33 results about "Pose estimation algorithm" patented technology

A method for pose and gait classification and motion prediction using spatio-temporal relationships between body joints includes capturing a sequence of images or video frames of a subject; applying a neural network-based pose estimation algorithm to the sequence of images or video frames to detect landmark positions of anatomical joints; constructing a spatio-temporal graph from the detected landmark positions of the one or more anatomical joints, wherein nodes of the spatio-temporal graph correspond to the anatomical joints and the landmark positions, spatial edges of the spatio-temporal graph represent anatomical connections between the anatomical joints within a single image or frame, and temporal edges connect the one or more anatomical joints across successive images or frames of the sequence of images or video frames; and inputting the constructed spatio-temporal graph into a spatio-temporal graph convolutional network (ST-GCN) to classify pose and gait patterns and predict motion or stability states.

The invention provides a portrait animation generation method and system based on a reversible encoder and dynamic optimization, and relates to the technical field of video portrait reconstruction, high-precision portrait animation generation is realized through multi-module collaborative innovation, multi-scale appearance features are extracted by using a VAE encoder, identity vectors extracted by ArcFace and global appearance embedding of CLIP are combined, and the portrait animation generation method and system based on the reversible encoder and dynamic optimization are realized. Constructing a multi-modal characteristic system; a high-precision attitude estimation algorithm is adopted to extract fine joint points, and an error between a generated attitude and an original attitude is calibrated in real time through a dynamic optimization strategy, so that the attitude driving accuracy is remarkably improved. A reversible residual block is designed to realize lossless two-way transmission of features, and the problem of information compression of a traditional encoder is avoided; a high-frequency detail injection mechanism is introduced, high-frequency features such as facial textures are extracted through a pre-training model and fused with shallow features, an attention mechanism guided by identity features is introduced based on the generation process of a diffusion model, and the consistency of character identities in the generation process is ensured.

The invention discloses an intelligent recognition method and system for grabbing scattered parts, and relates to the technical field of disordered grabbing. In order to solve the problems of high data labeling cost, low occlusion scene pose precision, difficulty in cross-domain generalization and the like in the prior art, a full-process closed loop of'simulation data synthesis-multi-modal segmentation-pose estimation-semi-supervised training 'is constructed: a field adaptive simulation data set is generated through PyBullet + PyRender, and a MaskNet algorithm is obtained based on ViT improved YOLO11 to realize multi-modal instance segmentation; a MaskPosenet two-stage pose estimation algorithm is designed, and cross-domain generalization is realized in combination with pseudo-label semi-supervised learning of a cosine scheduling threshold. According to the method, the data cost is greatly reduced, the complex scene capturing precision and generalization ability are improved, and the method is adaptive to industrial scenes such as logistics sorting and flexible manufacturing and has extremely high application value.

The invention relates to the field of computer vision and intelligent manufacturing, in particular to a vehicle charging port image marking method and device, and the method comprises the steps: obtaining a vehicle charging port image, and carrying out the local frame selection processing; sequentially carrying out graying processing, contrast-limited adaptive histogram equalization processing and median filtering processing; carrying out edge detection by using a Canny edge detection algorithm; adopting a Hough circle detection algorithm to generate an initial position parameter of the circular hole; or, adopting a pre-trained convolutional neural network semantic segmentation algorithm to carry out pixel-level prediction, and extracting corresponding edge points as initial position parameters of the circular holes; if it is judged that the precision of the initial position parameter does not meet the preset labeling requirement, click operation is executed, and a correction point set is formed; and based on the correction point set and the edge points corresponding to the initial position parameters, fitting an ellipse to obtain a marking result of the charging port circular hole. The method can improve the labeling efficiency, guarantees the labeling precision and stability, and is adaptive to subsequent deep learning segmentation and pose estimation algorithms.

The invention relates to the technical field of warehouse logistics equipment, and discloses a monocular camera-based pallet pose estimation method, a storage medium and related equipment. The method comprises the following steps: acquiring a fork entrance surface image of a pallet through a monocular camera; performing feature matching on the fork entrance surface image and a pre-stored template image to obtain a pallet area in the fork entrance surface image; the pallet area is preprocessed; detecting the edge in the pallet area through an edge detection algorithm; detecting angular points in the pallet area through angular points, and screening projection angular points corresponding to preset control angular points of the fork entrance surface of the pallet; obtaining projection coordinates of the projection angular points in a first coordinate system of the monocular camera through a sub-pixel angular point detection algorithm; and according to the 3D world coordinates and the projection coordinates of the control angular points, solving attitude parameters and position parameters of the pallet in the first coordinate system through a pose estimation algorithm. According to the method, the calculation amount of pallet pose estimation can be reduced, and the pose estimation response speed is increased.

This invention discloses a fault-tolerant method for quay crane pose estimation based on multi-sensor fusion, belonging to the technical field of port automation equipment. When a quay crane loses connection due to a network failure, the system automatically detects the pose loss by comparing and analyzing pre-registered quay crane IDs and placement order information with real-time received data. Upon detecting a missing quay crane position, the system immediately reports an error and records the disconnection information. Different intelligent inference strategies are employed based on the quay crane's position order within the area: for quay cranes in the middle position, the median value of the adjacent quay cranes is used as the initial pose; for quay cranes at the edge position, the approximate pose is inferred based on the positions of adjacent quay cranes and the overall trend. This invention employs a point cloud matching-based pose estimation algorithm, achieving high-precision relative pose estimation through multi-sensor fusion of LiDAR and visual cameras, significantly improving the fault tolerance and operational continuity of the port automation system.

The two-stage object six-degree-of-freedom pose estimation method based on a monocular camera solves the problem of how to achieve high-precision and high-efficiency pose estimation of weak-texture industrial parts, and belongs to the technical field of monocular camera pose estimation. The present invention is a two-stage pose estimation method from coarse to fine. The coarse pose estimation network roughly estimates the six-degree-of-freedom pose value based on the fusion of mask features and 2D position features. The coarse pose estimation stage provides the initial value for precise pose estimation, avoiding the optimization process of the precise pose estimation algorithm from falling into the local optimal value. The precise pose estimation network introduces a dynamic snake-shaped convolution layer to construct an object contour feature extraction network, improves the quality of contour feature generation, and filters out background contours. Through the differentiable Gauss-Newton optimization method, iterative optimization is performed at multiple scales to minimize the reprojection error between the projection contour and the object contour, thereby achieving accurate pose estimation.

The invention provides a laser radar position identification method based on point cloud structure features. Comprising the following steps: S1, acquiring three-dimensional point cloud data acquired by a laser radar, and generating a bird's-eye view BEV image reflecting point cloud spatial distribution characteristics; s2, extracting two-dimensional key points from the bird's-eye view BEV image, and dividing the corresponding three-dimensional key points into point features, line features and surface features; s3, constructing a global hash descriptor reflecting the overall structural characteristics of the point cloud and a local probability descriptor reflecting the local structural characteristics; s4, performing rapid candidate retrieval based on the global hash descriptor to obtain a candidate frame set for position identification; and S5, carrying out relative pose solving on the candidate frame and the current frame by adopting a robust pose estimation algorithm, and determining a final position identification result. According to the invention, through BEV image conversion, double descriptor cooperation and multi-dimensional innovation of a two-stage strategy, core bottlenecks of efficiency and robustness imbalance and incomplete feature representation in laser radar position identification are systematically solved.

A human pose estimation system 100 for a motor vehicle 130, the system includes an imaging device 102 and a processor 104, the processor further comprises a base model 110 operable to execute a pose estimation algorithm 112. The pose estimation algorithm is operable to determine a pose of the at least one vehicle occupant 132, 132’ contained in the at least one video image frame 120 captured by the imaging device. The processor further comprises a visibility predictor 114 operable to execute a visibility prediction algorithm 116 in parallel with the pose estimation algorithm. The visibility algorithm is operable to determine a visibility state of at least one keypoint relative to the pose of the at least one vehicle occupant contained in the at least one video image frame captured by the imaging device. Preferably, the processor is operable to generate a value of total loss function of the human pose estimation system and input the total loss function generated to an artificial neural network 106 to train the artificial neural network.

The invention relates to a rigid body target pose measurement method fusing multi-view semantic key points, and the method comprises the following steps: constructing a sample set, and training a target detection network and a semantic key point positioning network based on the sample set; detecting all target objects in the multiple views based on the trained target detection network, obtaining bounding boxes of the target objects and generating a clipped image set; performing key point prediction on the target object in the image set based on a semantic key point positioning network; key points under different views are matched based on multi-view set constraints, and a multi-view semantic key point set is formed; and inputting the obtained multi-view semantic key point set into a three-stage asymptotic pose estimation algorithm to solve the pose of the target object. According to the method, the semantic key points of multiple views are fused, so that the performance bottleneck in a special / complex scene in a traditional scheme can be effectively solved, and the accurate measurement of the position and the attitude of the rigid body target is realized.

The invention provides an object attitude estimation method and system based on three-dimensional feature mapping and visual angle aggregation, and belongs to the technical field of attitude estimation, and the method is applied to a system comprising an image processing module, a three-dimensional feature mapping module, a visual angle aggregation module and an attitude estimation module. The method specifically comprises the steps of performing feature extraction on a template image and a target image, performing three-dimensional space mapping on template image features, determining template three-dimensional features and view angle information, determining the view angle information of each target image feature and the corresponding template three-dimensional feature through feature matching, and determining the target image feature according to the view angle information. And obtaining a matching relation between target image features and template three-dimensional features through view angle aggregation, selecting an attitude estimation algorithm to solve an alternative rough attitude estimation result, and performing iterative optimization by taking the alternative rough attitude estimation result as an initial value of an attitude optimization network to obtain an object attitude estimation result. According to the method, the error transmission risk in attitude estimation can be effectively reduced, the accuracy is guaranteed, the calculation complexity can be reduced, and the estimation efficiency is improved.

In view of the problems of poor real-time performance and low recognition rate of existing head pose estimation algorithms in complex scenes, the application provides a new lightweight head pose estimation method. The method has a multi-level output structure, uses three different types of branch networks to extract features from the input image, and each branch has three stages, and each stage only needs to refine the features of the previous stage. The features extracted by each branch at the same stage are fused by a feature fusion module to generate a feature map, effectively avoiding the problem of feature loss. The feature extraction module selects the Ghost module as the feature extraction network, uses model compression to reduce the network parameters and computational complexity while ensuring the accuracy of the network; in order to extract more important effective features, an efficient channel attention module ECA-Net is introduced, thereby improving the accuracy of head pose estimation. The method of the application can achieve excellent head pose estimation efficiency and has wide applicability.

The invention provides a mechanical arm control method for tracking a dynamic target in real time, and belongs to the technical field of mechanical arm control, and the mechanical arm control method comprises the steps that a vision measurement system is constructed through a binocular infrared camera and a visible light camera, and a marker image on the dynamic target is captured; based on a visual detection algorithm, two-dimensional code data corresponding to the marker image are extracted for tracking identification detection, and when it is detected that the dynamic target is a to-be-tracked target, real-time position information and real-time attitude information of the to-be-tracked target are calculated through a perspective attitude estimation algorithm; performing fusion processing on low-frequency vision measurement data constructed by the real-time position information and the real-time attitude information and the high-frequency motion signal based on an inter-frame difference method to obtain real-time pose data; and the visual servo controller controls the mechanical arm to move according to a motion optimal solution obtained based on the mechanical arm kinematics model. According to the method, real-time tracking of the dynamic target is achieved, and the method has high operation grabbing efficiency and is suitable for various task scenes.

Images of a plurality of objects in a bin from multiple views are received from two image sensors. An initial pose estimation for an object is generated using a pose estimation algorithm and an image of the images. A depth map for the object is generated using a stereo matching algorithm and the images. A depth for the object is rendered using the initial pose estimation for the object and camera intrinsics of the two image sensors. A value that represents an accuracy of the depth of the object in the bin is determined based on comparing the depth map for the object and the rendered depth for the object. A robot arm is instructed to grasp the object based on comparing the value with a first threshold, the robot arm being a proximal distance from the two image sensors, the plurality of objects, and the bin.

This invention discloses a machine vision-based method, system, device, and medium for monitoring learning attention, specifically relating to the field of object pose determination in image analysis. It addresses the perspective distortion problem caused by changes in the distance between the learner's face and the camera in existing technologies. The method involves acquiring continuous video sequences using a monocular camera, detecting the learner's facial region, and extracting the two-dimensional coordinates of facial feature points. Based on these two-dimensional coordinates, topological persistence features are analyzed to determine the distance change state. When a distance change occurs, the motion trajectory of the facial feature points is constructed, and the curvature distribution is calculated. Curvature entropy features are obtained through information entropy analysis. Consistency judgment is performed in conjunction with the topological persistence features, and the estimated distance change is adaptively corrected. Finally, the projection model parameters of the three-dimensional head pose estimation algorithm are dynamically adjusted based on the distance change, and the head pose angle is calculated by combining the facial feature point coordinates, thus achieving accurate judgment of the learner's attention state.

The invention relates to a skeleton point-based shielding scene behavior recognition method and system, and the method comprises the steps: firstly extracting time sequence skeleton point data through a human body detection and posture estimation algorithm, and constructing five types of structured inputs through a mask module; then, the multi-stream GCN learns local spatio-temporal features of different body areas respectively; in the fusion process, the model can adaptively reinforce the feature contribution of a region which is highly related to the current action or is not shielded, and meanwhile, the interference caused by a weak correlation region and a missing part is effectively inhibited. Meanwhile, three types of regularization items of diversity, sparsity and consistency are designed, so that the weights are kept different among the types, are kept compressed on branches and are kept stable in the types. The scheme provided by the invention is obviously superior to the existing method in complex scenes such as random shielding, arm shielding, leg shielding and the like, has high robustness and strong generalization ability, and is suitable for the fields of intelligent monitoring, human-computer interaction, edge calculation behavior analysis and the like.

The invention discloses a skeleton action recognition method based on a partition space-time collaborative graph convolutional network, and relates to the technical field of computer vision and pattern recognition, and the method comprises the steps: firstly obtaining human skeleton sequence data through a depth sensor or a posture estimation algorithm, and carrying out the preprocessing; then, constructing a partition association graph, dividing part partitions based on a human body natural structure, and establishing in-part stable topology and cross-part semantic topology; then introducing a part dynamic collaborative convolution (PDC-GC) module, and cooperatively extracting global semantic and local dynamic features of the action through part association modeling and dynamic neighborhood convolution; introducing a space-time sub-domain coding (STSE) module, capturing multi-scale space-time dependence by dividing four sub-domains, and reducing calculation complexity in combination with sparse coding; and finally, constructing a partitioned space-time collaborative graph convolutional network (PC-GCN) network, fusing a PDC-GC module, an STSE module and multi-level features, and realizing high-precision skeleton action recognition.

This invention discloses a method and system for deep learning-based human behavior recognition using ZYNQ, comprising the following steps: 1) acquiring a video dataset of human behavior; 2) performing data preprocessing, processing the human estimation information collected from the video dataset in step 1) using a human pose estimation algorithm, and dividing the collected data into a training set and a test set; 3) constructing a human behavior recognition model, using a designed deep learning neural network to extract and process features from the dataset in step 2), thereby training the model; 4) setting training parameters and training the designed deep learning network; 5) according to actual needs, using ZYNQ as the core platform, transferring the trained model, and finally deploying it to achieve online human behavior recognition. This invention is beneficial for recognizing complex, long-term reasoning behaviors in video scenarios and has practical application value.

The present application relates to the field of stage lighting, and specifically discloses a method, device and medium for realizing intelligent stage performance, which comprises the following steps: establishing a pose prediction model for predicting the pose data of performers after a first time length by using a target detection algorithm, a target tracking algorithm, a human pose estimation algorithm and a time series prediction model; establishing and training a lighting effect model for calculating stage lamp control data from the pose data of performers by using a machine learning algorithm and a lighting parameter formula; and calculating and obtaining stage lamp prediction control data according to the real-time pose data of performers obtained by a sensor node, so as to control the stage lamp. The present application realizes intelligent control of the stage lamp, realizes real-time monitoring and tracking of the performers, and ensures that the light and shadow can accurately follow the motion changes of the performers by predicting the trajectories of the performers in advance.

This application provides a method, apparatus, computer device, and storage medium for pose estimation of a mobile robot. The method includes: acquiring a target image through a vision system; estimating the pose information of a target object from the target image using a preset pose estimation algorithm; determining the orientation information of the target object, the distance between the mobile robot and the target object, and the pose category of the target object based on the pose information; searching a preset prior knowledge base for the target observation angle of the vision system associated with the pose category; determining the target observation position of the mobile robot based on the current position of the mobile robot, the target observation angle of the vision system, the orientation information of the target object, and the distance between the mobile robot and the target object; controlling the mobile robot to move to the target observation position, and simultaneously executing the step of acquiring the target image through the vision system. This method can improve the accuracy of pose estimation.

The application relates to a lightweight robust face alignment method and system based on multi-task learning, which comprises the following steps: S1, collecting face images, and performing translation, scaling and rotation preprocessing on the face images to obtain a training set; S2, obtaining face key points L of the face images in the training set; using a pose estimation algorithm to mark 3D head pose Euler angles Theta of the face images in the training set, and converting the 3D head pose Euler angles Theta into a head pose rotation matrix R Θ ; S3, inputting the training set into a face alignment network to output face key point prediction values P and head pose rotation matrix prediction values R Φ , and respectively constructing a head pose loss function, a face key point loss function and a total loss function for training the face alignment network. The method provided by the application simultaneously aligns face key points and head poses using a lightweight model, and uses a head pose alignment task to assist the positioning of the face key points, thereby enhancing the robustness of the lightweight model to large-pose faces.

The application relates to the technical field of intelligent unmanned cluster systems, and discloses a cooperative positioning method for an unmanned cluster in a denial environment, which considers any pair of neighboring individuals except for both of which having no relative measurement capability, constructs observation equations among relative measurement information of both parties, displacement measurement information of both parties' odometers and relative poses of both parties, introduces a data-driven adaptive estimation technology, breaks through the deficiency of traditional relative pose estimation algorithms of unmanned systems that require robots to keep sustained excitation relative motion, thereby causing the control performance to be degraded, designs a distributed pose coupling estimator to realize cooperative estimation of relative poses of a leader individual by any individual in the cluster, realizes cooperative pose estimation of the unmanned cluster system under a weakly connected measurement topological condition, and solves the problem that existing cooperative positioning methods have poor scalability due to high requirements on measurement topologies.

The invention discloses a quay crane pose estimation fault tolerance method based on multi-sensor fusion, and belongs to the technical field of port automation equipment. According to the invention, when a quay crane is disconnected due to a network fault, the system compares and analyzes the pre-registered quay crane ID and quay crane placement sequence information with real-time received data, and automatically detects the pose disconnection condition. When a missing quay crane position is detected, a system immediately reports an error and records disconnection information, and different intelligent inference strategies are adopted based on the position sequence of the quay cranes in an area: for the quay crane in the middle position, an intermediate value of front and back adjacent quay cranes is used as an initial pose; for the quay cranes at the edge positions, the approximate poses are inferred based on the positions and overall trends of the adjacent quay cranes. According to the method, a pose estimation algorithm based on point cloud matching is adopted, high-precision relative pose estimation is realized through multi-sensor fusion of the laser radar and the visual camera, and the fault-tolerant capability and operation continuity of a port automation system are remarkably improved.

This invention provides a method, apparatus, and related equipment for driving a virtual human. The method includes: receiving a video to be identified; identifying the video using a target 2D pose estimation model and a target 3D pose estimation model to obtain the coordinates of multiple human body 3D key points; calculating the rotation angles of all frames in the video based on the coordinates of the multiple human body 3D key points; storing the coordinates of the multiple human body 3D key points and the rotation angles as a preset format file according to the parent-child relationship between the coordinates of the multiple human body 3D key points; and inputting the preset format file into a preset engine to drive a virtual human model in the preset engine. Based on the pose estimation algorithm, the optimized target 2D pose estimation model and target 3D pose estimation model are used to extract the coordinates of multiple human body 3D key points in the video to be identified, improving the speed and accuracy of pose estimation. Using the pose estimation algorithm to drive the virtual human model reduces production costs and facilitates the expansion of application scope.

The invention discloses a learning concentration degree monitoring method, system and device based on machine vision and a medium, particularly relates to the technical field of object posture determination in image analysis, and is used for solving the problem of perspective distortion caused by distance change between the face of a learner and a camera in the prior art. The method comprises the following steps: collecting a continuous video sequence through a monocular camera, detecting a learner face region, extracting two-dimensional coordinates of face feature points, analyzing topological persistence features based on the two-dimensional coordinates to judge a distance change state, constructing a motion track of the face feature points when distance change occurs, and calculating curvature distribution; curvature entropy features are obtained through information entropy analysis, consistency judgment is carried out in combination with topological persistence features, a distance variation estimation value is adaptively corrected, finally, projection model parameters of a three-dimensional head posture estimation algorithm are dynamically adjusted according to the distance variation, and a head posture angle is calculated in combination with facial feature point coordinates. The concentration degree state of the learner can be accurately judged.

The invention discloses a construction personnel illegal behavior detection method and system based on machine vision, and the method comprises the steps: intercepting a local image block from a coordinate set of a potential illegal region, analyzing the joint point distribution of the local image block through employing a posture estimation algorithm, and obtaining a limb posture vector group of construction personnel in high-altitude operation; obtaining the limb posture vector group, carrying out vector similarity matching on the limb posture vector group and a pre-established violation behavior database, and if the similarity exceeds a preset threshold value, marking as a violation posture type to obtain a classified violation label set; backtracking video sequence data according to an event starting frame index, extracting interference factor characteristics such as light change and object shielding of related frames, and correcting the interference factor characteristics by adopting a filtering algorithm to obtain an optimized clear image sequence; and summarizing a violation label set and a dynamic action trajectory curve from the optimized clear image sequence, generating real-time monitoring report data, and outputting the real-time monitoring report data to a field management system to trigger an alarm response.

The application provides an object pose estimation method and system based on three-dimensional feature mapping and perspective aggregation, and belongs to the technical field of pose estimation.The method is applied to a system comprising an image processing module, a three-dimensional feature mapping module, a perspective aggregation module and a pose estimation module, and specifically comprises the following steps: performing feature extraction on a template image and a target image, performing three-dimensional space mapping on the template image features, determining the template three-dimensional features and perspective information, determining the perspective information of each target image feature and the corresponding template three-dimensional features through feature matching, obtaining the matching relationship between the target image features and the template three-dimensional features through perspective aggregation, selecting a pose estimation algorithm to solve an alternative rough pose estimation result, and iteratively optimizing the initial value of a pose optimization network to obtain an object pose estimation result.The application can effectively reduce the error transmission risk in pose estimation, guarantee the accuracy, reduce the calculation complexity and improve the estimation efficiency.

The present application belongs to the field of machine vision, and relates to an object detection and pose estimation method. The method is based on the adjustment of YOLO-6D algorithm, the YOLO V2 detection network in the original algorithm is changed into YOLO V3 network, and an attention mechanism is added to enhance the detection ability of the model to the object with complex background and occlusion. And the pose estimation method is adjusted, the cell group is selected for RANSAC-based EPnP pose estimation to improve the estimation accuracy. The algorithm well overcomes the problems of weak anti-background interference ability and poor recognition accuracy of the occluded target of the traditional pose estimation algorithm, and the algorithm runs fast and has real-time processing capability, and the comprehensive performance exceeds other CNN-based algorithms.

The application discloses a pose data acquisition method and device, equipment and medium, and relates to the field of computer vision and robots. The method comprises the following steps: determining a target sampling viewpoint on a spherical surface with the center of a target object as the spherical center and a preset sampling distance as the radius; determining a target motion trajectory of a mechanical arm carrying a camera at the end based on the target sampling viewpoint and by using a preset trajectory planning method; controlling the mechanical arm to move in the target motion trajectory, and recording a video of the target object by using the camera carried at the end of the mechanical arm to obtain a target video; extracting a target image frame from the target video, and determining pose information corresponding to the target image frame by using a preset visual SLAM algorithm. The acquired pose data has the characteristic of uniform distribution, so that the pose information of the target object extracted finally is more complete, and great help can be provided for the development, test, verification and improvement of a 6D pose estimation algorithm.

The invention discloses an improved YOLOv8n-pose lightweight human body pose estimation method for solving the problems that an existing pose estimation algorithm is large in calculated amount, low in detection speed and the like, and aims at solving the problems that an existing model is poor in adaptability to small target detection and pose change, high in computing resource consumption and the like. The method comprises the following specific steps: firstly, acquiring a real-time video stream by deploying a camera, and performing down-sampling to extract a series of single-frame images; secondly, a human body detection model is used for obtaining a large number of human body pictures from video frames, joint point labeling is conducted on human bodies, and then the human bodies are divided into a training set, a verification set and a test set according to the proportion; thirdly, inputting the training set into a human body posture model for training to obtain a pre-trained human body posture estimation model; and finally, analyzing an input human body picture through the model, and accurately obtaining a key point position of each human body target. According to the attitude estimation method, a GhostNet module is introduced into a backbone network, so that the model parameter quantity and the calculation cost are remarkably reduced; an OREPA online convolution re-parameterization strategy is adopted, so that the training efficiency is improved; a task alignment dynamic detection head TADDH is designed, interaction between classification and positioning tasks is enhanced, and multi-scale target adaptability and positioning precision are improved. According to the lightweight structure, the model can be efficiently deployed in resource-limited equipment, and meanwhile, the model is excellent in detection of multiple persons, shielding and small targets in complex scenes, and has a wide application prospect.