Attitude detection method based on lifting platform

Abstract

The present invention relates to the technical field of posture detection, and in particular to a posture detection method based on a lifting platform. It comprises the following steps: S1, switching to a learning mode, wherein the user adopts different sitting and standing postures for sitting and standing; S2, collecting depth information, then obtaining character features, and marking the obtained character features as right or wrong; S3, forming a sitting and standing posture model library; S4, switching to a monitoring mode, S5, collecting depth information of the user's posture, obtaining character features, and inputting the obtained character features into the posture model library obtained in step S3 for analysis, and judging whether the collected user posture is correct or wrong; S6, if the posture is judged to be wrong for multiple consecutive times, then judging it is wrong and giving a reminder; if the posture is not judged to be wrong for multiple consecutive times, then returning to step S5. This detection method can detect whether the user's standing and sitting posture is correct, and then give a reminder by controlling the lifting table.

Description

Technical Field

[0001] The present invention relates to the technical field of posture detection, and in particular to a posture detection method based on a lifting platform. Background Art

[0002] A lift table is a type of lifting platform. It is an adjustable height desk that can be adjusted according to a person's height, sitting posture, or standing posture. Although the lift table can adjust the height to match the user's standing or sitting posture to a certain extent, making it more comfortable for the user to use, the lift table in the existing technology cannot help users adjust their standing or sitting posture. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a posture detection method based on a lifting table, by which it is possible to detect whether the user's standing or sitting posture is correct, and then to give a reminder by controlling the lifting table.

[0004] The technical solution adopted by the present invention is: a posture detection method based on a lifting platform, which comprises the following steps:

[0005] S1, switch to learning mode, the user adopts different sit-stand postures;

[0006] S2, the depth sensor collects the depth information of the different sitting and standing postures, then obtains the character features through the collected depth information, and marks the obtained character features as right or wrong;

[0007] S3, forming a sitting and standing posture model library according to the character features marked in step S2;

[0008] S4, switch to monitoring mode,

[0009] S5, the depth sensor collects the depth information of the user's posture, and then obtains the character features through the collected depth information, and inputs the obtained character features into the posture model library obtained in step S3 for analysis to determine whether the collected user posture is correct or wrong;

[0010] S6. If the posture is judged to be wrong for multiple consecutive times, the user's posture is judged to be wrong, and then the lifting platform is controlled to remind; if the posture is not judged to be wrong for multiple consecutive times, return to step S5 to continue collecting and analyzing.

[0011] Preferably, the character features include facial features and skeleton features.

[0012] Preferably, the depth sensor is a binocular camera, a two-way ranging sensor or an ultrasonic sensor.

[0013] Preferably, the data in the sitting and standing posture model library formed in step S3 is quantified, and the analysis in step S5 is to quantify the character features obtained in step S5, and then compare the quantified character features with the model library to see whether they exceed the set threshold. If the threshold is exceeded for multiple consecutive times, the posture is judged to be wrong. If the threshold is not exceeded for multiple consecutive times, return to step S5 to continue collecting and analyzing.

[0014] Preferably, after collecting the depth information in step S5, it is necessary to perform depth information analysis and obtain whether the user is standing or sitting at this time, and controlling the lifting platform to give a reminder in step S6 specifically means that if step S5 analyzes that the user is in a standing position, the lifting platform is controlled to descend; if step S5 analyzes that the user is in a sitting position, the lifting platform is controlled to ascend.

[0015] Preferably, after entering the monitoring mode, it is first necessary to determine whether the user is a real person, which includes the following steps:

[0016] S41, collecting continuous frame images of the person;

[0017] S42, extracting RGB information of the skin area of ​​each frame of the collected image, and then obtaining three matrices according to the extracted RGB three channel information;

[0018] S43, reducing the dimension of the three matrices obtained for each frame of image in step S42, and obtaining three new matrices for each frame of image;

[0019] S44, average calculation is performed on the three new matrices obtained for each frame of image in step S43, and each new matrix of each frame of image is respectively obtained with an average value, and then the first waveform is obtained by taking the average value of the new matrix of the R channel in each frame of image as the ordinate and the time as the abscissa, the second waveform is obtained by taking the average value of the new matrix of the G channel in each frame of image as the ordinate and the time as the abscissa; the third waveform is obtained by taking the average value of the new matrix of the B channel in each frame of image as the ordinate and the time as the abscissa;

[0020] S45, filtering the three waveforms obtained in step S44 through a filter;

[0021] S46, merging the three waveforms filtered in step S45;

[0022] S47, extracting the actual physiological data in the merged waveform diagram in step S46, and then comparing it with the standard physiological data to determine whether there is physiological data output.

[0023] Preferably, in step S44, the three new matrices need to be weighted averaged, and the weighted average calculation method is: arrange the front and rear frame differences of the dimension reduction matrix in sequence, filter out the pixels whose absolute value of change is greater than the set threshold, and calculate the average value of the remaining pixel values, which is the weighted value of the channel in the current frame.

[0024] Preferably, image brightness detection is required before acquisition in step S41. If the detected image brightness is insufficient, exposure compensation is required until the detected image brightness meets the standard.

[0025] Compared with the prior art, the present invention has the following advantages: a sitting and standing posture model library is established through a learning mode, and then the collected depth information is input into the sitting and standing posture model library for analysis to determine whether the posture is correct, and then a reminder is given through the lifting platform when the posture is wrong, which can help the user maintain a correct standing and sitting posture, and the reminder also needs to be based on whether the person is standing or sitting, and then the lifting platform is controlled to rise and fall to give a reminder, so the reminder effect is better.

[0026] Before monitoring, physiological data is used to detect whether the person is a real person. This can prevent users from placing dummies or photos to impersonate real people, thereby enabling the lifting platform to have accurate identity authentication functions and more diverse functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the signal after weighted averaging of the G channel when judging a real person in the present invention.

[0028] Figure 2 for Figure 1 Schematic diagram of the signal after filtering.

[0029] Figure 3 It is a signal schematic diagram of the time domain information after the three channels are merged when judging a real person in the present invention. DETAILED DESCRIPTION

[0030] The present invention is further described below through specific implementation modes, but the present invention is not limited to the following specific implementation modes.

[0031] A posture detection device based on a lifting platform, which includes a lifting mechanism of the lifting platform itself, a depth sensor and a camera arranged on the lifting platform, and a controller for receiving data collected by the depth sensor and the camera, processing the data and generating a feedback signal.

[0032] The controller may only include a microprocessor arranged on the lifting table. If there is only one microprocessor, all calculations are performed on the microprocessor. In this way, the model library can be formed by parameterized control technology. Compared with machine learning technology, parameterized control technology is relatively simple. It only needs to focus on a few fixed parameters to determine whether it belongs to the set range. In this way, the amount of data processing is small.

[0033] It can also include a microprocessor set on the lifting table and a server set at another location. The two are communicatively connected, and a large amount of calculation is performed on the server. In this way, machine learning technology can be used to form a model library. Most of the processing is also performed on the server, and then the feedback signal obtained after the calculation is transmitted to the microprocessor. The microprocessor only needs to control the lifting mechanism of the lifting platform and serve as a transmission medium, so the cost will be relatively low.

[0034] A posture detection method based on lifting platform:

[0035] S1, switch to learning mode, the user adopts different sit-stand postures;

[0036] S2, the depth sensor collects the depth information of these different sitting and standing postures, and then obtains the character features through the collected depth information, and marks the obtained character features as right or wrong; wherein the character features include facial features and skeleton features, and the depth sensor is a binocular camera, a two-way ranging sensor or an ultrasonic sensor;

[0037] S3, forming a sitting and standing posture model library according to the character features marked in step S2; the data in the formed sitting and standing posture model library is quantized;

[0038] The complete implementation process of the learning mode is as follows: the user's sitting and standing posture is collected through a depth sensor, and then transmitted to a microprocessor. The microprocessor encodes the collected data through a protocol and then sends it to the server in the form of pulses. The server receives the data sent by the microprocessor, and then decodes it through the same protocol, and then reorganizes and reduces the noise of the decoded data (the noise reduction method is mainly Kalman filtering and other methods), and extracts super-dimensional features from the noise-reduced data (mainly Fourier transform, the extracted features are mainly: frequency, peak, trough, variance, standard deviation, etc.), and then reduces the dimension of the super-dimensional features (the main method is linear discriminant analysis LDA and principal component analysis PCA and other dimensionality reduction methods), and then trains through a variety of machine learning models (such as random forests, neural networks, support vector machines, and Bayesian, etc.), and then obtains the model with the highest accuracy according to the cross-validation method and saves it to complete the model establishment. This protocol is specially designed by the applicant for the present invention, but it can also be realized by applying other existing protocols with some minor improvements.

[0039] S4, switch to monitoring mode,

[0040] S5, the depth sensor collects the depth information of the user's posture, and determines whether the user is standing or sitting at this time according to the collected depth information, and then obtains the character features through the collected depth information, and inputs the obtained character features into the posture model library obtained in step S3 for analysis to determine whether the collected user posture is correct or wrong; the analysis is to quantify the character features obtained in step S5, and then compare the quantified character features with the model library to see whether they exceed the set threshold. If the threshold is exceeded for multiple consecutive times, it is determined that the posture is wrong. If the threshold is not exceeded for multiple consecutive times, it returns to step S5 to continue collecting and analyzing; after the depth information is collected in step S5, it is necessary to perform depth information analysis to determine whether the user is standing or sitting at this time;

[0041] S6. If the posture is judged to be wrong for multiple consecutive times, the user's posture is judged to be wrong, and then the lifting table is controlled to remind. When the user is standing, the lifting table is controlled to be lowered, and when the user is sitting, the lifting table is controlled to be raised. If the posture is not judged to be wrong for multiple consecutive times, return to step S5 to continue collecting and analyzing. And controlling the lifting platform to remind in step S6 specifically means that if step S5 analyzes that the user is standing, the lifting platform is controlled to be lowered; if step S5 analyzes that the user is sitting, the lifting platform is controlled to be raised.

[0042] After switching to monitoring mode, it is also necessary to determine whether the person is a real person. The specific algorithm is:

[0043] First, exposure compensation adjustment is required: the system calibrates and locks the exposure value during initialization. The selection of exposure value includes but is not limited to the following methods: adjusting by comparing the numerical histogram of the overall picture with the empirical histogram. If the histogram is relatively dark, increase the overall numerical brightness; adjusting by detecting the brightness of the face area; adjusting exposure compensation by comparing the face area with other areas; during the algorithm detection process, the changes in ambient light can be tracked by monitoring the numerical value of the background brightness, and the changes in ambient light can be fed back to the face numerical value as compensation (especially when there are continuously moving objects in the background).

[0044] In an application example, the image histogram value can be provided to the user as feedback to indicate whether the office lighting is appropriate: for example, when the ambient light is detected to be weak, the user is reminded to increase the light source and light intensity.

[0045] Then, continuous frame images of the face are collected through the camera;

[0046] Then, by running the face detection algorithm, we get the color matrix of the corresponding area range of the three RGB channels.

[0047] After the face area is locked, the values ​​of the three RGB channels of the image are extracted and stored in an M*N matrix, where M is the width of the face area and N is the height (the values ​​of M and N can vary depending on the distance of the person. In actual applications, the system can perform image correction by changing the values ​​of M and N. In this application, M can be set to 640 and N to 480).

[0048] Moreover, assuming that the changes between the upper and lower frames of the image are not drastic, that is, there is no strong displacement of the object being measured, the average value in a small area (such as 5*5 gaussian kernel) can be regarded as the filtered single pixel value, ensuring that the algorithm will not be interfered by the noise generated by the external environment or the capture device hardware. In this way, by comparing the changes in the same pixel point, blood pulsation information can be obtained.

[0049] Afterwards, in order to reduce the computational intensity and noise interference, it is necessary to perform fuzzy dimensionality reduction on the color matrix (moving average methods such as Gaussian pyramid or Gaussian blur) to obtain three relatively small matrices, that is, mainly reduce the size, such as reducing the 640*480 matrix to 160*120. Afterwards, it is necessary to further perform weighted averaging on the signal of each matrix per second. In this specific embodiment, the weighted averaging method is: arrange the front and back frame differences of the dimensionality reduction matrix in sequence, filter out a certain proportion of pixels with relatively large absolute values ​​of change, and calculate the average value of the remaining pixel values, which is the weighted value of the channel in the current frame.

[0050] In addition, due to the movement of the object being measured and the facial movements, the values ​​of the corresponding pixels in consecutive frames may jump. If all the pixels are weighted, it will cause the signal baseline to jump. In the algorithm framework of this paper, the calculation of outlier removal is introduced: the absolute value of the difference between the corresponding pixels of consecutive frames is taken, and the distribution of the absolute value in the area is compared with the empirical template (assuming that in an ideal case, the value obeys the normal distribution), and the fitting parameters of the empirical distribution are found through distribution fitting, and the excess parts of some areas are treated as outliers.

[0051] like Figure 1 The figure shows the original weighted average signal of the G channel for 10 seconds (the camera uses a frequency of 30FPS). Due to the instability of the face area, we can see that the background noise of the original average signal is very large.

[0052] Considering the basic noise of the hardware and the jitter of face area tracking, Figure 1 Perform low-pass filtering. Considering that the normal heart rate range is 0.5-3Hz, we set the corresponding band-pass filter for filtering. Figure 2 For the effect after filtering, or for breathing detection, if the range of interest is 0.1-0.5Hz, you can also set the corresponding bandpass filter for filtering;

[0053] Due to the difference in the receiving intensity of the three RGB channels for the color of facial blood flow, we can merge the three channels: the three channels contain the same heart rate information, and the intensity of other interference signals is different in each channel. The three RGB channels can be merged in the frequency domain according to the Fourier transform, and the inverse Fourier transform can be performed to obtain the merged time domain information, such as Figure 3 As shown, the periodic signal is the heart rate signal and the envelope is the blood signal.

[0054] The facial dimensionality reduction image obtained by the camera includes three channels: R, G, B (red, green, blue). Due to the difference in wavelengths of the three colors, the depth of penetration into the skin is also different, which reflects different information: red reflects more and more accurate blood flow information, and because of its large penetration depth, it also contains a lot of noise information such as muscle activity; green is the most common heart rate detection color light, and the blood flow information obtained is the most stable (anti-sports interference, anti-physiological noise, etc.); blue is the light with the shallowest penetration, and it is most effective in resisting sports interference. By merging the signals of the three channels, the heart rate / blood flow information can be effectively amplified and extracted.

[0055] Channel merging methods include but are not limited to optical channel projection, entropy calculation, etc.

[0056] Entropy calculation: The main idea of ​​channel entropy merging is to obtain the probability weight of each channel by performing signal transformation on the three-channel information.

[0057] Optical channel projection: The signal obtained by channel merging is the original blood flow signal. The signal within the effective heart rate range can be extracted through bandpass filtering (0.6-3Hz). The heart rate can then be calculated through methods such as peak detection in the time domain or Fourier transform of frequency. In one example, the filtered signal is fast Fourier transformed, and the obtained frequency domain peak is tracked and the most likely heart rate frequency is identified. If the maximum peak exceeds twice the second peak, it can be considered that the confidence level of the heart rate calculation is high and it is a valid output. If the proportion of the maximum peak of the heart rate is continuously lower than a certain threshold, the detection is considered unstable, and the user can be reminded to improve indicators such as ambient light.

[0058] After the heart rate is calculated, the confidence index can be combined to further screen out valid heart rates with high confidence indexes. The user terminal can display the user's long-term heart rate trend based on the heart rate calculation results. If the continuous heart rate is further analyzed, advanced parameters such as tension, heart health status, and heart rate changes can be obtained.

[0059] The combination of face recognition algorithm and facial heart rate algorithm can effectively display detailed information about the person's work on the user side: the time period when the person appears in front of the workbench, and the change pattern of the heart rate during the corresponding period. Furthermore, when the heart rate is high, the user can be reminded to relax and exercise.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A posture detection method based on a lifting platform, characterized in that: It includes the following steps: S1, switch to learning mode, the user adopts different sit-stand postures; S2, the depth sensor collects the depth information of the different sitting and standing postures, then obtains the character features through the collected depth information, and marks the obtained character features as right or wrong; S3, forming a sitting and standing posture model library according to the character features marked in step S2; S4, switch to monitoring mode, S5, the depth sensor collects the depth information of the user's posture, and then obtains the character features through the collected depth information, and inputs the obtained character features into the posture model library obtained in step S3 for analysis to determine whether the collected user posture is correct or wrong; S6. If the posture is judged to be wrong for multiple consecutive times, the user's posture is judged to be wrong, and then the lifting platform is controlled to remind; If the posture is not judged to be wrong for multiple consecutive times, return to step S5 to continue collecting and analyzing; After entering the monitoring mode, you first need to determine whether the user is a real person, which includes the following steps: S41, collecting continuous frame images of the person; S42, extracting RGB information of the skin area of ​​each frame of the collected image, and then obtaining three matrices according to the extracted RGB three channel information; S43, reducing the dimension of the three matrices obtained for each frame of image in step S42, and obtaining three new matrices for each frame of image; S44, average calculation is performed on the three new matrices obtained for each frame of image in step S43, and each new matrix of each frame of image is respectively obtained with an average value, and then the first waveform is obtained by taking the average value of the new matrix of the R channel in each frame of image as the ordinate and the time as the abscissa, the second waveform is obtained by taking the average value of the new matrix of the G channel in each frame of image as the ordinate and the time as the abscissa; the third waveform is obtained by taking the average value of the new matrix of the B channel in each frame of image as the ordinate and the time as the abscissa; S45, filtering the three waveforms obtained in step S44 through a filter; S46, merging the three waveforms filtered in step S45; S47, extracting the actual physiological data in the merged waveform diagram in step S46, and then comparing it with the standard physiological data to determine whether there is physiological data output; The character features include facial features and skeleton features; After the depth information is collected in step S5, it is necessary to perform depth information analysis and determine whether the user is standing or sitting at this time. In step S6, the lifting platform is controlled to provide a reminder. Specifically, if step S5 analyzes that the user is standing, the lifting platform is controlled to descend; if step S5 analyzes that the user is sitting, the lifting platform is controlled to ascend.

2. The posture detection method based on the lifting platform according to claim 1 is characterized in that: The depth sensor is a binocular camera, a two-way ranging sensor or an ultrasonic sensor.

3. The posture detection method based on the lifting platform according to claim 1 is characterized in that: The data in the sitting and standing posture model library formed in step S3 is quantified, and the analysis in step S5 is to quantify the character features obtained in step S5, and then compare the quantified character features with the model library to see whether they exceed the set threshold. If the threshold is exceeded for multiple consecutive times, the posture is judged to be wrong. If the threshold is not exceeded for multiple consecutive times, return to step S5 to continue collecting and analyzing.

4. The posture detection method based on the lifting platform according to claim 1 is characterized in that: In step S44, the three new matrices need to be weighted averaged, and the weighted average calculation method is: arrange the front and back frame differences of the dimension reduction matrix in sequence, filter out the pixels whose absolute value of change is greater than the set threshold, and calculate the average value of the remaining pixel values. This value is the weighted value of the channel in the current frame.

5. The posture detection method based on the lifting platform according to claim 1 is characterized in that: In step S41, image brightness detection is required before acquisition. If the detected image brightness is insufficient, exposure compensation is required until the detected image brightness meets the standard.

Images