analysis and management method based on multi-dimensional perception and artificial intelligence

Abstract

The present invention discloses a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence, which relates to the field of smart elderly care and is mainly used for community smart elderly care. Key points of the solution: Collect real-time community videos, obtain various behavior categories of the target in the real-time video; mark the target pixel points in the real-time video, obtain the target image, and identify it; obtain images of various areas of the target, obtain the real-time expression complexity of the target, build a scoring model, and obtain the daily knowledge graph of the target according to the scoring model; set scoring weights, obtain the updated daily knowledge graph, obtain the feature vector of the daily knowledge graph of the target, and measure the similarity of the daily knowledge graph of the target; judge the adaptability of the target in the current community according to the results of the similarity measurement. The technical means proposed by the present invention scores the expressions of the elderly according to the complexity of the expression, and optimizes the expression scoring model according to the data fluctuation characteristics of the elderly, so that the model evaluation results are more accurate.

Description

Technical Field

[0001] The present invention relates to the field of smart elderly care, and in particular to a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence. Background Art

[0002] With the advent of the knowledge economy and network economy, smart elderly care has become the fundamental guarantee for a new lifestyle in the new century. With the development of society, not only has the aging problem in my country become increasingly prominent, but also the home-based elderly care services in my country have problems such as insufficient supply, low proportion, and low quality, which cannot meet the growing needs of the elderly. Therefore, with the continuous increase of China's aging population, elderly care has become an increasingly prominent social issue, and various elderly care models are being explored.

[0003] Existing technologies mainly construct knowledge graphs through the semantic information of community data, sensor data and data text, and obtain abnormal status information of the elderly through knowledge graph matching and analysis, thereby realizing smart elderly care.

[0004] However, the problem with the existing technology is that it does not take into account the image semantic information of the elderly's daily behavior when constructing and updating the knowledge graph, and it relies too much on semantic text information and sensor information. It cannot actually express the dynamic semantic information of the elderly in the retirement community and dynamically update the knowledge graph. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the present invention proposes a community elderly care analysis and management method based on multidimensional perception and artificial intelligence, including: collecting real-time community videos, obtaining various behavior categories of the target in the real-time video; marking the target pixels in the real-time video, obtaining the target image, and identifying it; obtaining images of various areas of the target, obtaining the complexity of the target's real-time expression, constructing a scoring model, and obtaining the target's daily knowledge graph according to the scoring model; setting scoring weights, obtaining an updated daily knowledge graph, obtaining the feature vector of the target's daily knowledge graph, and performing similarity measurement on the target's daily knowledge graph; judging the target's adaptation in the current community according to the result of the similarity measurement.

[0006] Through the technical means proposed in the present invention, the image convolution kernel size and feature extraction area are supervised by the image behavior category label, and the expression complexity is constructed through the fully connected network output. The multi-scale asymmetric convolution kernel image feature information is extracted under the supervision of the behavior category label. The network converges faster and the result is more accurate. The expression complexity of the elderly is used to score the expression, and the expression scoring model is optimized according to the data fluctuation characteristics of the elderly. The model evaluation result is more accurate and the robustness is better.

[0007] The present invention adopts the following technical solution, a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence, including:

[0008] Collecting real-time community videos; extracting target pixels from the real-time community videos and clustering the target pixels to obtain video frames of segments corresponding to a variety of behavioral features of the target.

[0009] The video segments are combined into a hypergraph; and a two-dimensional convolution is performed on the hypergraph to obtain a first eigenvector in the hypergraph.

[0010] The first feature vector is input into a first fully connected network to obtain a real-time behavior category of the target in the real-time video.

[0011] The target pixel points in the real-time video are marked to obtain the target image and perform recognition.

[0012] The target image is segmented to obtain a segmented image, and the segmented image is region-labeled to obtain images of various regions of the target.

[0013] Extract the second feature vectors in the images of each region of the target, input the second feature vectors into a second fully connected network, and obtain the real-time expression complexity of the target.

[0014] A scoring model is constructed according to the real-time expression complexity and the real-time behavior category of the target, and a daily knowledge graph of the target is obtained according to the scoring model.

[0015] Perform state fluctuation analysis on the historical data of the target, set scoring weights to update the scoring model, and obtain an updated daily knowledge graph of the target.

[0016] Obtain a third eigenvector of the daily knowledge graph, and perform similarity measurement on the target daily knowledge graph using the third eigenvector.

[0017] The adaptability of the target in the current community is judged according to the result of the similarity measurement, and then corresponding management is performed according to the adaptability.

[0018] Furthermore, a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence, the scoring model includes:

[0019]

[0020] in, Indicates the complexity of expression, represents the target behavior category score, Indicates the real-time status score of the target.

[0021] Furthermore, a community pension analysis and management method based on multi-dimensional perception and artificial intelligence performs state fluctuation analysis on the historical data of the target, sets scoring weights and updates the scoring model, including:

[0022] The expression of the scoring weight includes:

[0023]

[0024] in, Indicates the frequency of daily state fluctuations of the elderly, Indicates the fluctuation range of the elderly’s daily state.

[0025] Furthermore, a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence obtains the daily knowledge graph of the target according to the scoring model, including:

[0026] The real-time status score of the elderly is obtained through the scoring model.

[0027] Data is integrated according to the real-time status score and the behavior category to form a daily text.

[0028] A target daily knowledge graph is obtained according to the daily text.

[0029] Furthermore, a community pension analysis and management method based on multi-dimensional perception and artificial intelligence obtains the third feature vector of the daily knowledge graph, and performs similarity measurement on the target daily knowledge graph through the third feature vector, including:

[0030] The text information is input into the DSSM network, and the network encoder is used for convolution and update to output the feature vector of the current text information.

[0031] The similarity between the output feature vector and the original text feature vector is measured.

[0032] Furthermore, a community pension analysis and management method based on multi-dimensional perception and artificial intelligence extracts the second feature vectors in the images of each region of the target, inputs the second feature vectors into a second fully connected network, and obtains the real-time expression complexity of the target, including:

[0033] Convolution and pooling are performed on the images of each region of the target to extract their feature vectors.

[0034] The second feature vector is input into a second fully connected network and output through a normalization function. The size of the normalization function output value corresponding to the regional image represents the complexity of the target's expression.

[0035] Furthermore, a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence extracts target pixels and clusters the target pixels to obtain video frames of segments corresponding to multiple behavioral features of the target, including:

[0036] The target key pixel points in the real-time video are marked, and the density radius is set to cluster the key points until they are clustered into a certain area.

[0037] The real-time video is sliced ​​according to the clustering areas to obtain segmented video frames corresponding to a plurality of behavioral features of the target.

[0038] The beneficial effects of the present invention are as follows: through the technical means proposed by the present invention, the image convolution kernel size and feature extraction area are supervised by the image behavior category label, and the expression complexity is constructed through the fully connected network output, and the multi-scale asymmetric convolution kernel image feature information extraction under the supervision of the behavior category label is adopted, the network converges faster and the result is more accurate; the expression of the elderly is scored by the expression complexity, and the expression scoring model is optimized according to the data fluctuation characteristics of the elderly, and the model evaluation result is more accurate and the robustness is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0040] Figure 1 This is a flow chart of a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence in an embodiment of the present invention;

[0041] Figure 2 The present invention is a flowchart of a community elderly care analysis and management method based on multi-dimensional perception and artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Example 1

[0044] like Figure 1As shown, a community pension analysis and management method based on multi-dimensional perception and artificial intelligence in an embodiment of the present invention is provided, including:

[0045] 101. Collect real-time community videos; extract target pixels from the real-time community videos and cluster the target pixels to obtain video frames of segments corresponding to multiple behavioral features of the targets.

[0046] Real-time video of the community is collected through cameras in various areas of the community, wherein the collected video includes at least one target.

[0047] In this embodiment, the target takes the elderly in the community as an example. In this embodiment, the target can refer to the elderly.

[0048] A priori, facial expression images of all elderly people in the community are collected from all directions and stored in the retirement community face database for use in elderly face recognition and matching.

[0049] The target key pixel points in the real-time video are marked, and the density radius is set to cluster the key points until they are clustered into a certain area.

[0050] The real-time video is sliced ​​according to the clustering areas to obtain segmented video frames corresponding to a plurality of behavioral features of the target.

[0051] Clustering method implementation: mark the key points of the elderly head in the video frame, use the key points of the head as a reference, and set the density radius =20, which is the prior density radius. Randomly select head key points for clustering until they are clustered into one area. After clustering one area, randomly select other key points that are not involved in the clustering to continue clustering until there are no key points that are not involved in the clustering.

[0052] The video frames are sliced ​​according to the clustering areas, and then the sliced ​​videos are superimposed according to the changes in the slice positions of the continuous video frames, and finally the fragmented video frames corresponding to the target with multiple behavioral features are obtained.

[0053] 102. Merge the video frames into a hypergraph; perform two-dimensional convolution on the hypergraph to obtain a first eigenvector in the hypergraph.

[0054] The first feature vector is a feature vector in the video frames after being merged into a hypergraph.

[0055] The hypergraph is a triple G=<V,E,W> , which contains a vertex set V, a hyperedge set E, and a diagonal matrix W that records the weights of each hyperedge.

[0056] 103. Input the first feature vector into a first fully connected network to obtain a real-time behavior category of the target in the real-time video.

[0057] In this embodiment, the behavior category of the target represents the daily behavior of the elderly, including: playing cards, chatting, playing chess, dancing and the like.

[0058] The daily behaviors are not limited to the above contents and can be increased according to actual activities.

[0059] 104. Mark the target pixel points in the real-time video to obtain the target image and perform recognition.

[0060] The target image represents an image of an elderly person, and face recognition is performed on the target image obtained here through the prior knowledge of the community.

[0061] 105. Segment the target image to obtain a segmented image, and perform region marking on the segmented image to obtain images of various regions of the target.

[0062] The segmented image is a regional image of the target image, that is, a regional image of the body part of the elderly image.

[0063] 106. Extract second feature vectors from the images of each region of the target, input the second feature vectors into a second fully connected network, and obtain the real-time expression complexity of the target.

[0064] Select a partial area of ​​the target according to the target's behavior category, perform convolution with a convolution kernel of corresponding size, and obtain its feature vector.

[0065] The second feature vector is a feature vector of the target area image.

[0066] When the elderly person's behavior is chatting, we only need to focus on the feature type of the mouth area and select the behavior category label. The convolution kernel of size is convolved on the face area image.

[0067] 107. Construct a scoring model according to the real-time expression complexity and the real-time behavior category of the target, and obtain the daily knowledge graph of the target according to the scoring model.

[0068] The real-time status knowledge graph of the elderly is constructed by scoring their facial expressions under different behavior categories.

[0069] The expression scoring model is:

[0070] in, Indicates the complexity of expression, represents the target behavior category score, Indicates the real-time status score of the target.

[0071] The real-time status score of the elderly is obtained through the scoring model.

[0072] Data is integrated according to the real-time status score and the described behavior categories to form daily text.

[0073] Obtain the target daily knowledge graph based on daily text.

[0074] 108. Perform state fluctuation analysis on the historical data of the target, set scoring weights to update the scoring model, and obtain an updated daily knowledge graph of the target.

[0075] By conducting fluctuation analysis on the historical data of the elderly’s daily status, we mainly obtain the fluctuation frequency and amplitude of the elderly’s daily status.

[0076] Rating Weight

[0077] in, Indicates the frequency of daily state fluctuations of the elderly, Indicates the fluctuation range of the elderly’s daily state.

[0078] 109. Obtain a third eigenvector of the daily knowledge graph, and perform similarity measurement on the target daily knowledge graph using the third eigenvector.

[0079] The text information is input into the DSSM network, convolved and updated through the network encoder, and the feature vector of the current text information is output.

[0080] The third feature vector is the feature vector of the text information currently output.

[0081] The similarity between the feature vector of the current text information output and the feature vector of the original text is measured.

[0082] 110. Determine the adaptability of the target in the current community according to the result of the similarity measurement, and then perform corresponding management according to the adaptability.

[0083] When the Euclidean distance between eigenvectors It indicates that the elderly are poorly adapted to the current community.

[0084] When the Euclidean distance between eigenvectors It indicates that the elderly are adapting well to the current community.

[0085] Through the technical means proposed in the present invention, the image convolution kernel size and feature extraction area are supervised by the image behavior category label, and the expression complexity is constructed through the fully connected network output. The multi-scale asymmetric convolution kernel image feature information is extracted under the supervision of the behavior category label. The network converges faster and the result is more accurate. The expression complexity of the elderly is used to score the expression, and the expression scoring model is optimized according to the data fluctuation characteristics of the elderly. The model evaluation result is more accurate and the robustness is better.

[0086] Example 2

[0087] like Figure 2 As shown, a community pension analysis and management method based on multi-dimensional perception and artificial intelligence in an embodiment of the present invention is provided, including:

[0088] 201. Collecting real-time community videos; extracting target pixels from the real-time community videos and clustering the target pixels to obtain video frames of segments corresponding to multiple behavioral features of the targets.

[0089] Real-time video of the community is collected through cameras in various areas of the community, wherein the collected video includes at least one target.

[0090] In this embodiment, the target takes the elderly in the community as an example. In this embodiment, the target can refer to the elderly.

[0091] A priori, facial expression images of all elderly people in the community are collected from all directions and stored in the retirement community face database for use in elderly face recognition and matching.

[0092] 202. Merge the video frames into a hypergraph; perform two-dimensional convolution on the hypergraph to obtain a first eigenvector in the hypergraph.

[0093] Merge N consecutive video frames into a 3N-channel hypergraph, which retains the local spatiotemporal information of the original video and uniformly samples The hypergraph preserves the global spatiotemporal information in the original video.

[0094] Then, a two-dimensional convolution is performed on the hypergraph to obtain local spatiotemporal features. The shape of a tensor in the network is The local spatiotemporal feature maps are used to extract global spatiotemporal information and output through the pooling layer feature vectors.

[0095] 203. Input the first feature vector into a first fully connected network to obtain a real-time behavior category of the target in the real-time video.

[0096] The output The feature vectors are subjected to time-series 1D convolution and finally the feature categories are output through a fully connected network to improve the accuracy of behavior recognition.

[0097] In this embodiment, the target behavior category represents the daily behavior of the elderly, including: playing cards, chatting, playing chess, and dancing. The implementer can expand the behavior research category according to the actual research content.

[0098] By identifying the behaviors in the elderly video segment, the density radius is updated to ensure that the video slice includes all the interactive elderly video images. A mapping table is established between the size of the density radius update and the behavior category to specifically determine whether the density radius is increased or decreased.

[0099] 204. Mark the target pixel points in the real-time video to obtain the target image and perform recognition.

[0100] The pixels in the elderly area are marked as 1, and the pixels in other areas are marked as 0.

[0101] The labeled elderly image is input into the face recognition network encoder, the image is convolved and pooled, the global feature vector of the face area is extracted, and then the feature vector is classified through the normalized activation function of the fully connected layer to obtain the identity label of each elderly person.

[0102] 205. Segment the target image to obtain a segmented image, and perform region marking on the segmented image to obtain images of various regions of the target.

[0103] Through the semantic segmentation encoder and decoder, all the segmented images of the elderly in the slice image are finally obtained, and the segmented images of the elderly are annotated to obtain the facial area, leg area images and other parts of the elderly.

[0104] 206. Extract second feature vectors from the images of each region of the target, input the second feature vectors into a second fully connected network, and obtain the real-time expression complexity of the target.

[0105] Convolution and pooling are performed on the images of each region of the target to extract its feature vector. A partial region of the target is selected according to the target's behavior category, and convolution is performed according to the convolution kernel of the corresponding size to obtain its feature vector.

[0106] The initial convolution kernel size is ,Different convolution kernels are selected according to different behavior categories ,requirements. The purpose of using multi-scale asymmetric convolution is to ,obtain the pedestrian image features of the elderly more efficiently and ,accurately.

[0107] When the elderly person's behavior is chatting, you only need to pay attention to the feature type of the mouth area to get the elderly person's current mood state, and select the behavior category label. The convolution kernel of the same size is used to convolve the face area image, and the number of image convolutions is supervised by the category label , where the number of convolutions is It has a one-to-one mapping relationship with the behavior category label, which reduces the size of the receptive field and only requires perceiving the lateral correlation features of the mouth area.

[0108] When the elderly are playing cards or chess, we only need to focus on the feature categories of the face area and select The convolution kernel of the same size is used to convolve the face area image, and the number of image convolutions is supervised by the category label , it only needs to perceive the longitudinal correlation features of the human eye and face (excluding the mouth) area.

[0109] When the elderly dance, we need to focus on the features of the elderly’s face and legs, and use multi-scale asymmetric convolution ( ) to extract facial and leg features respectively.

[0110] The second feature vector is input into a second fully connected network and output through a normalization function. The size of the normalization function output value corresponding to the regional image represents the complexity of the target's expression.

[0111] Complexity of the old man's expression:

[0112] Among them, represents the maximum value of the normalization function output.

[0113] 207. Construct a scoring model according to the real-time expression complexity and the real-time behavior category of the target, and obtain a daily knowledge graph of the target according to the scoring model.

[0114] The scoring model comprises:

[0115]

[0116] in, Indicates the complexity of expression, represents the target behavior category score, Indicates the real-time status score of the target.

[0117] The daily knowledge graph of the target is obtained according to the scoring model, including:

[0118] The real-time status score of the elderly is obtained through the scoring model.

[0119] Data is integrated according to the real-time status score and the described behavior categories to form daily text.

[0120] Obtain the target daily knowledge graph based on daily text.

[0121] The status score of the elderly in their current mood is obtained through expert scoring. The real-time status of the elderly is obtained through the scoring model, and data is integrated according to the real-time status score and status category to form daily text for building the daily knowledge graph of the elderly.

[0122] 208. Perform state fluctuation analysis on the historical data of the target, set scoring weights to update the scoring model, and obtain an updated daily knowledge graph of the target.

[0123] By conducting fluctuation analysis on the historical data of the elderly’s daily status, we mainly obtain the fluctuation frequency and amplitude of the elderly’s daily status.

[0124] The expression of the scoring weight includes:

[0125]

[0126] in, Indicates the frequency of daily state fluctuations of the elderly, Indicates the fluctuation range of the elderly’s daily state.

[0127] Through the historical data of the elderly, the state fluctuation analysis is carried out. If the frequency of fluctuation of the elderly's daily state category is Large and volatile Large, indicating that the elderly are a sensitive group, and the score weight for the elderly should be increased , to ensure the importance of the elderly’s daily behavior.

[0128] Correspondingly, if the daily status category of the elderly fluctuates with a small frequency and a small amplitude, it means that the elderly belong to a stable population, and the corresponding score weight for the elderly should also be reduced.

[0129] The updated scoring model is:

[0130] in, Indicates the complexity of expression, represents the target behavior category score, represents the updated target real-time status score, Represents the scoring weight.

[0131] Through the updated scoring model, an updated daily knowledge graph of the target is obtained.

[0132] 209. Obtain a third eigenvector of the daily knowledge graph, and perform similarity measurement on the target daily knowledge graph using the third eigenvector.

[0133] The text information is input into the DSSM network, convolved and updated through the network encoder, and the feature vector of the current text information is output.

[0134] The similarity between the output feature vector and the original text feature vector is measured.

[0135] Similarity measurement method: Euclidean distance between feature vectors accomplish.

[0136] 210. Determine the adaptability of the target in the current community according to the result of the similarity measurement, and then perform corresponding management according to the adaptability.

[0137] When the Euclidean distance between eigenvectors It indicates that the elderly are poorly adapted to the current community.

[0138] When the Euclidean distance between eigenvectors It indicates that the elderly are adapting well to the current community.

[0139] The community should promptly follow up with the elderly people corresponding to the problem texts, improve physical and mental care for the elderly, resolve some of their demands, and promptly update the original health texts of the elderly.

[0140] In this embodiment, further, the personal data of each elderly person is collected through data collection sensors such as smart bracelets, including the elderly person's heart rate , Sleep quality of the elderly , Exercise volume of the elderly , statistics are conducted on various data information of the elderly in the time series state to obtain data parameters.

[0141] Furthermore, by matching the above-mentioned scoring model of the elderly with each data parameter, anomaly detection is performed. When the scoring model is low and the data parameters are high, the elderly may have psychological problems; when the scoring model is high and the data parameters are low, the elderly may have health problems; when the scoring model is high and the data parameters are high, the elderly is in a normal state.

[0142] According to the different status problems reflected by the elderly, corresponding management improvement measures are taken:

[0143] When the elderly have psychological problems, they should receive timely psychological counseling, increase daily care for the elderly, improve their living comfort, and promptly address some of their demands.

[0144] When an elderly person has health problems, their health should be monitored more closely in a timely manner, and their family members should be informed in a timely manner and corresponding health checks should be conducted.

[0145] Through the technical means proposed in the present invention, the image convolution kernel size and feature extraction area are supervised by the image behavior category label, and the expression complexity is constructed through the fully connected network output. The multi-scale asymmetric convolution kernel image feature information is extracted under the supervision of the behavior category label. The network converges faster and the result is more accurate. The expression complexity of the elderly is used to score the expression, and the expression scoring model is optimized according to the data fluctuation characteristics of the elderly. The model evaluation result is more accurate and the robustness is better.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A community pension analysis and management method based on multi-dimensional perception and artificial intelligence, characterized in that: include: Collect real-time community videos; Extract target pixels from the real-time community video and cluster the target pixels to obtain video frames of multiple behavioral features corresponding to the target; Merging the video clip frames into a hypergraph; performing a two-dimensional convolution on the hypergraph to obtain a first eigenvector in the hypergraph; Inputting the first feature vector into a first fully connected network to obtain a real-time behavior category of the target in the real-time video; wherein the behavior category of the target represents the daily behavior of the target; Marking the target pixel points in the real-time video to obtain a target image and perform recognition; Segmenting the target image to obtain a segmented image, and marking regions of the segmented image to obtain images of various regions of the target; Extracting second feature vectors from images of each region of the target, inputting the second feature vectors into a second fully connected network, and obtaining real-time expression complexity of the target; Building a scoring model according to the real-time expression complexity and the real-time behavior category of the target, and obtaining a daily knowledge graph of the target according to the scoring model; Performing state fluctuation analysis on the historical data of the target, setting scoring weights to update the scoring model, and obtaining an updated daily knowledge graph of the target; Characterized in that the scoring model includes: ； in, Indicates the complexity of expression, represents the target behavior category score, Indicates the real-time status score of the target; Obtaining a third feature vector of the daily knowledge graph, and performing similarity measurement on the target daily knowledge graph by using the third feature vector; Determining the adaptability of the target in the current community according to the result of the similarity measurement, and then performing corresponding management according to the adaptability; Performing state fluctuation analysis on the historical data of the target and setting scoring weights to update the scoring model include: The expression of the scoring weight includes: ； in, represents the daily state fluctuation frequency of the target, Indicates the daily state fluctuation range of the target; The daily knowledge graph of the target is obtained according to the scoring model, including: Obtaining a real-time status score of the target through the scoring model; Integrate data according to the real-time status score and the behavior category to form a daily text; Obtain a target daily knowledge graph according to the daily text; Obtaining a third feature vector of the daily knowledge graph, and performing similarity measurement on the target daily knowledge graph by using the third feature vector, including: The text information is input into the DSSM network, and the network encoder is used for convolution and update to output the feature vector of the current text information. The similarity between the output feature vector and the original text feature vector is measured; Extracting the second feature vectors in the images of each region of the target, inputting the second feature vectors into a second fully connected network, and obtaining the real-time expression complexity of the target, including: Performing convolution and pooling on the images of each region of the target to extract its feature vector; Inputting the second feature vector into a second fully connected network, outputting the feature vector after being normalized, and representing the complexity of the target's expression according to the output value of the normalized function corresponding to the regional image; Extracting target pixel points and clustering the target pixel points to obtain segment video frames corresponding to multiple behavior features of the target, including: Marking the target key pixel points in the real-time video, and setting the density radius to cluster the key points until they are clustered into a certain area; Slicing the real-time video according to the clustering areas to obtain segmented video frames corresponding to multiple behavioral features of the target; When the Euclidean distance between eigenvectors When , it means that the target is poorly adapted to the current community; When the Euclidean distance between eigenvectors , it indicates that the target is well adapted to the current community.

Images