You are on page 1of 55

HUMAN RECOGNITION

USING GAIT

Rayhan Chaudhary - 2021UCD2118


Shubham Verma – 2021UCD2155
Introduction

Gait Recognition is a type of behavioral biometric authentication that


recognizes and verifies people by their walking style and pace. Advances
in gait recognition have led to the development of techniques for forensics
use since each person can have a gait defined by unique measurements
such as the locations of ankle, knee, and hip.
How Does Gait Recognition Work?

 The system analyzes the silhouette, height, speed, and walking


characteristics and identifies the individual from a database.
 This technique is more convenient than retinal scanners or fingerprints in
public places as it doesn’t attract attention.
 Moreover, gait recognition is unlikely to be outsmarted — every person's
gait has no duplicates.
 A gait recognition system uses the shape of the human body and the way
it moves in order to identify it. The software, using CV algorithms,
detects a human silhouette on video and analyzes its movements. These
data create a human behavioral model.
Applications of Gait Recognition
 Video surveillance; handy in crowded areas for security purposes.
 The systems can quickly spot a person who is wanted by the authorities and prevent
terrorist attacks or other forms of crime. The latest gait recognition systems can
recognize people even from a distance up to 50 meters
 Smart home ecosystem
 Nursing homes and hospitals
 to alert staff if a patient falls
 can help diagnose neurological disorders and plan rehabilitation therapy by analyzing
the posture and walking patterns of the recovering patients
 Athletes
 helpful in training to assess movements closely
Parameters of Gait Recognition
 Step Length is the length measured from the heel of the previous footfall to the heel of
the current opposing footfall.
 Stride Length is the distance measured between the Posterior Heel points of two
consecutive footprints of the foot in question
 Step Width is the distance measured between the left foot and the right foot.
 Step Angle is the angle between the line of progression and the foot axis.
How is Data Collected for Gait Recognition?

 Video cameras or wearable sensors can be used to capture a gait.


The most striking example of such sensors may well be the special
costumes that actors wear on set so that motion artists can later draw the
character based on their movements.
 Another method of gait capture involves the use of radar to identify
moving objects remotely.
The object of interest is irradiated with radio waves that reflect off their
body. The system recognizes the reflected waves and uses the data for
identification.
Challenges in Data Collection for Gait Recognition

 Change in the viewpoint


 People carrying an item
 People walking in arbitrary directions
 People wearing heavy coats
 Inadequate lighting
 Improper camera angle
 Bad video resolution of camera, etc.
Feature Extraction in Gait
 A feature vector is a method for representing an image feature, or a part
of an image (an object), by measuring a set of features. These
measurements are stored in a feature vector, which is an n-dimensional
vector. This vector could be utilized to classify an object or to give the
higher-level visual information.
Feature Extraction in Gait
 The features can be extracted from a silhouette. A silhouette is to be
obtained of the person walking in front of the camera.
 Background subtraction is an effective way of generating a silhouette, it
is also used in reducing noise.
 A bounding rectangle is used to isolate the silhouette.
Feature Extraction in Gait
 Silhouette is a useful feature to employ as:
 most body parts are captured, along with the motion
both structural and transitional information is captured
it is particularly unaffected by clothes, lighting, and materials, etc.
Different Approaches for Gait
 Kinematic-Based Approaches:
 Silhouette Analysis: Extracting gait features by analyzing silhouettes of individuals
obtained from video sequences.
 Skeleton Analysis: Extracting gait features by analyzing the motion of key skeletal
joints obtained from motion capture systems or depth sensors.
 Model-Based Approaches:
 Hidden Markov Models (HMMs): Modeling gait patterns as sequences of hidden
states and observations, capturing temporal dependencies.
 Gaussian Mixture Models (GMMs): Modeling the probability distribution of gait
features using a mixture of Gaussian distributions, often used in combination with
HMM’s
HMM Framework for Gait Recognition
 The Hidden Markov Model (HMM) framework involves using HMMs to
model and recognize human gait patterns.
 HMMs offer a powerful mathematical framework for capturing the
temporal dynamics inherent in gait sequences.
 The HMM framework offers several advantages for gait recognition:
 Ability to Model Temporal Dynamics: HMMs can capture the sequential nature of
gait patterns, making them suitable for handling time-series data.
 Robustness to Variability: HMMs can accommodate variability in gait patterns due
to factors such as walking speed, clothing, footwear, and environmental conditions.
 Adaptability: HMM parameters can be adapted or updated over time to
accommodate changes in gait patterns or environmental conditions.
Components of HMM Framework
 States (Hidden States):
 States represent the underlying dynamics or patterns in the gait sequence.
 In gait recognition, states could correspond to different phases of the gait cycle (e.g.,
stance phase, swing phase).
Components of HMM Framework
 Observations:
 Observations are the measurable features extracted from the gait data.
 In gait recognition, observations could include various kinematic or appearance
features such as joint angles, step length, silhouette characteristics, etc.
 Transition Probabilities:
 Transition probabilities represent the likelihood of transitioning from one state to
another in the HMM.
 Transition probabilities are typically represented by a transition matrix, where each
element denotes the probability of transitioning from one state to another.
Components of HMM Framework
 Emission Probabilities:
 In gait recognition, emission probabilities model the relationship between gait features
and the underlying gait phases.
 Emission probabilities are often represented by probability distributions (e.g., Gaussian
distributions for continuous features) or histograms (e.g., discrete features).
 Initialization Parameters:
 Initialization parameters include initial state probabilities and initial values for
transition and emission probabilities.
 These parameters are typically estimated from training data using methods such as
maximum likelihood estimation (MLE) or expectation-maximization (EM) algorithms.
Components of HMM Framework
 Inference Algorithms:
 Inference algorithms calculate the likelihood of observing a given sequence of gait
features and decode the most likely sequence of hidden states.
 Common inference algorithms include the forward algorithm, backward algorithm,
and Viterbi algorithm.
 Training Algorithm:
 Training algorithms are used to estimate the parameters of the HMM from training
data.
 In gait recognition, training algorithms adjust the transition and emission probabilities
to maximize the likelihood of observing the training data.
 The Baum-Welch algorithm, also known as the expectation-maximization (EM)
algorithm, is commonly used for training HMMs.
Inference Algorithms: Forward Algorithm

 The forward algorithm calculates the likelihood of observing a given


sequence of sensor data over time, given the parameters of the HMM
(transition probabilities, emission probabilities, initial state probabilities).
 It is used to evaluate the overall likelihood of a sequence of sensor data
given the model.
 This is crucial for tasks such as parameter estimation (e.g., training the
HMM using the Expectation-Maximization algorithm) and for comparing
the fit of different models to the observed data.
Inference Algorithms: Viterbi Algorithm

 The Viterbi algorithm finds the most likely sequence of hidden states (gait
states) given the observed sequence of sensor data.
 It is used to decode the underlying sequence of gait states from observed
sensor data.
 This can provide valuable information about the individual's gait pattern and
behavior, such as identifying different phases of gait (e.g., walking, running,
standing) or detecting abnormal gait patterns.
Inference Algorithms: Difference between the two

 Main difference between Forward algorithm and Viterbi algorithm lies in


their objectives:
 Forward algorithm computes the likelihood of observing a sequence of sensor data
given the model,
 While the Viterbi algorithm determines the most likely sequence of hidden states (gait
states) given the observed data.
 Both algorithms play important roles in gait analysis using HMMs, but
focus on different aspects:
 Forward algorithm is more focused on likelihood evaluation and parameter estimation,
 While Viterbi algorithm is focused on state decoding and inference.
Training Algorithm: Baum-Welch algorithm

 The Baum-Welch algorithm is a fundamental tool in Hidden Markov Model


(HMM) training, which can be applied to gait recognition.
 Here are the steps of the Baum-Welch algorithm for gait recognition:
Step 1: It starts with an initial guess for the parameters of the HMM, including
the transition probabilities, emission probabilities, and initial state
probabilities.
Step 2: Use the forward-backward algorithm to compute the forward and
backward probabilities for each observation sequence. This step involves
calculating the probability of being in each state at each time step given the
observations.
Training Algorithm: Baum-Welch algorithm

Step 3: Calculate the expected number of transitions between states and the
expected number of times each state emits a particular observation, based on
the forward and backward probabilities computed in the previous step
Step 4: Update the model parameters: transition probabilities, emission
probabilities, initial state probabilities, based on the expected counts calculated.
This step involves maximizing the likelihood of the observation sequences
given the current model parameters.
 Repeats these steps iteratively until the parameters of the HMM converge to
stable values.
 The parameters obtained after convergence represent the trained HMM,
which provides a good model of the underlying gait patterns captured by the
observed sensor data.
View Invariant Gait Recognition

 It refers to the ability to identify or recognize an individual's walking pattern


or gait regardless of the viewing angle or perspective from which it is
observed.
 In real-world scenarios, it's common for surveillance cameras to capture individuals
from various perspectives, distances, and lighting conditions. View-invariant gait
recognition addresses this challenge by developing algorithms and techniques that can
accurately identify individuals' gaits from different viewpoints.
 Goal is to create robust systems that can effectively recognize an
individual's gait regardless of factors such as camera angle, distance,
clothing variations, lighting conditions, and occlusions
View Invariant Gait Recognition:
Model-Based Approach
 Involves analyzing the underlying biomechanical structure and dynamics of
human gait to create a mathematical model that can accurately represent
walking patterns.
 Focuses on capturing the inherent characteristics of gait that remain
consistent across different viewing angles and perspectives.
 Typically involves representing the human body as a set of interconnected
segments (e.g., legs, torso, arms) and joints, with each segment
characterized by its position, orientation, and motion parameters.
View Invariant Gait Recognition:
Model-Based Approach
 Step 1: Extract features from the observed gait sequences that are
consistent across different viewing angles.
 Step 2: Transform the extracted gait features into a common reference
frame or normalizing them to remove viewpoint-specific variations.
 Step 3: Apply Machine learning or pattern recognition algorithms to classify
gait patterns based on the view-invariant features extracted from the model.
Classifiers may include techniques such as support vector machines (SVMs),
neural networks, or hidden Markov models (HMMs).
View Invariant Gait Recognition:
Appearance-Based approach
 It focuses on extracting and analyzing visual features directly from the
observed images or video sequences of human walking.
 Unlike model-based approaches, appearance-based methods do not require
explicit modeling of the underlying gait dynamics. Instead, they capture
discriminative features directly from the appearance of the walking subject,
such as silhouette shapes, texture patterns, or motion cues.
 It offers advantages such as simplicity, computational efficiency, and
flexibility in handling complex gait variations.
 However, it may be sensitive to changes in lighting conditions, occlusions,
and variations in clothing or appearance.
View Invariant Gait Recognition:
Appearance-Based approach
 Step 1: Extract features from the appearance of the walking subject
captured in the video frames or images.
 Step 2: Perform view normalization or transformation to mitigate the effects
of viewpoint variations. Effective view normalization techniques are crucial
for ensuring the approach's view-invariance capabilities.
Techniques such as image warping, geometric alignment, or camera calibration
may be employed to bring the observed gait sequences into a common
reference frame.
 Step 3: Employ machine learning algorithms or pattern recognition
techniques to classify or recognize gait patterns based on the appearance-
based features.
View Invariant Gait Recognition:
Hybrid approach
 It combines elements of both model-based and appearance-based methods
to leverage their respective strengths and overcome their limitations.
 Step 1: Extract features from the observed gait sequences using a
combination of model-based and appearance-based techniques.
 Model-based features may include biomechanical parameters, joint angles, or dynamic
characteristics derived from a skeletal model of human gait.
 Appearance-based features may include silhouette shapes, texture patterns, or motion
cues extracted directly from the video frames or images.
 Step 2: Perform view normalization or transformation to mitigate the effects
of viewpoint variations on the extracted features.
View Invariant Gait Recognition:
Hybrid approach
 Step 3: Combine the model-based and appearance-based features into a
unified feature representation, using techniques such as:
 Concatenation: Simply concatenate the feature vectors obtained from different sources
into a single vector.
 Weighted Fusion: Assign different weights to the features from each source based on
their importance or reliability and combine them accordingly.
 Step 4: Employ machine learning algorithms or pattern recognition
techniques to classify or recognize gait patterns based on the hybrid feature
representation.
ROLE OF SHAPE AND
DYNAMICS
IN
GAIT RECOGNITION
EARLY DAYS AND SHORTCOMINGS
Gait recognition, the art of identifying individuals by their walk, boasts a rich history. Early
methods laid the groundwork for understanding how human movement patterns can be
unique identifiers. This approach, however, primarily focused on capturing the body's static
shape.

Imagine capturing a person's walking outline, like a shadow puppet. This essentially
describes the initial approach in gait recognition. Techniques relied on silhouettes, the
body's outline without details.

These early methods, while groundbreaking, had limitations. Variations in clothing and
viewing angles significantly impacted their accuracy.
SILHOUETTE-BASED METHODS
Imagine a person walking past a brightly lit wall. Early gait recognition techniques mirrored this scenario,
relying on silhouettes, the body outline without details. These silhouettes offered a basic understanding of a
person's body proportions and posture during walking.

There were two main ways silhouettes were used:

1. Single Silhouettes: A simpler method captured a single silhouette at a specific point in the gait cycle,
typically when both feet were on the ground (stance phase). This provided a snapshot of the person's
posture at a particular moment, offering information about body proportions like leg length and torso
width.

2. Silhouette Sequences: As technology advanced, researchers captured a sequence of silhouettes


throughout the entire gait cycle, encompassing both stance and swing phases. This sequence offered a
more dynamic picture of the walking motion, revealing subtle variations in posture as the legs moved
forward.
Single Silhouette :
(Single snapshot for a single gait cycle)
Silhouette Sequences :
(Multiple snapshots for a single Gait cycle)
LIMITATIONS OF THE SILHOUETTE APPROACH
While these early methods were groundbreaking, they had limitations that hindered their
robustness:

• Clothing Variations: Changes in clothing significantly altered the captured silhouette. A bulky
jacket or loose pants could drastically affect the body outline, making identification challenging.

• Viewing Angle Dependence: Recognition performance dropped significantly when the view
wasn't ideal. Looking at someone from the side or back presented a completely different
silhouette compared to a head-on view, hindering accurate identification.

Due to these limitations, researchers began to explore the complexities of human movement itself,
leading to the development of dynamics-based gait recognition methods. The next section will delve
into how researchers shifted their focus to capture the essence of how a person walks, leading to
more robust identification systems.
Gait Energy Images (GEI): Capturing the Rhythm of
Movement
The GEI Concept:
Imagine stacking a sequence of silhouettes captured throughout a person's gait cycle, one on top of another.
Areas where the silhouettes frequently overlap will appear brighter in the resulting image. This cumulative
effect creates a "heat map" of gait patterns, where brighter regions highlight areas of the body that move more
frequently or with greater range during walking.

Benefits:
• Dynamic Glimpse: GEI captures some level of movement, highlighting frequently used body regions during
gait (e.g., pronounced hip swing).
• Data Efficiency: It compacts a silhouette sequence into a single image, potentially saving storage and
processing resources.

Drawbacks:
• Limited Dynamics: GEI only considers overlap frequency, not actual movement direction. It doesn't fully
capture the intricacies of gait.
• Noise Sensitivity: Clothing variations or changes in walking speed can affect the silhouette shape and
consequently the resulting GEI.
Gait Energy Images (GEI):
(Heatmap becomes more defined as the number of snapshots increase)
LIMITATIONS OF THE “SHAPE” APPROACH
While these initial methods provided a foundation, they faced limitations:

• Vulnerability to Clothing: Variations in clothing significantly altered the captured


silhouette, making it difficult to distinguish between individuals wearing bulky
jackets or loose pants.

• Viewing Angle Dependence: Recognition performance dropped when the view


wasn't head-on. Looking at someone from the side or back drastically changed
the apparent shape, hindering identification.
The Deep Learning Revolution: A Shape-Dynamics
Fusion
The emergence of deep learning,
particularly convolutional neural
networks (CNNs), marked a turning
point in gait recognition.

CNNs excel at extracting features


and learning complex relationships
from data, making them ideal for
Proposed CNN structure for gait recognition. There are two, 2D-convolutional layers with the
analyzing the intricacies of human following parameters: f = number Of filters. k kernel size, s stride. p padding, and n = number
of nodes. Please refer Pg-14 for working.
movement.
Unlike traditional methods that focused on either shape or
dynamics, deep learning models can analyze both aspects
simultaneously:

1. Shape Analysis: Deep learning can leverage:


• Silhouette representations: These capture the overall body
shape from video frames.
• Pose estimation: Deep learning models can estimate the
positions of key body joints (knees, ankles, hips) from
images or videos, even with partial occlusions.

2. Dynamics Extraction: By analyzing sequences of images or


videos, deep learning models can capture the flow of
movement:
• Temporal features: Deep learning can automatically learn
the timing and rhythm of gait cycles, including the duration
of stance and swing phases.
• Motion patterns: The model can extract features that
represent the specific ways a person moves their limbs
Illustration Of gait kinematics information
acquisition, including angles and velocity Of the (e.g., swing amplitude of arms).
joints.
BENEFITS OF SHAPE-DYNAMICS FUSION
By considering both static information (shape) and dynamic changes (movement), deep learning models
create a richer representation of an individual's gait, leading to several advantages:

1. Improved Accuracy: Combining shape and dynamics provides a more comprehensive and distinctive
signature for identification. This can significantly improve recognition accuracy compared to pre-deep
learning approaches.

2. Enhanced Robustness: Deep learning models become less susceptible to factors that can affect gait
recognition, such as:
• Viewpoint variations: CNNs can learn to recognize gait patterns even if the person is captured from
different angles.

• Clothing and carrying objects: Deep learning models can focus on the underlying movement patterns
that are less affected by clothing or carried objects.

• Limited training data: Deep learning models have the ability to learn more effective representations
with smaller datasets compared to traditional methods.
WORKING OF A CNN BASED GAIT RECOGNITION MODEL
FACTORIAL HMM AND
PARALLEL HMM FOR GAIT
RECOGNITION
Feature and Decision Level Fusion with
FHMM and PHMM
1. While deep learning offers a powerful solution for gait recognition, other techniques can be valuable
depending on the specific application and data availability. Here, we explore two methods for
information fusion:
• Factorial Hidden Markov Models (FHMMs)
• Parallel Hidden Markov Models (PHMMs).

2. As discussed earlier, information fusion refers to the strategy of combining information from multiple
sources (e.g., different gait features) to create a more robust and accurate representation of an
individual's gait pattern. In gait recognition, this typically involves fusing features like shape
information (body proportions) and dynamic information (joint angles). By combining these, we can
overcome limitations associated with relying on a single feature, leading to:
•Improved Accuracy: Fusing information provides a richer description of gait, potentially leading to
more accurate identification compared to single-feature approaches.
•Enhanced Robustness: Recognition becomes less susceptible to factors that can affect gait
analysis, such as viewpoint variations, clothing, and limited training data.
FIRST OF ALL, A DIFFERENCE :
HMM VS NEURAL NETWORKS
HIDDEN MARKOV MODELS NEURAL NETWORKS
1. Focus on States and Transitions: HMMs represent a 1. Learn from Data: Neural networks are machine
system as a series of hidden states that you can't learning models that learn complex patterns from
directly observe. The model captures the probability of data. They consist of interconnected layers of artificial
transitioning between these states and the probability neurons that process information and improve their
of emitting certain observations from each state. predictions through training.

2. An HMM would model: 2. Light Bulb Example (using a Simple Neural Network):
• The probability of the bulb being burnt out (State 1) Imagine you have a dataset that includes information
or working (State 2). about the bulb's age, usage history (on/off cycles),
• The probability of the bulb transitioning from burnt and whether it's currently on/off. A neural network
out to working (unlikely) or vice versa (more likely). could learn from this data to:
• The probability of the bulb emitting light (observed • Predict the bulb's current state (burnt out or
as "on") when it's working (State 2). working) based on its age and usage history.
• The probability of the bulb emitting no light • Predict the likelihood of the bulb burning out in the
(observed as "off") even when it's working (due to a future.
switch being off) or because it's burnt out (State 1).
FACTORIAL HIDDEN MARKOV MODELS(FHMM)
One approach to information fusion in gait recognition is through Factorial Hidden Markov Models
(FHMMs). Unlike deep learning models that can learn complex relationships directly from data, FHMMs
offer a structured approach to combine information from multiple gait features.

Working Principle :
Imagine a layered cake, where each layer represents a distinct gait feature. Within each layer, a hidden
Markov model operates. These hidden Markov models capture the underlying dynamics of the specific
feature. For instance, the silhouette layer might have hidden states representing different poses during
the gait cycle (stance, swing).These hidden states can have multiple hidden sub-states that capture the
independent influence of factors affecting the system. Similarly, the joint angle layer might have hidden
states corresponding to different angle ranges for knees and ankles.

The magic happens during the training process. Here, information exchange occurs between these layers.
Imagine frosting flowing between the cake layers. This allows the FHMM to learn the relationships
between features. For example, the model might learn that a specific silhouette pose (e.g., heel strike)
often coincides with a particular range of knee and ankle angles.
Question - HOW GAIT RECOGNITION IS PERFORMED
BY FHMM’S?
In gait recognition with Factorial Hidden Markov Models (FHMMs), hidden states play a crucial role in capturing the
complex and nuanced aspects of human walking patterns. Here's a deeper dive into their power:

1. Human gait is influenced by underlying factors like muscle movements, posture, and coordination, which are not
directly observable.

2. FHMMs introduce hidden states, which represent these underlying factors influencing the observed gait data.

3. Each hidden state captures a specific combination of sub-states representing independent factors like:
• Leg Movement: "Swing" or "Stance" for each leg.
• Upper Body Movement: "Bent Torso" or "Straight Torso.“

4. The true power lies in combining these sub-states into factorial hidden states. These combined states represent
specific moments in a person's gait cycle:
• Example: (Left Leg - Swing, Right Leg - Stance, Upper Body - Bent Torso) depicts a specific walking posture with
a left leg swing and a bent torso.
FACTORIAL HIDDEN MARKOV MODEL

HIDDEN LAYERS
WITHIN EACH HIDDEN LAYER, A HIDDEN
MARKOV MODEL OPERATES

OBSERVABLE LAYER/S
BENEFITS OF USING FHMM
1. Reduced Dimensionality: Compared to single-feature HMMs that concatenate features, FHMMs avoid the issue of
high dimensionality. By modeling features in separate layers, FHMMs can improve training efficiency and potentially
reduce overfitting, especially with limited training data.

2. Effective with Weakly Correlated Features: FHMMs are well-suited for scenarios where the gait features exhibit weak
correlations. By capturing the underlying relationships, FHMMs can potentially improve recognition accuracy
compared to single-feature models.

3. Capture Individuality: By modeling these combinations, FHMMs can capture the unique aspects of a person's gait,
even with variations in clothing, walking speed, or carrying objects.

4. Handle Asynchronous Movements: Human gait isn't always perfectly symmetrical. FHMMs can model independent
leg movements effectively, allowing for asynchronous leg swings.

5. Robustness to Variations: Variations in walking surface or carrying objects might influence observed data (video
frames) but the underlying gait pattern remains similar. FHMMs with rich hidden states can focus on these underlying
patterns for better recognition.
PARALLEL HIDDEN MARKOV MODELS(PHMM)
Another approach to information fusion in gait recognition is through Parallel Hidden Markov
Models (FHMMs). Unlike deep learning models that can learn complex relationships directly from
data, Parallel Hidden Markov Models (PHMMs) offer a more refined approach by considering gait
as a combination of independent factors analyzed in parallel.

Working :

1.Factorization: The initial step involves decomposing the gait data (e.g., video frames) into
independent factors influencing the overall walking pattern. These factors could include:
• Leg Movements: Each leg's movement can be considered a separate factor.
• Upper Body Posture: The body's posture (bent or straight torso) can be another factor.
• Additional factors (optional): Depending on the application, arm swing patterns or head
orientation could also be included.
2. Sub-models with Hidden States: Each independent factor is modeled by a separate Hidden Markov Model
(HMM). These sub-models have their own set of hidden states representing the unseen aspects influencing
that specific factor. For example:
• Leg movement sub-model might have hidden states like "Stance" or "Swing" for each leg.
• Upper body posture sub-model might have hidden states like "Bent Torso" or "Straight Torso.“

3. Parallel Analysis: Unlike standard HMMs, PHMMs analyze these sub-models simultaneously. This allows for
the consideration of how different factors influence each other during gait.

4. Interaction Modeling: PHMMs learn the transition probabilities between hidden states within each sub-
model. Additionally, they learn how the hidden states in different sub-models interact and influence each
other. For instance, the model might learn that a specific leg movement ("Heel Strike") often coincides with a
particular upper body posture ("Straight Torso").

5. Recognition: When presented with a new gait sequence, the PHMM analyzes the hidden states in each sub-
model and considers their interactions. By comparing these states and their interactions to the learned
patterns, the model can potentially identify the individual based on their unique gait signature.
BENEFITS OF USING PHMM
1.Robustness to Variations: PHMMs excel at handling variations in external factors like clothing,
carried objects, or even walking surfaces. This is because they focus on independent factors
influencing gait. Each factor (e.g., leg movements) is analyzed separately, allowing the model to
capture the underlying gait patterns that remain consistent despite external variations.

2.Detailed Feature Analysis: By analyzing each independent factor with its own hidden states,
PHMMs provide a more detailed understanding of gait compared to traditional models. This allows
for a richer picture of the walk, potentially capturing specific details about individual gait styles.

3.Simpler Model Structure: Compared to FHMMs (Factorial Hidden Markov Models), PHMMs have
a simpler model structure. They utilize separate HMMs for each factor, which can be easier to train
and interpret. This simplicity can be beneficial, especially when dealing with limited training data.
THANK YOU!!

You might also like