B.

Voc Data Science Main Project Batch 2022 - 2025

ABSTRACT
Driver activity recognition plays a pivotal role in intelligent transportation systems,
enhancing road safety through real-time monitoring and classification of driver behaviors.
Examines various machine learning and deep learning techniques, including Hidden Markov
Models (HMMs), Deep Belief Networks (DBNs), Long Short-Term Memory (LSTM) networks,
Gaussian Network Models (GNMs), and Fuzzy Rule-Based Systems, among others. These
methodologies enable the detection of inattentive and aggressive driving behaviors while improving
classification accuracy through hybrid models and real-time data processing. The findings
emphasize the necessity for large-scale datasets, improved real-time deployment strategies, and
hybrid frameworks integrating machine learning and rule-based models. By addressing the
limitations of individual techniques, this research aims to contribute to safer and more intelligent
vehicular environments.
Keywords: Driver Activity Recognition, Deep Learning, Intelligent Transportation Systems,
Machine Learning, Hidden Markov Models, Long Short-Term Memory, Gaussian Network Models.

DEPARTMENT OF DATA SCIENCE 1

[Link] Data Science Main Project Batch 2022 - 2025

INTRODUCTION
Intelligent transportation systems are becoming increasingly sophisticated, driven by the
need for enhanced road safety and efficient traffic management. Driver activity recognition (DAR)
is a fundamental aspect of these systems, aiming to detect and classify driver behaviors such as
drowsiness, distraction, and aggressive driving. With rising concerns over traffic accidents caused
by inattentive driving, developing robust, real-time recognition frameworks has become a priority.
DAR utilizes various machine learning and deep learning approaches to analyze driver behaviors
based on sensor and vision-based data. Traditional probabilistic models, such as Hidden Markov
Models (HMMs) and Gaussian Network Models (GNMs), provide structured frameworks for
identifying driving patterns based on probabilistic state transitions. Meanwhile, advanced deep
learning models, including Deep Belief Networks (DBNs) and Long Short-Term Memory (LSTM)
networks, leverage hierarchical learning and sequential data processing to improve recognition
accuracy. Additionally, Fuzzy Rule-Based Systems contribute to interpretability by integrating
expert knowledge into driver behavior classification.
Deng et al.[1] explored HMM-based behavior recognition, while Qu et al.[3] examined computer
vision-driven monitoring systems. Alkinani et al.[2] provided insights into deep learning
applications for detecting inattentive driving, and Wan et al.[5] focused on safety-driven
autonomous vehicle models. By consolidating findings from multiple sources, this paper presents a
comprehensive analysis of existing DAR approaches, highlighting their advantages, limitations,
and areas for future improvement.

DEPARTMENT OF DATA SCIENCE 2

[Link] Data Science Main Project Batch 2022 - 2025

RELATED WORKS
Advancements in intelligent transportation systems have led to extensive research on driver
behavior recognition and autonomous vehicle decision-making. Hidden Markov Models (HMMs)
have been widely explored for modeling driving behaviors, as discussed by Deng and Söffker .
Deep learning approaches, particularly Deep Belief Networks (DBNs), have significantly improved
the detection of inattentive and aggressive driving behaviors. The integration of Long Short-Term
Memory (LSTM) networks has further enhanced driver activity recognition and lane change
intention prediction in intelligent vehicles .

Machine learning-driven decision strategies have also been proposed for autonomous vehicle
navigation, focusing on improving path planning and decision-making in real-time environments.
Research on Gaussian Network Models (GNMs) and Fuzzy Rule-Based Networks has highlighted
their effectiveness in adaptive driver behavior modeling. Additionally, decision-making and
planning in intersection environments remain a critical area of study, where deep learning models
play a crucial role in optimizing vehicle navigation strategies

The field of driver activity recognition is examined, with a focus on the application of Deep Belief
Networks (DBN) to enhance driver behavior monitoring. The study highlights the significance of
recognizing driver activities to enhance road safety, prevent accidents, and contribute to the
development of intelligent transportation systems. Various research papers emphasize the
challenges associated with driver activity recognition, such as occlusions, varying illumination, and
the complexity of distinguishing between similar activities.

Traditional methods for driver activity recognition relied on Hidden Markov Models (HMM) and
handcrafted feature extraction techniques. However, these approaches often faced limitations in
handling complex patterns of driver behavior. The integration of deep learning, particularly DBNs,
has shown promising improvements in recognizing distracted driving and classifying driver actions
with high accuracy. CNNs effectively capture spatial dependencies in image data, making them
suitable for analyzing driver posture and hand movements. Meanwhile, DBNs leverage hierarchical
feature learning, improving the robustness of activity recognition models.

Research also emphasizes the importance of publicly available datasets, such as the StateFarm
Distracted Driver dataset, which has been widely utilized for training and testing deep learning
models. These datasets provide diverse driving scenarios, enabling models to generalize well to
real-world conditions. Several studies have proposed hybrid deep learning architectures that
combine CNNs with Recurrent Neural Networks (RNNs) or Long Short-Term Memory (LSTM),
Fuzzy Rule Based Networks to capture both spatial and temporal dependencies in driver behavior.

Moreover, recent advancements in transfer learning and data augmentation techniques have further
improved model performance. Some studies incorporate sensor fusion by integrating data from in-
vehicle sensors and cameras to enhance accuracy. The literature also explores the ethical and
privacy concerns of monitoring driver behavior, emphasizing the need for secure and transparent
[Link] studies suggest that deep learning approaches, particularly Fuzzy Rule-

DEPARTMENT OF DATA SCIENCE 3

[Link] Data Science Main Project Batch 2022 - 2025

Based Networks, GNMs, and DBNs, significantly enhance driver activity recognition. Future
research should focus on real-time deployment, computational efficiency, and integrating
multimodal data sources for enhanced accuracy and reliability.

Alkinani et al.[2] explores deep learning techniques for detecting inattentive and aggressive
driving behavior, highlighting the role of Deep Belief Networks (DBN) in driver activity
recognition .DBNs, a class of generative deep learning models, are emphasized for their ability to
extract high-level features from complex driving data. The paper discusses how DBNs are
particularly effective in recognizing sequential patterns in driver behavior, making them suitable for
identifying distracted or aggressive driving.

Chen et al.[10] and Qu et al.[3] reviews decision-making and planning strategies for autonomous
vehicles in intersection environments. In the context of Deep Belief Networks (DBN), the study
likely explores how DBNs can be used to enhance perception, prediction, and decision-making in
complex traffic scenarios. DBNs, as a type of deep learning model, are effective in extracting high-
level features from sensor data, such as camera feeds, LIDAR, and radar inputs. The study may
discuss how DBNs process sequential and spatial data to recognize vehicle behaviors, predict
traffic flow at intersections, and optimize navigation strategies. By leveraging DBN’s hierarchical
learning capabilities, the approach enables autonomous vehicles to differentiate between various
intersection scenarios, such as left turns, right turns, and yielding behaviors.

DBNs consist of multiple layers of Restricted Boltzmann Machines (RBMs) that learn hierarchical
representations of input data. In driver behavior analysis, DBNs process multimodal inputs,
including visual data from in-car cameras and sensor-based information from vehicle telemetry.
The paper highlights that DBNs excel in capturing latent features that distinguish between normal
and unsafe driving behaviors, such as erratic lane changes, sudden braking, and prolonged
distractions.

Compared to conventional machine learning methods, DBNs demonstrate improved accuracy in

feature learning and classification. The study highlights that DBNs, when integrated with
Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs), enhance
recognition performance by combining spatial and temporal features. However, challenges such as
computational complexity and large training datasets are noted, requiring optimization techniques
for real-time applications.

Alkinani et al. conclude that DBNs are a promising approach for driver activity recognition,
particularly in identifying aggressive and inattentive behaviors. Future research should focus on
improving DBN efficiency, integrating sensor fusion techniques, and addressing privacy concerns
in real-world deployments.

The paper by Ponti et al.[7] introduces a Generalized Learning Approach to Deep Neural
Networks, emphasizing the flexibility of Gaussian Network Model (GNM) in adapting to diverse
data distributions. GNMs extend conventional deep learning frameworks by incorporating
adaptable learning strategies that improve generalization in complex tasks. The study highlights
how GNMs enhance feature extraction and pattern recognition, making them well-suited for driver

DEPARTMENT OF DATA SCIENCE 4

[Link] Data Science Main Project Batch 2022 - 2025

behavior analysis. GNMs optimize learning efficiency by integrating multiple neural architectures,
such as Convolutional Neural Networks (CNNs) and Deep Belief Networks (DBNs), to refine
model accuracy and robustness. The paper suggests that GNMs can significantly improve driver
activity recognition by efficiently handling variations in driving styles and environmental
conditions.

You et al.[12] GNMs are indirectly referenced in the context of nonlinear driver parameter
estimation and steering behavior analysis for Advanced Driver Assistance Systems (ADAS). The
paper develops a data-driven approach to analyze driver steering behaviors using real-world field
test data. It emphasizes that nonlinear modeling techniques, including GNMs, enhance the accuracy
of driver behavior predictions by accommodating dynamic changes in driving patterns. GNMs
improve system adaptability by processing nonlinear relationships between driver inputs and
vehicle responses. The study also highlights the importance of integrating GNMs with adaptive
control mechanisms to refine real-time driver activity recognition and enhance ADAS functionality.

Both studies conclude that GNMs provide a promising framework for improving driver behavior
modeling by incorporating adaptive learning strategies and robust pattern recognition techniques.
Future research should focus on optimizing GNM-based models for real-time applications, sensor
fusion, and personalized driver assistance systems.

Deng et al.[1] focus on Hidden Markov Models (HMMs) for predicting and recognizing driver
behaviors, particularly in complex driving [Link] authors introduce a Fuzzy Logic-
Hidden Markov Model (FL-HMM) to enhance driving behavior [Link] HMMs are
effective in modeling sequential driving actions, but their performance is often limited by
uncertainty in driver behavior data. The integration of fuzzy logic helps handle uncertainties and
imprecise driving inputs, leading to more accurate and adaptive behavior recognition. The study
demonstrates that FL-HMM outperforms standard HMMs in predicting lane changes, acceleration
patterns, and other driver activities, making it a suitable approach for Advanced Driver Assistance
Systems (ADAS) and intelligent transportation systems.

Deng and Söffker present a comprehensive analysis of HMM-based approaches for driver behavior
recognition and prediction. They classify existing methods into discrete and continuous HMMs,
highlighting the advantages of each in modeling time-series driving data. The study emphasizes
that HMMs effectively capture sequential dependencies in driving activities, making them ideal for
real-time behavior monitoring. However, challenges such as scalability, computational complexity,
and adaptability to individual driving styles remain open research questions. The paper suggests
that combining HMMs with deep learning techniques (e.g., DBNs, CNNs) can enhance
performance by integrating both probabilistic and feature-learning [Link] research
should focus on hybrid models, integrating HMMs with deep learning for more robust and
personalized driver assistance systems.

Jatla et al.[9] explores a machine learning-driven decision strategy for autonomous vehicle
navigation, with a focus on enhancing decision-making capabilities. In the context of the Hidden
Markov Model (HMM) algorithm, the study likely utilizes HMM for sequential data analysis,
allowing autonomous vehicles to predict and adapt to changing driving environments.

DEPARTMENT OF DATA SCIENCE 5

[Link] Data Science Main Project Batch 2022 - 2025

HMM is well-suited for modeling time-series data, making it effective for recognizing driving
patterns, detecting lane changes, and anticipating vehicle maneuvers. The study may have applied
HMM to process sensor inputs, such as LIDAR, radar, and cameras, to classify different driving
states and make probabilistic decisions. By leveraging HMM’s ability to model uncertainties in
dynamic environments, the proposed strategy improves autonomous vehicle navigation, ensuring
smoother transitions between driving states and reducing errors in path planning.

Research by Xing et al.[4] , Wan et al.[5] , and Tang et al.[6] highlights the application of Long
Short-Term Memory (LSTM) networks in driver activity recognition, autonomous vehicle safety,
and lane change intention prediction.

Xing et al.[4] propose a deep learning framework for driver activity recognition, leveraging LSTM
networks to process sequential driving data. Emphasizes the advantages of LSTM in capturing
temporal dependencies and recognizing complex driver behaviors, such as distraction, drowsiness,
and hand movements. The model improves real-time behavior recognition, enhancing the
adaptability of intelligent vehicle systems in monitoring driver states.

Wan et al.[5] focus on autonomous vehicle safety using an LSTM-based Super-Resolution

Direction of Arrival (DOA) estimation. The study demonstrates that LSTM models can accurately
estimate driving trajectories and surrounding vehicle positions, significantly improving the
situational awareness of autonomous and assisted driving systems. The approach enhances real-
time decision-making capabilities in intelligent transportation.

Tang et al.[6] develop an LSTM-based lane change recognition model for intelligent vehicles.
Traditional lane change detection methods struggle with delays and inaccuracies in real-world
conditions. LSTM networks effectively model driver intent from sequential driving signals, such as
steering angle, speed, and acceleration, leading to improved prediction accuracy. The results
suggest that LSTM outperforms conventional machine learning approaches, making it a reliable
choice for lane change detection and adaptive driver assistance.

Findings highlight the effectiveness of LSTM networks in learning and predicting driver behaviors
from sequential data, making them well-suited for real-time intelligent vehicle applications. Future
research should focus on integrating LSTMs with attention mechanisms and sensor fusion for even
more accurate and robust driver activity recognition.

Deng & Söffker and Chong et al.[1] explore the use of Fuzzy Rule-Based Networks (FRBNs) in
modeling and predicting driver behavior, combining fuzzy logic with machine learning techniques
for more adaptive and interpretable decision-making in intelligent transportation systems. [8][11]

Deng & Söffker introduce a Fuzzy Logic-Hidden Markov Model (FL-HMM) for driving behavior
prediction, integrating fuzzy rule-based decision-making with probabilistic modeling. The study
leverages fuzzy logic to handle uncertainties in driver behavior, such as varying acceleration
patterns and reaction times. The HMM component captures temporal dependencies, enabling a
dynamic understanding of driver actions over time. The model demonstrates high accuracy in
predicting lane changes and braking behavior, making it useful for Advanced Driver Assistance
Systems (ADAS).

DEPARTMENT OF DATA SCIENCE 6

[Link] Data Science Main Project Batch 2022 - 2025

Chong et al.[8] present a rule-based neural network approach for modeling naturalistic driver
behavior. This hybrid framework combines fuzzy rule-based systems with neural networks to learn
and adapt to real-world traffic scenarios. The fuzzy rules provide interpretability, while the neural
network component refines decision-making based on continuous learning. The study emphasizes
the effectiveness of fuzzy rules in handling uncertainty and imprecise input data, such as driver
distractions, variations in driving habits, and external traffic conditions. The model is validated
using naturalistic driving datasets, proving its capability to accurately simulate human-like driving
behaviors

Research demonstrates that Fuzzy Rule-Based Networks improve driver behavior modeling by
combining human-like reasoning with adaptive learning techniques. Future work can explore the
implementation of deep learning with fuzzy logic to further improve prediction accuracy and real-
time adaptability in intelligent transportation systems.

DEPARTMENT OF DATA SCIENCE 7

[Link] Data Science Main Project Batch 2022 - 2025

OVERVIEW OF ALGORITHMS
Accurately recognizing driver activity is crucial for enhancing road safety and developing
intelligent driver assistance systems. Various machine learning and deep learning algorithms have
been employed to analyze driver behavior, each with distinct methodologies and advantages. This
section provides an overview of five key algorithms: Deep Belief Networks (DBN), Gaussian
Naïve Bayes Model (GNM), Hidden Markov Model (HMM), Long Short-Term Memory (LSTM),
and Fuzzy Rule-Based Networks.

Each algorithm is examined through a structured framework, covering essential aspects such as
region proposal, feature extraction, object classification, and limitations.

 Region proposal involves identifying areas of interest in sensor or image data related to
driver activity.

 Feature extraction focuses on deriving relevant information from raw data to enhance
recognition accuracy.

 Object classification applies learned models to categorize detected behaviors,

distinguishing between actions such as normal driving, distraction, or aggressive.

 Finally, Limitations of each method are discussed, addressing computational complexity,

adaptability, and accuracy challenges.

[Link] Belief Networks (DBN)

Deep Belief Networks (DBN) are hierarchical deep learning models consisting of multiple layers of
Restricted Boltzmann Machines (RBMs) . They are widely used in driver behavior recognition due
to their ability to learn complex, high-dimensional feature representations. DBNs have been applied
in detecting inattentive and aggressive driving behaviors by analyzing sequential driving patterns
and sensor data.

[Link] Proposal

 DBNs do not perform region proposal in the conventional sense but can be adapted for
spatial-temporal feature selection.

 Studies have utilized DBNs for detecting critical driver behavior patterns by analyzing
facial movements, hand gestures, and interaction with vehicle controls.

 This approach aids in identifying distractions and aggressive maneuvers by segmenting

relevant regions in sensor and video data.

DEPARTMENT OF DATA SCIENCE 8

[Link] Data Science Main Project Batch 2022 - 2025

Figure 1:Deep Belief Network (DBN) training process using

Restricted Boltzmann Machines (RBMs) and fine-tuning with Softmax
classification.

[Link] Extraction

 Feature extraction in DBNs follows a layer-wise unsupervised pretraining approach. Each

RBM layer progressively extracts more abstract features from raw sensor data, enabling
deep hierarchical representations. The key steps include:

 The first RBM learns low-level features such as speed variations, steering angles, and
braking patterns.

 Deeper RBMs capture higher-order features, including sequential dependencies in driving

behavior.

 Fine-tuning is performed using supervised learning to optimize feature representations for

classification.

[Link] Classification

 Once feature representations are learned, DBNs use classifiers such as Softmax to
categorize driving behavior.

 Studies have demonstrated their effectiveness in distinguishing normal, inattentive, and

aggressive driving patterns.

 The ability of DBNs to model temporal dependencies enhances their accuracy in

recognizing real-time behavioral shifts in drivers.

Limitations

While DBNs offer several benefits, they also come with certain limitations:

1. High Computational Cost: The layer-wise training process requires substantial

computational resources, making real-time deployment challenging.

2. Training Complexity: Pretraining and fine-tuning involve multiple optimization steps,

increasing the risk of slow convergence.

DEPARTMENT OF DATA SCIENCE 9

[Link] Data Science Main Project Batch 2022 - 2025

3. Limited Spatial Awareness: Unlike CNNs, DBNs are not inherently designed to capture
spatial dependencies, which may reduce their effectiveness in analyzing visual driving data.

4. Sensitivity to Hyperparameters: The performance of DBNs depends on careful selection of

hyperparameters such as learning rate, number of hidden units, and regularization
techniques.

5. Difficulty in Real-Time Applications: The sequential nature of training and inference in

DBNs poses challenges for real-time driver monitoring systems.

While these challenges persist, DBNs remain an essential tool for driver behavior recognition,
especially when combined with other deep learning models to enhance accuracy and efficiency.
Their ability to model complex hierarchical representations makes them well-suited for capturing
intricate driving patterns.

[Link] Naïve Bayes Model (GNM)

Gaussian Naïve Bayes Model (GNM) is a probabilistic classifier based on Bayes' theorem,
assuming that features follow a Gaussian (normal) distribution. It is widely used in driver activity
recognition due to its simplicity, efficiency, and ability to handle real-time classification tasks.

[Link] Proposal

 GNM does not explicitly generate region proposals like CNN-based object detection
methods.

 However, in the context of driver behavior recognition, feature selection techniques can be
applied to extract relevant sensor-based or visual data points, such as hand positions, eye
movements, and vehicle control interactions.

 These features act as input for probabilistic classification, allowing the model to predict
driver activities.

[Link] Extraction

Feature extraction in GNM involves computing statistical properties of input data, assuming each
feature follows a Gaussian distribution. The key steps include:

 Extracting numerical features such as speed variations, steering angle fluctuations, hand
movement patterns, and gaze direction.

 Computing mean and variance for each feature, modeling the likelihood distribution based
on training data.

 Assuming conditional independence between features, which simplifies computation and

speeds up inference.

[Link] Classification

DEPARTMENT OF DATA SCIENCE 10

[Link] Data Science Main Project Batch 2022 - 2025

 GNM applies Bayes' theorem to calculate the probability of a given driver activity class
based on extracted features. The classification process follows these steps:

 Compute the posterior probability of each class using the Gaussian probability density
function.

 Assign the class with the highest posterior probability as the predicted driver activity.

 Common applications include classifying normal driving, distracted driving, and aggressive
driving behaviors.

Limitations

Despite its advantages, GNM has several limitations:

1. Strong Independence Assumption: It assumes feature independence, which may not hold
true for complex driver behaviors where multiple factors are interdependent.

2. Sensitivity to Feature Distribution: GNM assumes a Gaussian distribution for all features,
which may not always be accurate in real-world driving data.

3. Limited Expressiveness: Compared to deep learning models, GNM lacks the ability to
capture hierarchical and spatial relationships in image-based driver activity recognition.

4. Performance on High-Dimensional Data: As the number of features increases, the

independence assumption can degrade classification performance.

5. Lower Accuracy for Complex Behaviors: While GNM is efficient for simple classification
tasks, it may struggle with highly dynamic or overlapping driver activities.

GNM remains a valuable model for lightweight and interpretable driver behavior classification,
especially in real-time applications where computational efficiency is a priority.

[Link] Markov Model (HMM)

Hidden Markov Model (HMM) is a statistical model that represents sequential data using hidden
states and observable outputs[1][11]. It is widely used in time-series analysis, including driver
behavior recognition, due to its ability to model temporal dependencies and predict state transitions.

[Link] Proposal

 HMM does not explicitly perform region proposal in the traditional sense. However, in
driver activity recognition, it can be applied to segment driving sequences into meaningful
states, such as normal driving, distracted driving, or aggressive maneuvers.

 Sensor-based data, including steering angle, speed variation, and eye movement, serve as
key inputs to determine transition probabilities between different driver behavior states.

[Link] Extraction

DEPARTMENT OF DATA SCIENCE 11

[Link] Data Science Main Project Batch 2022 - 2025

Feature extraction in HMM involves processing time-series data and representing it as a sequence
of observable states. The key steps include:

 Collecting sequential sensor data such as acceleration patterns, lane deviations, and hand
movements.

 Defining discrete or continuous observation states based on statistical analysis of driving

behavior.

 Estimating transition probabilities between states using training data, capturing temporal
relationships and behavior shifts.

[Link] Classification

HMM classifies driver activities by analyzing the probability distribution of state transitions. The
classification process includes:

 Training the model on labeled driving sequences to learn state transition probabilities.

 Using the Viterbi algorithm or forward-backward algorithms to determine the most likely
sequence of hidden states given observed features.

 Assigning a driver behavior class based on the inferred state sequence.

Figure 2: Real-time driver activity recognition based on lane-keeping and lane-changing observations.

Limitations

Effectiveness in modeling temporal dependencies, HMM has several limitations:

1. Limited Scalability: Training an HMM with a large number of states can be computationally
expensive and require extensive data.

2. Assumption of Stationary Transition Probabilities: HMM assumes that transition

probabilities remain constant over time, which may not hold true in dynamic driving
scenarios.

3. Sensitivity to Initial Parameters: The performance of HMM depends heavily on the

initialization of transition and emission probabilities.

4. Lower Performance for Complex Patterns: Compared to deep learning models, HMM
struggles with highly complex and non-linear driver activity patterns.

DEPARTMENT OF DATA SCIENCE 12

[Link] Data Science Main Project Batch 2022 - 2025

5. Need for Large Annotated Datasets: Effective training requires well-labeled sequential data,
which can be challenging to obtain in real-world driving conditions.

Challenges exist, HMM continues to be a valuable approach for modeling sequential driver
behaviors, especially when integrated with other deep learning methods to improve prediction
accuracy.

[Link] Short-Term Memory (LSTM)

Long Short-Term Memory (LSTM) is a specialized recurrent neural network (RNN) architecture
designed to capture long-range dependencies in sequential data [4][5][6]. It has been widely adopted in
driver behavior recognition due to its ability to process time-series data effectively and model
temporal patterns in driving activities.

[Link] Proposal

 LSTM does not utilize traditional region proposal techniques, as it primarily focuses on
sequential data rather than spatial features.

 However, in driver activity recognition, relevant regions in time-series sensor data, such as
speed variations, acceleration patterns, steering wheel movements, and eye-tracking data,
can be segmented and used as input to the model.

 This segmentation helps in identifying key behavior transitions such as lane changing,
braking, or distracted driving.

[Link] Extraction

Feature extraction in LSTM involves transforming raw time-series data into meaningful feature
representations by leveraging memory units and gates to capture dependencies. The key steps
include:

 Extracting sequential features from driving data, including vehicle dynamics, eye movement
patterns, and hand positions.

 Utilizing LSTM cells to process time-series inputs while preserving contextual

dependencies over time.

 Learning hidden state representations that encode temporal dependencies crucial for
differentiating driver activities.

[Link] Classification

LSTM classifies driver behaviors by analyzing sequential dependencies in the input data. The
classification process involves:

 Training the model on labeled sequential data, where each sequence corresponds to a
specific driving behavior.

DEPARTMENT OF DATA SCIENCE 13

[Link] Data Science Main Project Batch 2022 - 2025

 Using learned hidden states to predict the likelihood of different driver activity categories,
such as normal driving, distracted driving, or aggressive maneuvers.

 Applying Softmax or other classification layers to assign a final category to each observed
sequence.

Limitations

Although effectiveness in modeling temporal sequences, LSTM has several limitations:

1. High Computational Cost: LSTM networks require significant computational resources due
to their recurrent structure and gating mechanisms.

2. Training Complexity: The model is prone to vanishing and exploding gradient issues if not
properly optimized.

3. Memory Constraints: Storing long sequences in memory can lead to inefficiencies,

especially when dealing with large-scale driving datasets.

4. Sensitivity to Hyperparameters: The performance of LSTM depends on careful tuning of

parameters such as sequence length, number of hidden units, and learning rate.

5. Limited Spatial Awareness: LSTM focuses on temporal dependencies but lacks built-in
spatial feature extraction, which may reduce its effectiveness in visual-based driver
monitoring systems.

Regardless of these drawbacks, LSTM continues to be an essential tool for driver activity
recognition. Its performance significantly improves when integrated with convolutional neural
networks (CNNs) or attention mechanisms, allowing for better spatial-temporal feature learning
and enhanced accuracy.

[Link] Rule-Based Network

Fuzzy Rule-Based Networks are a class of intelligent systems that utilize fuzzy logic to handle
uncertainty and imprecision in data ][1][8]. These networks are particularly useful in driver behavior
recognition, where the distinction between different driving activities is not always clear-cut.

[Link] Proposal

 Fuzzy Rule-Based Networks do not follow traditional region proposal techniques but
instead rely on fuzzy logic to segment and analyze data.

 In driver activity recognition, input features such as speed, acceleration, and hand
movement can be mapped into fuzzy sets (e.g., low, medium, high) to identify potential
behavior patterns.

 This approach helps in defining decision regions for different driving activities.

[Link] Extraction

DEPARTMENT OF DATA SCIENCE 14

[Link] Data Science Main Project Batch 2022 - 2025

Feature extraction in Fuzzy Rule-Based Networks involves transforming raw sensor data into
linguistic variables and membership functions. The key steps include:

 Defining fuzzy sets for relevant driving parameters such as braking intensity, steering angle,
and throttle usage.

 Applying fuzzy membership functions to map sensor inputs to corresponding fuzzy values.

 Generating fuzzy rules that establish relationships between different driving behaviors
based on expert knowledge or data-driven approaches.

[Link] Classification

Classification in Fuzzy Rule-Based Networks is performed using a set of IF-THEN rules that infer
driver behavior based on fuzzy logic operations. The classification process includes:

 Evaluating input data against predefined fuzzy rules.

 Aggregating rule outputs using fuzzy inference mechanisms such as Mamdani or Sugeno
models.

 Defuzzifying the results to produce a crisp decision regarding driver activity (e.g., normal
driving, distracted driving, aggressive driving).

Figure 3: Fuzzy rule-based system architecture for driver activity recognition, illustrating the
fuzzification, inference, and defuzzification process.

Limitations

Flexibility in handling uncertain data, Fuzzy Rule-Based Networks have several limitations:

1. Rule Complexity: Defining an extensive set of fuzzy rules can be challenging, particularly in
complex driving scenarios.

2. Scalability Issues: As the number of input features increases, the number of required fuzzy
rules grows exponentially, leading to increased computational complexity.

3. Dependency on Expert Knowledge: The effectiveness of fuzzy systems relies heavily on

well-designed membership functions and rule bases, which may require domain expertise.

4. Limited Adaptability: Fuzzy rules are typically predefined, making it difficult for the model
to adapt dynamically to new driving conditions.

DEPARTMENT OF DATA SCIENCE 15

[Link] Data Science Main Project Batch 2022 - 2025

5. Interpretability vs. Performance Trade-off: While fuzzy systems are interpretable, they may
not achieve the same accuracy as deep learning models, which can learn more complex
patterns from data.

Fuzzy Rule-Based Networks continue to play a crucial role in driver activity recognition. Their
integration with deep learning techniques enhances decision-making and interpretability, making
them valuable in hybrid systems for improving the accuracy and robustness of driver behavior
analysis.

DEPARTMENT OF DATA SCIENCE 16

[Link] Data Science Main Project Batch 2022 - 2025

DATASET DETAILS

[Link] Description

This research utilizes a structured dataset designed for driver activity recognition, containing
1,000 records and 13 features. The dataset captures various aspects of driver behavior, including
vehicle dynamics, hand and eye movements, and lane deviation, all of which are crucial for
detecting distracted or inattentive driving activities. The dataset provides a well-balanced
representation of different driver activities to facilitate effective classification using deep learning
models such as Deep Belief Networks (DBN).

[Link] Description

The dataset consists of 12 numerical features and 1 categorical target variable (activity). These
features are categorized into three primary groups:

1. Vehicle Dynamics Features: These features capture the movement and control of the
vehicle.
 Speed (km/h) – The velocity of the vehicle at the recorded instance.
 Acceleration (m/s²) – The rate of change of speed, indicating sudden acceleration or
deceleration.
 Brake Pressure (kPa) – The amount of pressure applied to the brake pedal, which
reflects braking behavior.
 Steering Angle (degrees) – The angle of the steering wheel, showing directional
changes made by the driver.
 Lane Deviation (meters) – The lateral displacement of the vehicle from the center of
the lane, indicating whether the driver is maintaining a stable lane position.
2. Hand Movement Features: These features track the driver's hand position and behavior.
 Hand X, Hand Y, Hand Z (3D coordinates) – The spatial position of the driver’s
hands in relation to the steering wheel.
 Hand-off Wheel Time (seconds) – The duration for which the driver’s hands are not
on the steering wheel, which is a strong indicator of distracted driving.
3. Eye Movement Features: These features help in assessing the driver’s attention level.
 Eye Gaze Horizontal (float) – The driver’s eye gaze movement along the horizontal
axis, indicating attention shifts.
 Eye Gaze Vertical (float) – The driver’s eye gaze movement along the vertical axis,
useful for identifying head tilts or distractions.
 Eye Movement (float) – The overall magnitude of eye movements, helping detect
inattentiveness or distractions.

4.2Target Variable (Driver Activity Classification)

DEPARTMENT OF DATA SCIENCE 17

[Link] Data Science Main Project Batch 2022 - 2025

 Activity (Categorical Variable): Represents the specific activity performed by the driver at
the recorded moment. The possible activities include:

 Drinking – The driver is holding or consuming a beverage.

 Texting – The driver is using a mobile device for texting.
 Other activities (e.g., talking, reaching for objects, or adjusting controls) may also
be included, depending on the dataset distribution.

DEPARTMENT OF DATA SCIENCE 18

[Link] Data Science Main Project Batch 2022 - 2025

RESULTS AND DISCUSSION

The performance of five different algorithms—Fuzzy Rule-Based, Hidden Markov Model

(HMM), Gaussian Naive Bayes (GNM), Deep Belief Network (DBN), and Long Short-Term
Memory (LSTM)—was evaluated for the driver activity recognition task. The key performance
metrics considered were Precision, Recall, F1-Score, and [Link] Fuzzy Rule-Based model
achieved a precision of 0.608, which was the highest among all models. However, it struggled in
terms of recall (0.270) and F1-score (0.188), indicating that while it made some correct predictions,
it failed to generalize effectively to unseen test data. This suggests that the rule-based approach may
require further refinement or additional rules to improve recall.

The HMM model performed slightly better in terms of recall (0.254) but had a lower precision
(0.253) and an F1-score of 0.242. This implies that HMM had moderate success in capturing
sequential dependencies in the data but still faced challenges in classification [Link] GNM
and LSTM models exhibited similar performance, both with a precision of 0.114, recall of 0.254,
and an F1-score of 0.157. These low precision values indicate that these models frequently
misclassified instances, resulting in a high number of false [Link] DBN model demonstrated
relatively better performance with a precision of 0.229, recall of 0.262, and an F1-score of 0.209.
This suggests that DBN was able to capture meaningful patterns in the dataset but still requires
optimization to improve its classification capability.

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Fuzzy Rule HMM GNM DBN
Based

Precision Recall F1-Score Accuracy

Figure 4:Performance Comparison Of Different Algorithms

The performance across all models was lower than expected, indicating possible limitations in
feature representation, dataset quality, or model hyperparameter tuning. The Fuzzy Rule-Based
model performed the best in precision but lagged in recall, while the DBN model showed balanced
performance across all metrics. The LSTM and GNM models struggled with poor precision,
suggesting difficulties in learning effective feature representations from the dataset.

One possible reason for the lower performance could be the dataset distribution, where some
activities might have been underrepresented, leading to biased classification. Future improvements

DEPARTMENT OF DATA SCIENCE 19

[Link] Data Science Main Project Batch 2022 - 2025

could involve dataset augmentation, better feature engineering, and fine-tuning hyperparameters to
enhance model [Link] summary, while the Fuzzy Rule-Based and DBN models showed
some promising results, further improvements are required to achieve higher accuracy in driver
activity recognition.

Algorithm Precision Recall F1-Score Accuracy

Fuzzy Rule Based 0.608 0.270 0.188 0.270

HMM 0.253 0.254 0.242 0.254

GNM 0.114 0.254 0.157 0.254

DBN 0.229 0.262 0.209 0.262

LSTM 0.114 0.254 0.157 0.254

Figure 5:Result Of Each Algorithm

DEPARTMENT OF DATA SCIENCE 20

[Link] Data Science Main Project Batch 2022 - 2025

CONCLUSION

The results demonstrated that no single model significantly outperformed the others, with
all achieving relatively low performance metrics. The Fuzzy Rule-Based model exhibited the
highest Precision but suffered from a low F1-Score, indicating limitations in overall effectiveness.
The DBN model provided a more balanced performance across different metrics, making it a
promising candidate for further refinement. However, the GNM and LSTM models struggled with
poor precision, leading to frequent misclassification of driver activities.

One key challenge identified in this study is the limited ability of the models to generalize
effectively across different driver activities. This could be attributed to dataset constraints, feature
extraction limitations, or suboptimal hyperparameter tuning. Future research should focus on
improving dataset quality, employing advanced feature extraction techniques, and optimizing
model architectures to enhance classification [Link] approaches that combine statistical
and deep learning models may offer improved performance by leveraging the strengths of each
method. Additionally, techniques such as spatiotemporal feature analysis and attention mechanisms
could further enhance model robustness. Expanding the dataset to include a broader range of driver
behaviors with high-quality annotations will also be essential for real-world applicability.

By addressing these challenges, future work can contribute to the development of more reliable
driver activity recognition systems, ultimately improving road safety and advancing intelligent
vehicle technologies.

DEPARTMENT OF DATA SCIENCE 21

[Link] Data Science Main Project Batch 2022 - 2025

IMPLEMENTATION DETAILS

[Link] RULE BASED NETWORK

import numpy as np
import pandas as pd
import skfuzzy as fuzz
import [Link] as ctrl
from [Link] import LabelEncoder, StandardScaler
from [Link] import accuracy_score, precision_score, recall_score, f1_score
from tabulate import tabulate

# Load training data

train_data = pd.read_csv("driver_activity_final.csv")
X_train = train_data.iloc[:, :-1].values # Select first 12 numerical columns
y_train = train_data.iloc[:, -1].values # Select the last column (activity)

# Encode labels
label_encoder = LabelEncoder()
y_train_encoded = label_encoder.fit_transform(y_train)

# Scale features
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)

# Define fuzzy variables

speed = [Link]([Link](-3, 3, 0.1), 'speed')
steering_angle = [Link]([Link](-3, 3, 0.1), 'steering_angle')
eye_movement = [Link]([Link](-3, 3, 0.1), 'eye_movement')
activity = [Link]([Link](0, len(label_encoder.classes_), 1), 'activity')

# Define membership functions

[Link](3)
steering_angle.automf(3)
eye_movement.automf(3)
[Link](3)

# Define fuzzy rules

rule1 = [Link](speed['poor'] & steering_angle['good'], activity['average']) # Example rule
rule2 = [Link](eye_movement['good'] & speed['average'], activity['good'])
rule3 = [Link](steering_angle['average'] & eye_movement['poor'], activity['poor'])

activity_ctrl = [Link]([rule1, rule2, rule3])

activity_sim = [Link](activity_ctrl)

DEPARTMENT OF DATA SCIENCE 22

[Link] Data Science Main Project Batch 2022 - 2025

# Function to predict activities from uploaded test data

def predict_activity(test_file):
test_data = pd.read_csv("/content/test_data_120.csv")
X_test = test_data.iloc[:, :-1].values # Select first 12 numerical columns

if test_data.shape[1] == 13:
y_test = test_data.iloc[:, -1].values
y_test_encoded = label_encoder.transform(y_test)
else:
y_test = None
y_test_encoded = None

X_test_scaled = [Link](X_test)
y_pred_encoded = []

for i in range(len(X_test_scaled)):
activity_sim.input['speed'] = X_test_scaled[i, 0]
activity_sim.input['steering_angle'] = X_test_scaled[i, 1]
activity_sim.input['eye_movement'] = X_test_scaled[i, 2]
activity_sim.compute()

pred_activity = int(round(activity_sim.output['activity']))
y_pred_encoded.append(pred_activity)

y_pred = label_encoder.inverse_transform(y_pred_encoded)

result_table = [[i+1, pred] for i, pred in enumerate(y_pred)]

print("\nPredicted Activities:")
print(tabulate(result_table, headers=["Record No.", "Predicted Activity"], tablefmt="grid"))

if y_test is not None:

accuracy = accuracy_score(y_test_encoded, y_pred_encoded)
precision = precision_score(y_test_encoded, y_pred_encoded, average='weighted',
zero_division=1)
recall = recall_score(y_test_encoded, y_pred_encoded, average='weighted', zero_division=1)
f1 = f1_score(y_test_encoded, y_pred_encoded, average='weighted', zero_division=1)

evaluation_table = [["Accuracy", accuracy],

["Precision", precision],
["Recall", recall],
["F1 Score", f1]]
print("\nEvaluation Metrics:")
print(tabulate(evaluation_table, headers=["Metric", "Value"], tablefmt="grid"))

DEPARTMENT OF DATA SCIENCE 23

[Link] Data Science Main Project Batch 2022 - 2025

return y_pred, accuracy, precision, recall, f1

return y_pred

# Example usage
predict_activity("test_data.csv")

[Link]
import numpy as np
import pandas as pd
from [Link] import LabelEncoder, StandardScaler
from sklearn.naive_bayes import GaussianNB
from [Link] import accuracy_score, precision_score, recall_score, f1_score
from tabulate import tabulate

# Load training data

train_data = pd.read_csv("driver_activity_final.csv")
X_train = train_data.iloc[:, :-1] # Select first 12 numerical columns
y_train = train_data.iloc[:, -1] # Select the last column (activity)

# Encode labels
label_encoder = LabelEncoder()
y_train_encoded = label_encoder.fit_transform(y_train)

# Scale features
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)

# Define Gaussian Naive Bayes model

gnb = GaussianNB()
[Link](X_train_scaled, y_train_encoded)

# Function to predict activities from uploaded test data

def predict_activity(test_file):
test_data = pd.read_csv("/content/test_data_120.csv")
X_test = test_data.iloc[:, :-1] # Select first 12 numerical columns

if test_data.shape[1] == 13: # If activity column exists

y_test = test_data.iloc[:, -1]
y_test_encoded = label_encoder.transform(y_test)
else:
y_test = None

DEPARTMENT OF DATA SCIENCE 24

[Link] Data Science Main Project Batch 2022 - 2025

y_test_encoded = None

X_test_scaled = [Link](X_test)
y_pred_encoded = [Link](X_test_scaled)
y_pred = label_encoder.inverse_transform(y_pred_encoded)

# Display predictions in tabular format

result_table = [[i+1, pred] for i, pred in enumerate(y_pred)]
print("\nPredicted Activities:")
print(tabulate(result_table, headers=["Record No.", "Predicted Activity"], tablefmt="grid"))

if y_test is not None:

accuracy = accuracy_score(y_test_encoded, y_pred_encoded)
precision = precision_score(y_test_encoded, y_pred_encoded, average='weighted')
recall = recall_score(y_test_encoded, y_pred_encoded, average='weighted')
f1 = f1_score(y_test_encoded, y_pred_encoded, average='weighted')

evaluation_table = [["Accuracy", accuracy],

["Precision", precision],
["Recall", recall],
["F1 Score", f1]]
print("\nEvaluation Metrics:")
print(tabulate(evaluation_table, headers=["Metric", "Value"], tablefmt="grid"))

return y_pred, accuracy, precision, recall, f1

return y_pred

# Example usage
predict_activity("test_data.csv")

[Link]
import numpy as np
import pandas as pd
from hmmlearn import hmm
from [Link] import LabelEncoder, StandardScaler
from [Link] import accuracy_score, precision_score, recall_score, f1_score
from tabulate import tabulate

# Load training data

train_data = pd.read_csv("driver_activity_final.csv")
X_train = train_data.iloc[:, :-1].values # Select first 12 numerical columns
y_train = train_data.iloc[:, -1].values # Select the last column (activity)

DEPARTMENT OF DATA SCIENCE 25

[Link] Data Science Main Project Batch 2022 - 2025

# Encode labels
label_encoder = LabelEncoder()
y_train_encoded = label_encoder.fit_transform(y_train)

# Scale features
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)

# Define and train HMM model

num_classes = len([Link](y_train_encoded))
model = [Link](n_components=num_classes, covariance_type="diag", n_iter=2000)
[Link](X_train_scaled)

# Function to predict activities from uploaded test data

def predict_activity(test_file):
test_data = pd.read_csv("/content/test_data_120.csv")
X_test = test_data.iloc[:, :-1].values # Select first 12 numerical columns

if test_data.shape[1] == 13: # If activity column exists

y_test = test_data.iloc[:, -1].values
y_test = [label if label in label_encoder.classes_ else "Unknown" for label in y_test]
y_test_encoded = label_encoder.transform(y_test)
else:
y_test = None
y_test_encoded = None

X_test_scaled = [Link](X_test)
y_pred_encoded = [Link](X_test_scaled)
y_pred = label_encoder.inverse_transform(y_pred_encoded)

# Display predictions in tabular format

result_table = [[i+1, pred] for i, pred in enumerate(y_pred)]
print("\nPredicted Activities:")
print(tabulate(result_table, headers=["Record No.", "Predicted Activity"], tablefmt="grid"))

if y_test is not None and "Unknown" not in y_test:

accuracy = accuracy_score(y_test_encoded, y_pred_encoded)
precision = precision_score(y_test_encoded, y_pred_encoded, average='weighted',
zero_division=1)
recall = recall_score(y_test_encoded, y_pred_encoded, average='weighted', zero_division=1)
f1 = f1_score(y_test_encoded, y_pred_encoded, average='weighted', zero_division=1)

DEPARTMENT OF DATA SCIENCE 26

[Link] Data Science Main Project Batch 2022 - 2025

evaluation_table = [["Accuracy", accuracy],

["Precision", precision],
["Recall", recall],
["F1 Score", f1]]
print("\nEvaluation Metrics:")
print(tabulate(evaluation_table, headers=["Metric", "Value"], tablefmt="grid"))

return y_pred, accuracy, precision, recall, f1

return y_pred

# Example usage
predict_activity("test_data.csv")

[Link]
import numpy as np
import pandas as pd
from [Link] import LabelEncoder, StandardScaler
from sklearn.neural_network import BernoulliRBM
from [Link] import Pipeline
from sklearn.linear_model import LogisticRegression
from [Link] import accuracy_score, precision_score, recall_score, f1_score
from tabulate import tabulate

# Load training data

train_data = pd.read_csv("driver_activity_final.csv")
X_train = train_data.iloc[:, :-1] # Select first 12 numerical columns
y_train = train_data.iloc[:, -1] # Select the last column (activity)

# Encode labels
label_encoder = LabelEncoder()
y_train_encoded = label_encoder.fit_transform(y_train)

# Scale features
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)

# Define DBN using RBM layers and logistic regression

rbm1 = BernoulliRBM(n_components=128, learning_rate=0.01, n_iter=50)
rbm2 = BernoulliRBM(n_components=64, learning_rate=0.01, n_iter=50)
log_reg = LogisticRegression(max_iter=1000)

dbn_pipeline = Pipeline(steps=[

DEPARTMENT OF DATA SCIENCE 27

[Link] Data Science Main Project Batch 2022 - 2025

('rbm1', rbm1),
('rbm2', rbm2),
('logistic', log_reg)
])

dbn_pipeline.fit(X_train_scaled, y_train_encoded)

# Function to predict activities from uploaded test data

def predict_activity(test_file):
test_data = pd.read_csv("/content/test_data_120.csv")
X_test = test_data.iloc[:, :-1] # Select first 12 numerical columns

if test_data.shape[1] == 13: # If activity column exists

y_test = test_data.iloc[:, -1]
y_test_encoded = label_encoder.transform(y_test)
else:
y_test = None
y_test_encoded = None

X_test_scaled = [Link](X_test)
y_pred_encoded = dbn_pipeline.predict(X_test_scaled)
y_pred = label_encoder.inverse_transform(y_pred_encoded)

# Display predictions in tabular format

result_table = [[i+1, pred] for i, pred in enumerate(y_pred)]
print("\nPredicted Activities:")
print(tabulate(result_table, headers=["Record No.", "Predicted Activity"], tablefmt="grid"))

if y_test is not None:

accuracy = accuracy_score(y_test_encoded, y_pred_encoded)
precision = precision_score(y_test_encoded, y_pred_encoded, average='weighted')
recall = recall_score(y_test_encoded, y_pred_encoded, average='weighted')
f1 = f1_score(y_test_encoded, y_pred_encoded, average='weighted')

evaluation_table = [["Accuracy", accuracy],

["Precision", precision],
["Recall", recall],
["F1 Score", f1]]
print("\nEvaluation Metrics:")
print(tabulate(evaluation_table, headers=["Metric", "Value"], tablefmt="grid"))

return y_pred, accuracy, precision, recall, f1

DEPARTMENT OF DATA SCIENCE 28

[Link] Data Science Main Project Batch 2022 - 2025

return y_pred

# Example usage
predict_activity("test_data.csv")

[Link]
import numpy as np
import pandas as pd
import tensorflow as tf
from [Link] import Sequential
from [Link] import LSTM, Dense, Dropout
from [Link] import LabelEncoder, StandardScaler
from [Link] import accuracy_score, precision_score, recall_score, f1_score
from tabulate import tabulate

# Load training data

train_data = pd.read_csv("driver_activity_final.csv")
X_train = train_data.iloc[:, :-1].values # Select first 12 numerical columns
y_train = train_data.iloc[:, -1].values # Select the last column (activity)

# Encode labels
label_encoder = LabelEncoder()
y_train_encoded = label_encoder.fit_transform(y_train)

# Scale features
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_train_scaled = X_train_scaled.reshape((X_train_scaled.shape[0], X_train_scaled.shape[1], 1)) #
Reshape for LSTM

# Define LSTM model

model = Sequential([
LSTM(64, return_sequences=True, input_shape=(X_train_scaled.shape[1], 1)),
Dropout(0.2),
LSTM(32, return_sequences=False),
Dropout(0.2),
Dense(16, activation='relu'),
Dense(len([Link](y_train_encoded)), activation='softmax')
])

[Link](optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

[Link](X_train_scaled, y_train_encoded, epochs=50, batch_size=16, verbose=1)

# Function to predict activities from uploaded test data

DEPARTMENT OF DATA SCIENCE 29

[Link] Data Science Main Project Batch 2022 - 2025

def predict_activity(test_file):
test_data = pd.read_csv("/content/test_data_120.csv")
X_test = test_data.iloc[:, :-1].values # Select first 12 numerical columns

if test_data.shape[1] == 13: # If activity column exists

y_test = test_data.iloc[:, -1].values
y_test_encoded = label_encoder.transform(y_test)
else:
y_test = None
y_test_encoded = None

X_test_scaled = [Link](X_test)
X_test_scaled = X_test_scaled.reshape((X_test_scaled.shape[0], X_test_scaled.shape[1], 1))
y_pred_encoded = [Link]([Link](X_test_scaled), axis=1)
y_pred = label_encoder.inverse_transform(y_pred_encoded)

# Display predictions in tabular format

result_table = [[i+1, pred] for i, pred in enumerate(y_pred)]
print("\nPredicted Activities:")
print(tabulate(result_table, headers=["Record No.", "Predicted Activity"], tablefmt="grid"))

if y_test is not None:

accuracy = accuracy_score(y_test_encoded, y_pred_encoded)
precision = precision_score(y_test_encoded, y_pred_encoded, average='weighted')
recall = recall_score(y_test_encoded, y_pred_encoded, average='weighted')
f1 = f1_score(y_test_encoded, y_pred_encoded, average='weighted')

evaluation_table = [["Accuracy", accuracy],

["Precision", precision],
["Recall", recall],
["F1 Score", f1]]
print("\nEvaluation Metrics:")
print(tabulate(evaluation_table, headers=["Metric", "Value"], tablefmt="grid"))

return y_pred, accuracy, precision, recall, f1

return y_pred

# Example usage
predict_activity("test_data.csv")

DEPARTMENT OF DATA SCIENCE 30

[Link] Data Science Main Project Batch 2022 - 2025

OUTPUT

[Link] RULE BASED NETWORK

DEPARTMENT OF DATA SCIENCE 31

[Link] Data Science Main Project Batch 2022 - 2025

[Link]

DEPARTMENT OF DATA SCIENCE 32

[Link] Data Science Main Project Batch 2022 - 2025

DEPARTMENT OF DATA SCIENCE 33

[Link] Data Science Main Project Batch 2022 - 2025

[Link]

DEPARTMENT OF DATA SCIENCE 34

[Link] Data Science Main Project Batch 2022 - 2025

[Link]

DEPARTMENT OF DATA SCIENCE 35

[Link] Data Science Main Project Batch 2022 - 2025

[Link]

DEPARTMENT OF DATA SCIENCE 36

[Link] Data Science Main Project Batch 2022 - 2025

DEPARTMENT OF DATA SCIENCE 37

[Link] Data Science Main Project Batch 2022 - 2025

REFERENCES

[1] Deng, Qi, and Dirk Söffker. "A review of HMM-based approaches of driving behaviors
recognition and prediction." IEEE Transactions on Intelligent Vehicles 7, no. 1 (2021): 21-31.

[2] Alkinani, Monagi H., Wazir Zada Khan, and Quratulain Arshad. "Detecting human driver
inattentive and aggressive driving behavior using deep learning: Recent advances, requirements and
open
challenges." Ieee Access 8 (2020): 105008-105030.

[3] Qu, Fangming, Nolan Dang, Borko Furht, and Mehrdad Nojoumian. "Comprehensive study of
driver behavior monitoring systems using computer vision and machine learning techniques."
Journal of Big Data 11, no. 1 (2024): 32.

[4] Xing, Yang, Chen Lv, Huaji Wang, Dongpu Cao, Efstathios Velenis, and Fei-Yue Wang. "Driver
activity recognition for intelligent vehicles: A deep learning approach." IEEE transactions on
Vehicular Technology 68, no. 6 (2019): 5379-5390.

[5] Wan, Liangtian, Yuchen Sun, Lu Sun, Zhaolong Ning, and Joel JPC Rodrigues. "Deep learning
based autonomous vehicle super resolution DOA estimation for safety driving." IEEE Transactions
on Intelligent Transportation Systems 22, no. 7 (2020): 4301-4315.

[6] Tang, Liang, Hengyang Wang, Wenhao Zhang, Zhongyi Mei, and Liang Li. "Driver lane change
intention recognition of intelligent vehicle based on long short-term memory network." IEEE
Access 8
(2020): 136898-136905.

[7] Ponti, Francesca, Fabrizio Frezza, Patrizio Simeoni, and Raffaele Parisi. "A Generalized
Learning Approach to Deep Neural Networks." Journal of Telecommunications and Information
Technology (2024).

[8] Chong, Linsen, Montasir M. Abbas, Alejandra Medina Flintsch, and Bryan Higgs. "A rule-
based neural network approach to model driver naturalistic behavior in traffic." Transportation
Research Part
C: Emerging Technologies 32 (2013): 207-223.

[9] Jatla, Srikanth, and Sowmya Mandadi. "A MACHINE LEARNING-DRIVEN DECISION
STRATEGY FOR AUTONOMOUS VEHICLE NAVIGATION." Industrial Engineering Journal 53,
no. 4 (2024): 155.

DEPARTMENT OF DATA SCIENCE 38

[Link] Data Science Main Project Batch 2022 - 2025

[10] Chen, Shanzhi, Xinghua Hu, Jiahao Zhao, Ran Wang, and Min Qiao. "A review of decision-
making and planning for autonomous vehicles in intersection environments." World Electric Vehicle
Journal 15, no. 3 (2024): 99.

[11] Deng, Qi, and Dirk Söffker. "Improved driving behaviors prediction based on fuzzy logic-
hidden markov model (fl-hmm)." In 2018 IEEE Intelligent Vehicles Symposium (IV), pp. 2003-
2008. IEEE,2018.

[12] You, Changxi, Jianbo Lu, and Panagiotis Tsiotras. "Nonlinear driver parameter estimation and
driver steering behavior analysis for ADAS using field test data." IEEE Transactions on Human-
Machine
Systems 47, no. 5 (2017): 686-699.

DEPARTMENT OF DATA SCIENCE 39