Client risk early warning system based on behavior analysis

Abstract

The invention discloses a customer risk early warning system based on behavior analysis, which relates to the technical field of behavior analysis and comprises a motion capture module, a human body model module, a virtual monitoring module, a motion recognition module, a first motion judgment module, a second motion judgment module and a risk early warning module. The motion capture module is worn on a client, monitors the motion of the client and generates motion data; according to the customer risk early warning system based on behavior analysis, by arranging an action capture module, a human body model module, a virtual monitoring module, an action recognition module, a first action judgment module, a second action judgment module and a risk early warning module, the proportion of the judged user body and mind state as the negative state in overall judgment can be calculated; whether the negative state has adverse effects on the physical and psychological health of the customer is judged, so that early warning can be performed in time, the user is reminded to adjust the state or dredge, and the physical and psychological health safety of the customer is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of behavior analysis, and in particular to a customer risk early warning system based on behavior analysis. Background Art

[0002] Human behavior reflects, to a certain extent, their psychological and physical state. In order to ensure their physical and mental health and safety, promptly detect abnormal situations, and prevent possible risks, it is necessary to conduct regular physical and mental examinations.

[0003] The prior art with publication number CN117216650A discloses a method for early warning of risky behavior of a target object and related equipment. In the specific method, sentiment analysis is first performed based on a preset sentiment analysis model and multiple historical information released by the target object to determine whether the target object has abnormal emotions. Then, the preset risk behavior analysis model is used to determine the most likely type of risky behavior of the target object. Finally, information mining is performed on the historical information corresponding to this risky behavior type, so as to clearly present various key information related to this risky behavior type. Finally, risky behavior early warning of the target object is realized based on the key information related to the risky behavior. It can perform in-depth risky behavior information mining on objects with abnormal emotions, and use the mined risky behavior information for early warning, thereby improving the accuracy of risky behavior early warning of the target object.

[0004] The above-mentioned existing technologies mainly judge whether the customer's emotions are abnormal by analyzing the comments posted by the customer on the Internet. However, on the one hand, judging the customer's emotional state only from the comments posted by the customer on the Internet has certain limitations, which may cause certain deviations in the judgment results. For customers who seldom comment on the Internet, too little information will be collected, which further reduces the accuracy of the judgment results, and even makes it impossible to make a judgment due to too few samples. Summary of the invention

[0005] The purpose of the present invention is to provide a customer risk early warning system based on behavior analysis to solve the above-mentioned deficiencies in the prior art.

[0006] In order to achieve the above-mentioned object, the present invention provides the following technical solutions: a customer risk early warning system based on behavior analysis, comprising a motion capture module, a human body model module, a virtual monitoring module, a motion recognition module, a first motion judgment module, a second motion judgment module, and a risk early warning module;

[0007] The motion capture module is used to be worn on the customer, monitor the customer's movements, and generate motion data; wherein the motion capture module can be a portable motion capture device, which is convenient for the customer to wear and reduces the impact on the customer's normal life;

[0008] The human body model module is used to establish a human body skeleton model, obtain the motion data, and control the human body skeleton model to perform motions corresponding to the motion data; wherein the human body skeleton model is a three-dimensional model of a human body with skinned bones, so that the three-dimensional model of the human body can imitate the corresponding movable joints of the human body to perform motions by controlling the motion of the bones, thereby enabling the three-dimensional model of the human body to imitate the human body to perform various motions;

[0009] The virtual monitoring module is used to monitor the human skeleton model to obtain a model action video, wherein the model action video is composed of continuous image frames;

[0010] The action recognition module is used to set key action tags, and use the key action tags to mark the key action images in the historical model action video, wherein multiple key action tags can be set according to different key actions, and the marking can be achieved by associating the key action images with the corresponding key action tags;

[0011] The action recognition module is also used to train a first neural network model based on the model action video, the key action image and the key action label to obtain a key action recognition model for identifying the key action image from the input model action video and outputting the corresponding key action label, wherein the model action video is decomposed into continuous image frames during training, and the key action recognition model also first decomposes the input model action video into continuous image frames, identifies the key action image from the image frames, and outputs the key action label corresponding to the identified key action image; preferably, other action labels can also be set during training and associated with non-key action images, so that when the key action recognition model identifies an image frame that is not a key action image during the recognition process, it outputs other action labels; and the present invention does not limit the specific neural network model. When training the model, the neural network model can be selected according to actual needs and training conditions, so that the accuracy of the trained neural network model reaches the set standard, and the CNN model can be selected as a reference;

[0012] The first action judgment module is used to set a frequency threshold of each key action tag, obtain the number of times each key action tag appears within a set first time period, obtain a first frequency, and judge the customer status based on the relationship between the first frequency of the key action tag and the corresponding frequency threshold, wherein the customer status includes a positive status and a negative status;

[0013] The second action judgment module is used to set a plurality of action duration thresholds corresponding to key action labels one by one;

[0014] The second action judgment module is further used to extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video, obtain the corresponding key action steps, and obtain the duration of the key action steps, and judge the customer status based on the relationship between the duration of the key action steps corresponding to the same key action labels and the action duration threshold;

[0015] The risk warning module is used to obtain the judged customer status, calculate the proportion of each status, and judge whether it will affect the customer's physical and mental health based on the proportion.

[0016] Furthermore, the system also includes a third judgment module, which is used to mark image frames of non-critical action images in the model action video as other action images, and judge the customer status based on the analysis of feature correlation between the current other action image and the historical other action images in the positive state.

[0017] Furthermore, the third judgment module judges the customer status based on the analysis of the feature correlation between the current other action image and the historical other action images in the positive state, including the following steps:

[0018] Obtain historical other action images in which the adjacent key action images are all in positive states from the historical model action video, and obtain historical normal other action images; for example, if a group of historical other action images has a group of key action images in positive states in front of it and a group of key action images in bland states behind it, then the group of historical other action images is obtained; if a group of historical other action images has a group of key action images in positive states in front of it and a group of key action images in negative states behind it, then the group of historical other action images is not obtained;

[0019] According to the type of other actions corresponding to the normal history other action images, setting the corresponding other action distinguishing label to be associated with the normal history other action images of the same type as the corresponding other action;

[0020] Based on the historical other action images, the normal historical other action images and their corresponding other action distinguishing labels, and the other action abnormal labels, the second neural network model is trained to obtain the abnormal recognition model, which is used to identify and select the normal historical other action images and their corresponding other action distinguishing labels based on the input other action images, and associate the remaining other action images with the other action abnormal labels, and then output the other action abnormal labels and the other action images associated with them, wherein the abnormal recognition model, after identifying and selecting the normal historical other action images and their corresponding other action distinguishing labels, keeps them in the database and does not display them directly, and the staff can view the identified and selected normal historical other action images and their corresponding other action distinguishing labels by consulting the database;

[0021] If the anomaly recognition model does not output, the customer is judged to be in a positive state.

[0022] Furthermore, the third judgment module judges the customer status based on the analysis of the feature correlation between the current other action image and the historical other action images in the positive state, and further includes the following steps:

[0023] Acquire other action images associated with other action abnormal labels, obtain other action images to be judged, and calculate the similarity between the other action images to be judged and normal historical other action images corresponding to each other action distinguishing label;

[0024] Setting a similarity judgment threshold to determine whether the highest similarity calculated is greater than the similarity judgment threshold;

[0025] If yes, the type of the other action image to be determined is the same as the type of the other action corresponding to the other action distinction label corresponding to the other action image with the highest similarity, and the type of the other action corresponding to the other action image to be determined can be determined, and the customer is judged to be in a negative state;

[0026] If not, the other action images to be determined are marked as unrecognizable images.

[0027] Furthermore, the first action judgment module is used to set a frequency threshold of each key action tag, obtain the number of times each key action tag appears within a set first time period, obtain a first frequency, and judge the customer status based on the relationship between the first frequency of the key action tag and the corresponding frequency threshold, including the following steps:

[0028] Set the frequency threshold for each key action tag;

[0029] Obtain the key action labels and their output time output by the key action recognition model;

[0030] Mark key action labels on the timeline;

[0031] A sliding frame having a first duration is set on the time axis, and the sliding frame is made to slide on the time axis, wherein the right end of the sliding frame lags behind the second duration set at the current time during the sliding, and the number of the same key action labels in the current sliding frame is the current first frequency of the corresponding key action label, that is, the first frequency of each key action label is updated in real time with the sliding of the sliding frame, wherein the second duration can be customized according to the delay time output by the key action recognition model, so that the second duration is slightly greater than the delay time output by the key action recognition model;

[0032] Determine whether the first frequency of each key action label is greater than the corresponding frequency threshold;

[0033] If so, it is determined that the customer status of the time period in which the sliding box is located at this time is a negative state. If not, it is determined that the customer status of the time period in which the sliding box is located at this time is a positive state.

[0034] Furthermore, the second action judgment module is further used to extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video, obtain the corresponding key action steps, and obtain the duration of the key action steps, and judge the customer status based on the size relationship between the duration of the key action steps corresponding to the same key action labels and the action duration threshold, including the following steps:

[0035] Extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video to obtain corresponding key action steps;

[0036] Check whether there are any unrecognizable images adjacent to each key action step before and after it;

[0037] If yes, then update the unrecognizable image to the corresponding position in the key action step, wherein updating to the key action step means adding the unrecognizable image to the key action step and modifying its mark to the corresponding key action step, and updating to the corresponding position means that if the unrecognizable image starts in front of the key action image, then after updating to the key action step, it is still in front of the key action image;

[0038] Calculate the duration of the key action step, and determine whether the duration of the key action step is greater than the corresponding key action duration threshold;

[0039] If yes, the customer is judged to be in a negative state, if no, the customer is judged to be in a positive state.

[0040] Furthermore, the risk warning module is used to obtain the determined customer status, calculate the ratio of each status, and determine whether it will affect the customer's physical and mental health based on the ratio, including the following steps:

[0041] Obtain the number of negative states and positive states of the customer determined by the first action determination module, the second action determination module, and the third action determination module;

[0042] Setting a status judgment threshold to determine whether the proportion of negative status is greater than the status judgment threshold;

[0043] If so, an early warning will be issued, judging that the negative state will have an adverse effect on the customer's physical and mental health, and the state needs to be adjusted in time to ensure physical and mental health and safety. If not, it means that the negative state is within the normal range and will not affect the customer's physical and mental health.

[0044] 1. Compared with the prior art, the customer risk warning system based on behavior analysis provided by the present invention can capture and record customer movements in real time by setting a motion capture module, a human body model module, a virtual monitoring module, a motion recognition module, a first motion judgment module, a second motion judgment module, and a risk warning module, and judge the user's physical and mental state based on the frequency, type, and duration of the movements. The number of reference samples is sufficient, which improves the judgment accuracy.

[0045] 2. Compared with the prior art, the customer risk warning system based on behavior analysis provided by the present invention, by setting a motion capture module, a human body model module, a virtual monitoring module, a motion recognition module, a first motion judgment module, a second motion judgment module, and a risk warning module, can calculate the proportion of the user's physical and mental state that is judged to be a negative state in the overall judgment, and judge whether the negative state will have an adverse effect on the customer's physical and mental health, so as to issue a warning in time, remind the user to adjust his state or conduct counseling, and ensure the physical and mental health and safety of the customer. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0047] Figure 1 A system structure block diagram provided for an embodiment of the present invention;

[0048] Figure 2 A diagram of the first action determination module determination steps provided by an embodiment of the present invention;

[0049] Figure 3 A diagram of the determination steps of the second action determination module provided in an embodiment of the present invention;

[0050] Figure 4 A diagram of the judgment steps of the third action judgment module provided in an embodiment of the present invention;

[0051] Figure 5 The following is a diagram of the overall working steps of the system provided in the embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0053] The embodiments described herein may be described with reference to plan views and / or cross-sectional views by means of ideal schematic diagrams of the present disclosure. Therefore, the example illustrations may be modified according to manufacturing techniques and / or tolerances. Therefore, the embodiments are not limited to the embodiments shown in the drawings, but include modifications of configurations formed based on manufacturing processes. Therefore, the areas illustrated in the drawings have schematic properties, and the shapes of the areas shown in the drawings illustrate the specific shapes of the areas of the elements, but are not intended to be restrictive. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those of ordinary skill in the art.

[0054] See also Figure 1-Figure 5 , a customer risk warning system based on behavior analysis, including a motion capture module, a human body model module, a virtual monitoring module, a motion recognition module, a first motion judgment module, a second motion judgment module, and a risk warning module;

[0055] The motion capture module is used to be worn on the customer, monitor the customer's movements, and generate motion data; wherein the motion capture module can adopt a portable motion capture device, which is convenient for the customer to wear and reduces the impact on the customer's normal life;

[0056] The human body model module is used to establish a human body skeleton model, obtain motion data, and control the human body skeleton model to perform motions corresponding to the motion data; wherein the human body skeleton model is a three-dimensional model of the human body with skinned bones, so that the three-dimensional model of the human body can imitate the corresponding movable joints of the human body by controlling the movement of the bones, so that the three-dimensional model of the human body can imitate the human body to perform various movements;

[0057] The virtual monitoring module is used to monitor the human skeleton model to obtain the model action video, which is composed of continuous image frames;

[0058] The action recognition module is used to set key action labels, and use the key action labels to mark the key action images in the historical model action video, wherein multiple key action labels can be set according to different key actions, and the marking can be achieved by associating the key action images with the corresponding key action labels;

[0059] The action recognition module is also used to train the first neural network model based on the model action video, the key action image and the key action label to obtain the key action recognition model, which is used to recognize the key action image from the input model action video and output the corresponding key action label, wherein the model action video is decomposed into continuous image frames during training, and the key action recognition model also first decomposes the input model action video into continuous image frames, recognizes the key action image from the image frames, and outputs the key action label corresponding to the recognized key action image; preferably, other action labels can also be set during training and associated with non-key action images, so that when the key action recognition model recognizes an image frame that is not a key action image during the recognition process, it outputs other action labels; and the present invention does not limit the specific neural network model. When training the model, the neural network model can be selected according to actual needs and training conditions, so that the accuracy of the trained neural network model reaches the set standard, and the CNN model can be selected as a reference;

[0060] The first action judgment module is used to set a frequency threshold of each key action tag, obtain the number of times each key action tag appears within a set first time length, obtain a first frequency, and judge the customer status based on the relationship between the first frequency of the key action tag and the corresponding frequency threshold. The customer status includes a positive status and a negative status, including the following steps:

[0061] A1. Set the frequency threshold of each key action tag;

[0062] A2. Obtain the key action labels and their output time output by the key action recognition model;

[0063] A3. Mark the key action labels on the timeline;

[0064] A4. A sliding frame with a first duration is set on the time axis, and the sliding frame is made to slide on the time axis. When sliding, the right end of the sliding frame lags behind the second duration set at the current time. The number of the same key action labels in the current sliding frame is the current first frequency of the corresponding key action label, that is, the first frequency of each key action label is updated in real time with the sliding of the sliding frame. The second duration can be customized according to the delay time output by the key action recognition model, so that the second duration is slightly larger than the delay time output by the key action recognition model.

[0065] A5. Determine whether the first frequency of each key action label is greater than the corresponding frequency threshold;

[0066] A6. If yes, then determine that the customer status of the time period in the sliding box is a negative status. If no, then determine that the customer status of the time period in the sliding box is a positive status.

[0067] The second action judgment module is used to set a plurality of action duration thresholds corresponding to the key action labels one by one; the second action judgment module is also used to extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video, obtain the corresponding key action steps, and obtain the key action step duration, and judge the customer status based on the size relationship between the key action step duration corresponding to the same key action label and the action duration threshold, including the following steps:

[0068] B1. Extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video to obtain the corresponding key action steps;

[0069] B2. Check whether there are any unrecognizable images adjacent to each key action step before and after it;

[0070] B3. If yes, then update the unrecognizable image to the corresponding position in the key action step, wherein updating to the key action step means adding the unrecognizable image to the key action step and modifying its mark to the corresponding key action step, and updating to the corresponding position means that if the unrecognizable image starts in front of the key action image, then after updating to the key action step, it is still in front of the key action image;

[0071] B4. Calculate the duration of the key action step and determine whether the duration of the key action step is greater than the corresponding key action duration threshold;

[0072] B5. If yes, the customer is judged to be in a negative state; if no, the customer is judged to be in a positive state.

[0073] The system also includes a third judgment module, which is used to mark the image frames of non-key action images in the model action video as other action images, and judge the customer status based on the analysis of the feature correlation between the current other action image and the historical other action images in the positive state. The third judgment module judges the customer status based on the analysis of the feature correlation between the current other action image and the historical other action images in the positive state, including the following steps:

[0074] C1. Obtain historical other action images in which the adjacent key action images are all in positive states from the historical model action video, and obtain historical normal other action images; for example, if a group of historical other action images has a group of key action images in positive states in front of it and a group of key action images in bland states behind it, then this group of historical other action images is obtained; if a group of historical other action images has a group of key action images in positive states in front of it and a group of key action images in negative states behind it, then this group of historical other action images is not obtained;

[0075] C2. according to the type of other actions corresponding to the normal historical other action images, setting the corresponding other action distinguishing label to be associated with the normal historical other action images of the same type as the corresponding other action;

[0076] C3. Based on the historical other action images, the normal historical other action images and their corresponding other action distinguishing labels, and the other action abnormal labels, the second neural network model is trained to obtain an abnormal recognition model, which is used to identify and select the normal historical other action images and their corresponding other action distinguishing labels based on the input other action images, and associate the remaining other action images with the other action abnormal labels, and then output the other action abnormal labels and the other action images associated with them. After the abnormal recognition model identifies and selects the normal historical other action images and their corresponding other action distinguishing labels, it keeps them in the database and does not display them directly. The staff can check the identified and selected normal historical other action images and their corresponding other action distinguishing labels by consulting the database;

[0077] C4. If the anomaly recognition model does not output, the customer is judged to be in a positive state.

[0078] C5. Acquire other action images associated with other action abnormal labels, obtain other action images to be judged, and calculate the similarity between the other action images to be judged and the normal historical other action images corresponding to each other action distinguishing label;

[0079] C6. Setting a similarity judgment threshold to determine whether the highest similarity calculated is greater than the similarity judgment threshold;

[0080] C7, if yes, then the type of other action corresponding to the other action distinguishing label corresponding to the other action image to be judged is the same as that of the other action image to be judged with the highest similarity, and the type of other action corresponding to the other action image to be judged can be determined, and the customer is judged to be in a negative state;

[0081] C8. If not, mark the other action images to be judged as unrecognizable images.

[0082] The risk warning module is used to obtain the judged customer status, calculate the proportion of each status, and judge whether it will affect the customer's physical and mental health based on the proportion, including the following steps:

[0083] D1. Obtain the number of negative states and positive states of the customer determined by the first action determination module, the second action determination module, and the third action determination module;

[0084] D2. Set the status judgment threshold to determine whether the proportion of negative status is greater than the status judgment threshold;

[0085] D3. If yes, then an early warning will be issued, judging that the negative state will have an adverse effect on the customer's physical and mental health, and the state needs to be adjusted in time to ensure physical and mental health and safety. If no, it means that the negative state is within the normal range and will not affect the customer's physical and mental health.

[0086] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. Customer risk early warning system based on behavioral analysis, characterized by: It includes a motion capture module, a human body model module, a virtual monitoring module, a motion recognition module, a first motion judgment module, a second motion judgment module, and a risk warning module; The motion capture module is used to be worn on the customer, monitor the customer's movements, and generate motion data; The human body model module is used to establish a human body skeleton model, obtain the action data, and control the human body skeleton model to perform actions corresponding to the action data; The virtual monitoring module is used to monitor the human skeleton model to obtain a model action video, wherein the model action video is composed of continuous image frames; The action recognition module is used to set key action labels, and use the key action labels to mark key action images in the historical model action video; The action recognition module is also used to train a first neural network model based on the model action video, the key action image and the key action label to obtain a key action recognition model, which is used to recognize the key action image from the input model action video and output the corresponding key action label; The first action judgment module is used to set a frequency threshold of each key action tag, obtain the number of times each key action tag appears within a set first time period, obtain a first frequency, and judge the customer status based on the relationship between the first frequency of the key action tag and the corresponding frequency threshold, wherein the customer status includes a positive status and a negative status; The second action judgment module is used to set a plurality of action duration thresholds corresponding to key action labels one by one; The second action judgment module is further used to extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video, obtain the corresponding key action steps, and obtain the duration of the key action steps, and judge the customer status based on the relationship between the duration of the key action steps corresponding to the same key action labels and the action duration threshold; The risk warning module is used to obtain the judged customer status, calculate the proportion of each status, and judge whether it will affect the customer's physical and mental health based on the proportion.

2. The customer risk early warning system based on behavior analysis according to claim 1 is characterized by: The system also includes a third judgment module, which is used to mark image frames of non-critical action images in the model action video as other action images, and judge the customer status based on the analysis of the feature correlation between the current other action image and the historical other action images in the positive state.

3. The customer risk early warning system based on behavior analysis according to claim 2 is characterized by: The third judgment module judges the customer status based on the analysis of the feature correlation between the current other action image and the historical other action images in the positive state, including the following steps: Obtain historical other action images in which the preceding and following adjacent key action images are all in a positive state from the historical model action video, and obtain historical normal other action images; According to the type of other actions corresponding to the normal history other action images, setting the corresponding other action distinguishing label to be associated with the normal history other action images of the same type as the corresponding other action; Based on the historical other action images, the normal historical other action images and their corresponding other action distinguishing labels, and the other action abnormal labels, a second neural network model is trained to obtain an abnormality recognition model, which is used to identify and select the normal historical other action images and their corresponding other action distinguishing labels based on the input other action images, and associate the remaining other action images with the other action abnormal labels, and then output the other action abnormal labels and the other action images associated with them; If the anomaly recognition model does not output, the customer is judged to be in a positive state.

4. The customer risk early warning system based on behavior analysis according to claim 3 is characterized by: The third judgment module judges the customer status based on the analysis of the feature correlation between the current other action image and the historical other action images in the positive state, and further comprises the following steps: Acquire other action images associated with other action abnormal labels, obtain other action images to be judged, and calculate the similarity between the other action images to be judged and normal historical other action images corresponding to each other action distinguishing label; Setting a similarity judgment threshold to determine whether the highest similarity calculated is greater than the similarity judgment threshold; If so, the other action image to be determined is of the same type as the other action corresponding to the other action distinction label corresponding to the highest similarity, and the customer is determined to be in a negative state; If not, the other action images to be determined are marked as unrecognizable images.

5. The customer risk early warning system based on behavior analysis according to claim 1 is characterized by: The first action judgment module is used to set a frequency threshold of each key action tag, obtain the number of times each key action tag appears within a set first time length, obtain a first frequency, and judge the customer status based on the relationship between the first frequency of the key action tag and the corresponding frequency threshold, including the following steps: Set the frequency threshold for each key action tag; Obtain the key action labels and their output time output by the key action recognition model; Mark key action labels on the timeline; A sliding frame having a first duration is set on the time axis, and the sliding frame is made to slide on the time axis, and the right end of the sliding frame lags behind the second duration set by the current time during the sliding, and the number of the same key action labels in the sliding frame is the current first frequency of the corresponding key action label; Determine whether the first frequency of each key action label is greater than the corresponding frequency threshold; If so, it is determined that the customer status of the time period in which the sliding box is located at this time is a negative state. If not, it is determined that the customer status of the time period in which the sliding box is located at this time is a positive state.

6. The customer risk early warning system based on behavior analysis according to claim 4 is characterized by: The second action judgment module is further used to extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video, obtain the corresponding key action steps, and obtain the duration of the key action steps, and judge the customer status based on the size relationship between the duration of the key action steps corresponding to the same key action labels and the action duration threshold, including the following steps: Extract key action images with the same corresponding key action labels and continuous in the model action video from the model action video to obtain corresponding key action steps; Check whether there are any unrecognizable images adjacent to each key action step before and after it; If so, updating the unrecognizable image to the corresponding position in the key action step; Calculate the duration of the key action step, and determine whether the duration of the key action step is greater than the corresponding key action duration threshold; If yes, the customer is judged to be in a negative state, if no, the customer is judged to be in a positive state.

7. The customer risk early warning system based on behavior analysis according to claim 3 is characterized by: The risk warning module is used to obtain the determined customer status, calculate the ratio of each status, and determine whether it will affect the customer's physical and mental health based on the ratio, including the following steps: Obtain the number of negative states and positive states of the customer determined by the first action determination module, the second action determination module, and the third action determination module; Setting a status judgment threshold to determine whether the proportion of negative status is greater than the status judgment threshold; If so, issue a warning.

Images