Sit and reach evaluation method based on machine vision
By collecting and segmenting specific images in the forward flexion evaluation of seating bodies, establishing the correlation relationship between the scale value and the coordinate set, and using the human posture key point detection model to determine the image with the longest distance of the hand displacement, thereby evaluating the results, solving the calculation power consumption problem caused by frame-by-frame segmentation of hand features in the prior art, and achieving efficient and accurate evaluation results.
Patent Information
- Application Number
- CN202311356708.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-19
- Publication Date
- 2025-04-22
AI Technical Summary
The existing machine vision-based seat forward bending evaluation method needs to be segmented from multiple images frame by frame when identifying marking scales and hand features, resulting in large computing power consumption and low efficiency.
By collecting images containing the combination features and the features of the motion part, multiple coordinate sets are divided from the target image using the segmentation model, the correlation relationship between the scale value and the coordinate set is established, and the results are then evaluated. The specific steps include collecting the first target image to obtain a coordinate set that identifies the combined features, continuously collecting the image at a fixed point to analyze the second target image at the longest distance of the hand displacement, obtaining the wrist bone key points through the human posture key point detection model, determining the second target image, and segmenting and obtaining the coordinate set of the moving part features from the image.
This method greatly saves computing power consumption, reduces hardware costs and layout costs, and simplifies environmental requirements and usage processes, improving the efficiency and accuracy of evaluation.
Smart Images
Figure CN119863831A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of sports evaluation, and in particular to a sit-and-reach evaluation method based on machine vision. Background Art
[0002] The training and assessment of some sports have been realized intelligently, and the image or video recognition solution based on machine vision does not need to use wearable devices or sensor devices, which not only reduces the hardware cost, but also reduces the layout cost. The environmental requirements are relatively simple. At the same time, it is also a non-sensing use for users, so it is highly respected. However, in the implementation process of the existing machine vision-based solutions, especially for the sit-and-reach exercise, the identification of the scale requires the hand features to be segmented frame by frame from multiple images until the image used to determine the score is found. The frame-by-frame segmentation of hand features consumes a lot of computing power and needs to be improved urgently. Summary of the invention
[0003] In order to solve the technical problems existing in the background technology, the present invention proposes a sit-and-reach evaluation method based on machine vision, comprising:
[0004] Acquire a first target image of a sitting forward flexion body that contains at least a combination of identification features, segment and obtain multiple first coordinate sets of multiple scale value lines in the combination of identification features from the first image through a segmentation model, and establish an association relationship between the multiple first coordinate sets and the scale values;
[0005] Continuously collecting images of the sitting forward flexion that at least include movement features and identification combination features, analyzing a second target image corresponding to the longest hand displacement distance from multiple frames of images, and obtaining a second coordinate set of the movement features from the second target image by segmentation using a segmentation model;
[0006] Get the key point coordinates from the second coordinate set;
[0007] The results are evaluated based on the positional relationship between the key point coordinates and the multiple first coordinate sets and the correlation relationship between the multiple first coordinate sets and the scale values.
[0008] Furthermore, the motion part feature is specifically a hand feature, and the key point coordinates are specifically the farthest point coordinates in the second coordinate set in the hand moving direction.
[0009] Furthermore, the motion part feature is specifically a cursor feature, and the key point coordinate is specifically one of the nearest point coordinates in the second coordinate set in the hand movement direction.
[0010] Furthermore, “analyzing from the multiple frames of images the second target image corresponding to the longest hand displacement” specifically includes:
[0011] Acquire multiple wrist bone key points corresponding to multiple frame images through a human posture key point detection model;
[0012] The positions of multiple wrist bone key points are compared to obtain the wrist bone key point that is farthest in the hand movement direction, and the image corresponding to the wrist bone key point is the second target image.
[0013] Furthermore, after the second target image in the multiple frames of images, there is an image in which the position of the key points of the wrist bones in the hand movement direction is closer to the starting position of the hand than the position of the key points of the wrist bones in the second target image.
[0014] Furthermore, it is also necessary to meet the following requirements:
[0015] In the continuous multiple frames of images including the second target image, the image spacing of the wrist bone key point coordinates corresponding to any two frames does not exceed the preset value, and the acquisition time difference between the first and last two frames of the above multiple frames is greater than or equal to the first preset time length.
[0016] Furthermore, “analyzing from the multiple frames of images the second target image corresponding to the longest hand displacement distance” specifically includes:
[0017] Acquire multiple wrist bone key points corresponding to multiple frame images through a human posture key point detection model;
[0018] When the coordinate distance between the wrist bone key points corresponding to any two frames in the continuous multi-frame images does not exceed the preset value, and the acquisition time difference between the first and last two frames in the above multi-frame images is greater than or equal to the second preset time length, any frame in the continuous frame images will be used as the second target image.
[0019] Furthermore, the first target image also includes human features, and the human features and the identification combination features conform to a preset positional relationship.
[0020] Furthermore, the segmentation model is trained with a batch of first-type image samples, and the first-type image samples at least include identification combination features and calibration features of the identification combination features.
[0021] Furthermore, the segmentation model is trained with batches of second-type image samples, and the second-type image samples at least include hand features and calibration features of the hand features.
[0022] The motion evaluation method proposed in the present invention has the following technical effects: the present invention not only reduces the hardware cost, but also reduces the layout cost, and the environmental requirements are relatively simple. At the same time, it is also non-sensitive for users to use. The movement of the key points of the hand is equivalent to the movement of the entire hand for judgment. Compared with the prior art method of segmenting the entire hand frame by frame along the hand contour and then making an overall comparison of the hand between multiple images, it naturally saves a lot of computing power. The key points of the hand are defined as the key points of the wrist bones, and the existing mature human posture key point detection model that has relatively low computing power requirements and can autonomously learn and continuously improve the judgment accuracy through sample training is cleverly used to complete it, which naturally has a better effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 This is a flow chart of the sit-and-reach evaluation method based on machine vision proposed by the present invention;
[0024] Figure 2 It is a partial flow chart of the sit-and-reach evaluation method based on machine vision proposed by the present invention;
[0025] Figure 3 This is a partial flow chart of the sit-and-reach flexion evaluation method based on machine vision proposed in the present invention. DETAILED DESCRIPTION
[0026] Example 1
[0027] A sit-and-reach evaluation method based on machine vision is proposed, including:
[0028] S1. Collect a first target image of a sitting forward flexion body that contains at least a combination of identification features, obtain multiple first coordinate sets of multiple scale value lines in the combination of identification features from the first image through segmentation model, and establish an association relationship between the multiple first coordinate sets and the scale values.
[0029] Furthermore, the first target image also includes human features and the human features and the identification combination features conform to a preset positional relationship, so that the relative position between the identification combination features in the first target image and the second target image and the camera remains unchanged.
[0030] The reason is that some sit-and-reach equipment is placed and movable at any time. Only when the exerciser is on the equipment and is in a ready state can it be considered that the equipment will not move during the sit-and-reach exercise. During the sit-and-reach exercise, the relative positions of the identification combination features must remain unchanged in multiple frames of images.
[0031] It should be pointed out that the segmentation model is trained with a batch of first-type image samples. The first-type image samples at least contain identification combination features and calibration features of the identification combination features. Through continuous learning and training, the segmentation output of the segmentation model can be made more accurate.
[0032] S2. Continuously collect the sitting forward flexion images that at least contain the movement part features and the identification combination features, analyze the second target image corresponding to the longest hand displacement distance from the multiple frames of images, and obtain the second coordinate set of the movement part features from the second target image by segmentation through the segmentation model.
[0033] Of course, when the sit-and-reach equipment has a cursor, since the hand moves synchronously with the cursor when it is extended, the longest distance of hand displacement is also acceptable.
[0034] Specifically, the identification combination features are generally printed on the plate of the sit-and-forth bending equipment. The identification combination features generally include multiple scale lines arranged in parallel and spaced apart, and the multiple first coordinate sets corresponding to the identification combination features are multiple first coordinate sets of the multiple scale lines, and the multiple scale lines correspond to the above-mentioned multiple scale values.
[0035] S3. Analyze the second target image corresponding to the longest hand displacement distance from the multiple frames of images, and obtain the second coordinate set of the hand feature / cursor from the second target image through segmentation model.
[0036] Specifically include:
[0037] S31, analyzing from the multiple frames of images to obtain a second target image corresponding to the longest hand displacement distance.
[0038] Specifically include:
[0039] S311. Acquire multiple wrist bone key points corresponding to multiple frame images through a human posture key point detection model.
[0040] In S311, the movement of the key points of the hand is judged to be equivalent to the movement of the entire hand. Compared with the existing technology of segmenting the entire hand frame by frame along the hand contour and then comparing the hand between multiple images, S311 naturally saves a lot of computing power. In S311, the key points of the hand are defined as the key points of the wrist bones, and the existing mature human posture key point detection model with relatively low computing power requirements and self-learning through sample training to continuously improve the judgment accuracy is cleverly used to complete it, which naturally has a better effect.
[0041] S312. Acquire a second target image corresponding to the longest hand displacement distance according to the positional relationship between the image coordinates of multiple wrist bone key points.
[0042] It should be pointed out that although multiple wrist bone key points correspond to multiple frames of images, the multiple frames of images are collected at fixed points, so multiple wrist bone key points can be put into any frame of image for reference and comparison.
[0043] The second target image is selected in the following two ways:
[0044] The first is to compare the positions of multiple wrist bone key points to obtain the wrist bone key point with the farthest position in the direction of hand movement, and the image corresponding to the wrist bone key point is the second target image. Moreover, after the second target image in the multi-frame image, there is also an image in which the position of the wrist bone key point is closer to the starting position of the hand than the position of the wrist bone key point in the second target image in the direction of hand movement. It can be understood here that the moving individual extends his hand to the front end and then returns. The image corresponding to the front end is the second target image. In the continuous multi-frame image including the second target image, the image spacing of the wrist bone key point coordinates corresponding to any two frames does not exceed the preset value, and the acquisition time difference between the first and last two frames in the above multi-frame image is greater than or equal to the first preset time length. Here, it can be limited whether the moving individual makes a pause action at the front end.
[0045] The second type is that when the coordinate distance between the key points of the wrist bones corresponding to any two frames in the continuous multi-frame image does not exceed the preset value, and the acquisition time difference between the first and last frames in the above multi-frame image is greater than or equal to the second preset duration, any frame in the continuous frame image is used as the second target image. In the actual implementation process, the individual athlete will always keep his hand at the same position as in the second target image and will not withdraw it for fear of losing the score, that is, he will not meet the requirement of retracting his hand in the first case until the score is announced, so that he will not retract his hand if the score is not released, and he will not meet the score if he does not retract his hand, thus entering an infinite loop. Therefore, here, when the hand stays for a long enough time, that is, greater than or equal to the second preset duration, it can be determined that this image is the second target image.
[0046] S32. Segment the second target image using a segmentation model to obtain a second coordinate set of motion features, where the motion features are specifically hand features or cursor features.
[0047] When the sit-and-reach equipment has a cursor, the cursor is used as the feature of the moving part. If not, the hand feature is used as the feature of the moving part.
[0048] Only the second target image is segmented, which greatly saves computing power.
[0049] Furthermore, the segmentation model is trained with batches of second-type image samples, and the second-type image samples at least include hand features and calibration features of the hand features.
[0050] S4. Obtain the coordinates of the farthest point / closest point in the hand moving direction from the second coordinate set of the hand feature / cursor feature as the key point coordinates.
[0051] S5. Evaluate the results based on the positional relationship between the key point coordinates and the multiple first coordinate sets and the correlation relationship between the multiple first coordinate sets and the scale values.
[0052] 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 sit-and-reach evaluation method based on machine vision, characterized in that: include: Acquire a first target image of a sitting forward flexion body that contains at least a combination of identification features, segment and obtain multiple first coordinate sets of multiple scale value lines in the combination of identification features from the first image through a segmentation model, and establish an association relationship between the multiple first coordinate sets and the scale values; Continuously collecting images of the sitting forward flexion that at least include movement features and identification combination features, analyzing a second target image corresponding to the longest hand displacement distance from multiple frames of images, and obtaining a second coordinate set of the movement features from the second target image by segmentation using a segmentation model; Get the key point coordinates from the second coordinate set; The results are evaluated based on the positional relationship between the key point coordinates and the multiple first coordinate sets and the correlation relationship between the multiple first coordinate sets and the scale values.
2. The evaluation method according to claim 1, characterized in that: The motion part feature is specifically a hand feature, and the key point coordinates are specifically the farthest point coordinates in the second coordinate set in the hand moving direction.
3. The evaluation method according to claim 1, characterized in that: The motion part feature is specifically a cursor feature, and the key point coordinate is specifically one of the nearest point coordinates in the second coordinate set in the hand movement direction.
4. The evaluation method according to claim 1, characterized in that: "Analyzing from multiple frames of images the second target image corresponding to the longest hand displacement" specifically includes: Acquire multiple wrist bone key points corresponding to multiple frame images through a human posture key point detection model; The positions of multiple wrist bone key points are compared to obtain the wrist bone key point that is farthest in the hand movement direction, and the image corresponding to the wrist bone key point is the second target image.
5. The evaluation method according to claim 1, characterized in that: After the second target image in the multiple frames of images, there is an image in which the position of the key point of the wrist bone in the hand moving direction is closer to the starting position of the hand than the position of the key point of the wrist bone in the second target image.
6. The evaluation method according to claim 1, characterized in that: Also need to meet: In the continuous multiple frames of images including the second target image, the image spacing of the wrist bone key point coordinates corresponding to any two frames does not exceed the preset value, and the acquisition time difference between the first and last two frames of the above multiple frames is greater than or equal to the first preset time length.
7. The evaluation method according to claim 1, characterized in that: “Analyzing from multiple frames of images the second target image corresponding to the longest hand displacement distance” specifically includes: Acquire multiple wrist bone key points corresponding to multiple frame images through a human posture key point detection model; When the coordinate distance between the wrist bone key points corresponding to any two frames in the continuous multi-frame images does not exceed the preset value, and the acquisition time difference between the first and last two frames in the above multi-frame images is greater than or equal to the second preset time length, any frame in the continuous frame images will be used as the second target image.
8. The evaluation method according to claim 1, characterized in that: The first target image also includes human features, and the human features and the identification combination features conform to a preset positional relationship.
9. The evaluation method according to claim 1, characterized in that: The segmentation model is trained by batches of first-type image samples, and the first-type image samples at least include identification combination features and calibration features of the identification combination features.
10. The evaluation method according to claim 1, characterized in that: The segmentation model is trained with batches of second-type image samples, where the second-type image samples at least include hand features and calibration features of the hand features.