International Journal of Recent Advances in Multidisciplinary Topics 119

Volume 3, Issue 6, June 2022

https://www.ijramt.com | ISSN (Online): 2582-7839

Implementation of Geriatric Agility Detection

Using MediaPipe Pose
Gagandeep Kaur1, Gunjan Jaju2*, Devansh Agarwal3, Krrish Iyer4, C. M. Prashanth5
1,2,3,4
Student, Department of Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India
5
Professor, Department of Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India

Abstract: It is crucial to identify frailty in the elderly population We propose using media-pipe's blaze pose machine learning
early and accurately to prevent falls and muscle movement loss model for this application, which can run on a small number of
This makes early intervention possible as determined appropriate
processing-powered machines. Human posture estimation has
by the medical professionals and prolongs muscular activity.
Following the Covid-19 outbreak, an enormous. The ageing improved significantly as a result of the rapid development of
population has been restricted to their dwellings, which has deep convolutional neural networks, which helps to accurately
significantly impacted their physical health extent. The elderly is assess the exercise the user is doing. Human pose estimation
more vulnerable to these impacts mostly as a result of lengthy locates body vital points in order to accurately distinguish
lockdowns and travel restrictions also a decline in social support. people's postures in photos or videos. This is a necessary step
The research aims to develop a technique to aid older people in
before beginning the workout analysis. We offer a method for
determining their level of frailty Geriatric Agility Detection with
Early and Accurate 30 second chair-stand test. Medical care is aiding in the resolution of this challenge by sensing a user's
described as geriatric. Agility, a term used to describe the ability body position throughout a workout and comparing it to a
to move quickly and freely is directly related to the reduction of computed angle between limbs while keeping a good form for
falls among older people. the above said exercise. We represent the human body as a
collection of limbs and calculate angles between them to detect
Keywords: Geriatric, Agility, Frailty, Reduction of falls, 30
incorrect posture and give accuracy to the user, based on the
second chair stand test.
current breakthrough in deep learning for human body posture
1. Introduction estimation. Other elements of the blazpose model [5] include
repetition counting and real-time body movement
According to WHO, the number and proportion of people identification.
aged 60 years and older will increase to 1.4 billion by 2030 and The objective of our study was to apply the current strategy
2.1 billion by 2050 [1]. Therefore, proper diagnosis and to detect frailty and make it accessible to the larger population
rehabilitation programs need to be designed to prevent of old people. The approach is hassle free, self-accessible and
immobility and postural instability, decreasing the risk of falls allows a user make comparative study without the need of a
and downstream decline on a frequent and consistent basis. technician making them more independent. The software can be
With the ongoing pandemic, the likelihood of age-related used on a laptop or camera and can be viewed as a form of
cognitive and functional problems have increased significantly cumulative solution to achieve better physique and reduce
due to decrease in physical activity and lack of accessing potential risks of injury.
services [2]. Thus, we introduce an economical and easy to use
method to help the older and frail population self-assess and 2. Background of the Work
deal with such problems without the requirement of a healthcare
All traditional approaches involve the use of expensive
provider.
The method revolved around the results obtained after hardware to capture and evaluate the results of the 30 Chair
Stand Test. These hardware components do provide a deep
performing the 30 second chair stand test [3] to help determine
the level of agility of a user according to their age and gender. understanding of the results obtained when the geriatric test is
This test is part of the Senior Fitness Test Protocol [4], and is performed by various participants. But their major drawback is
that it makes these tests difficult to perform without technical
designed to test the functional fitness of seniors. The user sits
in the middle of the seat, with their feet shoulder width apart, assistance and in a clinical environment. Below is the flow-
flat on the floor. The arms are to be crossed at the wrists and diagram explaining the traditional Geriatric-agility 30 second
chair stand test (30 CST) detection using hardware components.
held close to the chest. From the sitting position, the subject
stands completely up, then completely back down, and this is
repeated for 30 seconds. We count the total number of complete
chair stands (up and down equals one stand).

*Corresponding author: gunjanjajubkn@gmail.com

Kaur et al. International Journal of Recent Advances in Multidisciplinary Topics, VOL. 3, NO. 6, JUNE 2022 120

recent model, and it works well on lightweight devices like a

browser or a mobile device. As a result, BlazePose can be used
to estimate a single pose or a set of poses.
Below diagram shows the methodology we adopted to
accurately determine the number of sit-up counts while
performing 30 CST.

Fig. 1. Flow diagram of traditional geriatric test using pressure plates and
sensors

Later, the concept of human detection using Artificial

Intelligence and Machine Learning came into existence. Human
pose estimation is the process of inferring poses from
photographs or videos, and it may be done in both 2D and 3D.
We found a number of techniques for recognizing poses from
previous years. We looked at a number of studies from 2017
and noticed that the topic of machine learning is attracting an Fig. 2. Flow diagram of geriatric agility test using computer vision
increasing amount of interest from academic’s communities.
Many of the researchers advocated that the new system work in What is BlazePose?
conjunction with previous technologies to provide the state-of- Blaze pose is a real-time posture identification approach for
the-art results. Deep convolutional neural networks are now the detecting human poses in images and videos. It only works in
most popular solutions. There are two common approaches: one mode (single human pose detection). In simple terms, the
Regression the placement of key-points and the estimation of flame posture is a deep learning model that estimates human
key-point heat maps, then selecting the sites with the highest pose by identifying body parts such as elbows, hips, wrists,
heat values as candidates for the main points. knees, and ankles, and then combining these points to construct
The paper produced by Nora Millor and others discusses the a skeletal structure of your pose.
relevance of the 30-s Chair Stand Test as an important It's a simple model that uses depth-wise separable
parameter to discriminate among healthy, prefrail and frail convolution to deepen the network while lowering parameters,
populations using sensors such as accelerometers and increasing accuracy, and lowering computing costs. From the
gyroscopes [6]. nose to the left foot index, the Blaze position provides us with
In the year 2019, Xie et al introduced Open Pose, a human a total of 33 essential points that we can exploit as shown in the
pose estimation method. However, its efficiency is really low. below diagram [10].
To solve this, deep learning algorithms for recognizing human
body positions based on TensorFlow came into picture [7].

3. Methodology
Pose estimation is a machine learning challenge that guesses
the geographical locations of various body components to
estimate a person's pose from an image or video (key points).
Pose estimation is a computer vision technique for tracking a
person's or an object's movements. This is normally
accomplished by locating critical spots for the things in
question. We may compare different actions and postures and
draw conclusions based on these essential points. Later in this
section, we will be discussing how media-pipe works in the
background and how we exploited its power to boost our Fig. 3. 33 pose landmarks
application design.
Deep learning has been shown to outperform traditional Working of BlazePose:
computer vision approaches in a variety of tasks, including Detector and Estimator are two machine learning models
image segmentation and object detection, thanks to the rapid used by Blaze pose. The detector is used to extract the human
growth of deep learning solutions in recent years. As a result, portion of an image. The images supplied by the detector are
deep learning techniques have resulted in considerable fed into an estimator, which produces output.
advancements and improvements in posture estimation jobs. Human body Modelling:
Openpose [8], movenet, deeppose, posnet, bodynet [9], and The placement of human body parts is utilized to generate a
more deep learning estimation techniques are available. In our human body representation (such as a body skeleton posture)
instance, we're utilizing a Blazepose since it's Google's most from visual input data in human pose estimation. As a result,
Kaur et al. International Journal of Recent Advances in Multidisciplinary Topics, VOL. 3, NO. 6, JUNE 2022 121

human body modelling is an important part in estimating human necessary for videos, where detection runs once followed by
position. It's used to represent extracted characteristics and key landmark tracking. The default value is False.
points from visual input data. To define and infer human body MODEL_COMPLEXITY: Default value is 1. It can be 0, 1,
postures and generate 2D or 3D poses, a model-based technique or 2. If higher complexity is chosen, the inference time
is typically utilized. We employ an N-joints stiff kinematic increases.
model in which a human body is represented as an entity with ENABLE_SEGMENTATION: If set to True, the solution
joints and limbs that contains body kinematic structure and generates a segmentation mask along with the pose landmarks.
shape information. The default value is False.
The skeleton-based model, also known as the kinematic MIN_DETECTION_CONFIDENCE: Ranges from [0.0 –
model, is used for both 2D and 3D pose estimation. To reflect 1.0]. As the name suggests, it is the least confidence value for
the human body structure, this adaptable and intuitive human the detection to be considered valid. The default value is 0.5.
body model comprises a variety of joint configurations and limb MIN_TRACKING_CONFIDENCE: Ranges from [0.0 –
orientations. To record the relationships between different body 1.0]. It is the minimum confidence value for a landmark to be
sections, skeleton pose estimation models are used. However, considered tracked. The default value is 0.5.
as demonstrated in figure, kinematic models have limitations Step 3: Calculation of Body Pose angles - We use the
when it comes to capturing texture or shape information. shoulder-elbow-wrist and shoulder-hip-knee body points
marked by MediaPipe to first calculate the offset distance and
4. System and Algorithm Design then inclination to check for the right form during the exercise.
We create a user guide application that uses the cv2 module Offset is calculated using the distance formula:
[11] to capture and detect body poses in order to run the
programme. The MediaPipe framework and Python
programming language are being used to create an application where x and y are x-coordinate and y-coordinate of the given
user guide. body landmark.
There have been numerous exciting research We use the vector approach to find the inner angle of three
accomplishments using OpenCV as a module for articulating points. The angle between two vectors P₁2 and P₁3 is given by:
human body tracking, position, and even recognition systems.
We capture a real-time image with the laptop's camera, which 0= arctan(P1P2/P1P3)
has a resolution of 640x240 pixels, to process using the
MediaPipe framework. Objects for video capture and video
writer are constructed. We get the video metadata that will be
utilized to make the video capture object. MediaPipe will read
the image from OpenCV, do body detection, and create body
landmarks, which will return 3D body key points and joints,
resembling a skeleton. 3D key points in the body that have been
marked in the image will be computed and initialized as a tool
for reading pose body and recognition based on body pose that
has been initialized before. Below flowchart shows how we
have developed an effective algorithm that suits our needs to its
best.
Fig. 5. Vector approach to calculate angle
Pose detection algorithm:
Input- Picture frames captured from the default camera of the The angle formed by the shoulder, elbow, and wrist lines in
device relation to the y-axis is used. The important location in this case
Output- Sit Up count, performance feedback and exercise to is the elbow. The hip line, likewise, connects the knee and the
improve agility. shoulder, with the hip serving as a crucial point.
Step 1: The frames from the camera are converted to RGB Step 4 - Conditionals to check for correct body form - On the
format as Python captures the frame in BGR format. RGB basis of the posture, whether good or bad, sit up count is
format is required for pre-processing of the images. calculated and results are displayed. We compare the angle we
Step 2: Getting Key points - We choose the default values for calculated in Step 2 and compare it with threshold angles we
detecting pose landmarks and set the ENABLE calculated based on different researches.
SEGMENTATION flag to true in the Pose (adjustable)’s API We know knee height is approx. 49.6 cm in both the genders
as we want the utility to generate a segmentation landmark and and 30-CST is administered using a folding chair without
default values do a good job. armrest, with seat height of 17 inches (43.2 cm) [12]. We use
The following are some of the MediaPipe pose solution's these two heights to determine the best predictor angle using the
configurable APIs. formula,
STATIC_IMAGE_MODE: It’s a boolean. If set to True, O = arccos(AC.AB)/(|AC||AB|)
person detection runs for every input image. This is not O is the Approx Threshold Angle which is 110 degrees.
Kaur et al. International Journal of Recent Advances in Multidisciplinary Topics, VOL. 3, NO. 6, JUNE 2022 122

Table 1 Table 2
Male test subjects Female test subjects
Age Below average Average Above average Age Below average Average Above average
60-64 < 14 14 to 19 > 19 60-64 < 12 12 to 17 > 17
65-69 < 12 12 to 18 > 18 65-69 < 11 11 to 16 > 16
70-74 < 12 12 to 17 > 17 70-74 < 10 10 to 15 > 15
75-79 < 11 11 to 17 > 17 75-79 < 10 10 to 15 > 15
80-84 < 10 10 to 15 > 15 80-84 <9 9 to 14 > 14
85-89 <8 8 to 14 > 14 85-89 <8 8 to 13 > 13
90-94 <7 7 to 12 > 12 90-94 <4 4 to 11 > 11

Table 3
Agility based training routine
Session 1 to 5 Exercises Walking based: change of directions, cuts, obstacle crossing, bench balancing.
Double leg line jumps, bench step ups

Duration ~30s per exercise, 2 rounds, 5 exercises

Variations Elevated balancing, obstacle height, speed, plane of moment
Session 6 to 10 Exercises Walking based: change of directions, cuts, obstacle crossing, line balancing, swiss cross and combinations of these.
Unstable lunges, ball dribbles.

Duration ~30 s per exercise, 2 rounds, 7 exercises.

Variations Color coding of movement directions, all of the previous
Session 11 to 15 Exercises Walking based: change of directions, cuts, rotations, obstacle crossing, line balancing, swiss cross and combinations of
these.
Single leg line jumps, bench step ups, orientation games

Duration ~30 s per exercise, 3 rounds, 7 exercises

Variations Sound coding of movement tasks and directions, walking modes, all of the previous.
Session 16 to 20 Exercises Walking based: obstacle crawl and crossing, catch through the course, beam balancing, orientation games, hurdling
Sprints, lunges, lateral double leg jumps, bench step-ups.

Duration ~30 s per exercise, 3 rounds, 8 exercises

Variations All of the previous
Session 21 to 24 Exercises Walking based: Song with coded tasks on words, catch through the course, orientation reaction games.
Lateral and forward single leg jumps, skipping on unstable surfaces, lunges, ball throw and catch exercises.

Duration ~30 s per exercise, 3 rounds, 8 exercises

Variations Cognitive tasks, ball dribbling while performing the tasks, all of the previous.

Table 4
Traditional agility routine
Session 1 to 5 Exercises Static balance exercises in double leg stance
Squats, calf raises, supported split squats

Duration 30s exercise,30 s pause,3 rounds,8 exercises

Variations Perturbations
Session 6 to 10 Exercises Static balance exercises in step or tandem stance.
Squats, calf raises, side lunges, split squats, crunches

Duration 50s exercise,30 s pause,2 rounds,8 exercises

Variations Perturbations
Strength exercises o slightly unstable surface
Session 11 to 15 Exercises Static and dynamic balance exercises in step or tandem stance
Step up, calf raises, side lunges, split squats

Duration 35s exercise,30 s pause,3 rounds,8 exercises

Variations Cognitive tasks, unstable surfaces, arm balance
Strength exercises on slightly unstable surface
Session 16 to 20 Exercises Balance exercise in step on tandem stance
Planks, squats, single-leg calf raises, side lunges, split squats, Bulgarian

Duration 40s exercise,20 s pause,3 rounds,8 exercises

Variations Cognitive tasks, perturbations, arm balance
Session 21 to 24 Exercises Dynamic balance exercises in tandem or single leg stance
Squats, single-leg calf raises, step-ups, Bulgarian split squats

Duration 55s exercise,20 s pause,3 rounds,8 exercises

Variations Unstable surfaces, perturbations, arm balance
Kaur et al. International Journal of Recent Advances in Multidisciplinary Topics, VOL. 3, NO. 6, JUNE 2022 123

Step 5 - With every detection, the sit-up counters are

incremented for good posture and ignored for bad postures
respectively. A sit-up is made up of up and down phases so the
final count can be obtained by dividing the number of up-down
phases with 2.
Step 6 - Calculation of geriatric score - Once the total count
is obtained, we refer to Table 1 and Table 2 to display the result
of the test. [13]
Step 7 - Suggest exercise routine according to score - If the
user’s agility score is below average, the training routine is
displayed according to the table given below. [14]
If the user’s agility score is above average, a message saying
“You seem healthy” is presented. The suggested routine
depends on the total sit-up count and includes the exercises,
duration and its variations.

5. Results and Discussion Fig. 8. Test subject during up phase

To check the working of our algorithm we perform this test
on a 22-year-old female and a 23-year-old male subject. Since
this application is made for academic purposes and improved
and extension of our previous study [15], we performed the test
on young subjects only though the relevance of the application
will make more sense if test trials were performed on older
subjects.
As shown in the figure, the female subject enters her name,
age and gender and starts performing the chair stand test for 30
seconds.
Fig. 9. Result showing performance feedback

The capturing screen displays the timer to show the number

of seconds left and also the number of sit-ups performed. After
the test is over, the feedback of how the subject performed the
test is displayed as shown in the diagram. Since the subject’s
geriatric score is above average, we display the message - “You
seem healthy!”.
Fig. 6. Input from the user In another trial, a male subject performs this test. As shown
in the diagram, the application works similarly in this case as
well. Since the subject’s geriatric score is below average, the
screen also displays agility based routine exercises.
We can infer from our trials that Mediapipe can be
effectively used for Body Pose detection. The application works
fine even with dim light conditions. The output can be
interpreted by another user and does not require any technical
knowledge or knowledge about how agility detection works in
clinical settings.

Fig. 7. Test subject during down phase Fig. 10. Input from the user
Kaur et al. International Journal of Recent Advances in Multidisciplinary Topics, VOL. 3, NO. 6, JUNE 2022 124

activities, sign language recognition, and full-body gesture

control, human position estimation from video is crucial. It can
be used as the foundation for yoga, dance, and fitness
applications. For inference, most current state-of-the-art
algorithms rely on powerful desktop environments, whereas our
method delivers real-time performance on most recent mobile
phones, desktops/laptops, in Python, and even on the web. Our
application can be used for monitoring workouts without any
involvement of a trainer and can help elderly become
independent. This application offers features such as real-time
analysis, injury prevention and getting best results. The system
is limited for workout purposes with single-person
compatibility at a time.

References
[1] WHO, https://www.who.int/health-topics/ageing#tab=tab_1
Fig. 11. Test subject during down phase
[2] Frontiers in psychiatry
https://www.frontiersin.org/articles/10.3389/fpsyt.2020.577427/full
[3] CDC’s STEADI tools and resources “ASSESSMENT 30-Second Chair
Stand.”
[4] Birgitta Langhammer and all, The Senior Fitness Test, Journal of
Physiotherapy, June 2015.
[5] Valentin Bazarevsky and all, On-device, Real-time Body Pose Tracking
with MediaPipe Blaze Pose, Google AI Blog, 2020.
[6] N. Millor et al., "Automatic Evaluation of the 30-s Chair Stand Test Using
Inertial/Magnetic-Based Technology in an Older Prefrail Population," in
IEEE Journal of Biomedical and Health Informatics, vol. 17, no. 4, pp.
820-827, July 2013.
[7] Xie, Ling, and Xiao Guo. "Object Detection and Analysis of Human Body
Postures Based on TensorFlow." 2019 IEEE International Conference on
Smart Internet of Things (SmartIoT). IEEE, 2019. (Xie, 2019).
[8] Varol, G. et al. (2018). BodyNet: Volumetric Inference of 3D Human
Body Shapes. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y.
(eds) Computer Vision – ECCV 2018. ECCV 2018. Lecture Notes in
Computer Science, vol. 11211. Springer, Cham.
[9] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, Yaser Sheikh.
OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity
Fields. Cs.CV, 2018.
[10] Pose - mediapipe github repository (online)
https://google.github.io/mediapipe/solutions/pose.html
[11] Medium website (online)
Fig. 12. Test subject during up phase https://medium.com/@gowtham180502/motion-detection-and-tracking-
using-opencv-python-ea75ef927e72
[12] PhysioPedia website (online)
https://www.physio-pedia.com/30_Seconds_Sit_To_Stand_Test
[13] Toped sports website (online)
https://www.topendsports.com/testing/tests/chairstand.htm#:~:text=Cou
nt%20the%20total%20number%20of,chair%20stands%20in%2030%20s
econds
[14] Eric Lichtenstein, Agility-based exercise training compared to traditional
strength and balance training in older adults: a pilot randomized trial,
Fig. 13. Result showing performance feedback
Conference: Peer J. 2020; 8: e8781.
[15] G. Kaur, G. Jaju, D. Agarwal, K. Iyer, and C. M. Prashanth, “Geriatric
6. Conclusion Agility Detection Using Python and OpenCV”, in International Journal of
Recent Advances in Multidisciplinary Topics, vol. 3, no. 4, pp. 109–113,
Detecting and analyzing body language is gaining a lot of Apr. 2022.
attention lately. In applications like measuring physical