Posture detection technology based on RealSense

The rapid popularization of computers, so that people will be more and more work on the computer to complete, all walks of life, especially the program developers, writing workers, the working time on the computer more and more long, in this case the bad posture of the neck and shoulder lumbar spine will have a great impact, prone to the occurrence of a variety of diseases. The age of adolescent exposure to computers is getting smaller, and maintaining a good posture is beneficial to the growth and development of teenagers and to the protection of eyesight. Surveys show that the incidence of disease at the waist of the shoulder is getting higher, the age of onset is getting smaller, and it is related to sitting for a long time, and we also find that some peculiar office is standing office or even working on a treadmill. These questions all indicate that the correct posture is a great influence on health. Nowadays, it is necessary to maintain the correct posture when the working intensity is getting more and more and less time is being exercised.

This article first describes the structure of the intel® RealSense realsense device [1] and how the device obtains three-dimensional data from a realistic scenario. For this subject, the Intel RealSense official SDK extracts three-dimensional data from the user's face about 70 feature points. The three-dimensional data of these feature points is the basis of the posture Detection algorithm in this paper. After obtaining the three-dimensional data of the user's face over 70 points, the distance between the head rotation and pitch angle, the head and the computer display is calculated, and the training and recognition experiment is carried out using support vector machine (SVM) [2], which realizes the "Crooked Head", "Hump", "looking over large" and "looking over large" "The judgment of such bad posture, and after many experiments has obtained the good experimental result." Through the combination with MySQL database [3], the sitting characteristics, the length of the day different sitting position recorded, data analysis, generate reports to the user, so as to achieve a complete detection, reminders, corrective purposes.

1 real-time posture Detection algorithm

1.1 Hardware and development environment

The RealSense is the Deepin camera launched by Intel, the latest generation device is a full VGA resolution (640 x 480 60fps) depth camera, and 1080p resolution of the RGB camera, when shooting 1080p image transmission rate is 30fps. Detection range between 0.2 meters to 1.2 meters, depending on the use of the scene and the algorithm used, the detection distance is different. There are already a variety of laptops default with RealSense RealSense camera , this article uses the development version of the RealSense camera for developers to develop, 1.

The basic size of the development is 150mm x 30mm x 58mm, according to the official introduction, the required environment is configured as follows, need to have USB3.0 interface, according to the actual test, the desktop is best to use the rear USB3.0 interface, to ensure sufficient stable power supply. The official claims to use Intel's fourth generation and later core series CPUs to complete development, using the third-generation Core i5 CPU in the development of this article to work as normal.

It is important to note that the Intel RealSense SDK currently supports only Windows 8.1 and later Windows 64-bit operating systems and does not support Mac OS and Linux.

1.2 Posture Detection algorithm

Design of 1.2.1 Posture Detection system

The design flow of the whole set of sitting detection system 2.

Sitting detection system includes five main parts, namely user information input, device initialization and image preprocessing, real-time posture detection, posture information statistics, content results display and reminders. The real-time sitting posture Detection module is the focus of this paper, design flow 3.

In the real-time posture detection, if the eye tracking technology is turned on, the eye tracking technique is used to calibrate and assist judgment to enhance the stability and accuracy of the system.

1.2.2 Algorithm Determination

The user's head motion and head state is the main feature extraction space, through RealSense, in three-dimensional coordinates, the elevation of the head (PITCH, rotation around the x-axis), yaw angle (YAW, rotation around the y-axis), Tumbling angle (roll, rotation around the z axis), height information, The spatial position information is extracted by the coordinate transformation, the transformed data is Kalman filtered [4], the more accurate data is obtained for the sitting ratio, if the head information parameters meet certain indicators, then the user's posture is judged to be correct, such as 4.

When Euler angle [5] rolling angle roll absolute value is greater than 10 degrees, it can be considered that the user head around the head is too large, is the wrong posture, need to be reminded, more accurately, when the scroll value is positive 10 degrees above, the user head to the right tilt angle is too large, the duty is less than negative 10 degrees, The tilt angle of the user's head is too large, when the absolute value of Euler angle is greater than 18 degrees, it can be thought that the yaw of the user head is too large, so the wrong posture often appears in the programmer's group, because many programmers are using two monitors or more monitors, because the table is not big enough, The distance between the user and the monitor is too close, in the view of the monitor need a large turn, the neck and shoulder caused by the extra burden, also belong to the bad posture, more accurate description, when the yaw value is greater than 18 degrees, the user head to the right to rotate the amplitude is too large, when the yaw value is less than negative 18 degrees, The user's head to the left rotation is too large, when the Euler angle of the pitch value plus 5 after the absolute value of more than 25 degrees, you can determine the user's head pitch is too large, is not the correct posture, usually, the user when using the computer display will be slightly lower than the user's head height, And the most suitable posture is not the head must be heads-up display, but a little lower nod, so as to help relieve the pressure of the shoulder spine, more suitable for a long time work, more accurately, when the Euler angle of pitch value is greater than 20 degrees, the user is looking up at the display screen, The typical error in this situation is the hand-held chin look at the monitor, 5, in reading long paragraphs of text or browsing video content often appear such a posture, and when Euler angle of pitch value is less than negative 30 degrees, the user bow amplitude is too large, you can determine the user posture is not correct, In this case, the damage to the spine is more serious, the appearance of this bad posture is more for the wearer of the eyes of the user;

Another typical wrong posture is the distance monitor too close, through the RealSense depth information, can quickly get to the user's head, shoulder relative to the camera distance, and can accurately obtain the user's eyes to the camera distance information, so can quickly determine whether the user is close to the monitor, The display size of the experiment is 17 inches, the best distance between the user and the display is 70 cm to 80 cm, less than 70 centimeters away from the monitor too close, usually this is due to kyphosis, and in turn too close to the monitor, will not only cause serious damage to vision, will also aggravate the situation of kyphosis. Seated correctly-6.

Analysis of experimental results of 1.2.3

The number of experimental group was 4 groups, each group of one classmate, a total of four people, the experimental results of data such as Table 1-table 4.

Through the analysis of the test data of different students, we can find that the accuracy rate of the algorithm is more than 90%, compared to the traditional intelligent hardware recognition method 40% of the accuracy rate has been greatly improved, but in the experiment also found that the second group of detection accuracy is low, through the test of the height of students, eyesight and other information consulting, Also found another important factor, the user usually use the table size, the seat height of the user's posture will have a great impact, a suitable seat desk for health is also very necessary.

2 Summary

This paper designs a real-time posture detection algorithm using the Intel RealSense camera, which is suitable for the white-collar or programmer group that needs to sit in front of the computer for a long time, comprehensively analyzes the data of the user's head and eyes, and accurately determines whether the user is seated correctly. Compared to the traditional intelligent hardware [6] solution, eliminate the cumbersome to wear intelligent hardware equipment alone, and greatly improve the detection accuracy, reached more than 90% accuracy, while the very low false alarm rate also ensures that in the actual application scenario will not interfere with the user's normal work. In the design of the algorithm and the whole system software, there is still some optimization and improvement space, before the image information can be processed by a better image denoising algorithm, the user facial feature data can be processed in the filtering process, in order to reduce the acquisition of data errors and jumps on the impact of detection.

Posture detection technology based on RealSense