Table of Contents
1. Introduction............................................................................................................2
1.1. Academic Questions.......................................................................................3
1.2. Aim & Objectives.............................................................................................3
1.3. Artefact............................................................................................................4
1.4. Scope and limitations......................................................................................5
1.5. Report Structure..............................................................................................5
2. Literature review.....................................................................................................6
2.1. Human Pose Estimation..................................................................................6
2.2. Pose Estimation Frameworks.........................................................................6
2.2.1. OpenPose.................................................................................................6
2.2.2. Pose Net...................................................................................................7
2.3. Underlying math behind Pose Detection:........................................................7
2.3.1. Part Affinity Fields:....................................................................................8
2.4. Single Network Body to foot estimation..........................................................9
2.5. Data Collection with OpenPose....................................................................10
2.6. Algorithm for Pose Classification..................................................................10
2.7. Similar Systems.............................................................................................12
2.8. Comparision Table........................................................................................13
2.9. Analysis and findings.....................................................................................13
1. Introduction
Human posture estimation (HPE) is a noteworthy affair in Computer vision to find
joint positions. It has been in focal point of scientists basically due to its huge
applications in different significant fields like human PC collaboration, activity
acknowledgment, observation, picture understanding, danger expectation, and so
forth. Posture estimation alludes to CV methods that recognize human figures in
pictures and video, with the goal that one could decide, for instance, where
somebody's elbow appears in a picture. Existing datasets for learning of postures
are seen to be not testing enough as far as posture pose variety, object impediment
and view focuses. This makes the HPE process moderately basic what's more, limits
the use of the models that have been prepared on them. All things considered, this
innovation isn't perceiving the character in the image as there is no distinctive
detection of any data related to pose recognition. The calculation is just evaluating
where key body joints are. Posture estimation has numerous utilizations, from
intelligent establishments that respond to the body to expanded reality, activity,
fitness uses, and that's just the beginning.
Consistently a wide range of individuals are influenced from different sorts of
muscular skeletal disorders because of mishaps or maturing issue. Yoga can
promote positive physical change. Studies have indicated that yoga is effective in
overseeing side effects related with musculoskeletal clutters including osteoarthritis
carpal passage syndrome, hyper kyphosis and low back torment. Moreover, Yoga
shows significant upgrades in engine aptitudes and physiological methods that
incorporates pulse, heart speed rate and human body weight have been noted.
Exploration additionally suggests that cardiopulmonary benefits of yoga incorporate
improvement of cardiorespiratory wellbeing just as improved force expiratory volume
and expanded indispensable limit. Traditionally, yoga is done in a yoga place within
the sight of a yoga trainer who can manage the patient through helpful assistance.
This solution recommends an abstruse yoga preparing approach without the trainer
where the patient can remain in front of a device and perform yoga, presents
effectively without the need of coach or being present at the yoga centre. Ongoing
human pose recognition is a developing and significant region which specializes on
acknowledging and understanding the human posture from depth pictures.
Academic Questions
What are the optimum algorithm to detect human poses?
What can be done to classify poses in real-time?
What are the concern about data security when implementing data
demanding project like this?
Aim & Objectives
Aims
The aim of this project is to develop a system that helps in classifying yoga postures
in real time followed by text generated guidance to develop an accurate and fast
support system in making people more responsible toward their health.
Objectives:
The objectives set to achieve the aims of the project are:
To research on areas like Convolutional Neural Network, Computer Vision,
pose detection, Open source libraries
To understand how the algorithms works and which algorithms works best for
Yoga pose detection
Research on how the classification task carried out in Computer Vision and
using which algorithm best result can be achieved for Yoga dataset
Pre-processing the dataset as per using algorithms. Since the nature of
attributes needed for different algorithms may vary.
Visualize data in different form in order to get the correlation of used
attributes.
Should integrate prediction system with the web application to obtain the final
project.
To test the software for fulfilment of requirements
Artefact
This mobile friendly application is called “YogSathi” is your Virtual Yoga Instructor. It
is powered with Convolution Neural Network an algorithm based on Computer Vision
which helps on detecting and classifying various yoga postures. The application will
also give directions to perform various poses to user like a yoga session or
discourse. It will also assist user with text generated guidance. The mission of this
application is to reduce stress, anxiety and depression and make people focused on
healthy activities like yoga without having to worry to go on different places for the
same purpose.
The app can be divided into two major parts:
a. Web based system
b. Deep Learning model that detects human postures
Artefact 1: Web Based system
The system will have following features:
• User can register in the system
• User can login in the system
• User has a dashboard to navigate sub sections
• System has series of yoga called sessions
• User can customize yoga sessions
• User can see their progress
• System shows pictorial guide to user
• System detects various yoga poses
• System assists via generated text
Artefact 2: Deep Learning model that detects human postures
Scope and limitations
The scope of Yoga Pose classification system can is to help people to correct their
postures and other cardiovascular diseases that are especially present among old
people. Also, it is available any time without trainer or without having to go to a
specific yoga centre. To improve overall health, it also helps children to learn yoga,
improve their height and posture from a very young age which will blossom their
adulthood. “Healthy is wealth” so this app promotes health to every individual without
any money and time waste. It uses Pose estimation framework “PoseNet” as Human
Pose Estimation and datasets were collected via webscraping.
However, there are some limitations of this system that are discussed below:-
1. This system still cannot measure upto the level of a live tutor
2. In-order to master every human poses correctly, huge datasets is require
which this system lacks.
3. Audio sessions are missing which could make it more convenient to use
Report Structure
This chapter gives an introduction to the topic.
1. Introduction: This chapter further elaborates the Human Pose Detection, it’s
need in the world and explains how this system works as alternative solutions
to deliver yoga training for the problem faced in health. This chapter also
builds the theoretical basis of this project by elaborating Pose Net algorithms
considered for this project.
2. Literature Review: This chapter explains background of AI in healthcare.
Gives theoretical basis of this project by elaborating Different HPE
frameworks such as OenPose and PoseNet , its underlying maths and
techniques that are used for pose detection. Also, analysis of similar systems
is done in order to obtain comparative study of all the similar past project with
this system and analyse Yogasathi based on similar system along with its
findings.
3. Main Body: Implementation Prediction Engine – This chapter analyses the
datasets used for the development of the prediction engine, shows
development phase of the prediction engine. Results obtained by optimizing
then prediction algorithms have been presented in this chapter.
Implementation – Web Application - This chapter shows the step by step
development of the web application following feature driven development,
Process designs, database designs and system architecture.
4. Testing: Included in this chapter are the functional tests for the overall project
to verify that requirements have been implemented or not.
Analysis and Review – This chapter gives a critical evaluation of the overall
project, evaluates the project schedule and discusses wider implications of
the project.
5. Conclusion – This chapter provides conclusion of the overall project and lists
further enhancements that could benefit the project.
1. Literature review
1.1. Human Pose Estimation
Human Pose estimation techniques are key for understanding of human behaviour
in videos and images. The technology has been ballooning rapidly. The state of art
results shows machines now have exceeded human performance and generating
better results every day. The first people to introduce the method were Fishler and
Elchlager. Locating a position of person’s body part are useful for complex computer
vision tasks like tracking movement, action recognition, surveillance, threat
prediction and foundational technology for rising concepts like Human Computer
Interaction. The work in itself is demanding due to difficulties like elevated flexibility
of body parts, obstructing environment, and different camera angles.
Humans are capable to acknowledge stance by watching location and spot of body
parts. Similar concept is appealed in computer vision as the countered issue of HPE
is localization of person’s joints. People poses range of variations one easy to
understand and difficult at the same time. Predicting a spot of a person is strenuous
job due to diverse conditions including obstructions like clothes, lighting, many
people in single image etc. The latest techniques centres around the idea of training
a model, evaluating and comparing it to similar other systems using common
datasets like COCO, MPII which are collected keypoints data of people.
1.2. Hit-map
The research paper shades lights on how pose estimation revolves around
foundational concepts like heatmaps and offsets. With the rise of powerful deep
neural network like CNN in computer vision mission of evaluating body spot has
been achieved. The convolution neural network architectures like ResNET101 are
trained to produce outputs of heatmaps and offset for each keypoint from images or
video data.
The paper highlights the prediction of 17 keypoints. One of the strategy to counter
the spots would be using regressor on each point but as the number of person
increases the issue becomes more complex and next strategy comes to address it.
The next approach is by estimating activation maps that allows multiple prediction.
Although the accuracy of locating spot concerns the size of feature map. This limits
the performance and give rise to next challenge. Therefore, to counter it a
combination of classification and regression technique is used. At first the K spots or
keypoint is classified which is also known as heatmap and a two dimensional vector
is predicted known to be offset vectors that defines the parallel keypoints or relative
spots of the body part.
1.3. Pose Estimation Frameworks
1.3.1. OpenPose
OpenPose is a kind of pose estimating frameworks that foresees spots in human
body through images or videos. It utilizes spiral pose devices for distinguishing
various areas. A Convolutional Neural Network named CPM yields activation maps
for each and every focuses from picture. Dissimilar to other systems, Open Pose
comprises of a few features like face milestone estimation, body spot focuses
expectation, foot spot focuses forecast, and hand spot focuses discovery or entire
body position estimation. Open Pose Cafe, is created by specialists at Carnegie
Mellon University, was initially written in C++ and can be viewed as condition of
workmanship for human posture estimation.
It can utilize the real body-just models alternatively that are prepared on COCO and
MPII datasets. In view of the yield of the body identifier, facial bouncing box
recommendations can generally be evaluated from some point in body parts,
specifically ears, eyes, nose, and neck. Its library incorporates 3D keypoint present
location, by performing 3D triangulation with non-direct Levenberg-Marquardt
refinement over the aftereffects of various synchronized camera sees. The deduction
time of OpenPose beats all cutting edge techniques, while safeguarding top notch
results. It can run at around 22 FPS in a machine with a NVidia GTX 1080 Ti while
protecting high exactness. OpenPose has just been utilized by the exploration
network for some vision and mechanical technology subjects, for example, individual
re-recognizable proof, GAN-based video retargeting of human faces and bodies,
Human-Computer Interaction, 3D present estimation, and 3D human work model
age. Moreover, the OpenCV library has included OpenPose and our PAF-based
system engineering inside its Deep Neural Network (DNN) module.
1.3.2. Pose Net
PoseNet is another type of HPE that can be utilized to gauge either a solitary posture
or various stances, which means there is an adaptation of the calculation that can
identify just a single individual in a picture/video form different people in a
picture/video. For some reason are there two adaptations: The single individual
posture finder is quicker and less complex however requires just one subject present
in the picture (more on that later). For a single posture finder posture estimation
occurs in two stages:
1. An information RGB picture is taken care of through a convolutional neural
system.
2. Calculations are utilized to translate presents, present certainty scores,
keypoint positions, and keypoint certainty scores from the model yields. They
are discussed below:-
Posture : At the most significant level, PoseNet will restore a posture object
that contains a rundown of keypoints and an occasion level certainty score for
each identified individual.
Posture certainty score: This decides the general trust in the estimation of a
posture. It goes somewhere in the range of 0.0 and 1.0. It very well may be
utilized to shroud represents that are not esteemed sufficient.
Keypoint : A piece of an individual's represent that is assessed, for example,
the nose, right ear, left knee, right foot, and so forth. It contains both a position
and a keypoint certainty score.
Keypoint Certainty Score: This decides the certainty that an expected
keypoint position is exact. It extends somewhere in the range of 0.0 and 1.0. It
very well may be utilized to stow away keypoints that are not regarded
sufficient.
Keypoint Position: 2D x and y organizes in the first info picture where a
keypoint has been recognized.
1.4. Underlying math behind Pose Detection:
The engineering iteratively predicts proclivity handle that encode part-to-part
affiliation, appeared in blue, and recognition certainty maps, and appeared in beige.
The iterative forecast design, refines the expectations over progressive stages, t ∈
{1, . . . , T}, with halfway oversight at each stage.
The picture is first investigated by a convolutional arrange (instated by the initial 10
layers of VGG-19 and adjusted), producing a lot of highlight maps F that is
contribution to the primary phase of each branch. At the primary stage, the system
delivers a lot of discovery confidence maps S 1 = ρ 1 (F) and a lot of part proclivity
fields L 1 = φ 1 (F), where ρ 1 and φ 1 are the CNNs for surmising at Stage 1. In
each ensuing stage, the forecasts from the two branches in the past stage, alongside
the first picture highlights F, are connected and used to create refined expectations,
Where ρ t and φ t are the CNNs for deduction shows the refinement of the certainty
maps and fondness fields across stages. To manage the system to iteratively
foresee confidence maps of body parts in the main branch and PAFs in the
subsequent branch, two loss function is calculated toward the finish of each stage,
one at each branch separately. Between the losses assessed, prediction and the
ground truth maps and fields is utilized. Here, loss function is weighted spatially to
address a plain issue that some datasets don't totally mark all individuals. In
particular, the loss function at the two branches at stage t are:
Where S ∗ j is the ground truth part certainty map, L ∗ c is the ground truth part liking
vector field, W is a parallel cover with W (p) = 0 when the comment is absent at a
picture area p. The cover is utilized to abstain from punishing the genuine positive
expectations during preparing. The moderate management at each stage tends to
the evaporating inclination/gradient issue by renewing the gradient periodically. The
main target is
1.4.1. Part Affinity Fields:
Identified body parts (appeared as the red and blue) focuses on gathering them to
frame the full-body stances of an obscure number of individuals. A confidence
measurement is necessary conformation of body parts belonging to similar
individual. A possible approach to quantify the affiliation is to distinguish an extra
midpoint between each pair of parts on a limb and look for its occurrence between
client’s part detection discoveries. Midpoints are probably going to help bogus
affiliations. Such bogus affiliations emerge because of confinements in the portrayal:
Here, v = (xj2,k − xj1,k)/||xj2,k − xj1,k||2 is the unit vector toward the appendage. The
arrangement of focuses on the appendage is characterized as those inside a
separation edge of the line section, i.e., those focuses p for which
Where the appendage width σl is a separation in pixels, the appendage length is
Lc,k = ||xj2,k − xj1,k||2, and v⊥ is a vector opposite to v. The ground truth part
partiality field midpoints the fondness fields surprisingly in the picture,
Where nc(p) is the quantity of non-zero vectors at point p over all k individuals.
During testing, we measure relationship between competitor part recognitions by
registering the line vital over the comparing PAF along the line portion interfacing the
up-and-comer part areas. At the end of the day, we measure the arrangement of the
anticipated PAF with the competitor appendage that would be shaped by interfacing
the identified body parts. In particular, for two applicant part areas dj1 and dj2 , we
test the anticipated part fondness field, Lc along the line fragment to quantify the
trust in their affiliation:
Where p(u) inserts the situation of the two body parts dj1 and dj2 ,
By and by, we estimated the essential by examining and adding consistently
separated estimations of u.
1.5. Data Collection in HPV
One apparatus for removing human stance is OpenPose. The OpenPose Human
Body Demeanor Acknowledgment Task is an open source library created via
Carnegie Mellon College (CMU) in view of convolutional neural systems and directed
learning and created in the structure of caffe. Disposition estimation, for example,
human movement, outward appearance, and finger development can be
accomplished. Appropriate for single and multi-individual, with astounding
heartiness. OpenPose gives a base up way to deal with ongoing estimation of multi-
individual signals without the requirement for any character locators. Stacking the
OpenPose pre-prepared model, the calculation will separate 18 body joint and 17
lines associating joints. Figure shows a halfway datasets subsequent to removing
joint point data highlights utilizing OpenPose.
1.6. Algorithm for Pose Classification
Since the difference in body stance will bring about changes in places of body joints,
just the chart data inside the range between the base and greatest directions of x
and y will be safeguarded when the dataset is developed. The foundation is set dark
and the repetitive information is reduced. The entire dataset is handled before being
applied to the preparation procedure. The system is recorded beneath.
(a) Slice all pictures to a size of (60, 60) pixels, of which the edge parts are loaded
up with dark pixels.
(b) Name the pictures of right sitting stances with "0" and those of inaccurate sitting
stances with "1".
(c) Apply the information improvement to extend the dataset. Because of the
restricted trial conditions, the dataset built isn't sufficiently large. Along these lines,
the information improvement is essential for growing the dataset utilized for
preparing.
(d) Standardize the picture so pixels are inside the scope of 0 to 1. CNN model is
made out of 19 layers, the info layer is for contributing the handled pictures with (64,
64) pixels. The yield layer is the Grouping layer, utilizing the Softmax Classifier. The
Initiation layer applies ReLu work, and the Pooling layer has a step of (2, 2). The
SGD + energy Analyzer is utilized for preparing. The straight out cross entropy is
embraced as the misfortune work, thus, it is important to make vector of the grouping
names with One-Hot Encoding dependent on the quantity of characterizations. There
are just two classes so the names are gone to 2-dimensional. The scope of arbitrary
pivot of pictures during the information upgrade is set to 20. The scope of level move
of pictures is set to 0.2 (the proportion of width when the move) and pictures can do
flat inversion arbitrarily. The dataset is classified as preparing set, check set and
testing set, which takes up 80%, 10% and 10% individually. The built dataset is then
used to prepare the model to advance the boundaries and is put away in Tensor
Board day by day record.
Results:
The constant video picture is gained by utilizing the profound learning system model
finished via preparing, and the posture attributes of each edge are separated, and
afterward the accuracy of the sitting stance is judged. The acknowledgment results
are imprinted in the video interface, every one of picture is a screen capture of the
video show interface. Figure 5 shows the acknowledgment results for a lot of right
sitting stances, including four pictures of various sitting stances. From the
acknowledgment results, OpenPose can adequately extricate the joint data of the
understudies paying little mind to the table impediment, particularly the joint point
data of the chest area that can be utilized for sitting stance acknowledgment. Figure
6 shows the acknowledgment results for a lot of mistaken sitting stances. As can be
seen from the figure, when the understudy's cervical spine or body is inclined, it is
viewed as an inappropriate sitting stance. Albeit some joint focuses are feeling the
loss of, this doesn't influence the right acknowledgment results.
It very well may be seen from the figure that the precision of the preparation is
exceptionally low toward the start, and the exactness increments quickly as the
quantity of emphasess increments. When emphasizing to around multiple times, the
precision arrived at a greatest, in spite of the fact that there was some variance,
however it remained essentially above 90%.
It tends to be seen from the figure that when the system is simply starting to prepare,
the misfortune esteem is enormous, and as the quantity of emphasess expands, the
misfortune esteem diminishes quickly. At the emphasis of multiple times, the
misfortune esteem is limited and stays stable.
1.6. Similar Systems
1. “Towards Accurate Multi-person Pose Estimation in the Wild” In this
paper ResNet a CNN architecture with 101 layers is used for producing
heatmaps which is processed through single channel for solo keypoint and
two passage for offset due to prediction of offset vector in x and y direction.
This is done for 17 spot estimation. The model has three output passage
generating 3 K estimated result and has is configured to output stride of 8
pixels and 352x257 image size. Mathematically defining the process would
be:
Let, fk(xi) = 1…………(i)
If k is the keypoint at position (xi) and 0. Then,
K ∈ {1,…….N} similarly, the pixel spots of 357x257 image would be in range of i ∈
{1,…….N}
The task of generating a single localized function of combining both values is
burdensome. So the task is further divided. So for every single point a probability
function is calculated. The formula is
Hk(xi) = 1 if ||xi – lk|| ≤ R
“The position of keypoint xi is within a disk of radius R from the location l k of the k-th
keypoint. Thus, K number of heatmap is generated solving binary classification
problem for each position and keypoint independently.”
After the estimation of maps a 2D offset vector FK(xi) = lk – xi at every single
location and keypoint is predicted. K such vector fields are collected independently
and after accumulating both details fusion or aggregation is done in order to get a
highly focused activation maps fk(si) as shown in the diagram below.
"Where G (•) is the bilinear introduction piece. This is a type of Hough casting a
ballot: each point j in the picture crop lattice makes a choice with its gauge for the
situation of each keypoint, with the vote being weighted by the likelihood that it is in
the plate of impact of the relating keypoint. The normalizing factor approaches the
territory of the plate and guarantees that on the off chance that the heatmaps and
balances were great, at that point fk(xi) would be a unit-mass delta work trotted at
the situation of the k-th keypoint."
So crafted by the system is to return two principle highlights:
1. Produce roundabout heatmaps around body spots
2. Produce 2D vectors toward these spots
At that point an allotting activity is conveyed so as to return profoundly engaged
maps from information which will followed by gathered yield of the two significant
highlights.[ CITATION Geo17 \l 1033 ]
2. “Review Of Human Pose Estimation Through Single Image” In this paper
initial examination is carried out with the customary pictorial structure,
experience a conversation of the utilization of artificial Neural Systems that
improved the human posture estimation altogether and afterward the latest,
increasingly popular methodology to be specific “Stacked Hourglass” is used.
Current strategies depend on training, analysis and contrasting on some
regular datasets utilizing various techniques of ML modules. An extensive
investigation of probably the most popular ML techniques are given so as to
give a brief expository audit of these most powerful strategies. Methods such
as Pictorial Structures, Deep Neural net and Stacked Hourglass were tested
for HPS. Among all it was found that DNN took the estimation of human
posture to an unheard of level. For the techniques with DNN, Level of Right
Parts (PCP) is utilized, estimating level of accurately assessed body parts
where a body part is viewed as right if portion endpoints of the part and
ground-truth explained endpoints exist in the length of half from one another.
Evaluation of strategies with hourglass on the other hand is finished utilizing
Likelihood of Right Key point (PCK) which ascertains the pieces of picture
where the key focuses are effectively evaluated.
3. “Yoga Posture recognition By Detecting Human Joint Points In Real
Time Using Microsoft kinect ”To deal with more assortment in human
postures, this paper propose the idea of fine-grained progressive posture
characterization, in which posture estimation is detailed as a classification
task, and propose a dataset, “Yoga-82” with 82 classes of poses. For
classification, various levelled marks is given to yoga depending on the body
arrangement of the real time. The dataset contains a three-level pecking order
including body positions, varieties in body positions, and the real present
names. Arrangement of exactness of the best in class convolutional neural
system structures on Yoga-82. Additionally few progressive variations of
DenseNet is also presented.
Framework proposed here, perceives three major yoga, presented by recognizing of
human joint that focuses using Microsoft Kinect. Yoga was detected with 97%
exactness in each point between various body parts. The framework in this paper
can likewise be utilized to perceive different yoga poses from the reference model of
each posture
4. “PoseNet: A Convolutional Network for Real-Time 6-DOF Camera
Relocalization” This paper presents a robust and real-time monocular six
degree of freedom re-localization system. Their system trains a convolutional
neural network to regress the 6-DOF camera pose from a single RGB image
in an end-to-end manner with no need of additional engineering or graph
optimisation. The algorithm can operate indoors and outdoors in real time,
taking 5ms per frame to compute. It obtains approximately 2m and 3◦accuracy
for large scale outdoor scenes and 0.5m and 5◦accuracy indoors. This is
achieved using an efficient 23 layer deep convnet, demonstrating that
convnets can be used to solve complicated out of image plane regression
problems. This was made possible by leveraging transfer learning from large
scale classification data. We show that the PoseNet localizes from high level
features and is robust to difficult lighting, motion blur and different camera
intrinsic where point based SIFT registration fails.
Additionally, the nearest neighbour classifier as a feature vector representation is
compared to PoseNet, which demonstrates better performance of the PoseNet
2.8. Comparision Table
S.N Research Title Year Problem Tools /techniques Accuracy result
Domain
1. Towards 2017 Multi- -RestNet CNN to detect Training
Accurate Multi- people scale of People accuracy: 68%
person Pose Pose -Heatmaps for keypoints
Estimation in Estimation -COCO data set Test accuracy:
the Wild 67%
2. A Review of 2018 HPD from Three types of models DNN:91.5%
Human Pose single used to perform HPD
Estimation image -Deep Neural Network PS:90.9%
through Single -Pictorial Description
Image -Stacked Hourglass SH:92.1%
3. Yoga Posture 2017 Yoga -Microsoft Kinect to
recognition By Pose detect different parts of Accuracy: 97%
Detecting Detection joint
Human Joint
Points In Real
Time Using
Microsoft kinect
4. PoseNet: A 2016 Skin -Transfer Learning for Higher accuracy
Convolutional Cancer using deep ConvNet compared to K
Network for Detection -PoseNet to localize from nearest
Real-Time 6- high level features classification
DOF Camera algorithm.
Relocalization
2.9. Analysis: Why this system is different from others
While many variety pose estimation frameworks that are alternate to different
HPE domain tasks, have been publicly released, all require particular
equipment or potentially cameras, just as a lot of system arrangements and
set-up. When PoseNet is running on [Link] anybody with a tolerable
webcam-prepared in PCs or many mobile devices can explore this system
directly with the help of an internet browser. The accessibility of this model
inspires more developers and makers to experiment and apply pose detection
to any unique projects. Furthermore, since it is open sourced, a couple of lines
of code is enough to innovate this technology. Also, this can really help
protect client security. Since PoseNet on [Link] runs in the program,
no posture information ever leaves a client's PC.
Thus, yogasathi system uses Posenet fra
