Method for detecting bad postures for specific age groups applied to TV scenes

Abstract

The present invention discloses a method for detecting bad postures of people of a specific age group applied to television scenes, including detecting face data and limb key point data, wherein the face data includes age data, face posture angles (yaw, roll, pitch) and face frames, and the limb key point data includes limb key point positions and limb frames; matching, filtering and tracking and completing the face data and limb key point data, thereby detecting bad postures of people of a specific age group; triggering a prompt when the same bad posture is detected for multiple consecutive frames of images. The present invention detects bad sitting postures of people of a specific age group by matching face data and limb key point data, filtering face data and limb data, tracking and completing human body data, and continuous posture judgment, thereby filtering out bad postures of people of a specific age group based on face data and limb key point data.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for detecting bad postures of a specific age group in a television scene. Background Art

[0002] The vigorous development of artificial intelligence (AI) technology has promoted the development of intelligence in various industries. The joint development of hardware, algorithms and data has made AI technology more and more widely used in various industries. However, AI technology usually cannot directly output the results we need most. We need to perform relevant post-processing on the output results of AI to obtain results that meet our needs. Most of the existing sitting posture detection only involves the key points of the limbs, but is not linked to the age determined by the face, and no corresponding matching tracking is performed. It cannot be applied to the detection of bad postures for specific age groups in TV scenes. Summary of the invention

[0003] The purpose of the present invention is to provide a method for detecting bad postures of specific age groups in television scenes, so as to solve the problem that sitting posture detection in the prior art is not combined with face judgment and cannot be applied to the detection and correction of bad postures of specific age groups in television scenes.

[0004] The present invention solves the above problems through the following technical solutions:

[0005] Methods for detecting bad postures for specific age groups applied to TV scenes include:

[0006] Step S100, detecting face data and body key point data, wherein the face data includes age data, face posture angles (yaw, roll, pitch) and face frame, and the body key point data includes body key point positions and body frames;

[0007] Step S200, matching, filtering and tracking the human body data with the key point data of the limbs, thereby detecting bad postures of people in a specific age group; triggering a prompt when the same bad posture is detected in multiple consecutive frames of images.

[0008] In step S100, human body data is collected by a horizontally installed camera and transmitted to a TV. Two AI models are installed in the TV, which are used to obtain face data and limb key point data from images respectively.

[0009] Matching the face data with the limb key point data in step S200 specifically includes matching the face data and limb key point data that simultaneously satisfy two conditions: the face frame is within the limb frame, and the head key point among the limb key points is within the face frame, to obtain a human body data.

[0010] The filtering of the face data and the limb key point data in step S200 refers to filtering out erroneous data from the human body data, which specifically includes:

[0011] If the facial posture angle (yaw, roll, pitch) in the human body data is greater than a first set threshold, the human body data is filtered;

[0012] Determine the multiple relationship between the length of the spine of the image and the length of the face frame according to the position of the spine key point in the limb key points, and if the multiple relationship is less than a second set threshold, filter the human body data;

[0013] The shoulder length and spine length in the image are determined based on the point data in the limb key points. If the ratio of the shoulder length to the spine length is less than a third set threshold, the human body data is filtered.

[0014] The step S200 of tracking and completing human body data with face data and limb key point data specifically includes: if a frame of human body data obtained after matching and filtering lacks age data, the face data or limb key point data in the human body data are matched with the face data or limb key point data of the previous frame respectively; if the IOU value of the face frame or limb frame of the two is greater than a threshold, it is determined that the two match, and the age data of the previous frame is matched to the human body data to complete the age data, thereby screening out bad postures of people in a specific age group.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0016] (1) The present invention detects bad sitting postures of people in a specific age group by matching facial data with body key point data, filtering facial data and body key point data, tracking and completing body data, and continuously judging postures, thereby screening out and correcting bad postures of people in a specific age group based on facial data and body key point data.

[0017] (2) The hardware of this method only requires a monocular camera, no other sensing equipment is needed, and it only requires to be roughly horizontal when installed. The present invention not only utilizes limb data, but also facial data, so that it can detect bad postures for specific ages, and the data tracking effect is better.

[0018] (3) The present invention matches the results obtained by two common AI models (face and limbs), filters the age credibility through facial posture angles (yaw, roll, pitch), determines whether the upper body is not completely on the screen or is blocked, determines whether it is sideways, filters the data, and performs a human body data tracking and completion method. After processing, it can be applied to TV scenes to prompt bad sitting postures for specific age groups. The method is simple and effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described in detail below in conjunction with examples, but the embodiments of the present invention are not limited thereto.

[0021] Embodiment 1:

[0022] Combined with Figure 1 As shown, the method for detecting bad postures for a specific age group applied to a TV scene includes:

[0023] Step S100, detecting face data and body key point data, wherein the face data includes age data, face posture angles (yaw, roll, pitch) and face frame, and the body key point data includes body key point positions and body frames;

[0024] Step S200, matching, filtering and tracking the human body data with the key point data of the limbs, thereby detecting bad postures of people of a specific age group; triggering a prompt when the same bad posture is detected in multiple consecutive frames of images.

[0025] In step S100, human body data is collected by a horizontally installed camera and transmitted to a TV. Two AI models are installed in the TV, which are used to obtain face data and limb key point data from images respectively.

[0026] The matching of face data and limb key point data refers to finding the relationship between face data and limb key points obtained from a single frame image. Because some images may have face data but no corresponding limb key point data (only the face is shown in the image), and limb key point data but no corresponding face data (no face is shown in the image). The matching method is: match the face data and limb key point data that meet the two conditions that the face frame is in the limb frame and the head key point in the limb key point is in the face frame to obtain a human body data. There are three cases for this human body data, including both face and limb data, only face data, and only limb data.

[0027] The filtering of face data and limb key point data refers to filtering the face data and limb key point data in the human body data obtained in the previous matching step, which may be erroneous data. This includes judging whether to filter age by face posture angle (yaw, roll, pitch), judging whether to filter limbs by face frame and limb key point position, and judging whether to filter limbs by limb key point position relationship. Judging whether to filter age by face posture angle (yaw, roll, pitch) means that if the face posture angle (yaw, roll, pitch) is greater than the first set threshold, it means that the face is not facing the camera, and only the age data obtained by the front face is the most accurate, so the age data of the non-front face needs to be filtered. Judging whether to filter limbs by face frame and limb key point position means that when the upper body of the limbs does not appear completely on the screen or the upper body of the limbs is blocked, the multiple relationship between the spine length of the screen and the length of the face frame will become smaller when the spine key point position of the upper body of the limbs is judged. If its length multiple relationship is less than the second set threshold, the limb key point data is filtered. Judging whether to filter the limbs based on the position relationship of the limb key points means that when a person is facing the camera sideways, the body key point data may become inaccurate, and the shoulder length and spine length in the picture need to be judged through the point data in the limb key points. When facing sideways, the shoulder length will become smaller, and the spine length will remain unchanged, so the ratio will also become smaller. When it is less than the third set threshold, it can be determined that the person is facing sideways at this time, and the limb key point data is filtered.

[0028] Human body data tracking completion refers to a frame of filtered human body data obtained after the first two steps. It may be missing age data, so the frame data can be completed based on the previous frame human body data. The completion method is to match the face data and limb key point data in the human body data with the face data and limb key point data in the previous frame of human body data. The matching method is the intersection over union (IOU) matching method. IOU calculates the ratio of the intersection and union of the "predicted bounding box" and the "real bounding box". The IOU value is calculated to be greater than the threshold. If it is greater, it can be determined to be a match. The age data of the previous frame can be matched to the human body data.

[0029] Here are some examples:

[0030] Assume that the camera captures three people, A, B, and C (the camera only needs to be roughly level), the data in the previous time period is normal, and the data in the next time period is abnormal. Specifically, in the next time period, A moves his limbs out of the camera screen, B blocks his limbs (generating wrong limb key point data), and C moves his face out of the camera screen and tilts his limbs. Then the information that can be obtained by the three people changes from 6 pieces of information in the previous time period (face A, face B, face C, limb A, limb B, limb C) to 4 pieces of information in the next time period (face A, face B, abnormal limb B, abnormal limb C).

[0031] First, the face and limb data are matched. According to the matching rules mentioned in the technical solution, the information of the previous time period can be obtained as human body A (face A, limb A), human body B (face B, limb B), human body C (face C, limb C), and the information of the latter time period is human body A (face A, missing), human body B (face B, abnormal limb B), human body C (missing, abnormal limb C).

[0032] Secondly, the face data and limb data are filtered. After data filtering, since the limbs are occluded by B in the latter time period, the corresponding key point positions will also change. According to the filtering scheme mentioned in the technical solution, the abnormal limb B data in human body B (face B, abnormal limb B) will be filtered. Then the information of the latter time period is human body A (face A, missing), human body B (face B, missing), and human body C (missing, abnormal limb C).

[0033] Then, the human body data is tracked and completed. According to the completion plan, human body C is completed in combination with the information of the previous time period, and the information of the next time period can be obtained as human body A (face A, missing), human body B (face B, missing), and human body C (completed face C, abnormal limb C).

[0034] Finally, continuous posture judgment can determine the human body C (completed face C, abnormal limb C). If the age data in the completed face C is within the set specific age range, then the abnormal limb C can be detected. If abnormal limb C is detected in multiple consecutive frames, the detection will be confirmed to remind the user to correct the posture of watching TV.

[0035] Although the present invention is described herein with reference to the illustrative embodiments of the present invention, the above embodiments are only preferred embodiments of the present invention, and the embodiments of the present invention are not limited to the above embodiments. It should be understood that those skilled in the art can design many other modifications and embodiments, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A method for detecting bad postures for a specific age group applied to a television scene, characterized in that: include: Step S100, detecting face data and body key point data, wherein the face data includes age data, face posture angles (yaw, roll, pitch) and face frame, and the body key point data includes body key point positions and body frames; Step S200, matching, filtering and tracking the human body data with the key point data of the limbs, thereby detecting bad postures of people of a specific age group; triggering a prompt when the same bad posture is detected in multiple consecutive frames of images; The step S200 of matching the face data with the limb key point data specifically includes matching the face data and the limb key point data that simultaneously satisfy the two conditions that the face frame is within the limb frame and the head key point among the limb key points is within the face frame, to obtain a human body data; The filtering of the face data and the limb key point data in step S200 refers to filtering out erroneous data from the human body data, which specifically includes: If the facial posture angle (yaw, roll, pitch) in the human body data is greater than a first set threshold, the human body data is filtered; Determine the multiple relationship between the length of the spine of the image and the length of the face frame according to the position of the spine key point in the limb key points, and if the multiple relationship is less than a second set threshold, filter the human body data; The shoulder length and spine length in the image are determined based on the point data in the limb key points. If the ratio of the shoulder length to the spine length is less than a third set threshold, the human body data is filtered.

2. The method for detecting bad postures of a specific age group applied to a television scene according to claim 1, characterized in that: In step S100, human body data is collected by a horizontally installed camera and transmitted to a TV. Two AI models are installed in the TV, which are used to obtain face data and limb key point data from images respectively.

3. The method for detecting bad postures of a specific age group applied to a television scene according to claim 1, characterized in that: The step S200 of tracking and completing human body data with face data and limb key point data specifically includes: if a frame of human body data obtained after matching and filtering lacks age data, the face data or limb key point data in the human body data are matched with the face data or limb key point data of the previous frame respectively; if the IOU value of the face frame or limb frame of the two is greater than a threshold, it is determined that the two match, and the age data of the previous frame is matched to the human body data to complete the age data, thereby screening out bad postures of people in a specific age group.

Images