A SURVEY ON HUMAN POSE ESTIMATION AND CLASSIFICATION

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 2084
A SURVEY ON HUMAN POSE ESTIMATION AND CLASSIFICATION
Suvarna Nandyal1, Somashekhar S. Dhanyal*
1Professor, Department of Computer Science and Engineering, Poojya Doddappa Appa College of Engineering,
Kalaburgi, Karnataka, India,
*Research Scholar, Department of Computer Science and Engineering, Poojya Doddappa Appa College of
Engineering, Kalaburgi, Karnataka, India,
---------------------------------------------------------------------***---------------------------------------------------------------------
Keywords: Human pose estimation, Action recognition,
Activity recognition, Keypoint detection, yoga pose
recognition.
1. INTRODUCTION
Human Pose Estimation has aroused computer
vision community's interest. It is a challenging step
towards understanding object in images/ videos. For
decades, the estimation task of human pose has existed,
and it aims to determine the position of the human body
from given inputs. In recent years, vision-based algorithms
have attained rapid progress in a diversity of computer
version problems, including picture classification [1],
object detection [2], and semantic segmentation [3].Well-
designed networks with high estimate competence, better
datasets [4, 5], and more real-world investigation of body
models [6, 7] have made the estimation tool a lot easier.
Action/activity recognition [8], action detection [9], yoga
recognition [10], human tracking [11], films and
animations, human–computer interface, virtual reality,
video monitoring, medical assistance, self-driving vehicles,
and sports [12, 13, 14, 15], and so on are all applications of
human pose estimation.
There are systems that can distinguish activities from
images, however certain activities cannot be detected
using only a solitary stationary image from the present.
There is a requirement for additional information
pertaining to the events that occurred prior to and
following the occurrence. Videos are more accurate in
recognizing actions from videos in this situation.
Apart from many applications the Human Pose
approximation is needed and necessary in the arena of
yoga recognition also because an alternative and
complementary medicine for its beneficial effects on
physical and mental health. Numerous studies showed that
the properties of Yogasan improves health from different
disorders and reduce mental stress. There are many ways
to perform to practice yoga, by attending the classes
manually or online sessions or by self-learning through
reading books, watching videos through internet. Due to
many reasons people are attracted towards the self-
learning. Although self-learning is crucial in yoga practice,
incorrect postures can cause major injury to the body's
muscles and ligaments [44].Many practitioners believe
that the development of computer-assisted training
methods will help them improve their performance while
also protecting them from damage [16], thus requires an
automated yoga recognition system for self-learning .To
develop automated human/yoga recognition system there
are also many challenges in the domain. A good
recognition technique should be able to distinguish
between actions from different classes while also
generalizing across variants within a single class. This will
become more difficult as the number of action classes
grows, as the overlap between classes grows. Thesituation
in which the action takes place contributes significantly to
variation in the recording; changing lighting conditions
and dynamic backgrounds make camera issues even more
difficult. The paper intends to present a synoptic review of
human pose estimation and classification in order to
develop an automated yoga recognition system as the
initial step towards the development of self-training
system. We outline the human pose Estimation
approaches, human body models, methods and describe
the datasets of vision based human/yoga pose
classification from images/videos (A video is nothing more
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 03 | Mar 2022 www.irjet.net p-ISSN: 2395-0072
ABSTRACT: The recent development in estimation and
classification techniques in the field human activity
recognition has been important progress and outstanding
breakthrough. In computer vision, estimating and classifying
of human poses is a crucial and difficult task.This paper
extensively reviews the recent techniques with different
methods and summarizes the benchmark datasets, different
evaluation matrices for estimation and classification of
human pose. The reviewed technique helps the many
applicants to design automated action recognition, activity
recognition system and also provides a support for one of
the major application, yoga pose recognition system and
discuss some of promising future research directions.

than a sequence of images), classification methods, some
evaluation matrices and addresses the future directions.
2. APPROACHES FOR DETERMINING HUMAN POSE
The estimation of human posture is a crucial
stage in recognizing human pose, and it seeks to obtain the
human body's posture from given inputs (image/video).
Human Pose estimation approaches are divided into two
categories. Based on geometrical projection or a specific
image processing challenge, generative and discriminative
approaches are divided into two categories. Another
method to categorize the human pose estimate problem is
whether it is approached from a high-level abstraction and
worked down or from low-level pixel evidence and worked
up. Top-down procedures are those that work downwards,
whereas bottom-up approaches work upwards [17].
2.1 Generative: Creative strategies can be managed in
unique ways constructed on a variety of demonstrations of
human body prototypes, such as preexisting conceptions
near the body model's construction, geometric estimate
from dissimilar perspectives to 2D or 3D space, and so on.
This approach is not considered to be more
computationally efficient than the discriminative approach
[18] since the generative exemplary is given in terms of a
computer graphics representation of postures.
2.2 Discriminative: If you don't want to use human body
models, you may either directly train a mapping between
input sources and the space of human poses (learning-
based)or browse through examples that already exist
(example-based).Starting with image evidence,
discriminative techniques estimation pose via a mapping
or a search-based mechanism [18]. After training the
prototypical, test is substantially quickerthanreproductive
approaches since it delves into construction calculations or
limited search problems instead of maximizing a high-
dimensional parametric domain. Within their scope, the
method looks for the best solutions [19].
2.3 Top-down: To begin, use a high abstraction level to
recognize people and produce their positions in bounding
boxes. After that, each individual goes through the
problem-solving process of changing from high-level
semantics to lower-level video evidence, with high-level
semantics guiding low-level identification. The
computational cost of top-down techniques increases
dramatically as the number of persons in an image grows
[20].
2.4 Bottom–up: Predict all of each person's body
components in the input image first, then aggregate them
using human body model suitable or other techniques.
Depending on the procedure, little template patches, joints,
limbs may be used. In these techniques, fragments of visual
evidence are collected and labeled to provide descriptive
features. These traits are at times used to forecast human
postures directly, and they are now and then used to
detect body components whose manifestations in photos
are subsequently combined to generate a human existence
[18].The computing cost will remain steady as the number
of people in an image grows.
Bottom-up approaches, on the other hand, have difficulty
grouping matching body parts if there are certain people
with a lot of overlap [21].
3. HUMAN BODY MODELS
Human Pose estimation is a computer vision-
based system that recognizes and analyzes human posture.
Thedemonstrating of the human body is the most
significant aspect of human pose approximation. The
human body is a complex and flexible non-rigid object with
a variety of distinguishing characteristics, including
kinematic structure, body shape, and surface texture, as
well as location of body components or joints [1].The three
most prevalent types of human body models are shown in
Fig.1: skeleton-based, contour-based, and volume-based
models.
3.1 Skeleton-based Model: A human body's skeletal
system is made up of joints (key points) such ankles,
knees, shoulders, elbows, wrists, and limb orientations.
This model is used in both 2D and 3D human posture
estimation methods [22] due to its adaptability. A
skeleton-based model is a collection of joint positions and
limb orientations that follow the skeletal structure of the
human body (typically between 10 and 30).Also labeled as
a graph with vertices signifying joints and edges
expressing limitations or prior relationships of joints
inside the skeletal structure [23].This simple and flexible
human body topology is frequently used in both 2D and 3D
human pose estimates and human pose datasets [24].
3.2 Contour-based Model: Widely employed in human
posture estimate methods, this image comprises
approximate width and constructed by placing for the
body's limbs and torso, and depicts human body
components as rectangles or the person's silhouette's
borders. The model, which also includes cardboard models
and Active Shape Models (ASMs) [25], is a useful way to
portray a human posture.

3.3Volume-based Model: To depict 3D human body
silhouettes and positions, geometric shapes or meshes are
employed. Cylinders, conics, and other geometric shapes
were used to model bodily parts in the past. [26]. Modern
volume-based models, which are often obtained via 3D
scans, are represented as meshes. Popular volume-based
models include Shape Completion and Animation of People
(SCAPE) [27], Skinned Multi-Person Linear Model (SMPL)
[28], and a unified deformation model [29].
Figure 1.Human Body Models (a) skeleton-Based
Model (b) counter-based model
(c) volume-based Model
4. KEYPOINT DETECTION METHODS
Key point detection entails detecting people while
also locating their important points. The terms "key
points" and "interest points" are interchangeable. They’re
spatial positions, or points in an image, that describe
what’s fascinating or stand out in the image/video, and
they express a graphical representation of a person's
orientation. It is essentially a series of coordinates that
may be joined to define a person's pose. Each skeletal co-
ordinate is referred to component (joint or key point).Pair
is a valid connection between two pieces (or a
limb). Individual parts are generally identified first, then
connections are formed between them to produce the pose
in the manner listed below.
4.1 OpenPose: Like many previous bottom-up algorithms
for estimation of many human poses, Open Pose begins by
detecting parts (key points) corresponding to each person
in the image, then allocates parts to distinct human.
OpenPose model's architecture is depicted in Fig 2.Using
first few layers, OpenPose network collects features from
an image (VGG-19 in the above flowchart).After that, the
features are routed to two convolution layer branches that
run parallel to one another.
Figure 2: OpenPose architecture
The first branch generates 18 confidence maps, which each
represents a different aspect of the human posture
skeleton. According to the second branch, a collection of 38
Part Affinity Fields (PAFs) will be generated, which will be
used to establish relationships between parts in the model.
Figure 3.OpenPose steps for estimating human pose.
Each branch's forecasts are refined over successive stages.
Part confidence maps are used to create bipartite graphs
between pairs of parts, as seen in fig 3.Weaker linkages in
bipartite graphs are trimmed using PAF values. Using the
processes discussed above, human position skeletons can
be approximated and allocated to each person in the
image.
4.2 DeepCut: The following challenges were identified
using bottom-up method for multi-person human pose
approximation:
i. Make list of potential D body part candidates. For each
person in the image, this collection represents all

conceivable body part locations. Choose a subset of bodily
parts from the list of choices above.
ii. One of the C body part classes should be assigned to
each body part. The numerous types of body parts are
represented by the different body part classes, such as
"arm," "leg," "torso," and so on.
iii. Parts of the body that belongs to the same individual
but are separated.
Figure 4: The approach is depicted graphically. (a)
Initial detection (b) jointly clustered (c) Predicted
pose sticks
4.3 RMPE (AlphaPose): Pose Estimation from the top
down is a frequent method. The location of the subject is
utilized to calculate their pose. The posture extraction
approach may perform badly due to issues with
localization and duplicate bounding box forecasts.
Figure 5: Effect of redundant predictions (left)
and bounding boxes with a low level of
confidence (right)
The authors recommended that an erroneous bounding
box be used to abstract a high-quality solitary person
region using the Symmetric Spatial Transformer Network
(SSTN).Human posture skeleton for that person in this
extracted region is estimated using a Single Person
Position Estimator (SPPE).A Spatial De-Transformer
Network is utilized to remap the estimated human position
to the image's original coordinate system (SDTN).Finally,
to solve issue of duplicate pose deductions, a posture that
is parametric the technique of Non-Maximum Suppression
(NMS) is used, as illustrated in Fig 5.
4.4 Mask RCN: For semantic and instance segmentation,
this is a well-known design. The model predicts the
image's many elements' bounding box coordinates, as well
as a mask that conceptually segments the item. Fig 6 shows
how the basic design can be easily modified to estimate
human position.
Figure 6: Mask RCNN Architecture
To extract feature maps from an image, the basic design
starts with a CNN.A Region Proposal Network (RPN) uses
these feature maps to find bounding box contenders aimed
at the existence of objects. Candidates for the bounding
box select a region (area) from the CNN-extracted feature
map as their starting point. RoIAlign is a layer that is used
to minimize the size of the removed feature and make it
more consistent because the bounding box contenders can
be of different sizes. This mined feature is now sent
through parallel CNN branches, which predict the
bounding boxes and segmentation masks at the end. This
strategy is quite similar to the top-down method.
5. DATASETS
As we can see, there are a plethora of publicly
available datasets that allow us to compare diverse
approaches and gain perception into the (in) capabilities of
various approaches. The most often used sets are
discussed.

● COCO [4]: The COCO key - point database is a
multi-person two-dimensional pose estimation
dataset based on Flickr. COCO is the world's
largest 2D Pose Prediction dataset, which is used
to test algorithms for 2D Pose Estimation.
● MPII[30]: The MPII humans posture database is a
multi-person two-dimensional pose estimation
dataset gathered from YouTube videos. It covers
approximately 500 different human activities.
MPII was the first dataset to have such a huge
number of distinct positions.
● HumanEva [31]: Human Eva is a dataset for
single-person 3-dimensional Pose Estimation that
contains video sequences captured with a mixture
of RGB and grayscale camera. HumanEva was first
large-scale 3D Pose Estimation dataset. Marker-
based motion capture (mocap) cameras are used
to capture real-world 3D poses.
● KTH [32]: The KTH human motion dataset
features 25 distinct people performing six
different movements (walking, jogging, sprinting,
boxing, hand waving, and hand
clapping).Outdoors, outdoors with zooming,
outdoors with various attire, and interiors are the
four scenarios used. There is a lot of diversity in
the quality and length of the performance. The
backgrounds are quite consistent. There is only
minor camera movement aside from the zooming
scenario.
● Weizmann[33]: The Weizmann Institute's human
action dataset contains ten actions done by ten
people (walk, run, leap, gallop sideways, bend,
one-hand wave, two-hand wave,jump in place,
jumping jack, and skip). The backgrounds are
static, while the collection includes foreground
silhouettes. The point of view is fixed.
● INRIA XMAS [34]:Introduced the IXMAS dataset,
which covers behaviors from five different
perspectives.Around11 people participated in
fourteen acts (check watch, cross arms, scratch
head, sit down, get up, turn around, walk, wave,
punch, kick, point, pick up, throw overhead and
throw from bottom up).In terms of camera
configuration, the movements are carried out in
any direction. The camera perspectives are set in
stone, with a motionless background and lighting.
There are silhouettes and volumetric voxel
renderings in the collection.
● UCF [35]: Around 150 sequences of sport motions
in the UCF sports activity dataset. Bounding boxes
for human figures are part of the collection.For the
majority of action classes, there is considerable
variation in action execution, human appearance,
camera movement, perspective, illumination, and
background.
● YOGA-82 [36]:Existing pose estimate dataset’s
limits in terms of pose variety have prompted
researchers to dig deeper and offer a novel idea of
fine-grained hierarchical pose categorization,
together with a diverse dataset of 82 complicated
poses.
● YogaVidCollected [10]YogaVidCollected is a data
set of yoga videos that can be used to recognize
yoga poses. There are 6 asana in all, 88 videos, and
1.1 GB of data. At 30 frames per second, the total
time is 1 hour, 6 minutes, and 5 seconds.
6. HUMAN POSE CLASSIFICATION METHODS
Classification is the domain plays the most
important role of image analysis. Classification accepts the
given input images/videos and produces output in the
form of classifying the human or yoga pose. There are
many classification methods, some are listed below.
6.1 Support Vector Machine (SVM): Data analyzed using
a machine learning technique for categorization and
regression analysis. A supervised learning method is SVM.
The categorization is accomplished by constructing a
sequence of hyper planes in a multidimensional space with
a big margin between the elements of various classes. It
does quite well in terms of recognition and classification.
As a result, it is by far the most popular classifier.
6.2 Naive Bayesian classifier: Based on Bayes' theorem
[37], a supervised learning algorithm and probabilistic
classifier. It is a technique that involves counting the
instances in which a particular motion occurs in a video
system. To distinguish a novel action, the Bayes rule is
applied and activity with highest a posteriori probability is
picked.
6.3 K-Nearest Neighbor: The supervised learning domain
includes one of the most basic yet fundamental
classifications. It entails classifying objects based on votes
the majority of their neighbors [38].Features are

represented as location vector in a multi - dimensional
space of features to locate neighbors.
6.4 K means: It is data structure-sensitive clustering
technique that iteratively calculates the k-distances from
each class centroid to each datum [39].Unsupervised
classification algorithm K-means divides items into k
groups depending on their attributes. The sum of the
distances between each item and group or cluster centroid
is minimized during grouping. It is most suited for
classifying human positions.
6.5 Mean shift clustering: Image processing and
computer vision are two fields where it can be used. The
algorithm continually assigns each piece of data to the
nearest cluster centroid given a series of data points, with
direction to a nearest cluster centroid determined by the
location of the majority of nearby points.
6.6 Machines finite state: Static motions and postures are
represented by states, transitions are used to provide
temporal and/or probabilistic limitations, and arcs
between the beginning and final states are used to
describe dynamic gestures.
6.7 Hidden Markov Models: This model addresses the
keys to the segmentation problem. They're a type of
Markov chain [40], which are determinate state automata
with probabilities on each arc.
6.8 Dynamic time warping: Based on their related
characteristic values, this technique determines the
distances between every pair of potential locations in two
signals. It is utilized to overcome problems of rate
variations caused by the identification of human poses by
calculating and recognizing motion in a video stream [41].
6.9 Neural networks: Mathematical models whose
strategy is based on biological neuron functioning
[42].Probabilistic learning approaches, particularly
Bayesian, are commonly used to optimize them. Neural
networks make it possible to create quick classifications
that may be used in real-time systems. Neural networks
are also the best choice for recognizing poses.
7. EVALUATION METRICS
For human posture recognition, numerous
performance criteria used in various classification
disciplines have been modified and applied. We cite
frequently used measures like accuracy, precision, and
recall in this section based on [43].Before summarizing
these metrics, let's define True Positive, True Negative,
False Positive, and False Negative in the framework of
posture recognition (see fig 7):
⮚ True Positive actions are those in which the actual
and projected transactions coincide.
⮚ False Negative actions are those that belong to a
specific class but are predicted not to be from
that class.
⮚ True Negative actions are those in which the
actual and projected transactions do not match
the search class.
⮚ False Positives are activities in which the actual
transactions don’t match the examined class, but
are projected to do so.
7.1 Sensitivity: It's also known as recall, true positive
rate, or detection probability. It refers to positive cases
that were actually predicted to be positive. Sensitivity in
posture identification refers to the proportion of poses
predicted in each class.
Figure 7: Structure of the confusion matrix
Similarly, (1 - sensitivity) determines the system's
inability to detect pose.
This can be stated mathematically as:
7.2 Precision: It's also identified as the Positive
Prediction Value (PPV), and it refers to probability of a
detected example of pose occurring in real life. The
likelihood of the recognizer erroneously identifying a
detected posture is also determined by (1 - accuracy).
7.3 Specificity: It's sometimes referred to as the false
positive rate (FPR) or the true negative rate (TNR). It's
based on actual negative cases that were anticipated to be
negative. It assesses the sensitivity of the system to
negative classes. This can be stated mathematically as:

7.4 Negative Predictive Value (NPV): It's known as
"negative precision" since it calculates the probability
that a negative identification is correct when compared
to all previous negative identifications.
7.5 F_Measure: It provides details on the test's accuracy.
It calculates the precision and recall harmonic mean. As a
result, the F measure defines how precise and resilient
the classifier is at the same time. Its best value is 1 and its
worst value is 0.
7.6 Accuracy: When the classes are evenly sampled, the
accuracy produces good results. Calculated is the
proportion of valid predictions in respect to the total
number of samples. This can be stated mathematically as:
7.7 Likelihood Ratio: Compares the likelihood of an
action predicted correctly vs the likelihood of an activity
predicted incorrectly. It can be used to calculate both
true positive and true negative outcomes.
Mathematically, this can be expressed as:
7.8 Confusion Matrix: It is also known as the error
matrix, and it summarizes the prediction outcomes. It
also describes the model's overall performance. The
classifier's errors, as well as their types, are displayed in
the confusion matrix. The matrix's rows indicate
expected class instances, whereas the columns represent
actual class instances, or vice versa (Fig.7).
7.9 Intersection over Union (IoU): Also known as the
Jaccard similarity coefficient or the Jaccard index. It
assesses the detector's accuracy on a certain dataset. The
region of intersection between the predicted bounding
box and the ground-truth bounding box is referred to as
"Area of overlap," while the area contained by the
predicted bounding box and the ground-truth bounding
box is referred to as "Area of Union."
8. CONCLUSION
The purpose of this paper is to present the most recent
work in this field of study. It initially discuss the objective
of human pose estimation and then it presents the
human pose estimation approaches and key point
detection methods for pose representation and also
discussed about some of common datasets, different type
of classification methods provides an effective survey
study to design and develop the automated human
recognition system.In numerous computer vision
applications, the necessity to interpret human action has
become unavoidable. On the other hand, the major
application fields yoga pose recognition becoming a
prominent research work because of developed
techniques to perform yoga in front of system without
the help of trainer and become self-learner. We conclude
that our study helps to design and develop the automated
human recognition system/Yogasan recognition system
with different poses irrespective of many challenges.
REFERENCES
1. Ali S, Shah M. Human action recognition in videos
using kinematic features and multiple instance
learning. IEEE Transactions on Pattern Analysis and
Machine Intelligence. 2010,32(2),288-
303.DOI:10.1109/TPAMI.2008.284.
2. Anguelov D, Srinivasan P,KollerD,ThrunS,Rodgers
J, Davis J. Scape: shape completion and animation
of people. Journal of ACM Transactions on Graphics.
2005,24(3), 408–
416.https://dl.acm.org/doi/10.1145/1073204.10
73207.
3. Akansha A, Shailendra M, Singh N. Analytical
review on video-based human activity
recognition. 3rd International Conference on
Computing for Sustainable Global Development.
2016, 3839–3844.

4. Andriluka M, Iqbal U, Milan A, Insafutdinov E,
Pishchulin L, Gall J, Schiele B. Posetrack: A
benchmark for human pose estimation and
tracking. In Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition 2018,
5167–5176.https://arxiv.org/abs/1710.10000.
5. Andriluka M, Pishchulin L, Gehler P,Schiele B. 2d
human pose estimation: New benchmark and
state of the art analysis. In Proceedings of IEEE
Conference on Computer Vision and Pattern
Recognition. 2014, 3686–
3693.DOI: 10.1109/CVPR.2014.471.
6. Bai L, Efstratiou C,Ang S. WeSport: utilising wrist-
band sensing to detect player activities in
basketball games.IEEE International Conference
on Pervasive Computing and Communication
Workshops (PerCom Workshops).2016,1-6,DOI:
10.1109/PERCOMW.2016.7457167.
7. Ben M, Trabelsi I,Bouhlel M. Human action
analysis for assistance with daily activities.
International Journal on Human Machine
Interaction. 2016,1-10,
https://www.researchgate.net/publication/3058
48684.
8. Liefeng B, Cristian S. Twin Gaussian Processes for
Structured Prediction.International Journal of
Computer Vision. 2010, 87, 28–52.DOI
10.1007/s11263-008-0204-y.
9.Bhardwaj R, Singh K. Analytical review on human
activity recognition in video. 6th International
Conference Cloud System and Big Data
Engineering. 2016, 531–
536.https://ieeexplore.ieee.org/document/7724
978.
10. Cao Z, Simon T, Wei E, Sheikh Y. Real-time multi-
person 2d pose estimation using part affinity
fields. In Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition.
2017,1302-1310.DOI: 10.1109/CVPR.2017.143.
11. Chen T, He Z, Hsu C. Computer-assisted yoga
training system. Multimedia Tools and
Applications, 2018,77, 23969–
23991.https://doi.org/10.1007/s11042-018-
5721-2.
12. Chen T.He Z, Hsu C,ChouL,Lee Y, Lin P. Yoga
Posture Recognition for Self-training. In
Proceedings of the 20th Anniversary International
Conference on Multimedia Modeling 2014,8325,
496–505.https://doi.org/10.1007/978-3-319-
04114-8_42
13. Cootes F, Taylor J, Cooper H and Graham J. Active
shape models-their training and application.
Journal of Computer Vision and Image
Understanding, 1995,61(1),38-
59.https://doi.org/10.1006/cviu.1995.1004.
14. Felzenszwalb F, Huttenlocher P, Pictorial
structures for object recognition. International
Journal of Computer Vision.2005,61,55–
79.https://doi.org/10.1023/B:VISI.0000042934.1
5159.49.
15. Gorelick L, Blank M, Shechtman E, Irani M, Basri R.
Actions as space-time shapes .Tenth IEEE
International Conference on Computer Vision.
2005,1,1395-1402.doi: 10.1109/ICCV.2005.28.
16. Insafutdinov E, Andriluka M, Pishchulin L, Tang S,
Levinkov E, Andres B, Schiele B. ArtTrack:
Articulated Multi-Person Tracking in the
Wild. 2017 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR). 2017, 1293-
1301, doi: 10.1109/CVPR.2017.142.
17. Joo H, Simon T, Sheikh Y. Total capture: A 3d
deformation model for tracking faces, hands, and
bodies. In Proceedings of Conference on Computer
Vision and Pattern Recognition. 2018, 8320-
8329.DOI: 10.1109/CVPR.2018.00868.
18. Joo H, Simon T, Li X, Liu H, Tan L, Gui L, Banerjee S,
Godisart T.S, Sheikh Y. Panoptic studio: A
massively multiview system for social interaction
capture. IEEE Transactions on Pattern Analysis and
Machine Intelligence. 2017,190–204.
https://doi.org/10.1007/s41095-020-0171-y.
19. Ju X, Black J, YacoobY. Cardboard people: A
parameterized model of Articulated image motion.
In Proceedings of the Second International
Conference on Automatic Face and Gesture
Recognition. 1996, 38–44.
https://doi.org/10.1007/s11263-006-9781-9.
20. Kanazawa A, Black M.J, Jacobs D.W, Malik J. End-
to-end recovery of human shape and pose. In
Proceedings of IEEE Conference on Computer
Vision and Pattern Recognition. 2018, 7122–7131.
DOI: 10.1109/CVPR.2018.00744.
21. Krizhevsky A, Sutskever I, Hinton G.E. Imagenet
classification with deep convolutional neural
networks. In Proceedings of the 25th International
Conference on Neural Information Processing
Systems. 2012,1,1097–1105.
https://doi.org/10.1155/2020/7607612.
22. Liu Z, Zhu J, Bu J, Chen C. A survey of human pose
estimation: the body parts parsing based methods.
Journal of Visual Communication and Image
Representation. 2015, 32,10–
19.https://doi.org/10.1016/j.jvcir.2015.06.013.
23. Li M, Zhou.Z, Li J, Liu X, Bottom-up pose estimation
of multiple person with bounding box constraint,
24th International Conference on Pattern

Recognition. 2018,115-120. DOI:
10.1109/ICPR.2018.8546194
24. Long J,Shelhamer E, Darrell T, Fully convolutional
networks for semantic Segmentation. In
Proceedings of IEEE Conference on Computer
Vision and Pattern Recognition. 2015,3431–3440.
DOI: 10.1109/CVPR.2015.7298965.
25. Loper M, Mahmood N, Romero J, Pons-Moll G,
Black M.J. Smpl: A skinned multi-person linear
model. ACM Transactions on Graphics.Vol 34.Issue
6.pp.1–16,
https://doi.org/10.1145/2816795.2818013.
26. Luvizon D.C, Picard D, Tabia H, 2d/3d pose
estimation and action recognition using multitask
deep learning. In Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition.
2018, 5137–5146. https://hal.archives-
ouvertes.fr/hal-01815703.
27. Manisha V, Sudhakar K, Yuta N, Shanmuganathan
R. Yoga-82: A New Dataset for Fine-grained
Classification of Human Poses. IEEE/CVF
Recognition Workshops (CVPRW). 2020, 4472-
4479, DOI:10.1109/CVPRW50498.2020.00527.
28. Minnen D, Westeyn T, Starner T, Ward J, Lukowicz
P, Performance metrics and evaluation issues for
continuous activity recognition. Performance
Metrics for Intelligent Systems. 2006, 4.303-317,
DOI=10.1.1.129.5025.
29. Nguyen-Duc-Thanh N, Stonier D, Lee S, Kim H, A
new approach for human-robot interaction using
human body language. In Proceedings of the 5th
international conference on Convergence and
hybrid information technology, 2011, 762–769,
DOI: 10.1007/978-3-642-24082-9_92.
30. Nordsborg B, Espinosa G, Thiel V. Estimating
energy expenditure during front crawl swimming
using accelerometers. Procedia Engineering.
2014,132-137,
https://doi.org/10.1016/j.proeng.2014.06.024.
31. Pai F, ChangLiao H, Lin K P, Analyzing basketball
games by a support vector machines with decision
tree model, Neural Computing and Applications,
2017, 28,4159–4167.
https://doi.org/10.1007/s00521-016-2321-9.
32. Ren S, He K, Girshick R, Sun J. Faster R-CNN:
Towards real-time object detection with region
proposal networks. In Proceedings of the in
Neural Information Processing Systems. 2015,
28,91–99.DOI: 10.1109/TPAMI.2016.2577031.
33. Rodriguez D, Ahmed J, Shah M. Action mach a
spatio-temporal maximum average correlation
height filter for action recognition. IEEE
Recognition. 2008, 1–8.DOI:
10.1109/CVPR.2008.4587727.
34. Schuldt C, Laptev I, Caputo B. Recognizing human
actions: a local svm approach. In Proceedings of
the 17th International Conference on Pattern
Recognition, 2004,3.32–36. DOI:
10.1109/ICPR.2004.1334462.
35. Shan Z, Su E, Ming L. Investigation of upper limb
movement during badminton smash. In
Proceedings of 10th Asian Control Conference
(ASCC). 2015,1–6, DOI:
10.1109/ASCC.2015.7244605.
36. Sidenbladh.H, Black.M, Fleet D, Stochastic tracking
of 3D human figures using 2D image motion. In
Proceedings of 6th European Conf. Computer
Vision. 2000, 1843,702-718.
https://doi.org/10.1007/3-540-45053-X_45.
37. Sidenbladh H, De la T, Black M. A framework for
modeling the appearance of 3d articulated figures.
In Proceedings of Fourth IEEE International
Conference on Automatic Face and Gesture
Recognition. 2000, 368-375, DOI:
10.1109/AFGR.2000.840661.
38. Sigal L, Balan O, Black J, Humaneva. Synchronized
video and motion capture dataset and baseline
algorithm for evaluation of articulated human
motion. International Journal of Computer Vision,
2010, 87(1), 4-
27.https://doi.org/10.1007/s11263-009-0273-6.
39. Torres F,Kropatsch G.Top-down 3D Tracking and
Pose Estimation of a Die Using Check-Points”.
Structural, Syntactic, and Statistical Pattern
Recognition.2012.pp.492-500. DOI: 10.1007/978-
3-642-34166-3_54
40. Vrigkas M, Nikou C, Kakadiaris I, A review of
human activity recognition methods. Frontiers in
Robotics and AI. 2015, 2,1-28.
https://doi.org/10.3389/frobt.2015.00028.
41. Wenjuan G, Xuena Z, Jordi G, Andrews S, Thierry B,
Changhe T, El-hadi Z. Human Pose Estimation
from Monocular Images: A Comprehensive Survey,
Sensors (Basel, Switzerland). 2016, 16, 1-39. DOI:
10.3390/s16121966.
42. Weinland D, Ronfard R, Boyer E. Free viewpoint
action recognition using motion history volumes.
Computer Vision and Image Understanding. 2006,
104(2-3),249-257.
https://doi.org/10.1016/j.cviu.2006.07.013.
43. Xu K, Qin Z, Wang G. Recognize human
activities from multi-part missing videos. IEEE
International Conference on Multimedia and
Expo. 2016, 976–990. DOI:
10.1109/ICME.2016.7552941.

44. Yadav S.K, Singh A, Gupta A, Real-time Yoga
recognition using deep learning.
Neuralomputing and Applications. 2019,
9349– 9361.DOI: 10.1007/s00521-019-04232-7

A SURVEY ON HUMAN POSE ESTIMATION AND CLASSIFICATION

More Related Content

Similar to A SURVEY ON HUMAN POSE ESTIMATION AND CLASSIFICATION (20)

More from IRJET Journal (20)

Recently uploaded (20)

A SURVEY ON HUMAN POSE ESTIMATION AND CLASSIFICATION