MODULE 4

SUBMODULE 1
Question 109 (14 Marks)

Illustrate the steps involved in building an image classification system.

Answer

Introduction

Image classification is one of the fundamental tasks in computer vision where an image is
automatically assigned to a predefined category based on its visual content. The goal of an
image classification system is to enable computers to understand and categorize images
similar to how humans recognize objects.

Image classification systems are widely used in many real-world applications such as
medical image diagnosis, autonomous driving, object recognition, agriculture
monitoring, and security systems.

Steps Involved in Building an Image Classification System

Building an image classification system involves several sequential stages that allow the
computer to learn patterns from image data.

1. Data Collection

The first step is to collect a large dataset containing images belonging to different classes.

The dataset must include sufficient examples for each category so that the classification
model can learn distinguishing features.

Example:
For an animal classification system, the dataset may contain images of cats, dogs, horses,
and birds.

Important factors in data collection include:

• Large dataset size

• Diversity of images

• Different lighting and viewing conditions

A well-structured dataset improves the performance of the classification system.

2. Data Preprocessing
Raw images may contain noise, different sizes, or variations in brightness. Therefore
preprocessing is required to standardize the dataset.

Common preprocessing operations include:

• Resizing images to a fixed resolution

• Noise removal

• Normalization of pixel values

• Contrast adjustment
These steps ensure that all images have similar properties before feature extraction.

3. Feature Extraction

Feature extraction is the process of identifying important characteristics in an image that help
distinguish one class from another.

Examples of image features include:

• Edges

• Corners

• Shapes
• Texture patterns

• Color information

Feature extraction converts an image into a numerical feature vector that represents the
important visual information.

Common feature extraction methods include:

• SIFT (Scale Invariant Feature Transform)

• SURF (Speeded Up Robust Features)

• HOG (Histogram of Oriented Gradients)

These techniques capture structural patterns that are useful for classification.

4. Feature Representation

After extracting features, the next step is to represent these features in a structured form
suitable for machine learning algorithms.
Feature vectors are organized into a feature matrix, where:
 • Rows represent images

• Columns represent extracted features

This representation allows classifiers to process image data efficiently.

5. Model Training

During training, a machine learning classifier learns patterns from the labeled dataset.

Common classifiers used in image classification include:

• Support Vector Machine (SVM)

• K-Nearest Neighbors (KNN)

• Decision Trees

• Neural Networks
The classifier analyzes the feature vectors and learns to associate them with the correct class
labels.

6. Model Evaluation

After training, the model must be evaluated using a test dataset that was not used during
training.

Performance metrics used to evaluate classification systems include:

• Accuracy

• Precision

• Recall

• F1-score

Evaluation ensures that the system can correctly classify new unseen images.

7. Image Classification (Prediction Phase)

When a new image is given as input:

1. The image is preprocessed.

2. Features are extracted.

3. The trained classifier analyzes the feature vector.

4. The classifier assigns the image to the most appropriate category.
Example:

Input Image → Feature Extraction → Classifier → Output Label

Example output: Dog

Applications of Image Classification

Image classification systems are widely used in:

• Medical image diagnosis

• Autonomous driving systems

• Facial recognition systems

• Plant disease detection

• Security surveillance

Conclusion

Building an image classification system involves several stages including data collection,
preprocessing, feature extraction, feature representation, model training, evaluation,
and prediction. By learning patterns from large datasets, these systems can accurately
classify images and support many intelligent computer vision applications.

Question 110 (14 Marks)

a) Formulate a linear dynamical model for object motion. (7 Marks)

Answer

Introduction

In computer vision and object tracking, it is necessary to model how objects move over time.
A linear dynamical model is a mathematical model used to describe the motion of objects in
dynamic systems.

This model predicts the future state of an object based on its previous states and motion
parameters.

State Representation

The state of a moving object includes information such as:

• Position
 • Velocity

• Acceleration (in some cases)

For example, the state vector of a moving object can be represented as:

xₜ = [x, y, vx, vy]

Where:

• x, y → position coordinates

• vx, vy → velocity components

Linear Dynamical Model Equation

The motion of the object can be described using the state transition equation:

xₜ = A xₜ₋₁ + wₜ
Where:
• xₜ → state at time t

• xₜ₋₁ → previous state

• A → state transition matrix

• wₜ → process noise

The matrix A determines how the state changes over time.

Observation Model
In addition to motion prediction, observations from sensors are represented as:

zₜ = H xₜ + vₜ

Where:

• zₜ → observed measurement

• H → observation matrix

• vₜ → observation noise

Example

Consider tracking a car moving along a road.

The system observes the car's position in each frame. Using the linear dynamical model, the
system predicts the car's next position based on its current velocity and direction.

Importance

Linear dynamical models are widely used in:

• Video object tracking

• Robot navigation
• Autonomous vehicles

• Surveillance systems

b) Explain its use in prediction during tracking. (7 Marks)

Answer

Prediction is a key component in object tracking systems. The purpose of prediction is to

estimate the future position of a moving object before the next observation becomes
available.

Linear dynamical models are commonly used for this prediction process.

Prediction Process in Tracking

During tracking, the system performs the following steps:

1. Detect the object in the current frame.

2. Estimate the object's current position and velocity.

3. Use the motion model to predict the object's next state.

4. Compare predicted position with actual observations.

5. Update the model if necessary.

Example
Consider tracking a person walking in a video.

If the person moves steadily toward the right side of the frame, the linear dynamical model
predicts that the person will continue moving in the same direction in the next frame.
This predicted location helps the system locate the object quickly.
Advantages of Prediction

Prediction improves tracking performance by:

• Handling temporary occlusion

• Reducing tracking errors

• Maintaining object identity across frames

Prediction also helps reduce computational complexity by limiting the search area.

Applications

Prediction using linear dynamical models is widely used in:

• Kalman filter tracking

• Autonomous vehicle navigation
• Surveillance systems

• Human motion tracking

Question 111 (14 Marks)

a) Illustrate how linear dynamical models represent motion of objects. (7 Marks)

Answer

Linear dynamical models provide a mathematical framework for describing how an object's
state changes over time.

The model assumes that the motion of an object follows a linear relationship between
current and previous states.

Object State Representation

The motion of an object can be described using a state vector that includes position and
velocity.
Example:

xₜ = [x, y, vx, vy]

Where:
• x, y → object position
 • vx, vy → object velocity

This representation allows the system to describe both location and movement of the object.

Motion Equation
The change in the object's state over time is modeled as:

xₜ = A xₜ₋₁ + wₜ

Here:

• A is the state transition matrix

• wₜ represents random motion noise

The matrix A determines how the object moves from one frame to the next.

Example
Consider a ball rolling on the ground.

If the ball moves with constant velocity, the linear dynamical model predicts the next position
based on the previous position and velocity.

Importance

Linear dynamical models help represent motion patterns efficiently and are widely used in
visual tracking and motion estimation.

b) Explain how prediction using a linear dynamical model helps estimate the future
position of a moving object. (7 Marks)

Answer

Prediction using a linear dynamical model allows a system to estimate the future position of
an object based on its past motion.

Prediction Steps

1. Measure the current state of the object.

2. Estimate velocity and direction of motion.

3. Apply the motion model to predict the next state.
 4. Update predictions when new observations are available.

Example

If a vehicle moves at a constant speed along a road, the system predicts its future position
using the motion model.

Even if the object temporarily disappears due to occlusion, the predicted location helps the
tracking system maintain continuity.

Benefits
Prediction helps in:

• Maintaining smooth tracking

• Reducing tracking errors

• Handling missing observations

Application

Prediction using linear dynamical models is essential in:

• Kalman filter based tracking

• Robotics

• Autonomous driving systems

Question 112 (3 Marks)

State any three real-world applications of image classification.

Answer

Three real-world applications of image classification are:

1. Medical Image Analysis – Image classification helps doctors identify diseases from
medical images such as X-rays, MRI scans, and CT scans.

2. Autonomous Vehicles – Used to classify objects on the road such as pedestrians,

vehicles, traffic signs, and obstacles.

3. Face Recognition Systems – Used in security systems and smartphone authentication

to identify individuals.

Question 113 (3 Marks)

Differentiate between two-class and multiclass classification.

Answer

Two-Class Classification Multiclass Classification

Classifies data into two possible categories Classifies data into more than two
only. categories.

Also called binary classification. Called multiclass classification.

The output contains only two labels. The output contains multiple class labels.

Example: Spam vs Non-spam email, Face Example: Classifying images as cat, dog,
vs Non-face detection. horse, bird.

More complex since the model must

Simpler problem and easier to train models.
distinguish among several classes.

Question 114 (3 Marks)

Define classification in computer vision and state its purpose.

Answer

Classification in computer vision is the process of assigning an input image to one of

several predefined categories based on its visual features.

The purpose of classification is to automatically identify the type of object or scene present
in an image. It is widely used in applications such as object recognition, medical image
analysis, and facial recognition systems.

SUBMODULE 2
Question 115 (14 Marks)

Apply suitable techniques to control overfitting in image classification tasks.

Answer

Introduction
Image classification is a computer vision task in which images are assigned to predefined
categories such as animals, vehicles, or objects. Machine learning models are trained using
labeled image datasets to recognize patterns in images. However, during training, the model
may learn the training data too closely and fail to generalize to new unseen images. This
phenomenon is called overfitting.
Overfitting is a major problem in image classification because it reduces the model’s ability
to perform well on real-world data. Therefore, several techniques are used to control
overfitting and improve model generalization.

Understanding Overfitting in Image Classification

Overfitting occurs when the classification model becomes too complex and learns noise or
irrelevant details present in the training data. As a result:

• Training accuracy becomes very high

• Validation or testing accuracy becomes low

This indicates that the model is not learning the general characteristics of the objects.

Example:
A classifier trained to recognize cats and dogs may memorize specific training images instead
of learning general features like shape, ears, and fur patterns.

Techniques to Control Overfitting

Several techniques can be applied to reduce overfitting in image classification systems.

1. Increasing the Size of Training Data

A larger training dataset helps the model learn general patterns rather than memorizing
specific examples.
Advantages:

• Improves model generalization

• Reduces bias toward specific training samples

• Helps capture variations in images

Example variations include:

• Different object orientations

• Different lighting conditions

• Different backgrounds

2. Data Augmentation
Data augmentation increases the training dataset by creating modified versions of existing
images.

Common augmentation techniques include:

• Image rotation

• Flipping (horizontal or vertical)

• Scaling or zooming

• Cropping
• Brightness and contrast adjustment

Data augmentation improves robustness because the model learns to recognize objects under
different transformations.

3. Regularization

Regularization techniques add a penalty term to the model’s loss function to prevent overly
complex models.

Two common types are:

L1 Regularization

• Adds the absolute value of model weights to the loss function

• Encourages sparse models

• Some weights may become zero

L2 Regularization

• Adds the squared value of weights to the loss function

• Prevents weights from becoming excessively large

Regularization helps the model focus on meaningful features instead of noise.

4. Cross Validation

Cross validation divides the dataset into multiple subsets and repeatedly trains the model
using different subsets for training and validation.

Advantages:

• Helps evaluate model performance reliably

• Detects overfitting early
 • Assists in selecting the best classifier

A commonly used method is k-fold cross validation.

5. Dropout (Deep Learning Models)

Dropout is used in neural networks to prevent overfitting.

In dropout:

• Some neurons are randomly disabled during training

• The network cannot rely heavily on specific neurons

Benefits include:

• Reduces dependency on particular features

• Improves model generalization

6. Early Stopping

Early stopping monitors validation error during training.

If validation error starts increasing while training error continues decreasing, the training
process is stopped.

This prevents the model from learning noise in the training dataset.

7. Model Simplification

Reducing model complexity can also reduce overfitting.

Examples include:
• Reducing the number of features

• Using simpler models

• Reducing the number of layers in neural networks

Simpler models are less likely to memorize training data.

Applications

Techniques for controlling overfitting are used in many applications such as:
• Medical image classification
 • Face recognition systems

• Autonomous driving

• Plant disease detection

Conclusion

Overfitting is a common challenge in image classification tasks that reduces the model’s
ability to generalize. Techniques such as data augmentation, regularization, cross
validation, dropout, early stopping, and increasing training data help improve
classification performance and ensure reliable predictions on unseen data.

Question 116 (14 Marks)

a) Demonstrate how cross validation helps in selecting the best classifier model. (7
Marks)

Answer
Cross validation is a model evaluation technique used to assess how well a classifier performs
on unseen data. It is widely used in machine learning to compare different classification
models and select the most accurate one.

Concept of Cross Validation

In cross validation, the dataset is divided into multiple subsets called folds. The model is
trained on some folds and tested on the remaining fold.
This process is repeated several times so that each subset is used once for validation.

k-Fold Cross Validation Procedure

1. Divide the dataset into k equal subsets.

2. Use one subset as the validation set.

3. Train the model using the remaining subsets.

4. Evaluate the model on the validation set.

5. Repeat the process until every subset has been used as validation.
6. Compute the average performance across all folds.
Selecting the Best Classifier

Cross validation allows comparison between multiple classifiers such as:

• Support Vector Machine (SVM)

• Decision Tree
• K-Nearest Neighbors (KNN)

• Neural Networks

The classifier with the highest average validation accuracy is selected.

Advantages

Cross validation helps to:

• Prevent overfitting
• Provide reliable model evaluation
• Select optimal hyperparameters

b) Explain how training, validation and testing datasets are used to evaluate model
performance. (7 Marks)

Answer

To properly evaluate classification models, datasets are usually divided into three parts:

• Training dataset

• Validation dataset

• Testing dataset

1. Training Dataset

The training dataset is used to train the machine learning model.

The classifier learns patterns and relationships between input features and class labels.

Example:
Images used to train a system that identifies animals.

2. Validation Dataset
The validation dataset is used during training to tune the model.
It helps in:

• Selecting hyperparameters

• Choosing the best model

• Preventing overfitting
The validation dataset guides the training process.

3. Testing Dataset

The testing dataset is used after training is complete.

It evaluates the final performance of the model on unseen data.

This dataset provides an unbiased estimate of how well the classifier will perform in real-
world scenarios.

Importance

Using separate datasets ensures:

• Fair model evaluation

• Better generalization

• Reduced overfitting

Question 117 (14 Marks)

a) Illustrate different types of classification errors. (7 Marks)

Answer
Classification errors occur when the predicted class label differs from the actual class label.
These errors are commonly represented using a confusion matrix.

Components of Classification Errors

True Positive (TP)

The model correctly predicts a positive instance.

Example: A face detection system correctly identifies a face.

True Negative (TN)
The model correctly predicts a negative instance.

Example: A non-face image is correctly classified as non-face.

False Positive (FP)

The model incorrectly predicts a positive class for a negative instance.

Example: A background object is mistakenly classified as a face.

This is also called Type I error.

False Negative (FN)

The model incorrectly predicts a negative class for a positive instance.

Example: A face in an image is missed by the system.

This is called Type II error.

Importance of Understanding Errors

Analyzing classification errors helps improve model performance and reduce

misclassification.

b) Apply regularization to reduce overfitting. (7 Marks)

Answer

Regularization is a technique used in machine learning to reduce overfitting by limiting

model complexity.

It modifies the loss function by adding a penalty term for large parameter values.

Regularized Loss Function

Loss = Training Loss + Regularization Term

This prevents the model from assigning excessively large weights.

Types of Regularization

L1 Regularization
• Adds the absolute value of weights to the loss function
 • Encourages sparse models

• Helps in feature selection

L2 Regularization
• Adds the squared value of weights to the loss function

• Prevents extremely large weights

• Improves model stability

Benefits of Regularization

Regularization helps to:

• Reduce overfitting
• Improve model generalization
• Control model complexity

Example

In an image classification system, regularization prevents the classifier from memorizing

training images and encourages learning meaningful visual features.

Question 118 (3 Marks)

Explain the purpose of cross validation in model selection.

Answer
Cross validation is a technique used to evaluate and compare machine learning models by
dividing the dataset into multiple subsets for training and validation.

The purpose of cross validation is to:

• Estimate the performance of a classifier on unseen data

• Compare multiple models and select the best one

• Detect and reduce overfitting

It provides a more reliable estimate of model performance than using a single train-test split.

Question 119 (3 Marks)

What is regularization and why is it used?

Answer

Regularization is a technique used in machine learning to prevent overfitting by adding a

penalty term to the model’s loss function.

It limits the complexity of the model by restricting the magnitude of model parameters.

Regularization is used to:

• Improve model generalization

• Reduce overfitting

• Make the classifier more stable when predicting new data.

Question 120 (3 Marks)

Define overfitting in classification problems.

Answer

Overfitting occurs when a classification model learns the training data too closely, including
noise and irrelevant patterns.

As a result, the model performs very well on the training dataset but performs poorly on
unseen test data.

Overfitting reduces the ability of the classifier to generalize and decreases prediction
accuracy in real-world applications.

SUBMODULE 3
Question 121 (14 Marks)

a) Illustrate how tracking systems handle motion variation in practical videos. (7

Marks)

Answer
Introduction

Object tracking in computer vision refers to the process of continuously locating a moving
object across consecutive frames of a video. In real-world video scenes, objects rarely move
in a perfectly predictable way. Instead, they may change direction, speed, or trajectory due to
various environmental factors. These changes are called motion variations.

Tracking systems must handle such motion variations effectively to maintain accurate
tracking.
Motion Variation in Practical Videos

In practical video scenes, objects may exhibit different types of motion variations such as:

• Changes in speed
• Sudden direction changes

• Acceleration or deceleration

• Non-linear motion patterns

For example, a person walking in a surveillance video may suddenly turn or stop, causing
unpredictable motion.

Techniques Used to Handle Motion Variation

Tracking systems use several techniques to handle motion variations.

1. Motion Models
Motion models describe how objects move from one frame to the next.

Examples include:

• Constant velocity model

• Constant acceleration model

These models predict the next position of the object based on previous motion.

2. Kalman Filter
Kalman filters are widely used in tracking systems to estimate object motion.

Functions of Kalman filters:

• Predict the future position of objects

• Correct predictions using new observations

• Handle noisy measurements

3. Particle Filters
Particle filters are used when motion is highly unpredictable.
Advantages:

• Handle non-linear motion

• Track objects in complex environments

• Manage multiple hypotheses about object position

4. Optical Flow

Optical flow estimates motion between consecutive frames by analyzing pixel movement.

It helps detect small motion variations and track moving objects.

Example

In a traffic surveillance system:

• Vehicles may change lanes or speed
• Tracking algorithms predict their next position using motion models

• The system adjusts predictions based on actual observations

Conclusion

Handling motion variation is essential for robust tracking systems. Techniques such as
motion models, Kalman filters, particle filters, and optical flow help maintain accurate
tracking even when objects move unpredictably.

b) Explain how adaptive models help maintain tracking accuracy during scale and
appearance changes. (7 Marks)

Answer

Introduction

In video tracking, the appearance of objects may change due to factors such as scale
variation, lighting changes, or object deformation. Adaptive models help update the tracking
model to account for these changes and maintain accurate tracking.

Scale Variation

Scale variation occurs when the object moves closer to or farther from the camera.
Example:
 • A person walking toward the camera appears larger in the frame.

• The same person moving away appears smaller.

Adaptive tracking models update the size of the tracking window to match the object scale.

Appearance Changes

Object appearance may change due to:

• Lighting conditions

• Viewing angles

• Object rotation

• Partial occlusion

Example:
A person turning sideways may look different from a frontal view.

Adaptive Tracking Models

Adaptive models continuously update object features during tracking.

Common techniques include:

1. Online Model Update

The tracking system updates the appearance model using newly observed frames.
Benefits:

• Adapts to changing object appearance

• Maintains accurate tracking over time

2. Multi-scale Tracking

Multi-scale tracking uses different window sizes to detect objects at different scales.

This helps handle scale variations effectively.

3. Feature Adaptation
Tracking models update features such as:
 • Color histograms

• Texture descriptors

• Shape features

These updated features represent the current appearance of the object.

Example

In a surveillance system tracking a person:

• The person may move closer to the camera

• The tracking model updates object size and appearance

• This ensures continuous and accurate tracking

Conclusion
Adaptive tracking models improve robustness by updating object representations during
tracking. By handling scale variation and appearance changes, adaptive models maintain
accurate object tracking in dynamic environments.

Question 122 (14 Marks)

a) Apply suitable tracking strategies for monitoring a moving object in a scene. (7

Marks)
Answer

Tracking strategies are methods used to monitor moving objects across video frames.
Different strategies are used depending on the complexity of the scene and object motion.

Common Tracking Strategies

1. Feature-Based Tracking

Feature-based tracking uses distinctive features such as corners or edges to track objects.

Example features:

• SIFT features
• SURF features
 • Harris corner points

The tracker follows these features across frames.

2. Region-Based Tracking
Region-based tracking identifies a region or bounding box around the object and tracks that
region.

Example techniques include:

• Mean shift tracking

• CamShift tracking

These methods use color or texture information.

3. Motion-Based Tracking

Motion-based tracking uses motion information extracted from video frames.

Example methods:
• Optical flow

• Frame differencing

These techniques track moving objects based on motion patterns.

4. Model-Based Tracking

Model-based tracking uses a predefined model of the object.

Example:
Tracking a human body using body structure models.

Applications

Tracking strategies are used in:

• Video surveillance systems

• Autonomous vehicles

• Sports analysis
• Human activity monitoring
b) Explain how appearance and motion cues are used together to improve tracking
performance. (7 Marks)

Answer

Tracking performance improves when both appearance cues and motion cues are combined.

Appearance Cues
Appearance cues describe the visual characteristics of the object.

Examples include:

• Color information

• Texture features

• Shape descriptors

These cues help identify the object across frames.

Motion Cues

Motion cues describe how the object moves over time.

Examples include:

• Velocity of the object

• Direction of movement

• Motion trajectory

These cues help predict the object's future position.

Combining Appearance and Motion

Using both cues improves tracking accuracy.

Steps involved:

1. Appearance features identify the object.

2. Motion models predict the object's next position.

3. The tracker combines both sources of information.

4. The object is tracked more reliably.
Example

In surveillance videos:

• Appearance cues identify a person using clothing color.

• Motion cues track the person's movement direction.

Combining both ensures accurate tracking even in crowded scenes.

Question 123 (14 Marks)

a) Illustrate the complete tracking pipeline in video sequences. (7 Marks)

Answer

Object tracking systems follow a sequence of processing steps known as the tracking
pipeline.

Steps in the Tracking Pipeline

1. Video Acquisition

A video camera captures frames continuously.

These frames form the input to the tracking system.

2. Object Detection

The system detects objects in each frame.

Common detection techniques include:

• Background subtraction

• Feature-based detection

• Deep learning detectors

3. Feature Extraction

Important features of the detected object are extracted.

Examples include:
 • Color features

• Shape features

• Texture features

4. Motion Estimation

Motion estimation determines how the object moves between frames.

Techniques include:

• Optical flow

• Motion models

5. Object Association
The system associates detected objects across frames.
This ensures that the same object is tracked over time.

6. Trajectory Generation

The system records the object's movement path across frames.

This trajectory helps analyze object behavior.

Applications
Tracking pipelines are used in:

• Surveillance systems

• Traffic monitoring

• Sports analytics

• Human activity recognition

b) Demonstrate tracking by detection framework. (7 Marks)

Answer

Tracking by detection is a widely used framework in object tracking systems.

Instead of continuously tracking objects using motion alone, the system first detects objects
in each frame and then links the detections across frames.

Working of Tracking by Detection

1. Detect objects in every video frame.

2. Extract features from detected objects.

3. Match objects between consecutive frames.

4. Assign unique identities to each object.

5. Track the object's trajectory over time.

Advantages

• Handles appearance changes effectively

• Works well in crowded scenes

• Allows re-detection of objects after occlusion

Example

In pedestrian tracking:

• The system detects pedestrians in each frame.

• Detected bounding boxes are matched across frames.

• The system maintains the identity of each pedestrian.

Question 124 (3 Marks)

Explain the concept of tracking by detection.

Answer

Tracking by detection is a method in which objects are first detected independently in each
video frame and then associated across frames to maintain object identity.

Instead of relying only on motion prediction, the system repeatedly detects objects using an
object detector and links the detections to track objects over time.

This approach improves tracking accuracy and allows objects to be re-detected even after
temporary occlusion.
Question 125 (3 Marks)

State three simple tracking strategies used in video analysis.

Answer
Three commonly used tracking strategies in video analysis are:

1. Feature-based tracking – tracks distinctive features such as corners or edges across

frames.
2. Region-based tracking – tracks objects using bounding boxes or regions based on
color or texture information.
3. Motion-based tracking – tracks objects by analyzing motion patterns using
techniques such as optical flow.

These strategies help monitor moving objects in video sequences.

Question 126 (3 Marks)

Define object tracking in computer vision.

Answer

Object tracking is the process of locating and following a moving object across consecutive
frames of a video sequence.

The goal of object tracking is to maintain the identity and position of the object as it moves
through the scene.

Object tracking is widely used in applications such as video surveillance, autonomous

vehicles, sports analysis, and human activity monitoring.

SUBMODULE 4
Question 127 (14 Marks)

a) Apply linear dynamical models for motion prediction in video tracking systems. (7
Marks)

Answer

In video tracking systems, the position of a moving object must be predicted in future frames
based on its previous motion. Linear dynamical models (LDM) are mathematical models
used to describe and predict the motion of objects over time.
A linear dynamical model represents the motion of an object using a state vector that
includes parameters such as position and velocity.

Example state vector:

𝑥𝑡 = [𝑥, 𝑦, 𝑣𝑥 , 𝑣𝑦 ]𝑇

Where:

• 𝑥, 𝑦represent the position of the object

• 𝑣𝑥 , 𝑣𝑦 represent velocity components

The motion of the object between frames is described using the state transition equation:

𝑥𝑡 = 𝐴𝑥𝑡−1 + 𝑤𝑡

Where:

• 𝑥𝑡 = state at time 𝑡

• 𝐴= state transition matrix

• 𝑤𝑡 = process noise

In tracking systems, the model predicts the next position of the object based on its previous
position and velocity.

Example:
In a traffic monitoring system, a vehicle moving along a road can be tracked by predicting its
next location using the linear dynamical model.

The predicted position helps the tracking algorithm search for the object within a smaller
region, improving tracking efficiency and accuracy.

Thus, linear dynamical models are widely used in video surveillance, autonomous driving,
and robotics for motion prediction.

b) Explain how model parameters are estimated and updated during tracking. (7
Marks)

Answer

In object tracking systems, the parameters of the motion model must be continuously
estimated and updated as new video frames are processed.

The key parameters in a motion model include:

• Object position
 • Velocity

• State transition parameters

• Noise parameters

Parameter Estimation

Initially, the model parameters are estimated from observations in the first few frames.

Example:

1. Detect the object in the first frame.

2. Measure its position.

3. Estimate velocity using the displacement between frames.

This initial estimate forms the initial state of the tracking model.

Parameter Update During Tracking

As new frames are processed, the tracking system performs two steps:

1. Prediction Step

The motion model predicts the next object state using the state transition equation.

2. Correction Step

The predicted state is compared with the actual observed object position.
The difference between the prediction and observation is used to update the model
parameters.

Kalman Filter for Parameter Update

A commonly used technique for updating model parameters is the Kalman filter.

Kalman filter performs:

• Prediction of object motion

• Correction using observed measurements

• Reduction of noise in measurements

Importance
Updating model parameters helps tracking systems:
 • Adapt to changes in motion

• Improve prediction accuracy

• Maintain reliable tracking over time.

Question 128 (14 Marks)

a) Formulate a linear dynamical model for object motion. (7 Marks)

Answer

A linear dynamical model describes how the state of a moving object evolves over time
using linear equations. It is commonly used in object tracking and motion analysis.

State Representation

The motion of an object is represented using a state vector that includes position and
velocity.

Example state vector:

𝑥𝑡 = [𝑥, 𝑦, 𝑣𝑥 , 𝑣𝑦 ]𝑇

Where:

• 𝑥, 𝑦represent position coordinates

• 𝑣𝑥 , 𝑣𝑦 represent velocity components

State Transition Equation

The motion of the object between consecutive frames is described using the state transition
model:

𝑥𝑡 = 𝐴𝑥𝑡−1 + 𝑤𝑡

Where:

• 𝑥𝑡 = state at time 𝑡

• 𝐴= state transition matrix

• 𝑤𝑡 = process noise

The matrix 𝐴determines how position and velocity change over time.
Observation Model

The observed measurement is related to the object state using the equation:

𝑧𝑡 = 𝐻𝑥𝑡 + 𝑣𝑡

Where:

• 𝑧𝑡 = observed measurement

• 𝐻= observation matrix

• 𝑣𝑡 = observation noise

Example

In vehicle tracking:

• The system measures the vehicle position in each frame.

• The linear dynamical model predicts the next position based on previous motion.

Importance

Linear dynamical models are important because they provide a structured way to represent
object motion and enable efficient prediction in tracking systems.

b) Demonstrate prediction in tracking using the model. (7 Marks)

Answer

Prediction is a key component of object tracking systems. It estimates the future position of a
moving object using its previous motion information.

Prediction Using Linear Dynamical Model

The predicted state of the object is obtained using the equation:

𝑥̂𝑡 = 𝐴𝑥𝑡−1

Where:

• 𝑥̂𝑡 = predicted state

 • 𝐴= state transition matrix

This equation estimates the object's future position based on its previous state.

Steps in Prediction

1. Obtain the object's current state (position and velocity).

2. Apply the state transition equation.

3. Predict the object's next position.

4. Search for the object near the predicted location in the next frame.

Example

Consider tracking a car moving in a straight line.

• The system measures the car's position in the current frame.

• Using its velocity, the model predicts where the car will appear in the next frame.

This prediction reduces the search area and improves tracking efficiency.

Benefits
Prediction helps to:

• Maintain continuous tracking

• Handle temporary occlusion

• Reduce computational complexity.

Question 129 (14 Marks)

a) Illustrate how linear dynamical models represent object motion. (7 Marks)

Answer

Linear dynamical models represent object motion by describing how the object's state
changes over time using linear equations.

State Variables
The object's state includes parameters such as:
 • Position coordinates

• Velocity components

Example state vector:

𝑥𝑡 = [𝑥, 𝑦, 𝑣𝑥 , 𝑣𝑦 ]𝑇

Motion Representation

The motion of the object is represented by the equation:

𝑥𝑡 = 𝐴𝑥𝑡−1 + 𝑤𝑡

Where:

• 𝐴represents the motion model

• 𝑤𝑡 represents random noise

This equation describes how the object moves between frames.

Motion Characteristics

Linear dynamical models can represent:

• Constant velocity motion

• Smooth object movement

• Predictable motion patterns

Example

In pedestrian tracking:

• The system estimates the person's position and velocity.

• The model predicts how the person moves across video frames.

Importance

These models simplify motion representation and form the basis for many tracking
algorithms.
b) Explain the role of prediction in tracking the future position of a moving object. (7
Marks)

Answer

Prediction plays a crucial role in object tracking by estimating the object's future location
before the next frame is processed.

Purpose of Prediction

Prediction helps the tracking system:

• Estimate future object location

• Reduce the search region

• Maintain object identity across frames

Prediction Process

1. Determine the current state of the object.

2. Apply the motion model to estimate the next state.

3. Predict the object's position in the next frame.

4. Update prediction using actual observations.

Example

In a surveillance system tracking a pedestrian:

• The person's movement direction and speed are estimated.

• The model predicts where the person will appear in the next frame.

Advantages
Prediction improves:

• Tracking accuracy

• Tracking speed

• Robustness against temporary occlusion.

Question 130 (3 Marks)

Mention two applications of linear dynamical models.

Answer

Two applications of linear dynamical models are:

1. Object tracking in video surveillance systems – Used to predict the motion of

moving objects such as vehicles or pedestrians across video frames.

2. Robot navigation and autonomous vehicles – Used to estimate the future position
of moving objects and avoid collisions.

Question 131 (3 Marks)

State the role of state space representation in tracking.

Answer

State space representation describes the current state of a moving object using variables
such as position and velocity.

In tracking systems, the state vector helps:

• Represent object motion mathematically

• Predict the future position of the object

• Update the motion model using new observations

State space models are widely used in tracking algorithms such as Kalman filters.

Question 132 (3 Marks)

What is a linear dynamical system model?

Answer

A linear dynamical system model is a mathematical model that describes how the state of a
system changes over time using linear equations.

In computer vision, it is used to model the motion of objects in video sequences by

representing parameters such as position and velocity.

Linear dynamical models are commonly used for motion prediction, object tracking, and
control systems.

SUBMODULE 5
Question 133 (3 Marks)
Explain the difference between the prediction step and update step in the Kalman filter.

Answer

The Kalman filter estimates the state of a moving object through two main steps: prediction
and update (correction).

Prediction step:
In this step, the filter predicts the future state of the system using the motion model. It
estimates the next position of the object based on its previous state and motion dynamics.

Update step:
In this step, the predicted state is corrected using the actual measurement obtained from
sensors or observations. The filter combines the predicted value and measured value to
produce a more accurate estimate.

Thus, the prediction step estimates the future state, while the update step corrects the
prediction using new observations.

Question 134 (14 Marks)

a) Apply Kalman filtering framework in real-time tracking applications. (7 Marks)

Answer

Introduction

The Kalman filter is a recursive estimation algorithm widely used in computer vision and
control systems for tracking moving objects. It provides an efficient method to estimate the
state of a system over time by combining predictions from a motion model with
measurements obtained from sensors.
In real-time tracking applications, the Kalman filter helps predict and update the position of
moving objects accurately even when measurements are noisy.

Kalman Filter Framework

The Kalman filter operates in two main stages:

1. Prediction Stage

2. Update (Correction) Stage

Prediction Step
In the prediction stage, the filter predicts the next state of the object using the motion model.
State prediction equation:
 𝑥𝑡 = 𝐴𝑥𝑡−1 + 𝐵𝑢𝑡 + 𝑤𝑡

Where:

• 𝑥𝑡 = predicted state

• 𝐴= state transition matrix

• 𝑢𝑡 = control input

• 𝑤𝑡 = process noise

This step estimates where the object will appear in the next frame.

Update Step

When a measurement from the sensor becomes available, the filter updates the predicted state
using the measurement equation.

Measurement equation:

𝑧𝑡 = 𝐻𝑥𝑡 + 𝑣𝑡

Where:

• 𝑧𝑡 = observed measurement

• 𝐻= observation matrix

• 𝑣𝑡 = measurement noise

The Kalman filter combines the prediction and measurement to produce an improved
estimate.

Application in Real-Time Tracking

In video tracking:
1. Detect object position in the current frame.

2. Predict its next position using the Kalman filter.

3. Compare prediction with actual observation.

4. Update the estimate.

Example:
Tracking a moving car in a traffic monitoring system.
Advantages

• Works efficiently in real-time systems

• Reduces noise in measurements

• Improves tracking accuracy

Conclusion

Kalman filtering is widely used in real-time tracking applications such as surveillance

systems, robotics, and autonomous vehicles because it provides accurate estimation of object
motion.

b) Explain how process noise and measurement noise influence Kalman filter
performance. (7 Marks)
Answer

In Kalman filtering, two types of noise affect the accuracy of state estimation:

1. Process Noise

2. Measurement Noise

These noise components represent uncertainties in the system model and sensor
measurements.

Process Noise

Process noise represents uncertainty in the system dynamics.

It occurs because the motion model may not perfectly describe the object's movement.

Examples:

• Sudden acceleration of a vehicle

• Unexpected motion changes

• Environmental disturbances

If process noise is high, the Kalman filter relies more on measurements rather than
predictions.

Measurement Noise
Measurement noise represents errors in sensor observations.
Examples include:

• Camera measurement errors

• Sensor inaccuracies

• Lighting variations affecting detection

If measurement noise is high, the Kalman filter relies more on the prediction model.

Influence on Kalman Filter Performance

The Kalman filter balances prediction and measurement based on noise levels.

• High process noise → more reliance on measurements

• High measurement noise → more reliance on predictions

Proper estimation of noise parameters is essential for accurate tracking.

Conclusion

Process noise and measurement noise significantly influence Kalman filter performance.
Accurate modeling of these noises helps achieve reliable state estimation in tracking systems.

Question 135 (14 Marks)

Illustrate how the Kalman filter balances prediction and measurement to estimate
object position.

Answer

Introduction

The Kalman filter is a powerful algorithm used to estimate the position of moving objects by
combining predictions from a motion model with measurements obtained from sensors.

This balancing of prediction and measurement helps achieve accurate tracking even when
observations contain noise.

Prediction Step

In the prediction step, the filter estimates the object's future state using the previous state and
motion model.

Prediction equation:

𝑥𝑡 = 𝐴𝑥𝑡−1
This step provides an estimate of where the object is expected to appear in the next frame.

Measurement Step
The system then measures the object's actual position using sensors or detection algorithms.

However, measurements may contain noise or errors.

Correction Step

The Kalman filter combines the predicted state and the measured state using a weighting
factor called the Kalman gain.

Kalman gain determines how much the prediction and measurement influence the final
estimate.

If measurements are reliable → more weight is given to measurements.

If measurements are noisy → more weight is given to predictions.

Example

In vehicle tracking:

• The motion model predicts the vehicle's next location.

• The camera detects the vehicle position.

• The Kalman filter combines both values to estimate the accurate position.

Advantages

Balancing prediction and measurement helps:

• Reduce tracking errors

• Handle noisy observations

• Maintain continuous tracking

Conclusion

By combining predictions from the motion model with actual sensor measurements, the
Kalman filter produces accurate and stable estimates of object position in dynamic
environments.
Question 136 (14 Marks)

a) Illustrate Kalman filtering concept in object tracking. (7 Marks)

Answer
Kalman filtering is widely used in object tracking systems to estimate the position and
velocity of moving objects over time.

The algorithm recursively updates the estimated state of the object using predictions and
observations.

Key Components of Kalman Filtering

1. State Vector

Represents parameters such as:

• Position

• Velocity

Example:

𝑥𝑡 = [𝑥, 𝑦, 𝑣𝑥 , 𝑣𝑦 ]𝑇

2. Prediction Stage

The system predicts the next state using the motion model.

3. Measurement Stage

Sensors or detection algorithms provide measurements of object position.

4. Correction Stage

The predicted state is corrected using the measurement.

Example

Tracking a pedestrian in a surveillance video:

• The filter predicts the person's next position.
 • The detected position is used to update the estimate.

Applications

• Video surveillance
• Autonomous driving

• Robot navigation

b) Explain the assumptions of the Kalman filter in a practical tracking environment. (7

Marks)

Answer

The Kalman filter is based on several assumptions about the system and noise characteristics.

Assumption 1: Linear System Model

The system dynamics must be represented using linear equations.

This means object motion should follow a linear model.

Assumption 2: Gaussian Noise

Both process noise and measurement noise are assumed to follow a Gaussian distribution.

Assumption 3: Known System Model

The system matrices such as state transition matrix and observation matrix must be known.

Assumption 4: Independent Noise

Process noise and measurement noise are assumed to be independent.

Importance

These assumptions simplify the mathematical formulation and allow efficient computation of
optimal state estimates.

Question 137 (3 Marks)

Explain the roles of the state vector and covariance matrix in the Kalman filter.

Answer

In the Kalman filter, the state vector represents the current estimated state of the system,
such as position and velocity of a moving object.

The covariance matrix represents the uncertainty or error in the state estimate. It indicates
how confident the filter is about the predicted state.

During filtering, the covariance matrix is updated to reflect changes in uncertainty, helping
the filter determine how much to rely on predictions or measurements.

Question 138 (3 Marks)

What is meant by prediction and correction in Kalman filtering?

Answer
In Kalman filtering, prediction refers to estimating the future state of a system using the
previous state and motion model.
Correction (or update) refers to adjusting the predicted state using the actual measurement
obtained from sensors.
Prediction provides an estimate of the system state, while correction improves the estimate by
incorporating new observations.

SUBMODULE 6
Question 139 (14 Marks)

Demonstrate Kalman filtering algorithm flow with suitable tracking scenario.

Answer

Introduction

The Kalman filter is a recursive algorithm used for estimating the state of a dynamic system
from noisy observations. In computer vision and video tracking, it is widely used to estimate
the position and velocity of moving objects over time. The Kalman filter continuously
updates its estimates using two major steps: prediction and correction (update).
The algorithm is especially useful in tracking applications where measurements may contain
noise or uncertainty.

Kalman Filtering Algorithm Flow

The Kalman filtering algorithm operates through a sequence of steps for each time frame.

1. Initialization

Initially, the filter is initialized with:

• Initial state vector

• Initial covariance matrix

Example state vector:

𝑥0 = [𝑥, 𝑦, 𝑣𝑥 , 𝑣𝑦 ]𝑇

Where
𝑥, 𝑦= object position
𝑣𝑥 , 𝑣𝑦 = velocity components

The covariance matrix represents uncertainty in the initial estimate.

2. Prediction Step

The filter predicts the next state of the object using the state transition model.
State prediction:

𝑥𝑡 = 𝐴𝑥𝑡−1

Covariance prediction:

𝑃𝑡 = 𝐴𝑃𝑡−1 𝐴𝑇 + 𝑄

Where
𝐴= state transition matrix
𝑄= process noise covariance

This step estimates where the object will appear in the next frame.

3. Measurement Step

The system obtains a measurement from sensors or detection algorithms.

Example measurement vector:

𝑧𝑡 = [𝑥𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑 , 𝑦𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑 ]𝑇
However, measurements may contain noise.

4. Update (Correction) Step

The predicted state is corrected using the measurement.

Kalman gain:

𝐾𝑡 = 𝑃𝑡 𝐻 𝑇 (𝐻𝑃𝑡 𝐻 𝑇 + 𝑅)−1

Updated state estimate:

𝑥𝑡 = 𝑥𝑡 + 𝐾𝑡 (𝑧𝑡 − 𝐻𝑥𝑡 )

Updated covariance:

𝑃𝑡 = (𝐼 − 𝐾𝑡 𝐻)𝑃𝑡

Where
𝐻= observation matrix
𝑅= measurement noise covariance

Example Tracking Scenario

Consider tracking a vehicle in a traffic surveillance video.

1. The Kalman filter predicts the vehicle’s next position using previous motion.

2. The camera detects the actual vehicle position.

3. The filter combines prediction and measurement to produce an improved estimate.

4. The process repeats for every frame.

Advantages of Kalman Filter

• Handles noisy measurements effectively

• Provides real-time estimation

• Efficient for linear dynamic systems

• Improves tracking accuracy

Conclusion

The Kalman filtering algorithm provides a structured approach for estimating object motion
in video sequences. By continuously predicting and updating object states, it enables accurate
and efficient tracking in dynamic environments.

Question 140 (14 Marks)

Illustrate the role of state transition and measurement models in Kalman filter based
motion estimation.

Answer

Introduction

In Kalman filter based motion estimation, two important models describe the system
dynamics and observations:

1. State transition model

2. Measurement model
These models allow the Kalman filter to estimate the motion of objects and update the
estimates using observed data.

State Transition Model

The state transition model describes how the object's state changes over time.

State equation:

𝑥𝑡 = 𝐴𝑥𝑡−1 + 𝑤𝑡

Where
𝑥𝑡 = state at time 𝑡
𝐴= state transition matrix
𝑤𝑡 = process noise

Role of State Transition Model

The state transition model is used for prediction.

It estimates:
• Future object position
• Future velocity
Example state vector:

𝑥𝑡 = [𝑥, 𝑦, 𝑣𝑥 , 𝑣𝑦 ]𝑇

The transition matrix defines how position and velocity change between frames.

Measurement Model

The measurement model relates the predicted state to the observed measurement.

Measurement equation:

𝑧𝑡 = 𝐻𝑥𝑡 + 𝑣𝑡

Where
𝑧𝑡 = measurement vector
𝐻= observation matrix
𝑣𝑡 = measurement noise

Role of Measurement Model

The measurement model is used to:

• Incorporate sensor observations

• Correct prediction errors

• Improve accuracy of state estimation

Example

In tracking a pedestrian:
• State transition model predicts the next position.

• Camera measurements provide observed positions.

• Kalman filter combines both to estimate the true position.

Importance

Together, these models enable the Kalman filter to:

• Predict object motion

• Correct errors using measurements

 • Maintain accurate tracking over time.

Conclusion

The state transition model predicts object motion while the measurement model corrects the
prediction using observations. Both models are essential for accurate motion estimation in
Kalman filtering.

Question 141 (14 Marks)

a) Illustrate Kalman filter prediction step with equations. (7 Marks)

Answer

The prediction step estimates the future state of the system based on the previous state and
the motion model.

State Prediction Equation

𝑥̂𝑡 = 𝐴𝑥𝑡−1 + 𝐵𝑢𝑡

Where

𝑥̂𝑡 = predicted state

𝐴= state transition matrix
𝐵= control matrix
𝑢𝑡 = control input

Covariance Prediction

The uncertainty of the predicted state is calculated using:

𝑃𝑡 = 𝐴𝑃𝑡−1 𝐴𝑇 + 𝑄

Where

𝑃𝑡 = predicted covariance matrix

𝑄= process noise covariance

Purpose
The prediction step:
 • Estimates the next state of the object

• Updates uncertainty in prediction

• Provides a prior estimate before measurement arrives

Example

Tracking a moving car:

• The system predicts where the car will appear in the next frame.

b) Derive the Kalman filter update (correction) step using the measurement equations.
(7 Marks)

Answer

The update step corrects the predicted state using the actual measurement obtained from
sensors.

Measurement Equation

𝑧𝑡 = 𝐻𝑥𝑡 + 𝑣𝑡

Where

𝑧𝑡 = measurement vector
𝐻= observation matrix
𝑣𝑡 = measurement noise

Kalman Gain

The Kalman gain determines how much weight is given to the measurement.

𝐾𝑡 = 𝑃𝑡 𝐻 𝑇 (𝐻𝑃𝑡 𝐻 𝑇 + 𝑅)−1

Where

𝑅= measurement noise covariance

State Update Equation

𝑥𝑡 = 𝑥̂𝑡 + 𝐾𝑡 (𝑧𝑡 − 𝐻𝑥̂𝑡 )

Covariance Update

𝑃𝑡 = (𝐼 − 𝐾𝑡 𝐻)𝑃𝑡

Purpose

The update step:

• Corrects prediction errors

• Improves state estimation accuracy

• Incorporates new measurements.

Question 142 (3 Marks)

Mention the role of covariance matrix in Kalman filtering.

Answer

The covariance matrix represents the uncertainty or error in the estimated state of the
system.

In Kalman filtering, it is used to:

• Measure the confidence in the predicted state

• Determine how much weight should be given to predictions and measurements

• Update the uncertainty after each prediction and correction step

Thus, the covariance matrix helps the Kalman filter maintain accurate and reliable state
estimation.

Question 143 (3 Marks)

Define state vector and measurement vector.

Answer

The state vector represents the internal state of a system that needs to be estimated. In object
tracking, it may include parameters such as position and velocity of the object.
Example:
 𝑥𝑡 = [𝑥, 𝑦, 𝑣𝑥 , 𝑣𝑦 ]𝑇

The measurement vector represents the observations obtained from sensors or detectors.
Example:

𝑧𝑡 = [𝑥𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑 , 𝑦𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑 ]𝑇

Thus, the state vector describes the system state while the measurement vector provides
observed data used to update the estimate.

Question 144 (3 Marks)

What are the main steps in Kalman filter algorithm?

Answer
The Kalman filter algorithm consists of two main steps:

1. Prediction Step – The filter predicts the next state of the system using the motion
model.

2. Update (Correction) Step – The predicted state is corrected using the measurement
obtained from sensors.

These two steps are repeated continuously to estimate the state of the system over time.