A human posture behavior detection method based on deep learning

Abstract

The present invention relates to the field of electrical data processing technology, and in particular to a method for detecting human posture behavior based on deep learning. The present invention divides a motion process into a plurality of single posture processes and extracts the period of each type of joint motion data in each single posture process, analyzes the data difference between different periods of each type in each single posture process, and then determines the abnormal data points. The abnormal moment is determined according to the distribution of the abnormal data points, and the real abnormal moment is determined by the distribution of the abnormal degree of the corresponding period at the abnormal moment and the number of suspected abnormal data points. The label information of the corresponding human posture image is determined by the real abnormal moment, and then a trained posture recognition network is obtained. The present invention realizes the automatic labeling of the human posture image by accurately identifying the real abnormal moment in the motion process, thereby improving the training efficiency and accuracy of the posture recognition network.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical data processing, and in particular to a human body posture behavior detection method based on deep learning. Background Art

[0002] During exercise, people will make fixed postures and movements. During exercise, the correctness of the movement posture can be analyzed according to the posture and movement of the person. In order to realize automatic monitoring, the existing technology will use neural networks to analyze the human posture images of the person during the process, and then judge the state of the person's movement posture. However, the image labeling in the neural network training process is usually done manually, which makes the neural network training workload large, and also causes errors in the labeled label information, thereby affecting the accuracy of the neural network in detecting human posture behavior. Summary of the invention

[0003] In order to solve the technical problem in the prior art that the image label information is not accurately obtained, thereby affecting the accuracy of the neural network in detecting human posture behavior, the purpose of the present invention is to provide a human posture behavior detection method based on deep learning, and the technical solution adopted is as follows:

[0004] The present invention proposes a human body posture behavior detection method based on deep learning, the method comprising:

[0005] Acquire a human posture image during the motion process, and collect a human velocity sequence and different types of joint motion data sequences during the human posture image acquisition process;

[0006] Obtain a single posture process in the motion process according to the local data change characteristics in the human body speed sequence; obtain the cycle of each type of joint motion data sequence in each single posture process;

[0007] In each of the single posture processes, according to the difference between the data points at the same position between different cycles in the joint motion data sequence of each type, the suspected abnormality degree of each data point is obtained, and the suspected abnormal data points are screened out according to the suspected abnormality degree; according to the steps of the suspected abnormal data points at the same time between different types, the suspected abnormal moments are screened out;

[0008] In each of the single posture processes, the real abnormality of each suspected abnormal moment is obtained according to the distribution information of the suspected abnormality degree in each corresponding cycle and the number of the suspected abnormal data points, and the real abnormal moment is screened out according to the real abnormality;

[0009] The human body posture image corresponding to the real abnormal moment is assigned label information, training data of a posture recognition network is obtained according to the human body posture image and the corresponding label information, and the posture behavior in the real-time human body posture image is recognized according to the trained posture recognition network.

[0010] Furthermore, the different types of joint motion data sequences include at least a joint flexion sequence and a joint velocity sequence.

[0011] Furthermore, the process of obtaining a single posture in the motion process according to the local data change characteristics in the human body speed sequence includes:

[0012] The data value mean and data value variance within a preset neighborhood range of each element in the human body velocity sequence are respectively obtained to form a mean velocity sequence and a velocity variance sequence in turn; a forward difference sequence of the mean velocity sequence is obtained, and after removing the last element of the velocity variance sequence, each element is multiplied by the element at the corresponding position in the forward difference sequence and the absolute value is taken to obtain the local data change feature corresponding to each element, and a local data change feature sequence is formed; the local data change feature sequence is segmented using the maximum value as a segmentation point to obtain multiple subsequences, and the time process corresponding to each subsequence is a single posture process.

[0013] Further, the acquisition of each type of the joint motion data sequence in each single posture process includes:

[0014] The frequency corresponding to the maximum amplitude of each type of the joint motion data sequence in each single posture process is obtained by Fourier transform, and the reciprocal of the frequency is used as the period length to obtain the period corresponding to the joint motion data sequence.

[0015] Furthermore, the method for obtaining the suspected abnormality degree includes:

[0016] The suspected abnormality degree is obtained according to a suspected abnormality degree formula, and the suspected abnormality degree formula includes: Among them, Q j,k is the suspected abnormality degree of the kth data point in the jth cycle, n is the number of data points in one cycle of the joint motion data sequence of the corresponding type, q j,k is the data value of the kth data point in the jth period, is the evaluation data value of the kth data point of all cycles in the joint motion data sequence of the corresponding type, and norm() is the normalization function.

[0017] Furthermore, the method for obtaining the suspected abnormal moment includes:

[0018] If, during the single posture process, at a certain moment, corresponding data points of each type of the joint motion data sequence are all suspected abnormal data points, the corresponding moment is taken as the suspected abnormal moment.

[0019] Furthermore, the method for obtaining the real abnormality includes:

[0020] The real abnormality is obtained according to a real abnormality formula, wherein the real abnormality formula comprises:

[0021] Among them, W i is the true abnormality at the i-th suspected abnormal moment; n′ is the number of types of the joint motion data sequence; Δn i,L is the number of the suspected abnormal data points in the cycle where the i-th suspected abnormal moment is located in the joint motion data sequence of the L-th type; Q i,L Q′ is the suspected abnormality degree of the data point corresponding to the i-th suspected abnormal moment in the joint motion data sequence of the L-th type; i,L (max) is the maximum value of the suspected abnormality degree in the cycle where the i-th suspected abnormality moment is located in the joint motion data sequence of the L-th type; is the average value of the suspected abnormality degree in the cycle where the ith suspected abnormality moment is located in the joint motion data sequence of the Lth type, and norm() is a normalization function.

[0022] Further, the filtering out the real abnormal moment according to the real abnormality includes:

[0023] The suspected abnormal moment when the actual abnormality is greater than a preset actual abnormality threshold is taken as the actual abnormal moment.

[0024] Furthermore, the gesture recognition network includes:

[0025] Extract key point vectors from the human body posture image, and train the posture recognition network according to the key point vectors and the corresponding label information, wherein the input of the posture recognition network is the key point vector and the output is the label information.

[0026] Furthermore, the posture recognition network is a fully connected neural network.

[0027] The present invention has the following beneficial effects:

[0028] The present invention first utilizes the local data change characteristics in the human body velocity sequence to split the entire motion process into multiple single posture processes, which is convenient for subsequent data processing. Further considering the periodicity of the motion, the suspected abnormal data points are screened out based on the differences of the same data points in the period. Considering that the suspected abnormal data points may be noise points in the joint motion data sequence under a certain type, in order to avoid the influence of noise, the suspected abnormal moment is determined based on the distribution of suspected abnormal data points at the same time between different types. Further, by obtaining the real abnormality of the suspected abnormal moment, the accurate real abnormal moment is determined, and the human body posture image corresponding to the real abnormal moment is assigned label information. The posture recognition network is trained through accurate label information, thereby improving the accuracy of the posture recognition network in detecting human posture behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are 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 work.

[0030] Figure 1 A flow chart of a method for detecting human posture and behavior based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a flowchart of a human posture behavior detection method based on deep learning proposed by the present invention, in combination with the accompanying drawings and preferred embodiments, and its specific implementation method, structure, features and effects are described in detail as follows. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0032] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0033] The specific scheme of the flow chart of the method for detecting human posture and behavior based on deep learning provided by the present invention is described in detail below with reference to the accompanying drawings.

[0034] See also Figure 1 , which shows a flow chart of a method for detecting human posture behavior based on deep learning provided by an embodiment of the present invention, the method comprising:

[0035] Step S1: acquiring a human body posture image during the motion process, and acquiring a human body velocity sequence and different types of joint motion data sequences during the human body posture image acquisition process.

[0036] Because the embodiment of the present invention aims to pre-process the collected training data in the early stage of neural network training, while collecting the human body posture images during the movement process required for training, it is also necessary to collect the human body velocity sequence and different types of joint motion data sequences, and use this information to analyze the state of the corresponding human body posture image.

[0037] Preferably, in one embodiment of the present invention, different types of joint motion data sequences include at least joint flexion sequence and joint velocity sequence. In other embodiments, more types of joint motion information may be selected, such as joint motion direction.

[0038] In the embodiment of the present invention, an IMU device is used to collect joint flexion length information. The IMU device is a small sensor device that can be fixed on different joints on the body to measure the movement and angle of the joints, and then obtain the degree of joint flexion; an electric biomechanical instrument is used to obtain data such as the speed of the human body and the speed of the joints. The acquisition frequency is set to once per second.

[0039] It should be noted that, because there are multiple joints in the human body, and the motion process is often performed on one or some limbs, it is not necessary to consider all joints on the human body in the embodiments of the present invention. Sensors can be deployed to collect information based on the specific information of the motion process, thereby obtaining a joint motion data sequence of each joint. In one embodiment of the present invention, considering the existence of information on multiple joints, each joint is regarded as a type of data.

[0040] Step S2: obtaining a single posture process in the motion process according to the local data change characteristics in the human body velocity sequence; obtaining the period of each type of joint motion data sequence in each single posture process.

[0041] During the entire movement process, people will make many kinds of postures and movements. The corresponding human body speed performance is different under different posture behaviors. For example, walking slowly, standing, squatting, running and other postures have different human body speeds. Because the human body speed sequence is the data sequence of the entire movement process, the single posture process in the movement process can be obtained according to the local protective gear change characteristics in the human body speed sequence. In the subsequent data processing process, each single posture process is analyzed to improve the accuracy of periodic feature analysis, thereby improving the accuracy of abnormal data recognition.

[0042] Preferably, obtaining a single posture process in the motion process according to the local data change characteristics in the human body speed sequence includes:

[0043] In order to analyze the local information in the human body speed sequence, for each element, the data features of each element are analyzed with the neighborhood as a unit, and the mean value and variance of the data value within the preset neighborhood range of each element in the human body speed sequence are obtained respectively. Because each element corresponds to a collection moment, the mean speed sequence and the speed variance sequence can be sequentially formed. The forward difference sequence of the mean speed sequence is obtained. Because the human body movement speed between different postures in a movement process is different, but the speed information in the same posture process is the same or similar, the data features of the posture change moment can be highlighted through the difference sequence. Because there is no element after the last element in the acquisition process of the forward difference sequence, the last element cannot participate in the forward difference operation, so after removing the last element of the speed variance sequence, each element is multiplied by the element at the corresponding position in the forward difference sequence and the absolute value is taken to obtain the local data change feature corresponding to each element, and form a local data change feature sequence. The larger the local data change feature, the more drastic the human body speed change at the corresponding moment, and the greater the probability that the moment is the posture conversion time point. Therefore, the maximum value in the local data change feature sequence is used as the segmentation point for segmentation to obtain multiple subsequences. The time process corresponding to each subsequence is a single posture process.

[0044] It should be noted that the methods for obtaining technical features such as forward difference and maximum value acquisition are well known to those skilled in the art and will not be described in detail here. In one embodiment of the present invention, the neighborhood of each element is set to a range with a length of 21 centered on the corresponding element, that is, there are 10 other elements on the left and right sides of the corresponding element in a neighborhood.

[0045] For a single posture process, when the human body is performing the corresponding action, the joint motion data changes periodically. For example, when walking or running slowly, the swing amplitude and angle of each swing are basically the same. Therefore, in the subsequent process, the data between cycles in a single posture process can be analyzed to determine abnormal information.

[0046] Preferably, in one embodiment of the present invention, obtaining each type of joint motion data sequence in each single posture process includes:

[0047] The frequency corresponding to the maximum amplitude of each type of joint motion data sequence in each single posture process is obtained by Fourier transform, and the reciprocal of the frequency is used as the period length to obtain the period of the corresponding joint motion data sequence. It should be noted that Fourier transform is a technical means well known to those skilled in the art and will not be described in detail here.

[0048] It should be noted that because it is necessary to obtain the period of each type of joint motion data sequence in each single posture process, different types have different data performance characteristics, so the obtained period is not uniform, and it is necessary to obtain the corresponding period according to the data performance of the specific joint motion data.

[0049] Step S3: In each single posture process, according to the difference between the data points at the same position between different periods in each type of joint motion data sequence, the suspected abnormality degree of each data point is obtained, and the suspected abnormal data points are screened out according to the suspected abnormality degree; according to the distribution of suspected abnormal data points at the same time between different types, the suspected abnormal moments are screened out.

[0050] In each single posture process, the human body's behavior posture may be deformed and irregular at a certain moment due to unfamiliar movements, incorrect force points, etc., and the corresponding data at this time is an abnormal data point in the overall data performance. Because the data of the joints in a single posture process has periodic change characteristics, in order to filter out abnormal information, the difference between the data points at the same position between different cycles can be used to obtain the suspected abnormality of each data point.

[0051] Preferably, in one embodiment of the present invention, a method for obtaining the degree of suspected abnormality includes:

[0052] The suspected abnormality degree is obtained according to the suspected abnormality degree formula, and the suspected abnormality degree formula includes:

[0053] Among them, Q j,k is the suspected abnormality degree of the kth data point in the jth cycle, n is the number of data points in one cycle in the corresponding type of joint motion data sequence, q j,k is the data value of the kth data point in the jth period, is the evaluation data value of the kth data point of all cycles in the corresponding type of joint motion data sequence, and norm() is the normalization function. It should be noted that the normalization function can be implemented by a variety of basic mathematical operations such as maximum and minimum value normalization and function mapping, which are all technical means well known to those skilled in the art and are not limited or elaborated.

[0054] In the suspected abnormality degree formula, the evaluation data value of a certain data point in all cycles of a type of joint motion data sequence in a single posture process is obtained to represent the overall data value characteristics, and then the difference between each data point and the overall data value characteristics in each cycle is obtained. Because abnormal data should be low-probability data relative to normal data, the larger the difference, the more likely the corresponding data point is abnormal data caused by irregular movements caused by tissue injuries; considering that irregular or abnormal movements in a movement process should last for a period of time, rather than suddenly appearing at a single time point, the abnormality will be As the weight of the above difference, that is, It characterizes the data abnormality of the duration in the corresponding period. The larger the data is, the more the abnormal information generated at the corresponding moment is the abnormality caused by real motion, rather than the noise point information generated during the data acquisition process.

[0055] Suspected abnormal data points can be screened out according to the degree of suspected abnormality. Taking into account that the collected data may be affected by noise and data processing errors, the abnormal data obtained is only abnormal data. Considering that if an abnormal action is performed at a certain moment, the abnormal information should be reflected in different types of joint motion data. Therefore, in order to further determine the abnormal information, the suspected abnormal moments are screened out according to the distribution of suspected abnormal data points of different types at the same moment. That is, the more suspected abnormal data points are distributed between different types at a moment, the more abnormal the moment is.

[0056] Preferably, in one embodiment of the present invention, a method for obtaining a suspected abnormal moment includes:

[0057] If, during a single posture process, the corresponding data points of each type of joint motion data sequence at a certain moment are all suspected abnormal data points, the corresponding moment will be regarded as the suspected abnormal moment.

[0058] In one embodiment of the present invention, an abnormality threshold is set. If the suspected abnormality of a data point is greater than the abnormality threshold, the corresponding data point is marked as a suspected abnormal data point. Because the suspected abnormality in one embodiment of the present invention has a value range between 0 and 1, the abnormality threshold is set to 0.8.

[0059] Step S4: In each single posture process, the real abnormality of each suspected abnormal moment is obtained according to the distribution information of the suspected abnormality degree of each suspected abnormal moment in each corresponding cycle and the number of suspected abnormal data points, and the real abnormal moment is screened out according to the real abnormality.

[0060] Further analyze the suspected abnormal moment to determine whether the suspected abnormal moment is the moment when the posture abnormality actually occurs. As described in step S3, if a posture abnormality actually occurs at a certain moment, the corresponding joint motion data of each type will be abnormal, so when analyzing the true abnormality of the suspected abnormal moment, it is necessary to obtain it according to the distribution of the suspected abnormal degree and the number of suspected abnormal data points in each cycle corresponding to the moment, which means that the more the number of abnormal data points and the more prominent the distribution of the suspected abnormal degree, the greater the abnormal reaction of the human body in the current cycle and it is not caused by noise points.

[0061] Preferably, in one embodiment of the present invention, the method for obtaining the real abnormality includes:

[0062] The real abnormality is obtained according to the real abnormality formula, which includes:

[0063] Among them, W i is the true abnormality of the i-th suspected abnormal moment; n′ is the number of types of joint motion data sequence; Δn i,L is the number of suspected abnormal data points in the cycle where the i-th suspected abnormal moment is located in the L-th type of joint motion data sequence; Q i,L Q′ is the suspected abnormality degree of the data point corresponding to the i-th suspected abnormal moment in the L-th type of joint motion data sequence; i,L (max) is the maximum value of the suspected abnormality degree in the cycle where the i-th suspected abnormal moment is located in the L-th type of joint motion data sequence; is the average value of the suspected abnormality degree in the cycle where the i-th suspected abnormal moment is located in the L-th type of joint motion data sequence, and norm() is the normalization function.

[0064] In the true abnormality formula, using Indicates the distribution of suspected abnormality in the corresponding period. The larger the data is, the more prominent and intense reaction point there is in the corresponding period, which may be a data point where the human body reacts violently to stress. In order to avoid the influence of noise, the number of suspected abnormal data points needs to be combined as a weight, that is, the more suspected abnormal data points there are, the more abnormal actions are generated in the current period, not abnormal data caused by noise. Similarly, the abnormality of a suspected abnormal moment in a data type is obtained by combining the suspected abnormality degree of the corresponding suspected abnormal moment itself, and the real abnormality of the corresponding suspected abnormal moment is obtained by combining all types of data. If a suspected abnormal moment is the moment when an abnormal posture is actually generated, the abnormality reflected in different data types is a larger value, and the real abnormality obtained after accumulation is also a larger value.

[0065] The real abnormal moment can be screened out according to the real abnormality, which specifically includes: taking the suspected abnormal moment whose real abnormality is greater than a preset real abnormality threshold as the real abnormal moment. In one embodiment of the present invention, the real abnormality threshold is set to 0.8.

[0066] Step S5: assigning label information to the human posture image corresponding to the real abnormal moment, obtaining training data of the posture recognition network according to the human posture image and the corresponding label information, and identifying the posture behavior in the real-time human posture image according to the trained posture recognition network.

[0067] By performing the same data processing and analysis on each single posture process during a movement process, all real abnormal moments in the movement process can be obtained. By assigning label information to the human body posture image corresponding to the real abnormal moment, automatic labeling of the label information can be achieved, avoiding the errors and labor caused by manual labeling.

[0068] In one embodiment of the present invention, the human body posture image corresponding to the real abnormal moment is assigned an abnormal label, and the human body posture images at other moments are assigned a normal label. The label value of the abnormal label is set to 0, and the label value of the normal label is set to 1.

[0069] After obtaining the label information of the human posture image during the movement, the training data of the posture recognition network can be obtained according to the human posture image and the corresponding label information, and the posture behavior in the real-time human posture image can be identified according to the trained posture recognition network.

[0070] Preferably, in the training process of the posture recognition network, the key point vectors in the human posture image are extracted, and the posture recognition network is trained according to the key point vectors and the corresponding label information. The input of the posture recognition network is the key point vector, and the output is the label information. It should be noted that the method of extracting the key point information in the image and forming the key point vector is a technical means well known to those skilled in the art, and will not be described in detail or limited here. In order to expand the training set, image data of multiple motion processes can be obtained and the labels of the images can be obtained. The expansion and augmentation of the training set is a technical means well known to those skilled in the art, and will not be described in detail here.

[0071] In one embodiment of the present invention, the posture recognition network is a fully connected neural network, which has good classification function and can accurately classify according to the characteristics of the key point vectors learned during the training process. The specific structure and training method of the fully connected neural network are technical means well known to those skilled in the art and will not be described in detail here.

[0072] By using the posture recognition network to recognize real-time human posture images, it is possible to determine whether a person in real time has abnormal posture, and then issue an early warning or reminder to assist the corresponding personnel to adjust and learn their posture in a timely manner.

[0073] In summary, the embodiment of the present invention divides the motion process into multiple single posture processes and extracts the period of each type of joint motion data in each single posture process, analyzes the data differences between different periods of each type in each single posture process, and then determines the abnormal data points. The abnormal moment is determined according to the distribution of the abnormal data points, and the real abnormal moment is determined by the distribution of the degree of abnormality of the corresponding period at the abnormal moment and the number of suspected abnormal data points. The label information of the corresponding human posture image is determined by the real abnormal moment, and then a trained posture recognition network is obtained. The present invention realizes automatic labeling of human posture images by accurately identifying the real abnormal moment in the motion process, thereby improving the training efficiency and accuracy of the posture recognition network.

[0074] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0075] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A human body posture behavior detection method based on deep learning, characterized in that: The method comprises: Acquire a human posture image during the motion process, and collect a human velocity sequence and different types of joint motion data sequences during the human posture image acquisition process; Obtain a single posture process in the motion process according to the local data change characteristics in the human body speed sequence; obtain the cycle of each type of joint motion data sequence in each single posture process; In each of the single posture processes, according to the difference between the data points at the same position between different cycles in the joint motion data sequence of each type, the suspected abnormality degree of each data point is obtained, and the suspected abnormal data points are screened out according to the suspected abnormality degree; according to the distribution of the suspected abnormal data points at the same time between different types, the suspected abnormal moments are screened out; In each of the single posture processes, the real abnormality of each suspected abnormal moment is obtained according to the distribution information of the suspected abnormality degree in each corresponding cycle and the number of the suspected abnormal data points, and the real abnormal moment is screened out according to the real abnormality; Assigning label information to the human posture image corresponding to the real abnormal moment, obtaining training data of a posture recognition network according to the human posture image and the corresponding label information, and identifying the posture behavior in the real-time human posture image according to the trained posture recognition network; The process of obtaining a single posture in the motion process according to the local data change characteristics in the human body speed sequence includes: The data value mean and data value variance within a preset neighborhood range of each element in the human body velocity sequence are respectively obtained to form a mean velocity sequence and a velocity variance sequence in turn; a forward difference sequence of the mean velocity sequence is obtained, and after removing the last element of the velocity variance sequence, each element is multiplied by the element at the corresponding position in the forward difference sequence and the absolute value is taken to obtain the local data change feature corresponding to each element, and a local data change feature sequence is formed; the local data change feature sequence is segmented using the maximum value as a segmentation point to obtain multiple subsequences, and the time process corresponding to each subsequence is a single posture process.

2. A method for detecting human posture behavior based on deep learning according to claim 1, characterized in that: The different types of joint motion data sequences include at least a joint flexion sequence and a joint velocity sequence.

3. The method for detecting human posture behavior based on deep learning according to claim 1, characterized in that: The cycle of obtaining each type of joint motion data sequence in each single posture process includes: The frequency corresponding to the maximum amplitude of each type of the joint motion data sequence in each single posture process is obtained by Fourier transform, and the reciprocal of the frequency is used as the period length to obtain the period corresponding to the joint motion data sequence.

4. The method for detecting human posture behavior based on deep learning according to claim 1, characterized in that: The method for obtaining the suspected abnormality degree includes: The suspected abnormality degree is obtained according to a suspected abnormality degree formula, and the suspected abnormality degree formula includes: ;in, is the suspected abnormality degree of the kth data point in the jth cycle, n is the number of data points in one cycle in the joint motion data sequence of the corresponding type, is the data value of the kth data point in the jth period, is the average data value of the kth data point of all cycles in the joint motion data sequence of the corresponding type, is the normalization function.

5. The method for detecting human posture behavior based on deep learning according to claim 1, characterized in that: The method for obtaining the suspected abnormal moment includes: If, during the single posture process, at a certain moment, corresponding data points of each type of the joint motion data sequence are all suspected abnormal data points, the corresponding moment is taken as the suspected abnormal moment.

6. A method for detecting human posture behavior based on deep learning according to claim 1, characterized in that: The method for obtaining the real abnormality includes: The real abnormality is obtained according to a real abnormality formula, wherein the real abnormality formula comprises: ;in, is the actual abnormality at the i-th suspected abnormal moment; is the number of types of the joint motion data sequence; is the number of the suspected abnormal data points in the cycle where the i-th suspected abnormal moment is located in the joint motion data sequence of the L-th type; is the suspected abnormality degree of the data point corresponding to the i-th suspected abnormal moment in the joint motion data sequence of the L-th type; is the maximum value of the suspected abnormality degree in the cycle where the i-th suspected abnormality moment is located in the joint motion data sequence of the L-th type; is the average value of the suspected abnormality degree in the cycle where the i-th suspected abnormality moment is located in the joint motion data sequence of the L-th type, is the normalization function.

7. The method for detecting human posture behavior based on deep learning according to claim 1, characterized in that: The step of screening out the real abnormal moment according to the real abnormality comprises: The suspected abnormal moment when the actual abnormality is greater than a preset actual abnormality threshold is taken as the actual abnormal moment.

8. The method for detecting human posture behavior based on deep learning according to claim 1, characterized in that: The gesture recognition network comprises: Extract key point vectors from the human body posture image, and train the posture recognition network according to the key point vectors and the corresponding label information, wherein the input of the posture recognition network is the key point vector and the output is the label information.

9. A method for detecting human posture behavior based on deep learning according to claim 1 or 8, characterized in that: The posture recognition network is a fully connected neural network.

Images