10 VII July 2022

https://doi.org/10.22214/ijraset.2022.45215
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VII July 2022- Available at www.ijraset.com

Gym Trainer Application using Machine Learning.

Sathvik K Gatti1, Rohit P2, Sameer A3, Dr. DJ Ravi4
1, 2, 3, 4
Department of Electronics and Communication Engineering, Vidyavardhaka College of Engineering, Mysuru, India

Abstract: Human pose estimation localizes body key points to accurately recognizing the postures of individuals given an image.
This step is a crucial prerequisite to multiple tasks of computer vision which include human action recognition, human tracking,
human-computer interaction, gaming, sign languages, and video surveillance.
Fitness exercises are very beneficial to personal health and fitness; however, they can also be ineffective and potentially
dangerous if performed incorrectly by the user. Exercise mistakes are made when the user does not use the proper form, or pose.
In our work, we introduce Pose Trainer, an application that detects the user’s exercise pose and provides personalized, detailed
recommendations on how the user can improve their form. Pose Trainer uses the state of the art in pose estimation to detect a
user’s pose, then evaluates the vector geometry of the pose through an exercise to provide useful feedback. We record a dataset
of over 100 exercise videos of correct and incorrect form, based on personal training guidelines, and build geometric heuristic
and machine learning algorithms for evaluation. Pose Trainer works on four common exercises and supports any Windows or
Linux computer with a GPU.
Keywords: GPU, Machine learning algorithm, deep learning, human pose estimation, image processing, neural networks, blaze
pose, move net.

I. INTRODUCTION
Fitness is a trend today. Every year the revenue of the fitness industry grows by 8.7%, according to the Wellness Creatives report,
and fitness apps have not spared this field.
There are many cases of how technologies might help to improve your body - from tracking exercise activity to adjusting nutrition.
The question is how much better can such apps help improve the performance of physical exercises compared to human coaches?
Artificial Intelligence (Al, a broad name for a group of advanced methods, tools, and algorithms for automatic execution of various
tasks) has invaded practically all functional areas of business over the years. Pose estimation is among the most popular solutions
that AI has to offer; it is used to determine the position and orientation of the human body given an image containing a person.
Unsurprisingly, such a useful tool has found many use cases, for instance it can be used in avatar animation for Artificial Reality,
marker less Motion Capture, worker pose analysis, and many more.
With the arrival of human pose estimation technology, the fitness technology market has been filling up with AI-based personal
trainer apps. Being powered by computer vision, human pose estimation and natural language processing algorithms, these
technologies lead end-users through a number of workouts and give real-time feedback.
Exercises such as squats, deadlifts, and shoulder presses are beneficial to health and fitness, but they can also be very dangerous if
performed incorrectly. The heavy weights in-volved in these workouts can cause severe injuries to the muscles or ligaments. Many
people work out and perform these exercises regularly but do not maintain the proper form (pose). This could be due to a lack of
formal training through classes or a personal trainer, or could also be due to muscle fatigue or using too much weight. For our
course project, we seek to aid people in performing the correct posture for exercises by building Pose Trainer, a software application
that detects the user’s exercise pose and provides useful feedback on the user’s form, using a combination of the latest advances in
pose estimation and machine learning. Our goal for Pose Trainer is to help prevent injuries and improve the quality of people’s
workouts with just a computer and a webcam. The first step of Pose Trainer uses human pose estimation, a difficult but highly
applicable domain of computer vision.

II. LITERATURE REVIEW

1) Alexander Toshev Christian Szegedy, from google had published this paper. In this approach, pose estimation is formulated as
a CNN-based regression problem towards body joints. The model consisted of an Alex Net backend (7 layers). The model is
trained using L2, loss for regression. This approach reasons about pose in a holistic fashion, i.e., even if certain joints are
hidden, they can be estimated if the pose is reasoned about holistically. The paper argues that CNNs naturally provide this sort
of reasoning and demonstrate strong results.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 978
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VII July 2022- Available at www.ijraset.com

2) Joao Carreira, Pulkit Agrawal, Katerina Frigidaria, Jitendra Malik. This is an interesting paper that follows method that predicts
the estimates and then correct them iteratively. Instead of directly predicting the outputs in one go, they use a self-correcting
model that progressively changes an initial solution by feeding back error predictions, and this process is called Iterative Error
Feedback (IEF). It employs a standard Conv Net architecture pre-trained on ImageNet: the very deep google net.
3) Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun, Christoph Bregler New York University. This approach generates
heatmaps by running an image through multiple resolution banks in parallel to simultaneously capture features at a variety of
scales. The output is a discrete heatmap instead of continuous regression. A heatmap predicts the probability of the joint
occurring at each pixel. A multi-resolution CNN architecture (coarse heatmap model) is used to implement a sliding window
detector to produce a course heatmap output.
4) Mihai Fieraru Anna Khoreva Leonid Pishchulin Bernt Schiele in their paper “Learning to Refine Human Pose Estimation” have
given insights on refining the human pose estimation where It takes the help of RGB image and body pose estimate as input to
render out the suitable output and Exploiting the dependencies between the image and the predicted body pose makes it easier
for the model to identify the errors in the initial estimate and how to refine them along with methods to achieve state-of-the-art
results on the challenging MPII Human Pose and Pose Track benchmarks.
5) Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh in their paper “Open-Pose: Realtime Multi-Person 2D
Pose Estimation using Part Affinity Fields” suggested the usage of 2D pose of multiple people in an image where
“OPENPOSE” 2D human pose estimation as an input for their systems, where they used COCO human pose datasets for
training data sets and getting the desired output.
6) Simen Thys, Wiebe Van Ranst,Toon Goedem's in their paper “Fooling automated surveillance cameras: adversarial patches to
attack person detection” Object detection of this paper they target the popular YOLOv2 object detector, An example of real-
world adversarial attack. Sharif et al. demonstrate the use of printed eyeglasses that can be used to fool facial recognition
systems for making the patches more robust. Thus, vulnerability of certain neural networks has been exposed.
7) Manisha Patel*, Dr Nilesh Kalani, basic understanding and elementary concept of CNN, deep learning approaches for human
pose estimation. Open pose is a real-time multi-person system to detect the human body key points on a single image. CNN
based regression problem towards body joints in a holistic fashion. To resolve the problem of vanishing gradient Intermediate
supervision was applied at every stage in stacked hourglass
8) Bruno Artacho Andreas Savakis Rochester Institute of Technology Rochester, NY. The UniPose pipeline utilizes the WASP
module that features a waterfall flow with a cascade of atrous convolutions and multi-scale representations. The large FOV of
WASP obtains a better interpretation of the contextual information in the frame, and contributes to more accurate pose
estimation.

III. METHODOLOGY: DEEP LEARNING EQUIPPED FITNESS TRAINER

In general, our project aims to give a personal fitness training experience. It is powered with computer vision that captures the real
time video feed of the person and processes and analyses it using deep learning techniques to estimate the body posture of the
person and provides suitable and corrective feedback to the user to perform exercises efficiently. We achieve this by using
estimating the human pose and then marking the landmarks on it. Once we have the landmarks, we then form vectors in order to
analyze the posture of the person. We employ simple geometric heuristic approach in order to evaluate each exercise. We currently
provide exercise module for bicep curl, front raise, squats, shoulder shrugs, lateral raise.

Figure 1: General operating flow chart.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 979
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VII July 2022- Available at www.ijraset.com

For the pose estimation component, we utilize a pre-trained real-time system, called Blaze Pose, that can detect human body key
points in videos. (More detail on BlazePose and our reasoning for choosing this system can be found in Technical Approach.) This
model is functional out of the box, and thus is very simple to install for users of our application. Using BlazePose allows us to take
advantage of the state of the art in pose estimation algorithms for our task, and lets us focus on the actual evaluation of exercise
posture. For the posture evaluation (pose training) component, we have recorded videos of ourselves performing exercises. Our
videos include our best effort to correctly perform the exercise, as well as intentionally incorrect examples. The evaluation of our
posture identifier is dependent on the performance of the pose estimator. We work under the assumption that the pose estimator is
accurate a majority of the time, with small measurement deviances due to noise, which we correct for. We evaluate our posture
identifier in different ways depending on the algorithm: for heuristic. algorithms, we feed in all videos for evaluation, while for
machine learning algorithms, we evaluate by splitting our video dataset into train and test sets, and report results on the test set.
To design the basic Curl counter we have to select three points from the basic 33 key-points that are responsible for the movement
of the body.

Figure 2: Human Pose Estimation (HPE)

Here we select 3 points, shoulder joint X1 Y1, Elbow joint X2 Y2, Wrist joint X3 Y3

Figure 3: Key-points used in curl counter

The main process to recognise the curl counter is to check the angle shift within these three points and a change of these angles
from 180 degree to less than 45 degree and back from 45 degree to 180 degree will count as one curl.

There are two positions considered in this design:

1) UP State - 180 degree to 45 degree
2) DOWN State - 45 degree to 180 degrees
We select trigonometric function - tan for this recognition as it takes inputs opposite by adjacent. But during the process of
using tan, we’ll have drawbacks when the opposite and hypotenuse is one by one. When the sin is positive and cos is zero, it
becomes undefined.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 980
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VII July 2022- Available at www.ijraset.com

Figure 4: Atan function

Thus we added a new constant, ‘a - arctangent’ to remove this instability. We also got this function working from -pi/2 - +pi/2.

Figure 5: Atan function 2

In ‘a tan2’ we get a much bigger range of written values, so that it can work for a versatility process. Therefore we chose this
function. This function returns the value in range 0 - 2pi.

Figure 6: Atan2 function

We used Tlinker for building an interactive UI/UX for the application.

IV. RESULTS
1) Successfully implemented detection of human pose landmarks using media-pipe.
2) Able to extract the pose landmarks to cartesian coordinates for analysis.
3) Implemented a simple yet effective mathematical technique to find the angles between three joints.
4) Applied simple heuristic approach to figure out a way to recognise the states of exercises.
5) Integrated a simple User Interface to display exercise count and state information.

Figure 7: Curl down state

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 981
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue VII July 2022- Available at www.ijraset.com

Figure 8: Curl up state

Similarly, successful in implementing lateral raise and squat exercises in our application.

V. CONCLUSION
A. Successfully implemented a full-fledged basic gym trainer that can detect, analyse the human position.
B. Implemented a way to recognise the exercise and to count the repetitions in a more efficient way.
C. Further extended the work to implement even more advanced and capable techniques.

REFERENCES
[1] Alexander Toshev Christian Szegedy Google DeepPose: Human Pose Estimation via Deep Neural Networks (CVPR’14).
[2] Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun, Christoph Bregler New York University Efficient Object Localization Using Convolutional
Networks (CVPR’15)
[3] Joao Carreira,Pulkit Agrawal ,Katerina Fragkiadaki, Jitendra Malik.Human Pose Estimation with Iterative Error Feedback.
[4] Mihai Fieraru ,Anna Khoreva Leonid ,Pishchulin Bernt Schiele. ‘Learning to Refine Human Pose Estimation’, Max Planck Institute for Informatics,Saarland
Informatics Campus,Saarbrücken, Germany.
[5] Zhe Cao, Student Member, IEEE, Gines Hidalgo, Student Member, IEEE,Tomas Simon, Shih-En Wei, and Yaser Sheikh. ‘OpenPose: Realtime Multi-Person
2D Pose Estimation using Part Affinity Fields’.IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE.
[6] Simen Thys,Wiebe Van Ranst,Toon Goedem´e. ‘Fooling automated surveillance cameras:adversarial patches to attack person detection’.EAVISE, Technology
Campus De Nayer, KU Leuven, Belgium.
[7] Zhao Liu, Jianke Zhu, Jiajun Bu, Chun Chen. A survey of human pose estimation: The body parts parsing based Methods - Elsevier publications.
[8] Naimat Ullah Khan, Wanggen Wan - A Review of Human Pose Estimation from Single Image .
[9] Valentin Bazarevsky Ivan Grishchenko Karthik Raveendran Tyler Zhu

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 982