A human pose detection method and system based on detectability learning

Abstract

The application discloses a human posture detection method and system based on detectability learning, and the human posture detection method based on detectability learning comprises the following steps: acquiring image data to be detected for human posture; based on the image data to be detected for human posture, a pre-trained human posture detection model is used for detection to obtain position information of each human posture key point and detectability information; human posture key points with detectability information greater than or equal to a preset threshold are displayed to obtain a human posture detection result. Through the pre-trained model, the position of the human posture key point and the detectability are detected, the human posture detection demand of data such as occlusion, scale uncertainty, easily confused posture and complex background can be solved, and the method has the characteristics of high detection precision and high reliability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer digital image processing and pattern recognition, relates to the field of human posture video intelligent analysis, and particularly relates to a human posture detection method and system based on detectability learning. BACKGROUND

[0002] Human posture detection is a basic task for downstream tasks such as action recognition, pedestrian re-identification, human-computer interaction, and accurate human posture detection can bring great social value and commercial value; for example, the vision system of some service robots, posture teaching videos, animation and game production, medical health examination and other application scenarios provide convenience and services for human beings.

[0003] Different photo or video shooting methods, shooting angles and other factors will increase the uncertainty of the data, and some data that is difficult to carry out human posture detection tasks will inevitably appear, such as occlusion, confusion, complex background and the like; although some existing methods have deepened the network or adjusted the network structure, the effect on challenging posture detection data is limited, and has the defects of low detection precision and poor reliability; in summary, in view of the demand for human posture detection of data with occlusion, scale uncertainty, confusing posture and complex background, a new high-resolution human posture detection method and system are urgently needed. SUMMARY

[0004] The present application aims to provide a human posture detection method and system based on detectability learning to solve one or more of the above technical problems. The technical solution provided by the present application detects the key point position and detectability of human posture through a pre-trained model, which can solve the human posture detection demand of data with occlusion, scale uncertainty, confusing posture and complex background, and has the characteristics of high detection precision and high reliability.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solution:

[0006] The human posture detection method based on detectability learning provided by the present application comprises the following steps:

[0007] Obtain image data to be detected for human posture;

[0008] Based on the image data to be detected for human posture, a pre-trained human posture detection model is used for detection to obtain the position information and detectability information of each human posture key point;

[0009] Display the human posture key points with detectability information greater than or equal to a pre-set threshold value to obtain the human posture detection result;

[0010] The human body posture detection model comprises:

[0011] The human body posture detection deep neural network is configured to input image data to be detected for human body posture, and output a heat map.

[0012] The detectability learning supervision network is configured to input the heat map, and output detectability information of each human body posture key point.

[0013] The human body posture detection deep neural network is HRNet, ResNet or Stacked-Hourglass.

[0014] The detectability learning supervision network comprises:

[0015] The input layer is configured to input the heat map to be processed.

[0016] The down-sampling layer is configured to input the heat map and perform step-by-step down-sampling, and output a feature map with a preset ratio.

[0017] The average pooling layer is configured to input the feature map with the preset ratio, adjust the image resolution by using adaptive average pooling, and output a feature map with a resolution of 1*1.

[0018] The full connection layer is configured to adjust the number of channels to be the same as the number of human body posture key points based on the 1*1 feature map, and the feature values of the channels are respectively used to represent the detectability of the human body posture key points.

[0019] The human body posture detection model comprises:

[0020] The training sample set is obtained, wherein each training sample comprises a sample image, a human body posture key point position label and a human body posture key point detectability information label.

[0021] Based on the training sample set, a pre-constructed loss function is used to update the parameters by continuously reducing the error between the real attributes and the predicted attributes, and the pre-trained human body posture detection model is obtained after the preset convergence condition is reached.

[0022] The expression of the pre-constructed loss function is L=L1+αL2.

[0023] In the formula, alpha is a hyperparameter, L1 is a heat map loss, and L2 is a key point detectability loss.

[0024]

[0025] In the formula, N represents the number of key points; represents the heat map of the jth key point of the ith picture obtained by prediction; P ij represents the real heat map of the jth key point of the ith picture; n represents the number of samples selected in one iteration;

[0026]

[0027] In the formula, v ij represents the detectability of the jth key point of the ith picture labeled by the network. represents the detectability of the jth key point of the ith picture obtained by the network.

[0028] Further improvement of the present application is that in the step of obtaining the training sample set, the step of obtaining the human posture key point detectability information label comprises:

[0029] The maximum value of the detected heat map is used to distinguish the detectability of the human posture key point; wherein the detectability of the human posture key point that is easy to detect is labeled as 1, and the detectability of the human posture key point that is not easy to detect is labeled as 0.

[0030] Further improvement of the present application is that the step of obtaining the image data to be detected for human posture comprises:

[0031] The original image is obtained by shooting or video frame interception;

[0032] The original image is subjected to enhancement preprocessing and SSD target detection processing to obtain the image data to be detected for human posture.

[0033] Further improvement of the present application is that the step of performing enhancement preprocessing comprises:

[0034] The brightness of the image is calculated to obtain a brightness calculation result;

[0035] The brightness of the image is calculated to obtain a brightness calculation result;

[0036] Further improvement of the present application is that the step of performing SSD target detection processing comprises:

[0037] The image is subjected to SSD target detection to obtain a detection frame of each person's position, and a cropped image is obtained by cropping according to the detection frame.

[0038] The present application provides a human posture detection system based on detectability learning, comprising:

[0039] An image acquisition module is configured to acquire image data to be detected for human posture.

[0040] a prediction module, configured to detect, based on the image data to be human posture detected, by using a pre-trained human posture detection model, position information and detectability information of each human posture key point;

[0041] a screening module, configured to display the human posture key point with the detectability information greater than or equal to a preset threshold, and obtain a human posture detection result;

[0042] The human posture detection model comprises:

[0043] a human posture detection deep neural network, configured to input the image data to be human posture detected, and output a heat map; the heat map records position information of each human posture key point;

[0044] a detectability learning supervision network, configured to input the heat map, and output the detectability information of each human posture key point.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] In the human posture detection method based on detectability learning disclosed in the present application, the pre-trained model is used to detect the position of the human posture key point and the detectability, so that the human posture detection demand of the data with occlusion, scale uncertainty, easily confused posture and complex background can be solved, and the present application has the characteristics of high detection precision and high reliability. Specifically, the trained model is used to analyze and detect the position of the key point of the image to be recognized, and the position of the human posture key point and the detectability detection result are obtained; the human posture key point position is accurately analyzed by combining the human posture key point position and the detectability detection result, and finally the detection accuracy and reliability can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following briefly introduces the drawings needed to be used in the embodiments or prior art description; obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0048] Figure 1 is a flowchart of a human posture detection method based on detectability learning provided by an embodiment of the present application;

[0049] Figure 2 is a sample set diagram for model training in an embodiment of the present application;

[0050] Figure 3is a flowchart of a data preparation and video processing stage in an embodiment of the present application;

[0051] Figure 4 is a flowchart of the cooperation of a high-resolution human body pose key point detection model and a detectability learning supervision network branch in an embodiment of the present application;

[0052] Figure 5 is a schematic diagram of a detectability learning supervision network added in a detection model in an embodiment of the present application;

[0053] Figure 6 is a schematic diagram of a pose key point position detection result in an embodiment of the present application;

[0054] Figure 7 is a schematic diagram of a human body pose detection system based on detectability learning provided by an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0056] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0057] The present application will be described in further detail below with reference to the drawings:

[0058] Please refer to Figure 1 The human body pose detection method based on detectability learning provided by the embodiment of the present application comprises the following steps:

[0059] Step 1, obtaining image data to be human posture detection; specifically, the image can be obtained by shooting or intercepting a video frame to obtain an original image, and the original image can have defects such as occlusion, scale uncertainty, confusing posture, and complex background; the original image can be subjected to enhancement preprocessing and SSD target detection processing to obtain the image data to be human posture detection; wherein the enhancement preprocessing can be picture enhancement processing of the image for interference in different light sources, especially weaker light sources; the SSD target detection processing can be SSD target detection of the video image in the field of view to obtain a detection frame of each person's position; the cropped image to be processed is obtained by cropping according to the detection frame;

[0060] Step 2, based on the image data to be human posture detection, using a pre-trained human posture detection model for detection to obtain the position information of each human posture key point and the detectability information of each human posture key point;

[0061] Step 3, based on the position information of each human posture key point and the detectability information of each human posture key point, obtaining the final human posture detection result; specifically, the detection frame of all people in the field of view can be detected and analyzed respectively to obtain the key point position and detectability detection result of the human body in each detection frame; the detection result greater than the confidence threshold is displayed by using the threshold screening and the detectability analysis result.

[0062] In the technical solution provided in the above embodiment of the application, the human posture detection model comprises:

[0063] A human posture detection deep neural network is used to input image data to be human posture detection and output a heat map; the heat map records the position information of each human posture key point;

[0064] A detectability learning supervision network is used to input the heat map and output the detectability information of each human posture key point;

[0065] The human posture detection deep neural network can be HRNet, ResNet or Stacked-Hourglass.

[0066] The detectability learning supervision network comprises:

[0067] An input layer is used to input the heat map to be processed;

[0068] A multi-level down-sampling layer is used to input the heat map and perform step-by-step down-sampling to output a feature map with a resolution of 1x1;

[0069] A full connection layer is configured to adjust a channel number according to the human posture key point based on the 1*1 feature map, and a feature value of each channel is used to represent the detectability of the human posture key point.

[0070] In the technical scheme provided by the embodiment of the present application, the human posture attribute analysis deep neural network model trained is used to analyze and detect the key point position of the image to be recognized, to obtain the human posture key point position and the detectability detection result; the detectability analysis result is used to assist the key point position detection result, to obtain the detectability of the posture key point and make the key point position detection more accurate.

[0071] In the embodiment of the present application, when the human posture detection model is trained, the human posture key point position label in the data set and the human posture key point detectability label obtained by pre-labeling are used simultaneously.

[0072] In the embodiment of the present application, the step of performing the enhancement preprocessing can specifically include the following steps:

[0073] 1.1) Calculate the brightness of the image, and regard the image with brightness less than a predetermined threshold as a dim image;

[0074] 1.2) Perform brightness enhancement on the dim image and then perform detection, until the brightness is greater than or equal to the predetermined threshold, and then output, which is taken as the final result.

[0075] In summary, the high-resolution human posture detection method based on detectability learning provided by the embodiment of the present application processes the input video into pictures, manually labels to form a posture key point detectability learning sample set; uses the human posture attribute analysis deep neural network model trained to analyze and detect the key point position of the image to be recognized, to obtain the human posture key point position and the detectability detection result; combines the human posture key point position and the detectability detection result and the digital image processing technology to eliminate the interference of weak light sources, and accurately analyzes the human posture key point position. The present application can effectively recognize and analyze the key point position and the detectability of the human body in the image, and can be used for human behavior analysis in different application scenarios. Further, the present application proposes a high-resolution human posture detection system based on detectability learning by camera acquisition and combination with AI (artificial intelligence) analysis technology, simultaneously analyzes and counts the human posture in the video, uploads the result to the cloud in real time and displays it.

[0076] Please refer to Figures 1 to 6 The high-resolution human posture detection method based on detectability learning provided by the embodiment of the present application includes the following steps:

[0077] Step 1: based on the pre-selected training set, label to obtain a video image various posture key point detectability learning sample set;

[0078] Step 2: A human posture detection model is constructed, and a neural network is trained using data of a sample set; specifically, an architecture of the human posture detection model in the embodiment of the present application can adopt HRNet, ResNet or Stacked-Hourglass after adding a detectability supervision learning branch, and the trained human posture detection model is saved;

[0079] Step 3: Images are directly obtained from a camera or from a video file on a local hard disk and picture enhancement processing is performed, so as to reduce interference of a weak light source;

[0080] Step 4: An SSD target detection is performed on a video image in a field of view, so as to obtain a detection frame of a position of each person;

[0081] Step 5: A human posture detection model trained is used to perform posture key point position and detectability detection on a human image in the detection frame, so as to obtain position coordinates and a detectability score of each posture key point of each person;

[0082] Step 6: A detectability threshold is used to screen a detectability detection result of a human posture key point, and a display is performed on a detectability score greater than the threshold, so as to weaken interference of an interference target of a shielding and easily confused posture.

[0083] The embodiment of the present application provides a technical solution of the original explanation. At present, scale uncertainty, occlusion, illumination, and complex background are still difficulties in the human pose key point detection task. It is very important to improve the robustness of the human pose key point detection system to occlusion, scale uncertainty, and complex background. The existing key point detection method cannot evaluate whether the detected key point position is reliable. The detectability (reliability) analysis of the detected key point position will help the downstream task to reasonably and accurately use the detection result. The high-resolution human key point detection method based on detectability learning provided by the embodiment of the present application can not only improve the accuracy of human key point detection, but also evaluate the detectability of the key point. Specifically, the present application uses an SSD deep neural network target detector to perform target detection on an image to obtain a detection frame of each person; the SSD network is a target detection algorithm for directly predicting the coordinates and categories of a bounding box, which is proposed by Liu Wei et al. in 2016, and details are shown in the literature: Liu W, Anguelov D, Erhan D, et al. SSD: Single Shot MultiBox Detector [J]. 2016. The basic idea is to cancel the proposal generation process and use the feature maps of different convolution layers for comprehensive analysis to achieve the same effect. Deep neural networks have made great progress in the field of human pose detection. The present application can use various deep neural networks for video analysis, including HRNet, ResNet or Stacked-Hourglass. After multiple experimental tests, the HRNet with the added detectability supervision learning branch has achieved better results in the accuracy of pose key point position detection. The present application uses the HRNet with the added detectability supervision learning branch, but is not limited to the HRNet deep neural network. Explanatorily, the HRNet network is a detection algorithm for predicting the coordinates of human pose key point positions through a heat map, which is proposed by Ke Sun et al. in 2019, and details are shown in the literature: Sun K, Xiao B, Liu D, et al. Deep High-Resolution Representation Learning for Human Pose Estimation [J]. arXiv e-prints, 2019. The basic idea is to connect parallel networks of different resolutions in parallel, repeatedly fuse low-resolution feature maps of different scales, and enhance high-resolution features to obtain a more accurate heat map.

[0084] Please refer to Figure 2The embodiment of the present application is based on the MPII dataset, and is used for training a posture key point position attribute analysis deep neural network model. Specifically, according to the detection result of the HRNet on an image, the maximum value of a detected heat map is used to distinguish the detectability of a key point, and the maximum value of the heat map is selected as the distinguishing basis. The detectability of a key point that is obvious and easy to be detected by a network is marked as “1”, and the detectability of a key point that is not easy to be detected due to factors such as occlusion, illumination, and scale uncertainty is marked as “0”. The annotation of the MPII dataset is completed, and the MPII dataset is named as NMPII. Figure 2 The NMPII is a partial posture key point position analysis deep neural network detectability learning sample set. The sample must include a human body image to be detected, and the coordinate position and detectability annotation of a posture key point.

[0085] Please refer to Figure 3 In the data preparation and video processing stage, image enhancement processing is first performed. Mainly brightness adjustment, the specific steps can include:

[0086] 1) The brightness of the test image is calculated, and the image less than a certain threshold is regarded as a dim image;

[0087] 2) The dim image is subjected to brightness enhancement and then detection, and the output is taken as the final result.

[0088] Please refer to Figure 3 In the data preparation and video processing stage, the embodiment of the present application extracts a human detection frame in a video image by using the SSD, and cuts the video image according to the position of the detection frame to output the cut image containing the human body.

[0089] Please refer to Figure 4 The input RGB human body image is output by a high-resolution human key point detection model (HRNet), and the position of the key point is determined by using a heat map. Then, the heat map is input into a detectability supervision learning branch network to evaluate the detectability of the key point.

[0090] Please refer to Figure 5 , Figure 5 The detectability supervision learning branch network is a schematic diagram of the embodiment of the present application, and the data processing steps thereof include:

[0091] 1) The heat map output by the high-resolution human posture key point detection network is input, and the heat map has 16 channels, each of which represents a key point;

[0092] 2) In the down-sampling process, the image resolution is halved, and the channel number is doubled;

[0093] 3) When the resolution size of the feature map is reduced to 4x4, adaptive average pooling is adopted to adjust the resolution size of the feature map to 1x1.

[0094] 4) Through a full connection layer, the number of channels is changed to 16, corresponding to 16 key points, and the feature value of each channel represents the key point detectability.

[0095] In the embodiment of the present application, a sample set is learned through a deep neural network, the error between the real attribute and the predicted attribute is continuously reduced to make the prediction result of the network close to the real result, a deep posture key point position attribute analysis model is trained by using a deep learning algorithm; the specific training steps include:

[0096] 1) The deep neural network is trained for 200,000 times to generate a deep neural network target detector, and a model test is performed every 500 times to select a best performance model to save.

[0097] 2) The embodiment of the present application proposes LDLoss, a joint loss of location detectability, the heat map loss L1 of the HRNet key point detection of the present application is:

[0098]

[0099] In the formula, N represents the number of key points; represents the heat map of the jth key point of the ith picture obtained by prediction; P ij represents the real heat map of the jth key point of the ith picture; n represents the number of samples selected in one iteration;

[0100] The key point detectability loss L2 is:

[0101]

[0102] In the formula, v ij represents the detectability of the jth key point of the ith picture labeled by the network. represents the detectability of the jth key point of the ith picture obtained by the network.

[0103] The total loss function L is: L=L1+αL2;

[0104] In the formula, alpha is a hyperparameter. Through experimental test, the hyperparameter is set to 0.01.

[0105] The embodiment of the present application analyzes the posture key point position and detectability by using the detectability supervised learning auxiliary analysis mechanism, and the steps can include:

[0106] 1) All human body detection frame images in the field of view are cut to extract images containing human bodies;

[0107] 2) analyze each region containing a human body respectively to obtain a result of human body posture key point attribute detection of the region;

[0108] Referring to Figure 6 In the embodiment of the present application, finally, the target with a detectability score greater than a threshold value is displayed by threshold screening. The experimental results of the embodiment of the present application show that through the technical scheme provided by the embodiment of the present application, the position attribute analysis of the posture key points randomly appearing under the occlusion and easily confused posture can output the position coordinates and the detectability score of each key point; the key point position analysis reaches an accuracy of greater than 90% (the proportion of correct estimation of the key point position with the head diameter as the scale factor), and the accuracy in the occlusion and easily confused posture is improved more significantly; the detectability score is greater than 84; compared with the existing traditional detection method, the method provided by the embodiment of the present application not only has a high recognition accuracy, but also enhances the robustness of the algorithm, and realizes the accurate analysis of the position of the human body posture key point. The present application can also evaluate the detectability of the human body key point, which will help the downstream task to use the detection result reasonably and accurately.

[0109] In summary, the present application discloses a high-resolution human body posture detection method and system based on detectability learning, which comprises the following steps: labeling a variety of posture key point detectability learning sample sets of video images; constructing a posture key point position attribute analysis model adding a detectability supervised learning branch; training the neural network using the data of the sample set, and saving the trained network model to the local hard disk; directly obtaining images from the camera or from the video files on the local hard disk and performing picture enhancement processing to reduce the interference of weak light sources; using the SSD model to extract the human detection frame in the image and crop the image; using the trained network model to detect the key point position attribute of the newly obtained image to be recognized; using a confidence threshold to screen the detectability detection result to weaken the interference of the occlusion and easily confused posture target. The present application can effectively recognize and analyze human body video, accurately detect human body posture key point attribute, and be used for various application scenarios of human body posture analysis.

[0110] The following is an apparatus embodiment of the present application, which can be used to execute the method embodiment of the present application. For details not mentioned in the apparatus embodiment, please refer to the method embodiment of the present application.

[0111] Referring to Figure 7 In another embodiment of the present application, a human body posture detection system based on detectability learning is provided, which comprises:

[0112] An image acquisition module is configured to acquire image data to be detected for human body posture.

[0113] The prediction module is configured to detect, based on the image data to be detected, the human posture, by using a pre-trained human posture detection model, to obtain position information and detectability information of each human posture key point;

[0114] The screening module is configured to display the human posture key point whose detectability information is greater than or equal to a preset threshold, to obtain a human posture detection result.

[0115] The human posture detection model comprises:

[0116] The human posture detection deep neural network is configured to input the image data to be detected, and output a heat map. The heat map records position information of each human posture key point.

[0117] The detectability learning supervision network is configured to input the heat map, and output the detectability information of each human posture key point.

[0118] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0119] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure One The functions specified in one or more flows and / or blocks. Figure One The devices that implement the functions specified in one or more flows and / or blocks.

[0120] These computer program instructions can also be stored in a computer-readable memory that can guide the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure One The functions specified in one or more flows and / or blocks. Figure One The devices that implement the functions specified in one or more flows and / or blocks.

[0121] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide operational steps for implementing the functions of the flow Figure One One flow or multiple flows and / or the functions specified in one block or multiple blocks. Figure One Figure One One flow or multiple flows and / or the functions specified in one block or multiple blocks.

[0122] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that: the specific embodiments of the present application can still be modified or replaced by the equivalent, without departing from the spirit and scope of the present application, any modification or equivalent replacement, which should be covered within the protection scope of the claims of the present application.

Claims

1. A human pose detection method based on detectability learning, characterized in that, Includes the following steps: Acquire image data for human pose detection; Based on the image data of the human body posture to be detected, a pre-trained human posture detection model is used to perform detection, and the location information and detectability information of each human posture key point are obtained. Human posture key points whose detectability information is greater than or equal to a preset threshold are displayed to obtain human posture detection results. The human posture detection model includes: A deep neural network for human posture detection is used to input image data of the human posture to be detected and output a heat map; the heat map records the location information of each human posture key point; A detectability learning supervised network is used to take the heatmap as input and output detectability information for each human posture key point; The detectability learning supervised network includes: The input layer is used to input the heatmap to be processed. The downsampling layer is used to input the heatmap and perform step-by-step downsampling, outputting a feature map with a fractional ratio that meets the preset requirements. An average pooling layer is used to input a feature map whose fractionation reaches a preset requirement. Adaptive average pooling is used to adjust the image resolution, and the output feature map has a resolution of 1×1. A fully connected layer is used to adjust the number of channels to be the same as the number of human pose key points based on the 1×1 feature map. The feature values ​​of the channels are used to represent the detectability of human pose key points. The steps for obtaining the pre-trained human pose detection model include: Obtain the training sample set; each training sample includes a sample image, human pose key point location labels, and human pose key point detectability information labels. Based on the training sample set, a pre-constructed loss function is used to continuously reduce the error between the real attributes and the predicted attributes to update the parameters. After reaching the preset convergence condition, the pre-trained human pose detection model is obtained. The expression for the pre-constructed loss function is: ; In the formula, For hyperparameters; For heatmap loss; Loss of detectability at key points; ； In the formula, N represents the number of key points; This represents the heatmap of the j-th key point in the predicted i-th image. This represents the true heatmap of the j-th key point in the i-th image; n represents the number of samples selected in one iteration. ； In the formula, This indicates the detectability of the j-th keypoint in the labeled i-th image; This represents the detectability of the j-th keypoint in the i-th image obtained from the network.

2. The human pose detection method based on detectability learning according to claim 1, characterized in that, The deep neural network for human pose detection is HRNet, ResNet, or Stacked-Hourglass.

3. The human pose detection method based on detectability learning according to claim 1, characterized in that, The step of obtaining the training sample set, specifically the step of obtaining the human pose key point detectability information labels, includes: The detectability of human posture key points is distinguished by the maximum value of the detected heat map; the detectability of human posture key points that are easy to detect is marked as 1, and the detectability of human posture key points that are difficult to detect is marked as 0.

4. The human pose detection method based on detectability learning according to claim 1, characterized in that, The steps for acquiring image data for human posture detection specifically include: The original image is obtained by shooting or capturing video frames; The original image is enhanced and preprocessed with SSD target detection to obtain the image data for human pose detection.

5. The human pose detection method based on detectability learning according to claim 4, characterized in that, The steps for performing the enhancement preprocessing include: Perform brightness calculations on the image to obtain the brightness calculation results; Images whose brightness calculation results are less than a preset threshold are enhanced to obtain an enhanced preprocessed image.

6. The human pose detection method based on detectability learning according to claim 4, characterized in that, The steps of the SSD target detection processing include: Perform SSD object detection on the image to obtain the detection box of each person's location, and crop the image according to the detection box.

7. A human posture detection system based on detectability learning, characterized in that, include: The image acquisition module is used to acquire image data of the human body to be pose detected; The prediction module is used to perform detection based on the image data of the human body to be detected, using a pre-trained human body posture detection model, to obtain the location information and detectability information of each human body posture key point. The filtering module is used to display human posture key points whose detectability information is greater than or equal to a preset threshold, and obtain human posture detection results. The human posture detection model includes: A deep neural network for human posture detection is used to input image data of the human posture to be detected and output a heat map; the heat map records the location information of each human posture key point; A detectability learning supervised network is used to take the heatmap as input and output detectability information for each human posture key point; The detectability learning supervised network includes: The input layer is used to input the heatmap to be processed. The downsampling layer is used to input the heatmap and perform step-by-step downsampling, outputting a feature map with a fractional ratio that meets the preset requirements. An average pooling layer is used to input a feature map whose fractionation reaches a preset requirement. Adaptive average pooling is used to adjust the image resolution, and the output feature map has a resolution of 1×1. A fully connected layer is used to adjust the number of channels to be the same as the number of human pose key points based on the 1×1 feature map. The feature values ​​of the channels are used to represent the detectability of human pose key points. The steps for obtaining the pre-trained human pose detection model include: Obtain the training sample set; each training sample includes a sample image, human pose key point location labels, and human pose key point detectability information labels. Based on the training sample set, a pre-constructed loss function is used to continuously reduce the error between the real attributes and the predicted attributes to update the parameters. After reaching the preset convergence condition, the pre-trained human pose detection model is obtained. The expression for the pre-constructed loss function is: ; In the formula, For hyperparameters; For heatmap loss; Loss of detectability at key points; ； In the formula, N represents the number of key points; This represents the heatmap of the j-th key point in the predicted i-th image. This represents the true heatmap of the j-th key point in the i-th image; n represents the number of samples selected in one iteration. ； In the formula, This indicates the detectability of the j-th keypoint in the labeled i-th image; This represents the detectability of the j-th keypoint in the i-th image obtained from the network.

Images