Robot action analysis method and system based on image recognition

Abstract

The invention discloses a robot action analysis method and system based on image recognition, relates to the technical field of image processing, and comprises the steps of screening image frames, removing a plurality of approximately repeated redundant images, and reducing errors of subsequent image analysis. And carrying out foreground region and background region division on the screened image, and helping to divide the foreground region and the background region through stable image features corresponding to robot actions. Analyzing the matching information of the first key point, adding a second key point, forming a motion track based on the first key point and the second key point, supplementing and assisting the first key point by adding the second key point, thereby obtaining a high-quality motion track, extracting motion features through the motion track, and improving the motion accuracy. And the motion completion degree of the robot is evaluated. A robot action behavior evaluation standard system is unified, the accuracy and adaptability of robot action behavior analysis are improved, and stable operation of robot action behaviors is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a robot motion analysis method and system based on image recognition. Background Art

[0002] With the rapid development of artificial intelligence technology, robot motion analysis has shown great potential in many fields such as industrial automation, intelligent manufacturing, intelligent security, medical image analysis, and autonomous driving. However, the traditional human eye observation of robot motion has disadvantages such as strong subjectivity, easy fatigue, and difficulty in capturing high-speed or small movements. Therefore, the robot motion analysis solution based on image recognition came into being. It uses deep learning algorithms and computer vision technology to objectively, accurately, and in real time analyze and evaluate robot motion, providing strong support for the optimal design and efficient operation of robots. The development of this technology is of great significance to promoting the intelligent upgrading of robot technology and enhancing the robot's autonomous navigation, collaborative operation, and ability to adapt to complex environments.

[0003] In the prior art, the human eye is used to observe and evaluate the robot's motion behavior, but the motion capture is weak, and different people have different evaluation standards, resulting in inaccurate and inconsistent evaluation of the robot's motion behavior, making the robot's motion behavior analysis poorly accurate and low in adaptability.

[0004] Therefore, how to improve the accuracy and adaptability of robot motion behavior analysis is a technical problem that needs to be solved. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of poor accuracy and low adaptability of robot motion behavior analysis in the prior art, and to propose a robot motion analysis method based on image recognition, which includes: Obtain robot action videos within a period of time, convert the robot action videos into temporally continuous image frames, and filter the image frames; The filtered image is divided into a foreground area and a background area, and the first key point in the foreground area is detected, the first key point is matched, and the first key point matching information is obtained; Analyze the matching information of the first key point to add the second key point, and form a motion trajectory based on the first key point and the second key point; Motion features are extracted through motion trajectories, and the robot's motion completion is evaluated.

[0006] In some embodiments of the present application, the image frames are screened, including: A frame interval is set according to the frame rate of the robot action video, the image frames are initially screened by the frame interval, and the frames are arranged in chronological order to obtain a primary image frame sequence; Compare each pixel of the image between adjacent frames in the primary image frame sequence to obtain a pixel difference value of each pixel, integrate these pixel differences to generate a difference image, and calculate the total pixel difference value; The brightness, contrast and structure of the images between adjacent frames in the primary image frame sequence are compared to obtain the SSIM value; Converting the images between adjacent frames in the preliminary image frame sequence into a single histogram of a grayscale histogram or a color histogram, and calculating a joint histogram of two images in adjacent frames, and calculating a mutual information value through the single histogram and the joint histogram; The total pixel difference, SSIM value, mutual information value and their respective preset thresholds are compared to obtain the difference between the three, and the frame difference index is determined by combining the difference between the total pixel difference, SSIM value and mutual information value; The preliminary selected image frame sequence is traversed through the frame difference index, and the preliminary selected image frame sequence is screened for the second time.

[0007] In some embodiments of the present application, the screened image is divided into a foreground area and a background area, including: Analyze the robot action video to define the foreground object and background object, set the parameters of the Canny edge detector according to the foreground object and background object, and extract the fuzzy edge between the foreground object and background object in the image; Determine the stable image features in the image, integrate the stable image features to generate a stability index, adjust the blurred edge by comparing the difference in the stability index on both sides of the blurred edge, and obtain the foreground-background boundary, so as to divide the foreground area and the background area.

[0008] In some embodiments of the present application, the stability index is generated by integrating the stable image features, including: Normalize stable image features; ; in, For stability indicators, is the number of stable image features in the figure, For the The combined weights of stable image features, For the The size of the stable image features, , They are The maximum and minimum values ​​in is a preset constant.

[0009] In some embodiments of the present application, the first key point in the foreground area is detected, the first key point is matched, and the first key point matching information is obtained, including: Determine a first key point according to the foreground object, identify the first key point in the foreground area by using a preset key point detection algorithm, mark the first key point in each image, match the position and feature descriptor of the first key point in multiple images to obtain a first matching degree and a second matching degree respectively, determine a third matching degree by combining the first matching degree and the second matching degree, divide the first key point into two categories: matching key points and non-matching key points by using the third matching degree, and use the first matching degree, second matching degree, third matching degree and category of the first key point as the first key point matching information.

[0010] In some embodiments of the present application, the third matching degree is determined by combining the first matching degree and the second matching degree, including: ; in, For the The third matching degree of the first key point, , are the combined weights of the position and feature descriptors respectively, is the number of images - 1, For the The first key point The first match, is the first matching threshold, For the The first key point The second matching degree, is the second matching threshold, , Respectively The adjustment coefficients of the first matching degrees and the second matching degrees of the first key points are determined by the proportion of the first matching degrees less than the first matching degree threshold and the proportion of the second matching degrees less than the second matching degree threshold.

[0011] In some embodiments of the present application, analyzing the first key point matching information to add the second key point includes: A number of backup key points are randomly selected around the non-matching key point, the third matching degree of each backup key point is calculated, and the backup key point whose third matching degree exceeds a preset threshold is used as the second key point.

[0012] In some embodiments of the present application, motion features are extracted through motion trajectories, and the robot's motion completion is evaluated, including: Extract motion features and trajectory smoothness based on the motion trajectory, construct a curve of each motion feature changing over time, mark the standard baseline on the curve, and count the area of ​​the curve part outside the standard baseline as the additional area; Generate robot motion completion based on the additional area and trajectory smoothness of each motion feature; ; in, is the robot action completion degree, is the number of motion features, For the The evaluation weight of each motion feature, For the The additional area of ​​the motion feature, is the trajectory smoothness, is a preset constant.

[0013] Correspondingly, the present application also provides a robot motion analysis system based on image recognition, comprising: A screening module is used to obtain robot action videos within a period of time, convert the robot action videos into temporally continuous image frames, and screen the image frames; A detection module is used to divide the filtered image into a foreground area and a background area, detect a first key point in the foreground area, match the first key point, and obtain first key point matching information; An adding module, used for analyzing the matching information of the first key point to add the second key point, and forming a motion trajectory based on the first key point and the second key point; The evaluation module is used to extract motion features through motion trajectories and evaluate the robot's motion completion.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. Filter the image frames and remove multiple redundant images that are close to being repeated to reduce the error of subsequent image analysis. Divide the filtered images into foreground and background areas, and use the stable image features corresponding to the robot's movements to help divide the foreground and background areas, providing a basis for subsequent image analysis.

[0015] 2. Analyze the matching information of the first key point and add the second key point. Based on the first and second key points, form a motion trajectory. By adding the second key point to supplement and assist the first key point, a high-quality motion trajectory is obtained. The motion features are extracted through the motion trajectory, and the robot's action completion is evaluated. The robot action behavior evaluation standard system is unified, the accuracy and adaptability of the robot action behavior analysis are improved, and the stable operation of the robot action behavior is guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the flow of the robot motion analysis method based on image recognition proposed by the present invention; Figure 2 This is a schematic diagram of the structure of the robot motion analysis system based on image recognition proposed by the present invention. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0018] Reference Figure 1 , a robot motion analysis method based on image recognition, comprising the following steps: Step S101, obtaining a robot action video within a period of time, converting the robot action video into temporally continuous image frames, and screening the image frames.

[0019] In this embodiment, the robot action video is converted into temporally continuous image frames. This step is usually implemented using video processing software or programming libraries (such as OpenCV). By extracting frame by frame, we can get a series of static images representing the robot's action process. The frame rate of the video is how many images per second, which may result in a large number of repeated or very similar images. If each image is processed, it will cause the computing task to be too heavy, and may introduce some image errors, making the image analysis effect poor.

[0020] In some embodiments of the present application, the image frames are screened, including: A frame interval is set according to the frame rate of the robot action video, the image frames are initially screened by the frame interval, and the frames are arranged in chronological order to obtain a primary image frame sequence; Compare each pixel of the image between adjacent frames in the primary image frame sequence to obtain a pixel difference value of each pixel, integrate these pixel differences to generate a difference image, and calculate the total pixel difference value; The brightness, contrast and structure of the images between adjacent frames in the primary image frame sequence are compared to obtain the SSIM value; Converting the images between adjacent frames in the preliminary image frame sequence into a single histogram of a grayscale histogram or a color histogram, and calculating a joint histogram of two images in adjacent frames, and calculating a mutual information value through the single histogram and the joint histogram; The total pixel difference, SSIM value, mutual information value and their respective preset thresholds are compared to obtain the difference between the three, and the frame difference index is determined by combining the difference between the total pixel difference, SSIM value and mutual information value; The preliminary selected image frame sequence is traversed through the frame difference index, and the preliminary selected image frame sequence is screened for the second time.

[0021] In this embodiment, the image frames are initially screened by frame intervals, and the RGB value or grayscale value of each pixel in adjacent frames is directly compared. The difference between each pixel between the two frames is calculated to obtain a difference image. The total pixel difference is calculated. If most of the pixel values ​​in the difference image are close to 0, then the two frames are similar. In a digital image, each pixel has one or more numerical values ​​associated with it, which usually represent the color (such as RGB value) or grayscale of the pixel. Pixel values ​​are the basic units of image information, and they together constitute the visual content of the image. SSIM (structural similarity) is an algorithm that considers the similarity of image brightness, contrast and structure. It not only compares pixel values, but also considers the relative position and relationship between pixels. Therefore, SSIM can more accurately reflect the structural similarity of two images. The closer this value is to 1, the more similar the two frames are in structure. Mutual information is a method to measure the statistical dependence between two images. If the contents of two images are similar, then the mutual information between them will be high. Mutual information can capture the complex relationship between images, not just a simple comparison of pixel values. Calculate the grayscale histogram or color histogram of the two frames of images respectively. Calculate the joint histogram of the two frames of images, which represents the joint probability distribution of the two frames of images on the corresponding pixel values. Calculate the mutual information value between the two frames of images based on the joint histogram and the single histogram. The larger this value is, the higher the similarity in content between the two frames of images is.

[0022] It is understandable that the difference image focuses on the pixel-level differences and is very sensitive to subtle changes. Structural similarity focuses more on the overall structure and layout of the image and can capture the structural similarity between images. Mutual information can capture the complex relationships between images, including nonlinear and asymmetric relationships. Combining the difference image, structural similarity, and mutual information can comprehensively utilize their advantages to improve the accuracy and robustness of frame difference calculation.

[0023] Step S102, dividing the filtered image into a foreground area and a background area, detecting a first key point in the foreground area, matching the first key point, and obtaining first key point matching information.

[0024] In this embodiment, foreground-background separation refers to the process of separating the foreground area (usually the robot body area) from the background area (such as the environment and other non-target objects) in the image.

[0025] In some embodiments of the present application, the screened image is divided into a foreground area and a background area, including: Analyze the robot action video to define the foreground object and background object, set the parameters of the Canny edge detector according to the foreground object and background object, and extract the fuzzy edge between the foreground object and background object in the image; Determine the stable image features in the image, integrate the stable image features to generate a stability index, adjust the blurred edge by comparing the difference in the stability index on both sides of the blurred edge, and obtain the foreground-background boundary, so as to divide the foreground area and the background area.

[0026] In this embodiment, we need to have a certain understanding of the content of the foreground and background. For example, if the foreground is a robot and the background is a static indoor environment, we can expect the edges of the robot to be sharper and more obvious, while the edges of the background may be smoother or blurred. According to the characteristics of the foreground object and the background object, set the parameters of the Canny edge detector, such as high and low thresholds. These parameters will affect the sensitivity and accuracy of edge detection. Apply the Canny edge detector to the image to extract preliminary edges between the foreground object and the background object. These edges may not be accurate because the robot's motion behavior causes the image to change dynamically. After the first separation, we get a preliminary foreground and background separation result. However, this result may contain some misclassified pixels or areas. In order to further improve the accuracy of the separation, we can use stable features that are insensitive to motion changes for secondary separation. These stable features can be inherent properties of the robot, such as color, texture, shape, etc. (these features are less affected by the robot's motion behavior). These features are usually highly stable and consistent in the image, so they can be used to finely divide the foreground and background. Using stable features, we can fine-tune the boundaries of the first separation. For example, by comparing the differences in stable features between the foreground and background, we can more accurately determine the boundary between the foreground and background and adjust the separation boundary accordingly. The first separation is based on fast features such as edge information or depth information, which can roughly separate the foreground and background. However, these features may be affected by factors such as noise and lighting changes, resulting in inaccurate separation results. The second separation uses stable features for fine separation, which can correct errors in the first separation and improve the accuracy of separation.

[0027] In some embodiments of the present application, the stability index is generated by integrating the stable image features, including: Normalize stable image features; ; in, For stability indicators, is the number of stable image features in the figure, For the The combined weights of stable image features, For the The size of the stable image features, , They are The maximum and minimum values ​​in is a preset constant.

[0028] In this embodiment, stable image features such as color, texture, shape, etc. are quantized and standardized. express The average value of the stable image features is corrected, and the average value is taken as the stability index.

[0029] In some embodiments of the present application, the first key point in the foreground area is detected, the first key point is matched, and the first key point matching information is obtained, including: Determine a first key point according to the foreground object, identify the first key point in the foreground area by using a preset key point detection algorithm, mark the first key point in each image, match the position and feature descriptor of the first key point in multiple images to obtain a first matching degree and a second matching degree respectively, determine a third matching degree by combining the first matching degree and the second matching degree, divide the first key point into two categories: matching key points and non-matching key points by using the third matching degree, and use the first matching degree, second matching degree, third matching degree and category of the first key point as the first key point matching information.

[0030] In this embodiment, the first key points are some key points on the robot that can help with motion recognition, such as joints: such as shoulder joints, elbow joints, wrist joints, hip joints, knee joints and ankle joints. These joints are the hubs for the connection and movement of the robot's body, and their movement trajectories and speeds can directly reflect the robot's movements. End effectors: such as the robot's fingers, grippers or tools. These parts are the main parts for the robot to interact with the environment, and their position and posture are crucial to understanding the robot's operational tasks. Body center of gravity: Although it is not a specific physical point, the position of the robot's body center of gravity is of great significance for analyzing its stability and balance.

[0031] In this embodiment, a suitable key point detection algorithm is selected according to the characteristics of the robot and the complexity of the movement. OpenPose and HRNet are both excellent choices, which can detect multiple key points on the human body or robot. The detected key points are matched between consecutive frames. This is usually achieved by comparing the position, features, etc. of the key points. Ensure that the same key point is correctly identified in consecutive frames. For each detected key point, its feature descriptor is calculated. A feature descriptor is a vector that describes the image information around the key point, such as gradient direction, texture, etc.

[0032] In this embodiment, the position and feature descriptors are matched to obtain the first matching degree (position) and the second matching degree (feature descriptor), respectively. Generally speaking, the positions of the key points in the two images will differ within a certain range, but not particularly large (the time interval is short). If the positions of the key points that should be matched in the two images differ greatly, it means that the positions do not match. Although the positions are close, the differences between the feature descriptors are large, indicating that these key points may not be true matching pairs. The first matching degree (position) and the second matching degree (feature descriptor) are combined to determine the comprehensive matching degree, i.e., the third matching degree.

[0033] It should be noted that the first matching degree and the second matching degree here are both matching conditions between two images in adjacent frames.

[0034] In some embodiments of the present application, the third matching degree is determined by combining the first matching degree and the second matching degree, including: ; in, For the The third matching degree of the first key point, , are the combined weights of the position and feature descriptors respectively, is the number of images - 1, For the The first key point The first match, is the first matching threshold, For the The first key point The second matching degree, is the second matching threshold, , Respectively The adjustment coefficients of the first matching degrees and the second matching degrees of the first key points are determined by the proportion of the first matching degrees less than the first matching degree threshold and the proportion of the second matching degrees less than the second matching degree threshold.

[0035] In this embodiment, = is the number of images - 1, which refers to the total number of matching degrees (first matching degree or second matching degree). For example, if there are 100 images and the matching degree between two images in adjacent frames is calculated, there will be 99 matching degrees. The adjustment coefficient is determined by the proportion of the first matching degree less than the first matching degree threshold and the proportion of the second matching degree less than the second matching degree threshold. The proportion here is the proportion of unqualified matching degrees (describing the proportion of matching degrees that do not meet the requirements of a key point in the entire image sequence), to adjust the sum of the difference between the matching degree and the threshold.

[0036] Step S103: Analyze the first key point matching information to add the second key point, and form a motion trajectory based on the first key point and the second key point.

[0037] In this embodiment, those first key points that do not meet the requirements are recorded as non-matching key points, and key points are supplemented for these matching degrees, and based on the matching information of the first key point and the second key point, they are connected to form a motion trajectory. This can be achieved by tracking the position changes of the key points.

[0038] In some embodiments of the present application, analyzing the first key point matching information to add the second key point includes: A number of backup key points are randomly selected around the non-matching key point, the third matching degree of each backup key point is calculated, and the backup key point whose third matching degree exceeds a preset threshold is used as the second key point.

[0039] In this embodiment, other potential key points with stable positions are searched near the non-matching key points (such as in a circular area with a certain radius), and the search is performed among the several spare key points according to the third matching degree.

[0040] Step S104, extracting motion features through the motion trajectory and evaluating the robot's motion completion degree.

[0041] In this embodiment, the motion features include speed, acceleration, posture, etc. The extracted motion features are compared with the evaluation criteria pre-set for the robot's actions. The evaluation criteria should be formulated according to the specific tasks and action requirements of the robot (specific range of motion features).

[0042] In some embodiments of the present application, motion features are extracted through motion trajectories, and the robot's motion completion is evaluated, including: Extract motion features and trajectory smoothness based on the motion trajectory, construct a curve of each motion feature changing over time, mark the standard baseline on the curve, and count the area of ​​the curve part outside the standard baseline as the additional area; Generate robot motion completion based on the additional area and trajectory smoothness of each motion feature; ; in, is the robot action completion degree, is the number of motion features, For the The evaluation weight of each motion feature, For the The additional area of ​​the motion feature, is the trajectory smoothness, is a preset constant.

[0043] In this embodiment, the specific range of the motion feature is mapped to obtain a standard baseline on the curve. If the curve is within the standard baseline, it means that the motion feature is qualified. Otherwise, it is unqualified. The additional area refers to the area between the curve portion outside the standard baseline and the standard baseline. Represents the correction of the sum of the additional areas by trajectory smoothness. Trajectory smoothness refers to the degree to which the position, velocity, and acceleration of the robot change continuously over time during its motion, and the rate of change remains smooth. A smooth trajectory (which describes the continuity of the robot's movements) can reduce vibration and impact during the robot's movement, avoid damage to the robot's own structure and the surrounding environment, and also help maintain the robot's balance and improve the stability and accuracy of its motion.

[0044] Compared with the prior art, the present invention has the following beneficial effects: 1. Filter the image frames and remove multiple redundant images that are close to being repeated to reduce the error of subsequent image analysis. Divide the filtered images into foreground and background areas, and use the stable image features corresponding to the robot's movements to help divide the foreground and background areas, providing a basis for subsequent image analysis.

[0045] 2. Analyze the matching information of the first key point and add the second key point. Based on the first and second key points, form a motion trajectory. By adding the second key point to supplement and assist the first key point, a high-quality motion trajectory is obtained. The motion features are extracted through the motion trajectory, and the robot's action completion is evaluated. The robot action behavior evaluation standard system is unified, the accuracy and adaptability of the robot action behavior analysis are improved, and the stable operation of the robot action behavior is guaranteed.

[0046] Correspondingly, the present application also provides a robot motion analysis system based on image recognition, such as Figure 2 As shown, including, A screening module is used to obtain robot action videos within a period of time, convert the robot action videos into temporally continuous image frames, and screen the image frames; A detection module is used to divide the filtered image into a foreground area and a background area, detect a first key point in the foreground area, match the first key point, and obtain first key point matching information; An adding module, used for analyzing the matching information of the first key point to add the second key point, and forming a motion trajectory based on the first key point and the second key point; The evaluation module is used to extract motion features through motion trajectories and evaluate the robot's motion completion.

[0047] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present invention can be implemented by hardware, or by software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present invention.

[0048] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily required for implementing the present invention.

[0049] Those skilled in the art will appreciate that the modules in the system in the implementation scenario can be distributed in the system of the implementation scenario according to the implementation scenario description, or can be changed accordingly and located in one or more systems different from the implementation scenario. The modules in the above implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0050] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A robot motion analysis method based on image recognition, characterized in that: include, Obtain robot action videos within a period of time, convert the robot action videos into temporally continuous image frames, and filter the image frames; The filtered image is divided into a foreground area and a background area, and the first key point in the foreground area is detected, the first key point is matched, and the first key point matching information is obtained; Analyze the matching information of the first key point to add the second key point, and form a motion trajectory based on the first key point and the second key point; Motion features are extracted through motion trajectories, and the robot's motion completion is evaluated.

2. The robot motion analysis method based on image recognition according to claim 1, characterized in that: Filter the image frames, including: A frame interval is set according to the frame rate of the robot action video, the image frames are initially screened by the frame interval, and the frames are arranged in chronological order to obtain a primary image frame sequence; Compare each pixel of the image between adjacent frames in the primary image frame sequence to obtain a pixel difference value of each pixel, integrate these pixel differences to generate a difference image, and calculate the total pixel difference value; The brightness, contrast and structure of the images between adjacent frames in the primary image frame sequence are compared to obtain the SSIM value; Converting the images between adjacent frames in the preliminary image frame sequence into a single histogram of a grayscale histogram or a color histogram, and calculating a joint histogram of two images in adjacent frames, and calculating a mutual information value through the single histogram and the joint histogram; The total pixel difference, SSIM value, mutual information value and their respective preset thresholds are compared to obtain the difference between the three, and the frame difference index is determined by combining the difference between the total pixel difference, SSIM value and mutual information value; The preliminary selected image frame sequence is traversed through the frame difference index, and the preliminary selected image frame sequence is screened for the second time.

3. The robot motion analysis method based on image recognition according to claim 1, characterized in that: The filtered image is divided into foreground and background areas, including: Analyze the robot action video to define the foreground object and background object, set the parameters of the Canny edge detector according to the foreground object and background object, and extract the fuzzy edge between the foreground object and background object in the image; Determine the stable image features in the image, integrate the stable image features to generate a stability index, adjust the blurred edge by comparing the difference in the stability index on both sides of the blurred edge, and obtain the foreground-background boundary, so as to divide the foreground area and the background area.

4. The robot motion analysis method based on image recognition according to claim 3 is characterized in that: Integrate stable image features to generate stable indicators, including, Normalize stable image features; ； in, For stability indicators, is the number of stable image features in the figure, For the The combined weights of stable image features, For the The size of the stable image features, , They are The maximum and minimum values ​​in is a preset constant.

5. The robot motion analysis method based on image recognition according to claim 3 is characterized in that: And detect the first key point in the foreground area, match the first key point, and obtain the first key point matching information, include, Determine a first key point according to the foreground object, identify the first key point in the foreground area by using a preset key point detection algorithm, mark the first key point in each image, match the position and feature descriptor of the first key point in multiple images to obtain a first matching degree and a second matching degree respectively, determine a third matching degree by combining the first matching degree and the second matching degree, divide the first key point into two categories: matching key points and non-matching key points by using the third matching degree, and use the first matching degree, second matching degree, third matching degree and category of the first key point as the first key point matching information.

6. The robot motion analysis method based on image recognition according to claim 5, characterized in that: Determining a third matching degree by combining the first matching degree and the second matching degree includes, ； in, For the The third matching degree of the first key point, , are the combined weights of the position and feature descriptors respectively, is the number of images - 1, For the The first key point The first match, is the first matching threshold, For the The first key point The second matching degree, is the second matching threshold, , Respectively The adjustment coefficients of the first matching degrees and the second matching degrees of the first key points are determined by the proportion of the first matching degrees less than the first matching degree threshold and the proportion of the second matching degrees less than the second matching degree threshold.

7. The robot motion analysis method based on image recognition according to claim 5, characterized in that: Analyze the matching information of the first key point and add the second key point, including: A number of backup key points are randomly selected around the non-matching key point, the third matching degree of each backup key point is calculated, and the backup key point whose third matching degree exceeds a preset threshold is used as the second key point.

8. The robot motion analysis method based on image recognition according to claim 1, characterized in that: Extract motion features through motion trajectories and evaluate the robot's motion completion, including: Extract motion features and trajectory smoothness based on the motion trajectory, construct a curve of each motion feature changing over time, mark the standard baseline on the curve, and count the area of ​​the curve part outside the standard baseline as the additional area; Generate robot motion completion based on the additional area and trajectory smoothness of each motion feature; ； in, is the robot action completion degree, is the number of motion features, For the The evaluation weight of each motion feature, For the The additional area of ​​the motion feature, is the trajectory smoothness, is a preset constant.

9. A robot motion analysis system based on image recognition, characterized in that: include, A screening module is used to obtain robot action videos within a period of time, convert the robot action videos into temporally continuous image frames, and screen the image frames; A detection module is used to divide the filtered image into a foreground area and a background area, detect a first key point in the foreground area, match the first key point, and obtain first key point matching information; An adding module, used for analyzing the matching information of the first key point to add the second key point, and forming a motion trajectory based on the first key point and the second key point; The evaluation module is used to extract motion features through motion trajectories and evaluate the robot's motion completion.

Images