Motion posture analysis system based on artificial intelligence and image recognition

Abstract

The invention discloses a motion posture analysis system based on artificial intelligence and image recognition. The motion posture analysis system comprises a common RGB camera, a depth camera and a high-speed camera. According to the motion posture analysis system based on artificial intelligence and image recognition, the advantages of a common RGB camera, a depth camera and a high-speed camera are fused through a multi-view camera array, so that two-dimensional and three-dimensional and static and dynamic all-dimensional data acquisition is realized; the problem of data missing caused by limb shielding in single-person movement and personnel overlapping in multi-person movement is effectively avoided, the image quality is optimized through combination of Gaussian filtering and histogram equalization, simultaneous detection of multiple persons is achieved through algorithm fusion of OpenPose and HRNet, the key point positioning precision is improved through high-resolution feature map fusion, and the method is suitable for large-scale popularization and application. The system simultaneously covers static attitude analysis and dynamic attitude analysis, and adopts attitude modeling to quantify an action standard, so that the objectivity of motion attitude analysis is improved, the quantitative standard is clarified, and the analysis efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motion posture analysis, in particular to a motion posture analysis system based on artificial intelligence and image recognition. BACKGROUND

[0002] In the fields of sports training, medical rehabilitation, consumer fitness, etc., accurate analysis of human motion posture is the core prerequisite for optimizing action, reducing injury and improving effect. With the development of technology, motion posture analysis has evolved from subjective experience judgment to semi-automatic quantification, but traditional technical solutions always have many limitations and are difficult to meet the needs of high precision and high efficiency in complex scenarios.

[0003] In the past, whether a sports coach guides the movements of an athlete or a rehabilitation physician assesses the training posture of a patient, the core relies on manual visual observation, which is based on personal experience accumulation and lacks unified quantitative standards, and there are differences in the judgment dimensions and sensitivity of different professionals. To make up for the shortcomings of subjective judgment, traditional methods will assist in using simple tools, which can only measure static posture or a single dimension and cannot cover the whole process of dynamic movement, disrupting the continuity of movement and leading to a disconnection between the analysis results and the actual movement state. For multi-person movement or rapid movement, manual observation is difficult to track multiple targets or capture instantaneous action details at the same time. For example, a coach cannot simultaneously focus on the positioning posture and limb force mode of 11 players in football training. For a table tennis forehand stroke, the naked eye cannot distinguish the coordination between the wrist force timing and the elbow joint extension angle, resulting in analysis that can only stay at the surface level of whether the movement is smooth, reducing the objectivity of motion posture analysis, blurring the quantitative standard and reducing the analysis efficiency. SUMMARY

[0004] The purpose of the present application is to provide a motion posture analysis system based on artificial intelligence and image recognition to solve the problems of reduced objectivity of motion posture analysis, blurred quantitative standard and reduced analysis efficiency in the background art.

[0005] To achieve the above purpose, the present application provides the following technical solution: a motion posture analysis system based on artificial intelligence and image recognition, comprising an ordinary RGB camera, a depth camera, a high-speed camera, a Gaussian filtering algorithm, a target detection algorithm, histogram equalization, OpenPose and HRNet, comprising the following steps: Step 1: image acquisition; Step 2: data preprocessing; Step 3: human body detection and key point extraction; Step 4: posture modeling and analysis; Step five: the result feedback, the motion posture analysis system based on artificial intelligence and image recognition, through the multi-camera array fusion ordinary RGB camera, depth camera, high-speed camera advantages, realize two-dimensional plus three-dimensional, static plus dynamic all-round data acquisition, solve the single camera such as ordinary RGB camera lack of depth information, high-speed camera cannot capture the function limitation of space position, multi-angle shooting combined with triangulation and image matching algorithm, effectively avoid the data missing problem caused by body shielding in single person movement, multiple people movement personnel overlapping, even in complex motion scene still can guarantee the data integrity, anti-interference ability is obviously superior to traditional single camera acquisition scheme, data preprocessing stage, the combination of Gaussian filter and histogram equalization optimizes the image quality, lays the foundation for subsequent analysis, the algorithm fusion of OpenPose and HRNet, not only realizes the multi-person detection, but also improves the key point positioning accuracy through high-resolution feature fusion, the target detection algorithm based on deep learning such as YOLO realizes the rapid segmentation of human region, avoids the complex calculation of traditional image segmentation algorithm, the system covers static posture and dynamic posture analysis at the same time, and adopts posture modeling quantization action standard, breaks through the dimension limitation of traditional analysis technology, can fully reflect the space characteristics and time characteristics of motion posture, improves the objectivity of motion posture analysis, and improves the analysis efficiency.

[0006] Preferably, the image acquisition adopts a plurality of ordinary RGB cameras, depth cameras and high-speed cameras to form a multi-camera array, the ordinary RGB camera converts light signal into electrical signal by using light sensor, and then generates digital image, the depth camera obtains depth information of object by emitting and receiving infrared light through triangulation principle, which can provide three-dimensional space data, the high-speed camera can capture the details of fast motion by shooting moving pictures at high frame rate, through the multi-camera array, multiple cameras shoot from different angles at the same time, and the triangulation and image matching algorithm is used to solve the shielding problem.

[0007] Preferably, the data preprocessing includes three steps of denoising, image enhancement and target detection and segmentation, the Gaussian filter algorithm is used for denoising, the weighted average of each pixel point and its neighborhood pixel point in the image is carried out, the Gaussian noise in the image is effectively removed, the image is smoother, the histogram equalization method is used for image enhancement, the gray histogram of image is adjusted, the gray distribution of image is more uniform, the contrast and definition of image are improved, the quality of moving image in backlight or weak light environment is improved, the target detection and segmentation use the target detection algorithm based on deep learning to quickly process the input image, identify and segment the human target, exclude background interference, and provide pure target image for subsequent posture analysis.

[0008] Preferably, the human body detection and key point extraction use the combination of the OpenPose and the HRNet (High-Resolution Network) algorithms. The OpenPose algorithm detects the key points of multiple human bodies simultaneously through the Parts Affinity Fields (PAFs) algorithm. First, it extracts features from the input image, and then it predicts the positions of each key point and the relationships between the key points through a specific network structure. It can accurately identify the key points of each human body in a multi-person motion scene. The HRNet algorithm keeps high-resolution feature maps and continuously fuses feature information of different resolutions in the entire network structure, enabling the model to more accurately capture the positions of human body key points and improve the positioning accuracy of key points.

[0009] Preferably, after the system inputs the pre-processed image, the algorithm first extracts features from the image and converts the image into a series of feature vectors. Then, it analyzes these feature vectors through the trained model to predict the coordinate positions of each key point of the human body in the image.

[0010] Preferably, the pose modeling includes static pose analysis and dynamic pose analysis. The static pose analysis includes joint angle calculation, body center of gravity determination, and limb proportion and alignment evaluation. The joint angle calculation calculates the joint angle using trigonometric functions based on the coordinates of adjacent joint nodes. The body center of gravity determination calculates the position of the body center of gravity by weighted average of the coordinates of each key node. The limb proportion and alignment evaluation compares the coordinate differences of symmetric joint nodes on the left and right sides to determine whether the limbs are symmetrical.

[0011] Preferably, the dynamic pose analysis includes motion trajectory tracking, motion speed and acceleration calculation, and motion continuity and rhythm analysis. The motion trajectory tracking traces the coordinates of the key points in consecutive image frames to draw the motion trajectory of the limbs. The motion speed and acceleration calculation calculates the motion speed and acceleration of the key points based on the coordinate changes and time intervals of the key points in adjacent image frames. The motion continuity and rhythm analysis analyzes the smoothness of the position changes of the key points between adjacent frames and the time intervals of the motion to determine the continuity and rhythm of the motion.

[0012] Preferably, the result feedback involves the system organizing the analyzed pose data to generate a detailed data report, superimposing the visual annotations of the pose analysis results on the original motion video, and superimposing and comparing the user's motion pose data with the pre-stored standard templates. The system monitors and judges the motion pose data in real time based on the pre-set risk threshold.

[0013] Compared with the prior art, the beneficial effects of the present application are that: the motion posture analysis system based on artificial intelligence and image recognition combines the advantages of ordinary RGB cameras, depth cameras and high-speed cameras through a multi-camera array, realizes two-dimensional plus three-dimensional, static plus dynamic all-around data acquisition, solves the functional limitations of single cameras such as ordinary RGB cameras lacking depth information and high-speed cameras being unable to capture spatial positions, multi-angle shooting combined with triangulation and image matching algorithms effectively avoids the problem of data loss caused by body occlusion in single-person motion and personnel overlap in multi-person motion, even in complex motion scenes, the data integrity can still be guaranteed, the anti-interference ability is significantly better than that of the traditional single-camera acquisition scheme, in the data preprocessing stage, the combination of Gaussian filtering and histogram equalization optimizes the image quality, laying a foundation for subsequent analysis, the algorithm fusion of OpenPose and HRNet realizes multi-person detection at the same time and improves the key point positioning accuracy through high-resolution feature map fusion, the target detection algorithm based on deep learning such as YOLO realizes rapid segmentation of human regions, avoiding the complex calculation of traditional image segmentation algorithms, the system simultaneously covers static posture and dynamic posture analysis, and adopts posture modeling and quantized motion standards, breaking through the dimensional limitations of traditional analysis techniques, can comprehensively reflect the spatial and temporal characteristics of motion postures, improving the objectivity of motion posture analysis, and clarifying the quantitative standards and improving the analysis efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 It is a schematic diagram of the system structure of the present application. DETAILED DESCRIPTION

[0015] 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 part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] Please refer to Figure 1 The present application provides a technical solution: a motion posture analysis system based on artificial intelligence and image recognition, comprising an ordinary RGB camera, a depth camera, a high-speed camera, a Gaussian filtering algorithm, a target detection algorithm, histogram equalization, OpenPose and HRNet: comprising the following steps: Step one: image acquisition; Step two: data preprocessing; Step three: human body detection and key point extraction; Step four: posture modeling and analysis; Step five: the result feedback, the motion posture analysis system based on artificial intelligence and image recognition, through the multi-camera array fusion ordinary RGB camera, depth camera, high-speed camera advantages, realize two-dimensional plus three-dimensional, static plus dynamic all-round data acquisition, solve the single camera such as ordinary RGB camera lack of depth information, high-speed camera cannot capture the function limitation of space position, multi-angle shooting combined with triangulation and image matching algorithm, effectively avoid the data missing problem caused by body shielding in single person movement, multiple people movement personnel overlapping, even in complex motion scene still can guarantee the data integrity, anti-interference ability is obviously superior to traditional single camera acquisition scheme, data preprocessing stage, the combination of Gaussian filter and histogram equalization optimizes the image quality, lays the foundation for subsequent analysis, the algorithm fusion of OpenPose and HRNet, not only realizes the multi-person detection, but also improves the key point positioning accuracy through high-resolution feature fusion, the target detection algorithm based on deep learning such as YOLO realizes the rapid segmentation of human region, avoids the complex calculation of traditional image segmentation algorithm, the system covers static posture and dynamic posture analysis at the same time, and adopts posture modeling quantization action standard, breaks through the dimension limitation of traditional analysis technology, can fully reflect the space characteristics and time characteristics of motion posture, improves the objectivity of motion posture analysis, accurately quantizes the standard and improves the analysis efficiency.

[0017] Preferably, the image acquisition adopts a plurality of ordinary RGB cameras, depth cameras and high-speed cameras to form a multi-camera array, the ordinary RGB camera converts light signals into electrical signals by using a light sensor, and then generates a digital image, the depth camera obtains depth information of an object by emitting and receiving infrared light through the principle of triangulation, which can provide three-dimensional space data, the high-speed camera can capture the details of rapid motion by shooting moving pictures at a high frame rate, and the multi-camera array can solve the shielding problem by simultaneously shooting from different angles through multiple cameras and using triangulation and image matching algorithms.

[0018] Preferably, the data preprocessing includes three steps of denoising, image enhancement and target detection and segmentation, the denoising uses the Gaussian filter algorithm, which effectively removes Gaussian noise in the image by weighted average of each pixel point and its neighborhood pixel points, so that the image is smoother, the image enhancement adopts the histogram equalization method, which adjusts the gray histogram of the image to make the gray distribution of the image more uniform, improves the contrast and clarity of the image, and improves the quality of the moving image in backlight or weak light environment, the target detection and segmentation uses the target detection algorithm based on deep learning to quickly process the input image, identifies and segments the human target, and excludes background interference to provide pure target image for subsequent posture analysis.

[0019] Preferably, the human body detection and key point extraction use the combination of the OpenPose and the HRNet (High-Resolution Network) algorithms. The OpenPose algorithm detects the key points of multiple human bodies simultaneously through the Parts Affinity Fields (PAFs) algorithm. First, it extracts features from the input image, and then it predicts the positions of each key point and the relationships between the key points through a specific network structure. It can accurately identify the key points of each human body in a multi-person motion scene. The HRNet (High-Resolution Network) algorithm continuously fuses feature information of different resolutions in the entire network structure by maintaining high-resolution feature maps, enabling the model to more accurately capture the positions of human body key points and improve key point positioning accuracy.

[0020] Preferably, after the system inputs the pre-processed image, the algorithm first extracts features from the image, converts the image into a series of feature vectors, and then analyzes these feature vectors through the trained model to predict the coordinate positions of each key point of the human body in the image.

[0021] Preferably, the pose modeling includes static pose analysis and dynamic pose analysis. The static pose analysis includes joint angle calculation, body center of gravity determination, and limb proportion and alignment evaluation. The joint angle calculation calculates the joint angle using trigonometric functions based on the coordinates of adjacent joint nodes. The body center of gravity determination calculates the position of the body center of gravity by weighted average calculation of the coordinates of each key node of the human body. The limb proportion and alignment evaluation compares the coordinate differences of the symmetric joint nodes on the left and right sides to determine whether the limbs are symmetrical.

[0022] Preferably, the dynamic pose analysis includes motion trajectory tracking, motion speed and acceleration calculation, and motion continuity and rhythm analysis. The motion trajectory tracking traces the coordinates of the key points in consecutive image frames to draw the motion trajectory of the limbs. The motion speed and acceleration calculation calculates the motion speed and acceleration of the key points based on the coordinate changes of the key points in adjacent image frames and the time interval. The motion continuity and rhythm analysis analyzes the smoothness of the position changes of the key points between adjacent frames and the time interval of the motion to determine the continuity and rhythm of the motion.

[0023] Preferably, the result feedback is that the system organizes the analyzed pose data to generate a detailed data report, superimposes the visual annotations of the pose analysis results on the original motion video, and superimposes and compares the user's motion pose data with the pre-stored standard templates. The system monitors and judges the motion pose data in real time based on the pre-set risk threshold.

[0024] Finally, it should be noted that the above is only to illustrate the technical solutions of the present application, and is not a limitation on the scope of protection of the present application, and simple modifications or equivalent replacements of the technical solutions of the present application by those skilled in the art do not deviate from the essence and scope of the technical solutions of the present application.

Claims

1. A motion posture analysis system based on artificial intelligence and image recognition, comprising a common RGB camera, a depth camera, a high-speed camera, a Gaussian filtering algorithm, a target detection algorithm, histogram equalization, OpenPose, and HRNet, characterized in that: It comprises the following steps: Step one: image acquisition; Step two: data preprocessing; Step three: human body detection and key point extraction; Step four: pose modeling and analysis; Step five: result feedback. 2.The sports posture analysis system based on artificial intelligence and image recognition according to claim 1, characterized in that: The image acquisition adopts a multi-view camera array composed of multiple ordinary RGB cameras, depth cameras and high-speed cameras. The ordinary RGB camera converts light signals into electrical signals using a light sensor, and then generates a digital image. The depth camera obtains depth information of an object by emitting and receiving infrared light using the triangulation principle, and can provide three-dimensional spatial data. The high-speed camera shoots moving pictures at a high frame rate, and can capture details of fast motion. Through the multi-view camera array, multiple cameras shoot from different angles at the same time, and use triangulation and image matching algorithms to solve the occlusion problem.

3. The motion gesture analysis system based on artificial intelligence and image recognition according to claim 2, characterized in that: The data preprocessing includes three steps of denoising, image enhancement and target detection and segmentation. The denoising uses the Gaussian filter algorithm to effectively remove Gaussian noise in the image by weighted averaging of each pixel point and its neighborhood pixel points in the image, making the image smoother. The image enhancement uses the histogram equalization method to make the gray scale distribution of the image more uniform, improve the contrast and clarity of the image, and improve the quality of the moving image in backlight or weak light environment. The target detection and segmentation use the target detection algorithm based on deep learning to quickly process the input image, identify and segment the human body target, and exclude background interference to provide pure target image for subsequent pose analysis.

4. The motion posture analysis system based on artificial intelligence and image recognition according to claim 3, characterized in that: The human body detection and key point extraction uses the combination of the OpenPose and HRNet (High Resolution Network) algorithms. The OpenPose algorithm detects the key points of multiple human bodies at the same time through the Parts Affinity Fields (PAF) field. It first extracts features from the input image, and then predicts the position of each key point and the relationship between the key points through a specific network structure. It can accurately identify the key points of each human body in a multi-person motion scene. The HRNet algorithm constantly fuses feature information of different resolutions in the entire network structure by maintaining high-resolution feature maps, so that the model can more accurately capture the position of human body key points and improve the key point positioning accuracy.

5. The motion gesture analysis system based on artificial intelligence and image recognition according to claim 4, characterized in that: After the system inputs the preprocessed image, the algorithm first extracts features from the image, converts the image into a series of feature vectors, and then analyzes these feature vectors through the trained model to predict the coordinate position of each key point of the human body in the image.

6. The motion gesture analysis system based on artificial intelligence and image recognition according to claim 5, characterized in that: The posture modeling includes static posture analysis and dynamic posture analysis, the static posture analysis includes joint angle calculation, body center of gravity determination and limb proportion and alignment evaluation, the joint angle calculation is to calculate the joint angle according to the coordinates of the extracted adjacent joint nodes by using trigonometric function, the body center of gravity determination is to obtain the position of the body center of gravity by weighted average calculation on the coordinates of each key node of the human body, and the limb proportion and alignment evaluation is to compare the coordinate difference of the symmetrical joint nodes on the left and right sides to determine whether the limbs are symmetrical.

7. The motion gesture analysis system based on artificial intelligence and image recognition according to claim 6, characterized in that: The dynamic posture analysis includes motion trajectory tracking, motion speed and acceleration calculation and motion continuity and rhythm analysis, the motion trajectory tracking refers to drawing the motion trajectory of the limbs by tracking the coordinate change of the key points in the continuous image frames, the motion speed and acceleration calculation refers to calculating the motion speed and acceleration of the key points according to the coordinate change amount and time interval of the key points in the adjacent image frames, and the motion continuity and rhythm analysis refers to analyzing the smoothness of the position change of the key points between the adjacent frames and the time interval of the motion to determine the continuity and rhythm of the motion.

8. The motion gesture analysis system based on artificial intelligence and image recognition according to claim 7, characterized in that: The result feedback is that the system arranges the posture data obtained by analysis, generates a detailed data report, superimposes the visualized annotation of the posture analysis result on the original motion video, and superimposes and compares the motion posture data of the user with the pre-stored standard template, and the system monitors and judges the motion posture data in real time according to the preset risk threshold.

Images