See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/372823421

Deep Learning-Based Yoga Posture Specification Using OpenCV and Media

Pipe

Article in Applied and Computational Engineering · August 2023

DOI: 10.54254/2755-2721/8/20230085

CITATIONS READS

0 337

4 authors, including:

Anilkumar Chunduru Jyothsna Robbi

GMR Institute of Technology GMR Institute of Technology
40 PUBLICATIONS 127 CITATIONS 1 PUBLICATION 0 CITATIONS

SEE PROFILE SEE PROFILE

Vandana Sree Sattaru

GMR Institute of Technology
1 PUBLICATION 0 CITATIONS

SEE PROFILE

All content following this page was uploaded by Anilkumar Chunduru on 28 August 2023.

The user has requested enhancement of the downloaded file.

 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230085

Deep learning-based yoga posture specification using OpenCV

and media pipe

Chunduru Anilkumar1,3, Robbi Jyothsna1, Sattaru Vandana Sree1, E. Gothai2

1
Dept of Information Technology, GMR Institute of Technology, Rajam, Andhra
Pradesh-532127
2
Department of Computer Science and Engineering, Kongu Engineering College,
Erode, India

3
[email protected]

Abstract. Yoga is a year’s discipline that calls for physical postures, mental focus, and deep
breathing. Yoga practice can enhance stamina, power, serenity, flexibility, and well‐being.
Yoga is currently a well-liked type of exercise worldwide. The foundation of yoga is good
posture. Even though yoga offers many health advantages, poor posture can lead to issues
including muscle sprains and pains. People have become more interested in working online than
in person during the last few years. People who are accustomed to internet life and find it difficult
to find the time to visit yoga studios benefit from our strategy. Using the web cameras in our
system, the model categorizes the yoga poses, and the image is used as input. However, the
media pipe library first skeletonizes that image. Utilizing a variety of deep learning models, the
input obtained from the yoga postures is improved to improve the asana. On non-skeleton photos,
VGG16, InceptionV3, NASNetMobile, YogaConvo2d, and also InceptionResNetV2 came in the
order of highest validation accuracy. The proposed model YogaConvo2d with skeletal pictures,
which is followed by VGG16, reports validation accuracy in contrast, NASNetMobile,
InceptionV3, and InceptionResNetV2.

Keywords: deep learning, media pipe, classification, skeletonization, posture detection.

1. Introduction
People typically assume that yoga could be a kind of exercise that has stretching and folding of the piece
however Yoga is far quite a simple exercise. Yoga could be an approach to life or the Art of living
through the mental, religious and physical path. It permits us to attain stillness and to faucet into the
consciousness of the inner self. It additionally helps in learning a way to rise on top of the pull of mind,
emotions, and lower bodily wants and face the challenges of day-to-day life. Yoga works on the amount
of one's body, mind, and energy. Regular observation of yoga brings positive changes within the practice
– robust muscles, flexibility, patience, and physiological state. even as moving into the yoga cause
properly is vital, therefore starting up is the correct approach. get laid with awareness, coordination
every body movement with the breath as you gently begin the cause and enter the resting position Human
cause estimation has returned to a protracted approach within the last 5 years, astonishingly hasn't
surfaced in several applications simply however. this is often a result of additional focus has been placed
on creating cause models larger and additional correct, instead of doing the engineering work to form

© 2023 The Authors. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0
(https://creativecommons.org/licenses/by/4.0/).

80
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230085

them quickly and deploy-able all over. With movement, our mission was to style and optimize a model
that leverages the most effective aspects of progressive architectures, while keeping illation times as
low as doable. The result's a model which will deliver correct key points across a good kind of poses,
environments, and hardware setups. Using a depth camera makes it possible to detect disguised hands
and facial expressions as well as the entire articulation of mortal hands. Utilizing a portable moving
depth camera, low-cost, three-dimensional (3D) reconstruction of terrain has been investigated. Reason
of the sterility essential item in the operating room, touch-free, interfaces getting popular in medical
applications. Algorithms for estimation and shadowing of mortal disguises are usually learned from
sizable pool training data. Viscosity assessment of mortal disguise data reveals the data online
observance structure, which can then be used for shadowing and disguise estimation. Compared to the
data used to estimate the viscosity of mortal disguise, our stuff is different. Performance in reading is
impacted by biased training data. Using a weight retrogression to deleting these training datasets. There
are numerical methods for altering data distribution from established data viscosity.

2. Related work
Presented the YOGI data-set, which has roughly 5500 photos. Using the "tf-posture," Different angles
and points from these visual frames were fed into a number of machine learning models, such as Logistic
Regression, followed by Random Forest, Support Vector Machine (SVM), Decision Tree, followed by
Nave Bayes, and also K Nearest Neighbour (KNN). The accuracy of the given algorithm random-forest
classifier was 99.04%. [1].

Figure 1. Architecture of VGG-16.

In order to get beyond the drawbacks of being state-of-the-art, they proposed a vision-grounded system
in this work for the real-time recognition of yoga activities. To do this, we first create a sizable data-set
of ten Yoga positions that were photographed in challenging real-world settings so that the built system
might operate more effectively when placed in those settings. Twenty-seven people of all ages
participated in the creation of the internal data set of yoga poses (8 males and 19 ladies). The videos,
which feature acts in all their conceivable variants, were recorded using MI Max and also One Plus 5T
smartphones in both indoor and outdoor settings. Second, we suggested a lightweight 3D CNN armature
that detects Yoga positions by taking use of the crucial spatial-temporal link between them. Yoga
disguise sequences can be investigated to improve the planned system's subtlety in recognition. For the
real-time recognition of yoga poses, the developed method can be be improved and implemented on
moveable bedded bias.[2].
A key area of computer vision has been how the human body responds to visuals. The estimate
problem for mortal disguise is made significantly easier by the use of a depth camera. created a body
component identification algorithm for the commercial Kinect device. The recognition performance was
improved by using tentative retrogression timbers to incorporate knowledge and global variables, as the

81
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230085

stoner's height and also branch lengths. A data-driven strategy specifically keeps database of mortal acts
and seeks the best-matching acts at runtime to grease disguise rebuilding has been studied solely. [3].

Figure 2. Flowchart for pose recognition.

A yoga recognition system that makes use of an ordinary RGB camera. The data-set is intimately
available and was gathered using HD 1080p Logitech webcams on 15 people (10 men and 5 women).
To photograph the stoner and describe important details, Open Pose is employed. End-to-end deep
literacy-grounded frames do not require handmade features, therefore new asanas may be added by
simply again training t model with fresh data. For purpose of discovering yoga postures, the method of
utilizing both CNN and LSTM on data obtained from Open Pose has been designed to be largely
successful. For 12 people, the system can identify the six asanas both in real-time and on recorded films
(five males and seven ladies). Real-time testing and data collection have been done by many people.
Yoga acts on a videotape are successfully detected by the system with 99.04 frame-wise delicacy and
99.38 frame delicacy after polling 45 frames. The system successfully produced 98.92 dishes in real-
time.[4].
Both ANN and FCM classifier were trained using 30% of the subjects in this study, while the
remaining 70% were used for testing. Three posture cases were differently chosen from each type of
posture among all subjects during the training phase. Each data frame was fed into the associated ANN
classifier algorithm to determine the classification order and corresponding FCM classifier algorithm to
determine the identify outcome. The average degree of delicacy for posture detection was 89.34%, and
it ranged from 70% to 100%. The present data frame was quickly acknowledged as the first posture. The
ultimate recognition outcome in our recognition styles was calculated using accretive probability. As a
result, several instances position three was recognized as posture one, which decreased the identification
delicateness. [5]
With the introduction of Kinect, a computer-assisted self-training system has been developed. With
the extraction from the skeleton for pose recognition that adopted three postures: downward-facing dog
stance, warrior 3 posture, and tree posture, the devices typically contain RGB cameras, infrared
projectors, and detectors. During the experiments, an overall accuracy of 82.84% was attained. [6]

82
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230085

OPENPOSE is a real-time multi-person human posture recognition library that, for the first time,
identifies important facial and body components in a single photo. With the aid of underlying layers,
OpenPose has been used to recognize the introduction's major parts and extract specifics from the input
image.[7]
For self-training, a yoga posture recognition method using the Kinect camera has been created. 300
films of 12 yoga poses being performing five times each by five different yoga practitioners were
collected. Once the body has been extracted from the video clip, a skeleton is used to depict the yoga
positions. Produced accuracy was 99.33%. [8]

Table 1. Comparison table.

Author Title Techniques Advantages
Yoga pose
classification: Both The obtained results by both media pipe
Shubham CNN and also Media CNN, Media and non-media pipe versions are
Garg et.al Pipe inspired deep pipe, examined to see if skeletonization of the
(2022) learning approach Classification yoga pose image increases the model
for real-world performance or not.
application
Three-dimensional In this paper, poses are captured by a
CNN-inspired deep Convolution
camera as well as recorded videos of the
Shrajal jain learning architecture Neural networks,
poses feeding the sequence to the
et.al for yoga pose Conv 3D and max
proposed the 3D CNN neural network
(2020) recognition in the poling model, which was developed, had
real-world (Maxpool3D)
99.96% of training accuracy and 99.49%
environment of validation accuracy.
Three categories are created from the
Controllable data
Kyungyong Bayesian network subsampled motion capture data. There
sampling in the
Yang et.al model, PCA were 8 699 Type I postures, 7 300 Type II
space of
(2015) techniques poses, and 2 386 Type III positions as a
human pose
consequence of the classification.
Yoga posture recognition is quite good.
For 12 people, the system can identify the
Santosh
Real-time Yoga six asanas both in real time and on
Kumar
recognition using CNN, LSTM recorded films (five males and seven
Yadav et.al
deep learning females). Real-time testing and data
(2019)
collection have been done by many
people.
Yoga Posture
Bayesian network is trained by the
Recognition and
ANN classifier, database. Subject two and followed by
Quantitative
Jiwen Zhang FCM classifier five performed many standards given to
Evaluation with
Et.al (Futures the evaluation results by the Bayesian
Wearable Sensors
(2019) commission network while gaining on, FCM classifier
Based on Two-Stage
merchant) algorithm membership computed is very
Classifier and Prior
high
Bayesian Network

83
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230085

Table 1. Continue.
Author Title Techniques Advantages
Development Consideration of 12 poses that are
CNN,
of a yoga posture captured through webcams. The poses are
Chhaihyuo Logistic
coaching system skeletonized and application of feature
Long et.al Regression,
using an interactive extraction and classification through
(2021) SVM, KNN,
display based Machine Learning models with 96% of
naive Bayes
on transfer learning accuracy.
Yoga Pose Detection
Deepak By providing the input, the model
and Classification CNN, RNN,
Kumar et.al detects the accurate pose among the
Using Deep MLP, SVM
(2020) learned 6 poses.
Learning
Novel IoT-Based
Trained with 300 clips of 12 yoga poses
Privacy-Preserving
through the help of a Kinect camera (a
Yoga Posture
Munkhjargal DCNN, CNN, depth cam that sees in 3D and creates a
Recognition System
Gochoo et.al Adaboost skeleton image and detects the
Using Low-
(2019) algorithm, movements). The body contour is
Resolution Infrared
extracted first and then skeletonized and
Sensors and Deep
produces 99.33% of accuracy.
Learning

Table 1 explains the comparison over the different papers along with their advantages, title name and
respective authors.

3. Methodology
In this model, the main factor, to fete the asana, is the angle and distance between the joints. Using
Media pipe, milestones of different body corridors & joints will be detected and also arc tan (function
of NumPy) will calculate the angles between those milestones. According to the angles, it'll estimate the
posture and also recoup the instructions of that asana and it'll read out the instructions. An overall
architecture of the given system is shown in illustration below.

Figure 3. Flowchart of the procedure.

The visual is gains by a camera, it be divides camera module on a smartphone, now extensively
accessible, or a webcam, it is useful fashion to capture prints because nearly everybody has one of these
types of captured input results. The system's input element is the camera. A webcam, a mobile camera,
or a divides camera module can be used as the source.

84
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230085

Figure 4. Block diagram.

The camera is used to receiving images and give total data to the model After receiving the visualized
item, we developed a model using CNN[9][10][11]. The proposed system can fete a wide range of
positions. As a conclusion, exercising data sets to the topmost extent doable. Pose identification is done
using Media Pipe the stoner's input data is taken. Produce a correct shell exposure of the stoner using
this data. Milestones on the mortal body gestures identify the crucial points and places

4. Conclusion
We have created a system in this design that consists of a channel for disguise identification, point
localization on the mortal body, and an error identification method. This technique attempts to aid
individuals in correctly yoga practice on their own and assist with ailments that may result from
improper yoga poses. The approach is suitable for evaluating the stoner's disguise from the front and
providing feedback so that they can improve their yoga disguise using deep literacy techniques. The
designed model is mounted atop a dashboard that was likewise made with stoners in mind.

References
[1] Garg, Shubham Saxena, Aman Gupta, Richa 2022 2022/06/03 Yoga pose classification: a
CNNand MediaPipe inspired deep learning approach for real-world application Journal of
Ambient Intelligence and Humanized Computing https://doi.org/10.1007/s12652-022-03910-
0
[2] Jain, S., Rustagi, A., Saurav, S., Saini, R., & Singh, S. (2020). Three-dimensional CNN-inspired
deep learning architecture for Yoga pose recognition in the real-world environment. Neural
Computing and Applications.
[3] Yang, K., Youn, K., Lee, K., & Lee, J. (2015). Controllable data sampling in the space of
human poses. Computer Animation and Virtual Worlds, 26(3-4), 457–467.
doi:10.1002/cav.1662
[4] Yadav, S. K., Singh, A., Gupta, A., & Raheja, J. L. (2019). Real-time Yoga recognition using
deep learning. Neural Computing and Applications. doi:10.1007/s00521-019-04232-7
[5] Wu Z, Zhang J, Chen K, Fu C. Yoga Posture Recognition and Quantitative Evaluation with
Wearable Sensors Based on Two-Stage Classifier and Prior Bayesian Network. Sensors
(Basel). 2019 Nov 23;19(23):5129. DOI: 10.3390/s19235129. PMID: 31771131; PMCID:
PMC6929085.
[6] Long, Chhaihuoy Jo, Eunhye Nam, Yunyoung 2022 2022/03/01 Development of a yoga posture
coaching system using an interactive display based on transfer learning The Journal of
Supercomputing 5262 EP - 5284 https://doi.org/10.1007/s11227-021-04076-w
[7] D. Kumar and A. Sinha, "Yoga Pose Detection and Classification Using Deep Learning,"
International Journal of Scientific Research in Computer Science, Engineering and

85
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230085

Information Technology, pp. 160–184, Nov. 2020, DOI: 10.32628/cseit206623.

[8] M. Gochoo et al., "Novel IoT-Based Privacy-Preserving Yoga Posture Recognition System Using
Low-Resolution Infrared Sensors and Deep Learning," in IEEE Internet of Things Journal,
vol. 6, no. 4, pp. 7192-7200, Aug. 2019, DOI: 10.1109/JIOT.2019.2915095.
[9] Easwaramoorthy, S., Sophia, F., & Prathik, A. (2016, February). Biometric Authentication using
finger nails. In 2016 international conference on emerging trends in engineering, technology
and science (ICETETS) (pp. 1-6). IEEE.
[10] Easwaramoorthy, S., Moorthy, U., Kumar, C. A., Bhushan, S. B., & Sadagopan, V. (2018).
Content based image retrieval with enhanced privacy in cloud using apache spark. In Data
Science Analytics and Applications: First International Conference, DaSAA 2017, Chennai,
India, January 4-6, 2017, Revised Selected Papers 1 (pp. 114-128). Springer Singapore.
[11] Subramanian, M., Cho, J., Sathishkumar, V. E., & Naren, O. R. (2023). Multiple types of Cancer
classification using CT/MRI images based on Learning without Forgetting powered Deep
Learning Models. IEEE Access.

86

View publication stats