Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Yoga Postures Correction and Estimation using Open

CV and VGG 19 Architecture
Tanmay Hande1
Pranali Dhawas2
Bhargavi Kakirwar 3
Aaditya Gupta4
G.H. Raisoni College of Engineering, Nagpur, India

Abstract:- The COVID-19 epidemic has significantly been developed to ensure the safe and effective practise of
changed how we work out, with more people turning to yoga.
home fitness as a way to stay active during stay-at-home
orders. However, without access to professional trainers, One such technology that is becoming increasingly
beginners may struggle to perform exercises with proper popular is the use of pose estimation with artificial
form, increasing the risk of injury. Therefore, there is a intelligence (AI). Pose estimation is the process of using
need for systems to monitor exercise performance for computer vision algorithms to track and analyze the
both short- and long-term injury prevention. In this movements of an individual's body. By using this
study, we present an approach for accurately detecting technology, it is possible to estimate the position of various
and correcting yoga postures using pose estimation joints and body parts in real-time and provide feedback on
techniques with OpenCV and VGG-19 architectures the correctness of a yoga posture [3]. In this study, the use
with GPU transfer learning. To precisely measure and of pose estimation with AI has several advantages for
correct body posture during training sessions, the correcting and estimating yoga postures. Firstly, it provides
suggested solution combines deep learning-based a more objective and precise assessment of posture, as
algorithms and computer vision approaches. To confirm compared to relying on subjective human observation.
the effectiveness of the VGG-19 model on the utilised Secondly, it allows for real-time feedback, which can be
dataset, We conducted a large number of tests, incredibly useful for individuals who are learning yoga
comparing the performance of several machine learning postures or trying to improve their form. Lastly, it can be a
and deep learning strategies for estimating yoga helpful tool for yoga teachers who want to provide more
postures. With a precision of 98.11 percent, the findings personalised instruction and support to their students.
show the usefulness of the suggested technique in
precisely recognising and correcting exercise postures. Deep learning techniques' capacity to do end-to-end
The findings of this study have significant implications optimisation is one of its key benefits [4]. Pose estimation
for improving the effectiveness and safety of yoga and action recognition are often difficult to integrate to carry
sessions and could be extended to other domains that out a useful joint optimisation, sometimes requiring 3D
require precise human pose estimation. convolutions or heat map transformations [5]. The back-
propagation chain required for end-to-end learning is
I. INTRODUCTION disrupted when detection-based techniques are used, since
they need the non-differentiable argmax function as a post-
The popularity of yoga has surged in recent years, with processing step to recover the joint coordinates. To address
many individuals incorporating it into their daily routines this issue, we offer a new method for combined 2D and 3D
for its numerous physical and mental health benefits. posture estimation based on an extension of the
However, incorrect posture and breathing during yoga differentiable soft-argmax [6]. As a result, we can build a
sessions can lead to severe pain and chronic problems, fully trainable multi-task framework by layering action
highlighting the need for technologies that can help recognition on top of posture estimation. However, the
individuals correct and improve their yoga postures [1]. absence of the aforementioned tools and trainers may be a
Among these techniques, physical postures, known as significant deterrent to our home yoga. Our goal is to create
asanas, have become particularly popular in the western an AI-based trainer that will enable you to yoga more
world. As the pandemic continues, many people have turned effectively at home. The objective of this study is to create
to yoga as a means of keeping themselves physically and an AI system that supports exercise by estimating posture to
mentally fit. However, it is important to perform asanas determine the quality and quantity of repetitions. This
correctly, as improper stretching or performing research, which aims to make exercise easier and more
inappropriate asanas and breathing inappropriately can lead enjoyable, includes a non-destructive interface.
to severe pain and chronic problems. This highlights the
growing need for tools and technologies that can help There are a tonne of apps on the market that instruct
individuals correct and improve their yoga postures [2]. To the user on the exercises to complete [7]. However, with our
address this need, a scientific analysis of asana practise has application, we not only instruct the user on which exercise
to execute, but also on how to hold themselves correctly and

IJISRT23APR2280 www.ijisrt.com 2420

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
utilise computer vision to count the repetitions. This mobility. Relying solely on pose estimation for correcting
programme may be seen as a yoga helper that offers instant and estimating yoga postures could potentially neglect the
posture detection and dietary advice. By broadening its nuances of breath and mind awareness and may not cater to
usage, the programme may be utilised at gyms as smart the unique needs and goals of each practitioner. Therefore,
trainers, decreasing the need for human involvement while personalized guidance from a qualified yoga teacher who
still allowing users to use it at home. can provide hands-on adjustments, verbal cues, and
modifications should complement the use of pose
In our work, we present a fitness application powered estimation.
by AI that recognises the user's yoga stance, records the
number of repetitions of the prescribed exercise, and gives Researchers are attempting to develop a Deep
personalised, in-depth information about how to improve the Learning-based system that can accurately detect yoga
user's body posture. To help those who don't have access to positions and offer users with feedback in an effort to
a gym but are nonetheless ready to work out at home to substitute a teacher. The experimental investigation by
maintain their physique and fitness and keep their body in Kinger et al. [10] looked at machine learning and deep
excellent form, an AI-based exercise assistant and fitness learning strategies for identifying yoga positions. Support
guide has been developed. to assist them in doing the Vector Machine, Convolutional Neural Network, and
exercises properly and stop them from suffering any short- Convolutional Neural Network with Long Short-Term
term or long-term injuries. Along with a personalised daily Memory models were tested, and their results were
yoga calorie count, this also offers a personalised health compared. The hybrid CNN-LSTM models produced the
advise and food plan. The following parts of this essay are fewest misclassifications, according to the research.
structured as follows: Methodology strategies and tools are Microsoft Kinect was utilised by Kadbhane et al. [11] to
outlined in Section II. In Section III, we describe the record video data and identify 20 bodily joints. Using data
findings from the recommended investigation. Section IV collected from ten selected human joint locations, the
discusses the results and what should be done next. reference structure for each yoga pose was developed. By
computing the angles made by all the vectors joining any
II. LITERATURE SURVEY two joint locations, they were able to establish the cosine
similarity of the vectors. When the predicted divergence is
The development of posture estimation systems that too high, the posture is incorrect. The total accuracy attained
can recognise and track human body positions from photos was not mentioned in the article.
or videos has been made feasible by improvements in
computer vision and machine learning. These methods have S. Haque et al. [12] proposed ExNet, a multilayer
shown success in many contexts, including healthcare, convolutional neural network (CNN). Push-ups, pull-ups,
sports, and entertainment. In the field of yoga, pose cycling, swiss ball hamstring curls and walking were
estimation can be used to offer immediate feedback to included in a collection of 2000 images of individuals in
practitioners, enabling them to adjust their postures yoga poses. The model made use of the Adam optimizer and
accurately. an automatic learning rate reduction technique. ExNET was
able to identify a 2D human yoga stance from the dataset
Li et al.'s work [8] used a bottom-up method for with 82.68% accuracy after 50 iterations. The model has
segmenting a yoga practitioner in real-time and estimating overfitting issues and requires greater hyper-parameter
many people's poses using a single-shot method. They tweaking. The dataset used by Agarwal et al. [13] included
developed a multi-task CNN to accurately detect and 5500 photos of 10 distinct yoga positions. They suggested a
classify the key points, which helped in estimating the method in which the user's skeleton is initially identified
practitioner's pose. On the dataset of 82 postures, the using the tf-pose estimation technique. In the last stage, six
suggested approach has an accuracy of 90.1%. Another different machine learning models—Decision Tree, Random
study by Yadav et al. [9] aimed at creating YogNet, a Forest, Logistic Regression, Naive Bayes, SVM, and
lightweight CNN architecture optimized for mobile devices, KNN—were used and contrasted. With a 99.04% accuracy
to predict yoga postures accurately. The study used an rate, the Random Forest classifier had the best result.
ensemble of deep neural networks and achieved an accuracy
of 92.3% on the Yoga-82 dataset. Additionally, the proposed Anilkumar et al. [14] developed a system that makes
method was effective in correcting the practitioner's posture, use of the Mediapipe library to do geometric analysis using
with an accuracy of 93.7%. Part-based modeling was camera frame data in order to provide the user feedback on
utilized, which relied on the key point-level structure to train their yoga stance. For a particular yoga posture, the system
the real-time segmentation activity in the first study. computes the angles between different joints and compares
them to precise angles recorded in a database. If the angle
In the two investigations, significant locations were difference is greater than a certain threshold, the gadget
located using heatmaps and regression using deep neural gives the user feedback through text or speech. A system
networks to increase the precision of yoga posture created by Luvizon et al. [15] can recognise human motion
identification. However, there are limitations to pose in the second and third dimensions as well as postures
estimation, including its reliance on consistent input data, connected to four different types of typical human activity.
potential oversimplification of yoga postures, and lack of Using input from four separate sources, the system was
individualization for differences in body shape, size, and trained using a multitask convolutional neural network.

IJISRT23APR2280 www.ijisrt.com 2421

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
More than a hundred recordings of both proper and  Determine which features are most important to assess
improper exercise form were used to inform Chen et al.'s based on technique and body type differences in
[16] proposal for a posture correction system. OpenPose was weightlifting that are thought to have a significant risk
used to train the model, and 18 landmarks were determined of injury.
so that it could identify different body orientations. The  Produce fitness videos that include the desired
strategy was put to the test with four distinct exercises: the technique elements as well as ones without any
front raise, standing shoulder press, bicep curl, and shoulder technique elements. After that, utilise the movies to
shrug. train the human posture estimation algorithms, which
will provide datasets for testing and assessment.
A smart fitness trainer based on posture detection and  Create a web application that can recognise the exercise
estimate was developed by Zou et al. [17] using deep the subject is doing and the viewing angle being used in
learning algorithms. In order to compare the human joints the video, allowing for the automated testing of certain
with the ideal positions and provide users feedback, they technique elements.
employed a multi-person pose estimation framework and  Create universal formulae that have a high probability
AlphaPose. Yadav et al. [18] also used deep learning to of spotting technique-related risks for regular users.
develop a real-time model that accurately identified key  Review the data and contrast the results of the various
points from videos of six different yoga poses. The system posture estimators. Examine the system's ability to
achieved high accuracy both framewise and based on preset distinguish the filming perspective and method, as well
edges, and was tested on twelve individuals with a success as its capacity to spot certain flaws in the technique.
rate of 98.92%. Thoutam et al. [19] employed Keras multi-
purpose pose estimation to extract features and a Multilayer III. MATERIALS AND METHODS
Perceptron for classification of six yoga poses based on
angles between twelve keypoints, achieving a high accuracy A. Yoga Pose Estimation Dataset
of 99.58%. We use a publicly available online dataset, YogAI,
which consists of 16 yoga poses, which were performed by
A. Research Gap and Motivation 10 participants (5 males and 5 females) in a controlled
While there has been significant research done on the environment with consistent lighting and camera settings.
correction and estimation of yoga postures using pose Each participant performed the poses multiple times,
estimation with OpenCV, there are still some gaps in the resulting in a total of 1,506 images. The poses were selected
field. One of the main research gaps is the need for larger to represent a range of difficulty levels, including both static
and more diverse datasets to train the deep learning models. and dynamic poses. The dataset consists of 16 yoga poses,
Most of the current datasets are relatively small and limited which include both static and dynamic poses. Some of the
in terms of the range of yoga poses they cover. Additionally, static poses included in the dataset are downdog, tree, and
there is a lack of standardization in the evaluation metrics warrior. Some of the dynamic poses included in the dataset
used for comparing different models and techniques, which are plank and goddess.
makes it difficult to compare results across studies.
For each image in the YogAI dataset, annotations were
Another research gap is the need for more robust and provided for key points on the body, such as joints and body
accurate pose estimation algorithms that can handle parts. The annotations were created using a custom
variations in lighting, camera angles, and other annotation tool developed by the researchers, and were
environmental factors. Many of the current pose estimation verified by multiple annotators to ensure accuracy. In
algorithms rely on specific assumptions about the pose or addition to the key point annotations, the dataset also
the environment, which can limit their accuracy and includes pose labels for each image. We choose 18 body key
applicability in real-world scenarios. points, each of which is made up of the x and y coordinates
of a body point. One dictionary and a 2D array are produced
Finally, there is a need for more user-friendly by this. When the dictionary returns several values for a
interfaces and feedback mechanisms for users to interact given key, all of this data, together with the associated
with these systems. While many of the current systems confidence levels, is put into an array. The bodily
provide feedback in the form of text or audio, more user- components that the dictionary's keys and values represent
friendly and interactive interfaces are required so that users are their coordinates. The presence of the observed body
may get feedback and direction in real time while engaging points in the lexicon is predicated on a high degree of
in yoga. confidence.
B. Research Objectives The key point annotations in the dataset include joint
locations (e.g., elbow, knee, ankle) and body parts (e.g.,
 Gain knowledge of many cutting-edge techniques for head, torso, limbs). The annotations were created using a
estimating the stance of a person, then choose a few custom annotation tool developed by the researchers, and
possibilities to test out. were verified by multiple annotators to ensure accuracy. The
dataset also includes pose labels for each image. The YogAI
dataset contains 1,506 images, with an average of 94 images
per pose. To facilitate model training and assessment, the

IJISRT23APR2280 www.ijisrt.com 2422

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
dataset is split into a training set (consisting of 80% of the assessment and VGG-19 for move learning on a
pictures) and a validation set (consisting of 20% of the GPUplatform provides an accurate and efficient solution to
photographs). correct and estimate yoga postures. Figure II shows the
architecture for pose estimation and correction using
B. Dataset Preprocessing OpenPose.
We remove data points with low accuracy probabilities
to avoid false positives. The filtering process involved Pose OpenCV estimation is the initial part of the
removing points that were forecasted to be erroneous based suggested architecture. A popular library named OpenCV is
on human posture assessment methods and filteringout used for computer vision applications including object
trailing points with excessively high variability. This step identification, face recognition, and position estimation.
played a crucial role in determining the precision of the This structure makes advantage of OpenCV to instantly
result and avoiding method problems that do not exist. To estimate the user's 2D stance. To identify the important
filter out inaccurate estimations, each of the three human areas of the human body, including the joints, limbs, and
posture estimate methods provided a confidence score for other body components, OpenCV employs a pre-trained
each guess. The simplest method to weed out To make low- deep learning model. The user's stance is then estimated
confidence estimates was to ignore important details. A using these essential points. The accuracy of pose estimation
confidence score threshold was set at 70% to remove any is crucial for the accuracy of the overall system, and
confidence ratings with a probability lower than this value. OpenCV provides high accuracy in real-time.
However, a confidence score threshold of up to 90% was
needed for some calculations to account for noise that could The second component of the proposed architecture is
lead to false positives. transfer learning using VGG-19 on a GPU platform. The
ImageNet dataset was used to train the pre-trained deep
Human pose estimation and computer vision systems learning model VGG-19. Reusing a previously trained
may make inherentin important spots that scored highly on model and optimising it on a fresh dataset is what transfer
probability but had an excessive distance difference from learning entails in order to improve accuracy. In this
the previous frame. To avoid false data, these points were architecture, transfer learning is used to fine-tune the VGG-
filtered out. The key point information was organized into a 19 model on a new dataset of yoga postures. The use of
dictionary with the x and y coordinates and the likelihood transfer learning on a GPU platform ensures that the model
that it is true (z). This array's index was used as the frame, can be trained faster and more efficiently than on a CPU
and if an item was absent, a null value was inserted. The platform.
confidence score was in the [0, 1] range, and the x and y
coordinates were normalised to fall within that range. The
confidence score threshold had to be lowered to 60% for all
movies created from a side-on viewpoint because of the
belowconfidence score for side-view important points
compared to the rest of the data.

Fig 2 Architectural Flow

Convolutional layers make up the VGG-19

architecture, and each layer extracts progressively intricate
information from the input picture. After the convolutional
layers, the final classification is performed by three
completely coupled layers. There are a total of 19 layers in
Fig 1 Preprocessed Image with Key Point Annotations the design, with the first 16 serving as convolutional layers
and the last 3 as fully connected layers. Dropout layers and
C. Network Architecture max-pooling layers are used into the design to reduce the
The proposed network architecture for the Correction feature maps' spatial dimensionality and prevent overfitting.
and Estimation of Yoga Postures using Pose Estimation with
OpenCV and VGG-19 with GPU transfer learning is a For GPU transfer learning, the VGG-19 model is
combination of two powerful techniques: deep learning and loaded onto a GPU after it has already been trained. After
pose estimation The utilization of OpenCV for present the model has been trained, its weights are locked and a new

IJISRT23APR2280 www.ijisrt.com 2423

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
set of fully connected layers is added to the architecture. The TN
TNR =
whole network is trained using the new dataset, and the TN + FP
weights of the extra layers are initialised at random. The
pre-trained weights speed up the training process, while the In the research, the training data were folded 10 times,
GPU speeds up the computations. or split evenly, to maximise the efficiency with which the
mapping between inputs and outputs could be learned. The
model's hyperparameters were optimised using these folds,
and the deep learning system's performance was assessed
using a variety of measures. Scores for Fine Tuned
MobileNet and VGG-19 with GPU Transfer Learning
Architecture were tallied over 10 folds for accuracy,
specificity, sensitivity, and precision. Taking into
consideration a weight loss of 0.00001, the training
procedure used ten distinct learning rates: 0.01 for the first
100 epochs, 0.001 for the next 100, and so on.

Training accounted for 70% of the time in the

Fig 3 VGG-19 Architecture with GPU Transfer Learning research, with testing taking up 10% and validation taking
up 20% of the effort. The dataset was collected from a wide
The final component of the proposed architecture is range of fields, and it was partitioned into subject-based
the integration of pose estimation and transfer learning. The training and testing sets to prevent information from one
key points obtained from the OpenCV pose estimation group from influencing the other. Accuracy stabilised
model are used as input to the VGG-19 model to correct and between 800 and 1000 iterations, Consequently, A 10-fold
estimate the yoga postures of the user. The VGG-19 model cross-validation was carried out to gauge the efficacy of the
uses the key points as features to classify the yoga posture proposed layout. with 100 iterations counted for training in
and provides feedback to the user on the correctness of their each fold.
posture. The use of a deep learning model ensures that the
system can handle variations in posture, lighting conditions, To ensure a thorough examination of the system, some
and camera angles, providing accurate feedback to the user. procedures may have been modified. However, despite the
various methodologies used, the study's findings suggest a
IV. RESULT AND DISCUSSION satisfactory resolution. Nonetheless, the dataset used in the
research could benefit from including more individuals and
The classification accuracy, precision, specificity, and videos to address the research issue adequately. For this
sensitivity of the networks are measured during the testing software to be marketable, it needs to be acceptable to
phase. Calculating classification accuracy is as simple as consumers. Key points on a 2-dimensional video provide
dividing the number of accurate guesses by the total number crucial guidance on whether or not the user is using proper
of predictions: form throughout their yoga. However, it is also essential to
know which muscles to engage during a weightlifting
𝑇𝑃 + 𝑇𝑁 exercise. Although proper form may increase the likelihood
Accuracy = 𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
of using the right muscles, other sophisticated equipment or
user input may be necessary to determine this.
Where, TP = True Positive, FP = False Positive, TN =
True Negative, FN = False Negative. The study's findings suggest that 2D Human Pose
Estimation might give frontal feedback on weight training
To determine a model's accuracy, a statistician may approaches in healthy persons. However, further research is
apply the following formula, which compares the observed required to determine whether these results can be translated
and expected proportion of correct forecasts. to side viewing angles and to improve technique detection.
Pose Trainer's successful implementation of side views
TP
Precision = suggests that dynamic time warping is preferable for such a
TP + FP technique's specifics. Recognising tasks that require a lot of
rotation might be difficult, but the algorithm still delivers
Sensitivity, the fraction of positive instances that are
useful insights. The current approach produces similar
correctly categorised, is the metric by which this rate is
results to the Pose Trainer while gaining from a larger
measured. One may express the sensitivity in mathematical
dataset and giving more detailed feedback on harder
terms as
exercises. This research expands upon earlier efforts with
Pose Trainer by demonstrating how important parts of 2D
TP
TPR = Human Pose Estimation may be integrated with more
TP + FN fundamental methodologies to offer feedback on proper
weightlifting form.
We also build true negative values, which show the
proportion of false-negative occurrences that are properly
detected based on their class, to evaluate specificity:

IJISRT23APR2280 www.ijisrt.com 2424

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Table 1 Performance Metrics for VGG-19 with Transfer outcomes is essential. However, the method used in this
Learning Architecture research successfully predicted every vantage point by
Performance Metrics calculating the direction of the vector between the two
Folds Specificit Sensitivit Accurac Precisio shoulders. With a bigger dataset and a comparable
y y y n distribution of frontal and lateral views, this approach may
Fold-I 98.78 98.95 98.61 98.86 provide the same results.
Fold-II 98.41 98.45 98.52 98.22
Fold-III 98.99 98.96 98.39 98.54
Fold-IV 98.36 98.81 98.65 98.43
Fold-V 98.74 98.11 98.73 98.26
Fold-VI 98.90 98.86 98.84 98.72
Fold-VII 98.22 97.68 97.68 96.22
Fold-VIII 97.99 97.59 97.93 97.23
Fold-IX 98.70 97.95 98.65 96.56
Fold-X 97.76 98.16 98.46 96.74
Overlappe
NULL NULL NULL NULL
d Data
Average 98.11 97.42 98.37 97.75

Fig 5 Training Accuracy Curve for Fine Tuned MobileNet

Architecture

Fig 6 Training Loss Curve for Fine Tuned MobileNet

Architecture

It's important to keep in mind these limitations and

Fig 4 Training Accuracy and Loss Curve for VGG-19 with potential sources of error when using the angle detector.
GPU Transfer Learning Further testing and improvements to the algorithm may be
needed to increase its accuracy and ability to detect a wider
It is obvious that the use of 2D Human Pose range of angles. Regarding the technique evaluation system,
Estimation and machine learning approaches to the study of it is promising that the machine learning approach used by
weight training has the potential to be a useful tool for Pose Trainer was able to accurately identify correct and
lowering the risk of injury and enhancing overall exercise incorrect form for several exercises. It should be noted,
efficacy. The findings of this study suggest that a larger however, that the research relied on a small dataset
dataset and improvements to the technique detection could consisting mostly of healthy people. Additional research and
lead to more precise results, particularly for side view testing may be necessary to determine the system's
detection. Also, it's possible that the system's dynamic effectiveness for a wider range of exercises and individuals
distortion might be better used by an approach that exploits with varying levels of fitness and health.
it. Noting that the angle prediction is a vital part of assessing
the training films and that a bad prediction might lead to bad

IJISRT23APR2280 www.ijisrt.com 2425

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
REFERENCES

[1]. Islam, Muhammad Usama & Mahmud, Hasan &

Ashraf, Faisal Bin & Hossain, Iqbal & Hasan, Md
Kamrul. (2017). Yoga posture recognition by
detecting human joint points in real time using
microsoft kinect. 668-673. 10.1109/R10-
HTC.2017.8289047.
[2]. H. Tang, Q. Wang and H. Chen, "Research on 3D
Human Pose Estimation Using RGBD Camera," 2019
IEEE 9th International Conference on Electronics
Information and Emergency Communication
(ICEIEC), 2019, pp. 538-541, doi:
10.1109/ICEIEC.2019.8784591.
[3]. Zhe Cao et al. “OpenPose: Realtime Multi-Person 2D
Pose Estimation Using Part Affinity Fields”. In: IEEE
Transactions on Pattern Analysis and Machine
Intelligence 43.1 (2021), pp. 172–186. doi:
Fig 7 Yoga Pose Estimation using VGG-19 with GPU 10.1109/TPAMI.2019.2929257.
Transfer Learning [4]. Santiago Riofrio et al. “Gesture Recognition Using
Dynamic Time Warping and Kinect: A Practical
V. CONCLUSION Approach”. In: Nov. 2017, pp. 302–308. doi: 10 .
1109/INCISCOS.2017.36.
In conclusion, the Correction and Estimation of Yoga [5]. T. L. Munea, Y. Z. Jembre, H. T. Weldegebriel, L.
Postures using Pose Estimation with OpenCV is a promising Chen, C. Huang and C. Yang, "The Progress of
technology for helping individuals improve their yoga Human Pose Estimation: A Survey and Taxonomy of
practice. This study has demonstrated the effectiveness of Models Applied in 2D Human Pose Estimation," in
using 2D Human Pose Estimation to detect and correct yoga IEEE Access, vol. 8, pp. 133330-133348, 2020, doi:
postures by providing real-time feedback to users. The Pose 10.1109/ACCESS.2020.3010248.
Trainer developed in this study has shown to be successful [6]. Jinbao Wang, Shujie Tan, Xiantong Zhen, Shuo Xu,
in identifying correct and incorrect execution of various Feng Zheng, Zhenyu He, Ling Shao,Deep 3D human
yoga postures, and in providing accurate feedback to users. pose estimation: A review, Computer Vision and
Image Understanding, Volume 210, 2021,
The findings of this study also suggest that the Pose 103225,ISSN 1077-3142,
Trainer could be useful in reducing the risk of injury https://doi.org/10.1016/j.cviu.2021.103225.
associated with weight training exercises. By providing [7]. Valcarce-Torrente M, Javaloyes V, Gallardo L,
feedback on the proper technique and muscle engagement, García-Fernández J, Planas-Anzano A. Influence of
individuals could be able to perform weightlifting exercises Fitness Apps on Sports Habits, Satisfaction, and
with proper form and minimize the risk of injury. Intentions to Stay in Fitness Center Users: An
Experimental Study. Int J Environ Res Public Health.
However, there is still much room for improvement in 2021 Oct 2;18(19):10393. doi:
this technology. The dataset used in this study could benefit 10.3390/ijerph181910393. PMID: 34639692;
from being larger and more diverse, and more research is PMCID: PMC8507994.
needed to determine whether the results from front view [8]. Li, Miaopeng & Zhou, Zimeng & Li, Jie & Liu,
detection can be translated to side view detection. Xinguo. (2018). Bottom-up Pose Estimation of
Additionally, the angle detection system used in this study Multiple Person with Bounding Box Constraint. 115-
has limitations and may not be accurate in detecting angles 120. 10.1109/ICPR.2018.8546194.
that are not directly frontal or oblique. [9]. Yadav, Santosh & Agarwal, Aayush & Kumar,
Ashish & Tiwari, Kamlesh & Pandey, Hari & Ali
Overall, this study has contributed to the growing field Akbar, Shaik. (2022). YogNet: A two-stream network
of computer vision and pose estimation and has the potential for real time multi person yoga action recognition and
to provide valuable assistance to individuals looking to posture correction. Knowledge-Based Systems. 250.
improve their yoga practice or perform weightlifting 109097. 10.1016/j.knosys.2022.109097.
exercises safely. With further research and development, [10]. S. Kinger, A. Desai, S. Patil, H. Sinalkar and N.
this technology could become a valuable tool in the health Deore, "Deep Learning Based Yoga Pose
and fitness industry. Classification," 2022 International Conference on
Machine Learning, Big Data, Cloud and Parallel
Computing (COM-IT-CON), Faridabad, India, 2022,
pp. 682-691, doi: 10.1109/COM-IT-
CON54601.2022.9850693.

IJISRT23APR2280 www.ijisrt.com 2426

Volume 8, Issue 4, April – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[11]. Kadbhane,S., Datir,K., Jagdale,T., Dhongade,S., &
Jagtap,G.(2021). Yoga Posture Recognition.
International Journal of Advanced Research in
Computer and Communication Engineering, 10(1),
143–147. doi:10.17148/IJARCCE.2021.10128
[12]. Haque, S., Rabby, A. S., Laboni, M. A., Neehal, N.,
& Hossain, S. A. (2019). ExNET: Deep Neural
Network for Exercise Pose Detection.
Communications in Computer and Information
Science Recent Trends in Image Processing and
Pattern Recognition, 186–193. doi:10.1007/978-981-
13-9181-1_17
[13]. Agrawal, Yash & Shah, Yash & Sharma, Abhishek.
(2020). Implementation of Machine Learning
Technique for Identification of Yoga Poses. 40-43.
10.1109/CSNT48778.2020.9115758.
[14]. Anilkumar, Ardra & K.T., Athulya & Sajan, Sarath &
K.A., Sreeja. (2021). Pose Estimated Yoga
Monitoring System. SSRN Electronic Journal.
10.2139/ssrn.3882498.
[15]. Luvizon, Diogo C., David Picard, and Hedi Tabia.
"2d/3d pose estimation and action recognition using
multi task deep learning." Proceedings of the IEEE
conference on computer vision and pattern
recognition. 2018
[16]. Chen, Steven & Yang, Richard. (2018). Pose Trainer:
Correcting Exercise Posture using Pose Estimation.
[17]. Zou, J., Li, B., Wang, L., Li, Y., Li, X., Lei, R., &
Sun, S. (2019). Intelligent Fitness Trainer System
Based on Human Pose Estimation. Signal and
Information Processing, Networking and Computers,
593– 599. doi:10.1007/978-981-13-7123-3_69
[18]. Yadav, Santosh & Singh, Amitojdeep & Gupta,
Abhishek & Raheja, Jagdish. (2019). Real-time Yoga
recognition using deep learning. Neural Computing
and Applications. 31. https://link.springer.com
/article/10.1007/s00521-019. 10.1007/s00521-019-
04232-7.
[19]. Vivek Anand Thoutam, Anugrah Srivastava, Tapas
Badal, Vipul Kumar Mishra, G. R. Sinha, Aditi
Sakalle, Harshit Bhardwaj, Manish Raj, "Yoga Pose
Estimation and Feedback Generation Using Deep
Learning", Computational Intelligence and
Neuroscience, vol. 2022, Article ID 4311350, 12
pages, 2022. https://doi.org/10.1155/2022/4311350
[20]. S. K. Yadav, A. Singh, A. Gupta, and J. L. Raheja,
“Real-time Yoga recognition using deep learning,”
Neural Computing & Applications, vol. 31, no. 12,
pp. 9349–9361, 2019.

IJISRT23APR2280 www.ijisrt.com 2427