AN ABNORMALITY IN CLINICAL GAIT ANALYSIS :

A SURVEY

Dr . D. Chitra Akshaya.M Athira.M Indhumathi .S

Professor Final year CSE Final year CSE Final year CSE
Department of CSE Department of CSE Department of CSE Department of CSE
P.A. College of Engineering P.A.College of Engineering P.A.College of Engineering P.A.College of Engineering
and Technology and Technology and Technology and Technology
Pollachi -6442002 Pollachi-642002 Pollachi-642002 Pollachi-642002
Email: [email protected] Email:[email protected] Email:[email protected] Email: [email protected]
om

ABSTRACT and unsteady strides often indicate

underlying neuromuscular or skeletal
Gait analysis is the
issues requiring targeted interventions.
systematic study of human movement
The methods for analysing gait include
during walking and running with
systems that use markers, motion
applications in biomechanics, clinical
capture without markers, and machine
diagnostics and computer vision. Gait
learning techniques. Recent
recognition, a specialized area of this
advancements combine sophisticated
field identifies individuals through
sensors, computer vision, and deep
their distinctive walking patterns,
learning technologies to improve
providing a non invasive and reliable
accuracy and reliability when working
biometric solution for surveillance,
with complex datasets. The Gait and
authentication, and security systems.
Action Video Dataset provides labelled
Clinical gait analysis evaluates
video recordings of individuals in
abnormal walking patterns to assist in
varied environments and conditions,
diagnosing and treating conditions
enabling the development of gait
such as cerebral palsy, Parkinson's
recognition methods that address
disease, stroke, and disorders affecting
challenges related to clothing, viewing
the muscles or skeleton. The
angles, and environmental factors.
movements such as limping, shuffling,
Advancements in foundational
principles, modern methodologies, and their movement patterns with respect
comprehensive datasets continue to to disease progression or
drive progress in gait analysis and rehabilitation are limited, increasing
recognition technologies. the burden on healthcare to monitor
movements that could otherwise be
1.INTRODUCTION
self managed with objective
Clinical Gait Analysis is a feedback.
well-established method for
Notable advancements in
understanding and monitoring
computer vision and machine learning,
human health and performance by
video based human action recognition
identifying abnormalities in gait
has emerged as a powerful tool for gait
pattern movements. Traditionally,
analysis, providing clinicians with
the interpretation of human
enhanced usability and utility. Current
movement relies on clinical
computer vision research in gait has
experience and observational skills to
focused on action classification, human
identify abnormalities and provide
gait identification and gait pathology
explanations for deviations in
classification. Gait tasks are important
movement patterns. The
fall short in providing clinically
conventional approach is heavily
relevant information, particularly
dependent on the subjective
regarding the spatial and temporal
judgment of experts can vary
location of movement abnormalities.
between clinicians. Gait also poses
Existing technologies such as three
challenges in objectively measuring
dimensional motion capture do not
movement without the availability of
allow health and medical specialists to
resource intensive and expensive
retrospectively use data formats like
motion capture equipment, often not
single camera colour video for
available in many clinical settings.
capturing clinical movement for
Further, tools for patients to monitor
analysis.
The limited progress in computer units, or three dimensional motion
vision research for the task of capture systems are frequently
localizing spatial and temporal inaccessible or impractical in many
abnormalities for clinical gait analysis clinical scenarios such as monitoring
is attributed, in part, to a lack of an individual in their home
understanding regarding the clinical environment. Existing gait analysis
requirements for this information and techniques tend to be narrowly
the limited availability of clinically focused, concentrating on a limited set
annotated video data essential for of gait types and pathologies. Gait
training and testing models. models encounter challenges in
achieving generalizability across a
Many human action recognition
diverse range of gait and movement
tasks primarily focus on the
abnormalities. The limitations, current
classification of broad action
methods primarily focus on general
categories, such as cycling and yoga.
classification without providing explain
The intra class variations within these
ability via spatial or temporal location
actions play a critical role in
information. Clinical gait analysis
distinguishing typical and atypical
research is being advanced by
behaviours for a given activity. A finer
integrating recent computer vision
level analysis of actions, encompassing
techniques particularly action
both spatial and temporal dimensions,
recognition methods to identify
enables the identification of subtle
abnormal gait patterns from single
differences between actions.
camera video data
The current state of the art
Clinical gait analysis is an
methods in gait analysis employing
essential diagnostic tool employed in
deep learning have shown great
clinical settings to assess
promise. The methods often rely on
musculoskeletal and neurological
depth sensors, inertial measurement
conditions impacting human
movement. Trained specialists, These models includes approaches
including physiotherapists and medical utilizing appearance based and model
experts systematically evaluate gait based methods to classify human
parameters to identify abnormalities. actions in video. The methods such as
The evaluations include quantitative Convolutional Neural Networks based
and qualitative data on kinematics, approaches focused on extracting
kinetics, and muscle activity during spatial features from images. The
walking. Such data helps develop action recognition requires both
targeted treatment plans and monitor temporal and spatial features to
progress. Clinical gait analysis focuses model long range temporal structures.
on internal factors such as range of The development of large video
motion, muscle strength, flexibility and datasets like Kinetics 700 and
neurological and psychological ActivityNet, deep learning methods
influences on movement. Identifying such as Recurrent Neural Networks
deviations in gait patterns provides and Temporal Segment Networks
insights into underlying issues and have become increasingly popular.
guides treatment decisions, using both The Gait approaches allow for
specialized equipment and more cost efficient and accurate recognition of
effective tools such as visual actions by combining both temporal
observation, functional assessments and spatial features.
and portable sensors.
1.2 Vision Based Approaches for Gait
1.1 Overview of Human Action Recognition
Recognition in Videos
Computer vision has established
its effectiveness in gait recognition,
Human action recognition is a
person identification, and human
computer vision task that involves
action recognition. Gait recognition,
understanding human actions based
uses gait as a biometric feature, allows
on visual information and motion.
for person identification based on
walking patterns. There are two main pose-based approaches using skeletal
approaches: appearance-based and data from sensors like Microsoft
model-based. Appearance based Kinect have gained attention. The
methods utilize gait energy images Vision based approaches aim to
and temporal learning techniques to classify abnormal gait patterns and
extract features from silhouettes of provide more precise analysis of gait
gait sequences. Model based methods events through models such as
use skeleton based models like three Recuurent Neural Networks and
dimensional pose estimation, to graph convolutional networks. The
separate motion from appearance and integration of skeletal data and pose
address challenges like camera angles estimation techniques enhances the
and clothing influences. Hybrid accuracy and application of clinical
models combining both appearance gait analysis offering better
and motion information also offer assessment and classification of
promising results for gait recognition neurological conditions.
and person identification.

1.3 Vision Based Approaches for

2.TECHNIQUES FOR CLINICAL
Clinical Gait Analysis
GAIT ANALYSIS
Vision based methods are
increasingly utilized in clinical gait In 2010, Okumura et al. [22]
analysis due to their objectivity, cost constructed the world’s largest gait
effectiveness, and ability to automate database with 1,035 subjects (aged 2–
data collection. The Vision based 94), enabling statistically reliable
methods can analyse gait parameters evaluations of vision-based gait
for pathologies like Parkinson’s recognition. The study highlighted
Disease and cerebral palsy. This is performance variations across age
limited use of appearance-based and groups based on walking ability and
hybrid models in clinical settings, physical strength.
 In 2014, Muro-de-la Herran et al. [21]
In 2010, Hossain et al. [17] addressed reviewed wearable and non-wearable
clothing variations in gait recognition gait analysis systems, emphasizing the
using part-based clothing clinical applications in diagnosing
categorization and adaptive weight conditions like Parkinson’s disease and
control, significantly improving stroke.
identification accuracy under different
clothing conditions. In 2015, Zhang et al. [25] explored
biometric recognition advancements,
In 2011, Makihara et al. [19] developed comparing facial, gait, and fingerprint
a gait-based age estimation model, recognition with addressing challenges
demonstrating the walking styles like environmental and user
change with age, enhancing biometric variability.
and healthcare applications.
In 2016, Arora et al. [10] proposed
In 2012, Iwama et al. [18] proposed a using Gait Flow and Gaussian Images
gait-based person verification system with an Extension Neural Network,
for forensic applications, proving its improving accuracy in gait recognition.
effectiveness in distinguishing
individuals via surveillance video In 2018, Bei et al. [12] used Kinect
analysis. sensors for movement disorder
detection, leveraging adaptive gait
In 2012, Akae et al. [3] introduced a analysis for early diagnosis.
method to reconstruct motion data
from low frame-rate videos, improving In 2019, Abdelraouf et al. [1] studied
gait recognition in limited surveillance backward walking’s effects on leg
scenarios. muscle activation in young females
with patellofemoral pain syndrome,
revealing therapeutic benefits for pathological gait classification,
rehabilitation. demonstrating high accuracy in
identifying abnormalities.
In 2019, Amini and Banitsas [6]
employed Kinect v2 to improve In 2022, Dos Santos et al. [16] reviewed
mobility in Parkinson’s patients deep learning-based gait recognition,
experiencing freezing of gait, highlighting its superiority over
significantly reducing freezing episodes traditional methods in feature
through laser visual cues. extraction and classification.

In 2020, An et al. [7] evaluated model- In 2022, Mogan et al. [20] introduced
based gait recognition on a large multi- Vision Transformers for gait
view dataset, proving its robustness in recognition, demonstrating improved
diverse conditions. performance over convolutional neural
networks.
In 2020, Bhattacharya et al. [15]
explored emotion perception from gait In 2022, Benson et al. [13] reviewed
using Spatial Temporal Graph Inertial Measurement UnitMU -based
Convolutional Networks, opening new running gait analysis , comparing
possibilities in affective computing. laboratory and real-world settings to
highlight wearable sensors’
In 2021, Al-Jubouri et al. [5] developed capabilities.
a gait dataset of autistic children,
classified with a Rough Set Classifier to In 2022, Bertaux et al. [14] presented a
aid diagnosis and therapy. dataset on gait patterns before and
after total hip arthroplasty, providing
In 2021, Albuquerque et al. [4] insights into post-surgical recovery.
introduced a spatiotemporal deep
learning framework for automatic
 In 2022, Rattanasak et al. [24] affect their robustness and accuracy.
developed a real-time gait phase Variations in sensor characteristics
detection system for transtibial such as noise or calibration errors may
prosthesis users using wearable impact the reliability of gait
sensors, improving prosthetic recognition systems in real-world
functionality. settings.
2. Gait recognition accuracy can be
In 2024, Ahmedt-Aristizabal et al. [2] significantly affected by clothing, as
reviewed deep learning applications in changes in clothing such as bulk or
seizure video analysis, emphasizing flow of garments can alter gait features
their relevance in clinical gait studies. extracted from silhouettes or statistical
representations. Many methods may
In 2024, Ranjan et al. [23] introduced a only partially address these challenges,
clinical gait analysis dataset with 1,874 leading to reduced accuracy in real
sequences of normal, abnormal, and world scenarios.
pathological gaits, marking a 3. Many gait recognition approaches
significant contribution to gait perform poorly as the camera
research. viewpoint changes significantly
designed for specific angles or rely on
silhouette-based representations. The
viewpoint dependency reduces their
2.1 LIMITATIONS OF CLINICAL
GAIT ANALYSIS generalizability across different
settings.
The limitations are :
4. Advanced gait recognition methods

1. Gait recognition methods are often especially those involving deep learning

tested in controlled environments, but or Vision Transformers can be

real-world conditions such as lighting, computationally intensive. High

shadows, and background clutter may computational demands for both

training and inference may limit the
 applicability of these methods in real or are obstructed by other people.
time or resource constrained Partial occlusions are common in real-
environments. world environments and present a
5. Most gait recognition systems challenge for existing systems.
primarily focus on visual or silhouette 9. Gait patterns can change over time due
based gait analysis. Gait recognition to aging, injuries, or changes in
neglect the integration of additional physical activity levels. Most gait
data modalities, such as depth data, recognition studies do not address the
inertial sensors or physiological signals long-term stability of their models or
could improve robustness and techniques, limiting their effectiveness
accuracy. over extended periods.
6. Some gait recognition studies are 10.Gait recognition systems may raise
focused on specific tasks, such as privacy concerns, particularly in
movement disorder detection. This surveillance applications. Ethical issues
task-specific focus may limit the regarding surveillance and the
adaptability of the methods to other potential misuse of personal data may
gait-related applications, such as biric hinder the widespread adoption or
authentication or health monitoring. acceptance of gait recognition
7. Models trained on specific datasets or technologies.
in controlled conditions may not
3. DATASETS ON CLINICAL GAIT
generalize well to other datasets or ANALYSIS
real-world environments. Variations in
Gait datasets are commonly used in
gait, clothing, or background
computer vision tasks such as gait
conditions that were not part of the
identification. The significant need for
training data can reduce the
clinically suitable gait datasets that
performance of the model.
provide accessible video data for tasks
8. Many gait recognition approaches
such as gait abnormality identification.
struggle of individuals are partially
The current gap highlights the demand
occluded, for example, carrying objects
for datasets specifically designed to Gait recognition has been the
support clinical gait analysis, focusing primary focus of many of the current
on abnormalities that are crucial for gait related datasets. The datasets
diagnosis and treatment in healthcare typically consist of gait sequences in
field. silhouette format derived from video
data. The methods like closed circuit
Gait datasets are frequently
television or video to capture initial
utilized for tasks such as gait
gait sequences, many of these datasets
identification in Computer Vision. The
do not provide access to the original
significant gap in the availability of
video data due to issues with data
clinically suitable gait datasets that
storage and privacy concerns.
provide accessible video data for tasks
such as identifying gait abnormalities. Since the primary task is
Many video datasets have been cited in biometric identification based on gait
the literature for gait related activities, patterns, data collection often assumes
most of these datasets are not clinically the use of representative general
appropriate, large enough or populations exhibiting typical
specifically tailored for clinical gait variations of normal gait patterns.
analysis using visual data. The gap The datasets lack the representation
underscores the need for a new of abnormal or pathological gait types
clinically focused gait dataset namely and the necessary annotations for
the Gait Abnormality in Video Dataset. clinical gait analysis. The examples of
The proposed Gait Dataset aims to normative gait patterns provided
overcome the limitations of current multiple views and represented with
datasets and enable improved clinical two dimensional skeletons and three
gait analysis for identifying gait dimensional body meshes.
abnormalities.
The notable number of gait
datasets created for abnormality or
pathology detection. Many of these recognition have been driven by large
datasets suffer from several datasets of action classes, many
limitations, such as a small number of datasets either lack a dedicated
subjects and sequences, a limited walking class or contain a walking
variety of abnormal gaits, restricted class that is not suitable for Clinical
representation of pathologies, and Gait Analysis due to inconsistent gait
scarce availability of clinically cycles or noisy data. Multi object
annotated data. Privacy concerns tracking involves identifying people
exacerbate these limitations, with and objects from single or multiple
many datasets either being camera views, shares a use case with
unavailable or only accessible in gait recognition in surveillance and
reduced dimensional data formats person identification. Similarly,
such as key point coordinates. The pedestrian tracking, used in robotics
development of using rich features and autonomous vehicles, tracks a
extracted from video data is hindered. person’s path of walking. Many multi
object tracking and pedestrian datasets
Biomechanical analysis of
individual body parts to understand provide bounding box annotations for
human movement in sporting and people walking within a scene. The
everyday activities shares similarities
with abnormality/pathology detection. data has the advantage of capturing
The primary aim is to measure and subjects in natural environments,
describe human movement. Few
studies provide gait data for performing natural gaits. The absence
biomechanical analysis, privacy of clinical annotations and the presence
concerns limit the availability of these
datasets or restrict them to reduced of objects like bags carried by subjects
dimensional data types. can impact gait patterns.
Action recognition is a well-known
Current gait datasets exhibit
task in computer vision classifies
variability in several aspects, including
human actions in images and video.
action sequence lengths, the presence
Recent developments in human action
of subjects without pathological gait,
 variations in camera viewing angles, as gait speeds, clothing items, and
and a range of covariate factors such carried objects.

Table 1. Datasets with Normal Gait Actions and Activities

# Total Total Gait Original Data RGB

Dataset (year) Gait Type
Subjects Sequences Purpose Modalities Available
1 DAI (2016) 2 30 AP RGB, Sil Norm No
2 Ismail (2017) 11 N/S AP 3D-S Norm No
KTH (Class -
3 25 N/S AR RGB Norm No
Walking) (2004)
IXMAS (Class -
4 10 50 AR RGB Norm No
Walk) (2006)
Ogale (Class -
5 10 80 AR RGB Norm No
Walk) (2007)
Weizmann
6 (Class - Walk) 9 9 AR RGB Norm No
(2007)
Huang(Class -
7 8 48 AR RGB Norm No
Walk) (2008)
i3DPost (Class -
8 8 64 AR RGB Norm No
Walk) (2009)
HumanEva I
3D Mocap,
9 (Walking Class) 4 12 AR Norm No
RGB,
(2010)
10 MuHAVi (2010) 14 112 AR Sil Norm No
HMDB51 (Class
11 N/S 100(approx.) AR RGB Norm Yes
– Walk )(2013)
PKU-MMD 3D-S,
12 (Class - 66 N/S AR depth, Norm No
Walking) (2017) RGB
N-UCLA(Class Depth,
13 10 N/S AR Norm Yes
– Walk )(2014) RGB, 3D-S
NTU RGB+D 3D-S,
14 (Class - Walk) N/S N/S AR Depth, Norm Yes
(2016) RGB, TIR
NTU RGB+D 3D-S,
15 120 (Class - N/S N/S AR Depth, Norm Yes
Walk) (2016) RGB, TIR
 Datasets with abnormal or datasets with abnormal gaits have
pathological gait types are typically been produced using actors to simulate
less diverse with smaller number of or wear equipment to simulate an
subjects and sequence size are publicly abnormal gait. Many cases, authors
available due to medical regulations. have collected gait data specific to
The particular case for visual data are their objective and have not made this
challenges in acquiring gait data from public or have provided reduced
populations demonstrates abnormal dimensional data such as only
gait. The limiting factors include providing key point data instead of full
stringent ethical and regulatory RedED Blue Green images. Colour
processes as well as issues related to Representation data is important for
privacy. The barriers for people with current computer vision research, the
mobility issues, common in gait omission of this data from publication
abnormalities to travel and access is potentially limiting development in
institutions such as motion capture healthcare and related fields of human
labs. As a result, many existing gait movement analysis.s.

Table 2. Datasets with Abnormal Gait Actions and Activities

Total Gait Original Data

Total Subjects Gait Type RGB Available
Sequences Purpose Modality
Norm/Path (CP- No
OMally (1997) 156 N/S AD 3D Mocap
child)
Bauckhage Norm/
7 N/S AD RGB, Sil No
(2005) Abnorm(Msk)
Norm/Path (CP-
Wolf (2006) 64 N/S AD 3D Mocap No
child)
GEI, RGB, Norm/Abnorm
DAI2 (2015) 5 75 AD No
Sil (Msk)
Norm/
DGD (2015) 7 56 AD 3D-S No
Abnorm(Msk)
Prochazka2015 51 N/S AD 3D-S Norm /Norm No
 Variant
(2015) ( Elderly)/Path
(PD)
SPHERE2015 Norm/Path
10 40 AD 3D-S No
(2016) (PD)/Path (stroke)
Diplegia (2017) 156 N/S AD 3D Mocap Norm/Path (CP) No
Gholami (2017) 20 20 AD 3D-S Norm/Path(MS) No
Norm/Path (CP-
Laet (2017) 356 1719 AD 3D Mocap No
child)
Norm/
Bei(2018) 120 N/S AD 3D-S No
Abnorm(Msk)
INIT Gait Norm/Abnorm
10 160 AD Sil No
(2018) (Msk)
Norm/
Kozolow (2018) 28 168 AD 3D-S No
Abnorm(Msk)
Norm/Path
Li (2018) 42 N/S AD 3D-S, RGB No
(CP)/Path (PD)
Walking Gait 3D-S, PC, Norm/Abnorm
9 81 AD No
(2018) Sil (Msk)

3.2 Gait Abnormality in patterns in the Gait Abnormality in

Video Dataset Video Dataset. The largest collection
of online accessible abnormal gait
The response to the demand
videos annotated by expert clinicians
for a large and representative video
for clinical analysis.
dataset featuring abnormal gait

Table 3. GAVD Dataset Features Summary

3.2.1 Data collection and were manually sourced from public

annotation domain websites and video. Expert
Publicly available online clinicians screened and annotated
videos of normal and abnormal gaits
relevant videos to ensure clinically obstruction, walking in a singular
accurate identification of abnormal direction relative to the camera.
gaits.
4 . Experimental Studies
Abnormal gaits are defined as gait
An experimental study aimed to
patterns that significantly deviate from
validate the effectiveness of Gait
statistical or biomechanical norms
Abnormality Video Dataset, for
based on subjective appearance. The
Computer Gait Analysis tasks in
patterns consist of gaits resulting from
computer vision and identify potential
pathologies, simulated abnormalities,
areas for future research. The
and exercises such as toe walking
promising potential of region based
which consists of walking with
action recognition models in
modified posture. The importance to
identifying gait abnormalities from
note the term abnormal does not imply
video data, our focus shifted towards
negativity or inferiority but rather
utilizing such models for Computer
indicates deviation from typical gait
Gait Analysis. The SlowFast and
patterns. Subjective identification of
Temporal Segment Network models
abnormal gait is often the initial step
were finetuned for the binary
in diagnosing issues and pathologies.
classification task of recognizing
Clinical experts screen and abnormal gaits from normal gaits is
identify normal and abnormal gait crucial for the initial identification of
sequences from over Four hundred gait issues and potential pathologies.
and fifty online videos. Many video is SlowFast and Temporal Segment
further annotated frame by frame Network were selected for their
focusing on subjects demonstrating efficiency and effectiveness in learning
gait sequences consisting of at least from entire action video clips, such as
two gait cycles with minimal visual those collected in the Gait Abnormality
Video Dataset.
4.1 Fine-tuning Action Recognition Study Dataset. The models
with Gait Abnormality Video Dataset demonstrated strong performance on

The fine tuning of action these test sets, confirming their utility

recognition models involved selecting a for clinical gait analysis.

total of five hundred eighty two gait

4.1 RESULTS AND ANALYSIS
sequences, consisting of two hundred
The pre-trained Temporal
ninety one normal and two hundred
Segment Network and SlowFast
ninety one abnormal from the Gait
models demonstrated strong
Abnormality Video Dataset. Pre
performance in detecting abnormal
trained models from the Kinetics four
gait patterns achieving high accuracy
hundred dataset were fine tuned using
on the Gait Abnormality Video
the Mmaction two toolbox with
Dataset. The accuracy decreased when
preprocessing focused on the
tested on the Gait Pattern Joint Action
individual's actions by annotating
Task Knowledge dataset with the
bounding boxes and applying padding.
SlowFast model performing better
A ten fold cross validation protocol was
than the Temporal Segment Network
used, with models trained for up to two
for detecting normal gaits. The models
hundred epochs, demonstrating
showed improved performance when
consistent accuracy and loss during
subjects were viewed from the front or
validation. Temporal Segment
back but struggled with side views
Network and SlowFast models were
suggesting potential issues with
evaluated on unseen data from three
generalizability across different
distinct sources: abnormal gait from
camera perspectives. The Clinical
the Gait Abnormality Video Dataset,
Abnormality Study Dataset revealed
normal gait from the Gait Pattern
that both models frequently
Joint Action Task Knowledge gait
misclassified side view gait sequences
dataset, and a clinically controlled gait
as abnormal highlighting the need for
dataset from the Clinical Abnormality
further investigation into the impact of
camera angles on model performance, particularly in abnormality detection.

Table 4. Comparison of model accuracy in gait normality/abnormality classification

on test sets
TESTSET SLOWFAST TSN
GAVD test set 0.94 0.92
GPJATK test set 2 0.76 0.59
CASD test set 3 0.87 0.53

4.1.1 DISCUSSIONS investigations should focus on camera

lens distortion and angle influence
The pre-trained action
model performance.
recognition models on the Gait
Abnormality Video Database 4.1.2 ENHANCEMENTS OF
successfully identified abnormal gait CLINICAL GAIT ANALYSIS

patterns although the false positives

The availability of numerous
were observed in sequences from the
gait datasets such as the Chinese
left or right side of the individual. The
Academy of Sciences Institute of
models showed strong performance
Automation and Osaka University
when individuals walked towards or
Image and Speech Information
away from the camera but faced
Retrieval Multi View Large Scale Pose
challenges in detecting abnormal gaits
datasets, the lack of abnormal gait
when the person moved sideways
datasets that can be applied to a
relative to the camera. The model
variety of gait analysis tasks, especially
emphasizes the importance of
in real world clinical analysis. The
movement direction and camera angle
following contributions are presented:
in detection implementing view
1. A comprehensive summary
invariant methods is crucial for
encompassing over one hundred
improving accuracy and further
 and fifty existing gait related addressing barriers such as the
datasets for computer vision is lack of medically validated
provided. The Gait Abnormality movement data. The dataset
in Video Dataset is introduced as focuses on video data, aligning
an accessible video dataset with trending healthcare
annotated by clinical developments such as telehealth
professionals. It is designed to and self-monitoring for
support the development of healthcare.
computer vision methods for 2. Baseline results are provided
clinically validated gait analysis investigating the suitability of
(Figure 1). The Gait region-based action recognition
Abnormality in Video Dataset models for the task of gait
aims to advance the state of the abnormality classification.
art in gait analysis and
classification research by

Figure 1. Example frames representing a variety of subjects from GAVD

dataset

CONCLUSION gait analysis methodologies, with a

growing focus on gait and neuro
The last decade has seen degenerative diseases within the field
significant advancements in human of Clinical Gait Analysis. The
application of computer vision Females with Patellofemoral
technology in this domain has Pain Syndrome.
garnered attention, particularly in 2. Ahmedt-Aristizabal, D., Armin,
the context of leveraging spatial- M.A., Hayder, Z., Garcia-
temporal visual information to aid Cairasco, N., Petersson, L.,
clinical gait analysis. Further efforts Fookes, C., Denman, S.,
are required to provide clinically McGonigal, A. (2024): Deep
relevant insights into the spatial and Learning Approaches for Seizure
temporal nature of movement Video Analysis: A Review.
abnormalities. The introduction of Epilepsy & Behavior 154.
the Gait Abnormality in Video 3. Akae, N., Mansur, A., Makihara,
Dataset offers a validated, scalable Y., Yagi, Y. (2012): Video from
dataset with clinically annotated data Nearly Still: An Application to
by human movement experts. The Low Frame-Rate Gait
dataset aims to address challenges Recognition. Presented at the
such as camera view invariance, IEEE Conference on Computer
pathology adaptation, and frame- Vision and Pattern Recognition
wise classification, ultimately 4. Albuquerque, P., Verlekar, T.T.,
improving abnormality detection and Correia, P.L., Soares, L.D.
contributing to the enhancement of (2021): A Spatiotemporal Deep
clinical gait analysis using computer Learning Approach for
vision in healthcare. Automatic Pathological Gait
REFERENCES Classification. Sensors (Basel,
Switzerland) 21(18), 6202
1. Abdelraouf, O.R., Abdel-Aziem,
5. Al-Jubouri, A.A., Ali, I.H.,
A.A., Ahmed, A.A., Nassif, N.S.,
Rajihy, Y. (2021): Gait and Full
Matar, A.G. (2019): Backward
Body Movement Dataset of
Walking Effects on Activation
Autistic Children Classified by
Pattern of Leg Muscles in Young
Rough Set Classifier. Journal of
 Physics: Conference Series Individuals Walking in Front of
1818(1), 012201 a X-Box 360 Kinect Sensor.
6. Amini, A., Banitsas, K. (2019): 10.Arora, P., Srivastava, S.,
Using Kinect v2 to Control a Singhal, S. (2016): Analysis of
Laser Visual Cue System to Gait Flow Image and Gait
Improve the Mobility during Gaussian Image Using Extension
Freezing of Gait in Neural Network for Gait
7. An, W., Yu, S., Makihara, Y., Recognition. International
Wu, X., Xu, C., Yu, Y., Liao, R., Journal of Rough Sets and Data
Yagi, Y. (2020): Performance Analysis
Evaluation of Model-Based Gait 11.Baker, R., Esquenazi, A.,
on Multi-View Very Large Benedetti, M., Desloovere, K.
Population Database With Pose (2016): Gait Analysis: Clinical
Sequences. IEEE Transactions Facts. European Journal of
on Biometrics, Behavior, and Physical and Rehabilitation
Identity Science 2(4), 421–430 Medicine 52(4), 560–574
8. Anggraeni, N.D., Ferryanto, F., 12.Bei, S., Zhen, Z., Xing, Z.,
Atmojo, S., Mihradi, S., Taocheng, L., Qin, L. (2018):
Dirgantara, T., Mahyuddin, A. Movement Disorder Detection
(2015): Gait Parameters via Adaptively Fused Gait
Determination by 3D Motion Analysis Based on Kinect
Analyzer System for Initial 13.Benson, L.C., Räisänen, A.M.,
Indonesian Gait Database. Tech. Clermont, C.A., Ferber, R.
Rep., (2022): Is This the Real Life, or
9. Araujo, R.M., Andersson, V. Is This Just Laboratory? A
(2014): Kinect Gait Biometry Scoping Review of IMU-Based
Dataset - Data from 164 Running Gait Analysis. Sensors
22(5), 1722
14.Bertaux, A., Gueugnon, M., Gait-Based Person-Verification
Moissenet, F., Orliac, B., Martz, System for Forensics. Presented
P., Maillefert, J.F., Ornetti, P., at the IEEE Fifth International
Laroche, D. (2022): Gait Conference on Biometrics:
Analysis Dataset of Healthy Theory, Applications, and
15.Bhattacharya, U., Mittal, T., Systems (BTAS)
Chandra, R., Randhavane, T., 19.Makihara, Y., Okumura, M.,
Bera, A., Manocha, D. (2020): Iwama, H., Yagi, Y. (2011): Gait-
STEP: Spatial Temporal Graph Based Age Estimation Using a
Convolutional Networks for Whole-Generation Gait
Emotion Perception from Gaits. Database. Presented at the
Proceedings of the AAAI International Joint Conference
Conference on Artificial on Biometrics (IJCB)
Intelligence 34(02), 1342–1350 20.Mogan, J.N., Lee, C.P., Lim,
16.dos Santos, C.F.G., de S. K.M., Muthu, K.S. (2022): Gait-
Oliveira, D., de F.S. Santos, D. ViT: Gait Recognition with
(2022): Gait Recognition Based Vision Transformer.
on Deep Learning: A Survey. 21.Muro-de-la Herran, A., Garcia-
17.Hossain, M.A., Makihara, Y., Zapirain, B., Mendez-Zorrilla,
Wang, J., Yagi, Y. (2010): A. (2014): Gait Analysis
ClothingInvariant Gait Methods: An Overview of
Identification Using Part-Based Wearable and Non-Wearable
Clothing Categorization and Systems, Highlighting Clinical
Adaptive Weight Control. Applications. Sensors
Pattern Recognition 22.Okumura, M., Iwama, H.,
International Journal of Therapy Makihara, Y., Yagi, Y. (2010):
and Rehabilitation 26(1), 1–9 Performance Evaluation of
18.Iwama, H., Muramatsu, D.,
Vision-Based Gait Recognition
Makihara, Y., Yagi, Y. (2012):
 Using a Very Large-Scale Gait 25.Zhang, C., Zhang, S., Yang, J.,
Database. Cheng, W. (2015): Biometric
Parkinson’s Disease. Journal of Recognition. Volume 9428.
Healthcare Engineering 2019, 1–
8
23.Rahm Ranjan, David Ahmedt-
Aristizabal, Mohammad Ali
Armin, Juno Kim. (2024).
Computer Vision for Clinical
Gait Analysis: A Gait
Abnormality Video Dataset.
Journal of Medical Imaging and
Health Informatics, 154.
24.Rattanasak, A., Uthansakul, P.,
Uthansakul, M., Jumphoo, T.,
Phapatanaburi, K.,
Sindhupakorn, B., Rooppakhun,
S.(2022),Real-Time Gait Phase
Detection Using Wearable
Sensors for Transtibial
Prosthesis Based on a kNN
Algorithm.,10.3390/s22114242
Sensors. IEEE Sensors Journal
18(17), 7305–7314
Volunteers and Patients Before
and 6 Months After Total Hip
Arthroplasty. Scientific Data
9(1), 399