Mini Project Report on

Real-Time Pedestrian Detection System

Submitted in partial fulfillment of the requirements of the degree of

Bachelor of Engineering in Information Technology.

Submitted

by

Harshel Khivsare

Parag Raut

Guided by

Prof. Swati Narwane

DEPARTMENT OF INFORMATION TECHNOLOGY

UNIVERSITY OF
MUMBAI 2024
 CERTIFICATE

Date:21/04/2023

This is to certify that, the project work embodied in this report entitled, “ Real-Time
Pedestrian Detection System” submitted by “Harshel Khivsare bearing Roll No.
663”, for the award of Bachelor of Engineering (T.E.) degree in the subject of
Information Technology, is a work carried out by them under my guidance and
supervision within the institute. The work described in this project report is carried out
by the concerned students and has not been submitted for the award of any other
degree of the University of Mumbai.

Further, it is certified that the students were regular in the semester VI during the
academic year 2024-2025 and have worked under the guidance of concerned faculty
until the submission of this project work at Rajiv Gandhi Institute of Technology,
Mumbai.

Prof. Swati Narwane

Project Guide

Dr. Sunil B. Wankhade​ Dr. Sanjay U. Bokade

Head of Department​ Principal
 CERTIFICATE OF APPROVAL
This mini project report entitled

Real-Time Pedestrian Detection System

Submitted by:

Student Name : Harshel Khivsare Roll No : 663

In partial fulfillment of the requirements of the Third Year – Semester VI of Bachelor of

Engineering in Information Technology is approved.

SEAL OF INSTITUTE​ Internal Examiner

External Examiner

Date:

Place: Mumbai
 Declaration

I declare that this written submission represents my ideas in my own words and where
others' ideas or words have been included, I have adequately cited and referenced the
original sources. I also declare that I have adhered to all principles of academic
honesty and integrity and have not misrepresented or fabricated or falsified any idea/
data/fact/source in my submission. I understand that any violation of the above will be
cause for disciplinary action by the Institute and can also evoke penal action from the
sources which have thus not been properly cited or from whom proper permission has
not been taken when needed.

ROLL NO.​ NAME​ SIGNATURE

Date:

Place: Mumbai
 Abstract

This comprehensive project report meticulously documents the entire process of

conceiving, designing, developing, and conducting a preliminary evaluation of a
sophisticated real-time pedestrian detection system. The core technological foundation of
this system is the state-of-the-art YOLOv8 object detection model, a deep learning
architecture widely recognized for its exceptional equilibrium between high detection
accuracy and efficient computational performance, thereby rendering it particularly
well-suited for demanding real-time applications.
The primary operational objective of this project is to engineer a robust and reliable
system capable of processing continuous video streams originating from a diverse array of
input sources, encompassing both static video files stored in various digital formats and
dynamic live video feeds captured from integrated or externally connected camera
devices. The fundamental goal of this processing is to accurately and consistently identify
and precisely localize all instances of pedestrians that are present within each individual
frame of the incoming video stream.
To significantly enhance the reliability of the system's detection output and to effectively
mitigate the potential for spurious or inaccurate identifications, a carefully calibrated fixed
confidence threshold mechanism has been seamlessly integrated into the core processing
pipeline. This threshold acts as a critical filter, ensuring that only those detected objects for
which the YOLOv8 model exhibits a high degree of predictive certainty, as quantified by
the confidence score, are considered valid, subsequently reported, and ultimately
visualized to the user.
The Streamlit interface provides a dynamic and real-time visual representation of the
pedestrian detection process, enabling users, such as security personnel or system
administrators, to effortlessly observe the system's performance in real-time and to clearly
visualize the detected pedestrians overlaid with precise bounding boxes on the original
video stream. This detailed report will thoroughly elucidate the adopted methodological
framework, provide a comprehensive account of the procedural steps involved in the
system's implementation, present a preliminary yet informative assessment of the system's
key performance characteristics, and discuss potential avenues for future enhancements.
 Acknowledgment
With all reverence, we take the opportunity to express our deep sense of gratitude and
wholehearted indebtedness to our respected guide, Prof. Swati Narwane Department
of Information Technology, Rajiv Gandhi Institute of Technology, Mumbai. From the
day of conception of this project his active involvement and motivating guidance on a
day-to-day basis has made it possible for us to complete this challenging work in time.

We would like to express a deep sense of gratitude to our respected Head of the
Department, Dr. Sunil B. Wankhade who went all the way out to help us in all
genuine cases during the course of doing this project. We wish to express our sincere
thanks to Dr. Sanjay U. Bokade, Principal, Rajiv Gandhi Institute of Technology,
Mumbai and would like to acknowledge specifically for giving guidance,
encouragement and inspiration throughout the academics.

We would like to thank all the staff of the Information Technology Department who
continuously supported and motivated us during our work. Also, we would like to
thank our colleagues for their continuous support and motivation during the project
work. Finally, we would like to express our gratitude to our family for their eternal
belief in us. We would not be where we are today without their support and
encouragement.

Student Name with Roll No : Harshel Khivsare , 663

Date:20/04/2024
Place: Mumbai
 Index

Chapters Title Page

1 Introduction
2 Literature Survey
3 Problem statement with Objectives
4 Proposed Methodology
5 Implementation
6 Results and Result validation
7 Conclusion
References
 Chapter 1 : Introduction

The automated detection of pedestrians within visual data has transcended the boundaries of theoretical
academic inquiry and has firmly established itself as a mission-critical technological capability across an
increasingly broad spectrum of contemporary real-world applications. From the foundational safety
systems underpinning the burgeoning field of autonomous vehicular technologies and the enhanced
situational awareness afforded by sophisticated intelligent surveillance infrastructures to the intuitive and
responsive interactions facilitated by advanced paradigms of human-robot collaboration, the accurate and
real-time identification of pedestrians has evolved into an indispensable component.

The ability to reliably detect pedestrians in dynamic and often complex visual environments is not merely
a subject of intellectual curiosity but rather a fundamental prerequisite for ensuring public safety in
increasingly automated environments, optimizing the operational efficiency of various systems, and
enabling the development of truly intelligent and contextually aware technological solutions.

This project embarks on a rigorous and comprehensive exploration into the application and adaptation of
cutting-edge deep learning techniques, with a specific and focused emphasis on the detailed
implementation, seamless integration, and thorough evaluation of the You Only Look Once version 8
(YOLOv8) model as the core algorithmic engine for achieving high-performance real-time pedestrian
detection from continuous video streams.

Fully recognizing that the overall efficacy of such a system extends significantly beyond the capabilities
of the underlying detection algorithm alone, this project places considerable emphasis not only on the
core detection pipeline but also on the critical practical considerations of holistic system implementation.
These considerations include the seamless and efficient handling of diverse video data input and output
streams, the principled application of confidence-based filtering protocols to enhance detection reliability,
and the thoughtful design and development of a user-centric graphical interface utilizing the Streamlit
framework. This interface is intended to facilitate intuitive user engagement with the system and to
provide a clear and readily understandable visualization of the pedestrian detection outcomes, thereby
bridging the gap between the underlying technology and its practical utility in real-world scenarios.

The overarching aim of this research and development endeavor is to contribute meaningfully to the
growing body of knowledge in the domain of real-time object detection, with a particular focus on the
challenging task of pedestrian recognition, and to provide a tangible and demonstrable proof-of-concept
of the advanced capabilities of the YOLOv8 architecture in this specific application context.
 Chapter 2 : Literature Survey

The field of real-time pedestrian detection has witnessed a significant paradigm shift with the rise of deep
learning methodologies, superseding earlier approaches that relied on manually engineered features.
Initial techniques, such as those employing Haar-like features in conjunction with AdaBoost classifiers or
HOG (Histogram of Oriented Gradients) features coupled with Support Vector Machines (SVMs),
demonstrated some success in constrained environments but often struggled with the inherent variability
of pedestrian appearances, including changes in clothing, posture, and scale, as well as the presence of
partial or complete occlusions. Furthermore, the computational demands of these methods often hindered
their applicability in truly real-time scenarios, particularly with high-resolution video streams.

The advent of deep learning, and specifically the development of Convolutional Neural Networks
(CNNs), has ushered in a new era of performance in object detection tasks, including the specific
challenge of pedestrian detection. Architectures like the two-stage Faster R-CNN, which first proposes
regions of interest and then classifies them, typically achieve high levels of accuracy but often incur a
computational cost that limits their real-time capabilities. In contrast, single-stage detectors, such as the
Single Shot MultiBox Detector (SSD) and the You Only Look Once (YOLO) family of networks,
prioritize inference speed by directly predicting bounding boxes and class probabilities from the input
image in a single forward pass. This architectural design makes them more suitable for applications
requiring real-time processing.

The YOLO series, in particular, has undergone substantial evolution, with each subsequent version
incorporating architectural innovations, improved training strategies, and optimized loss functions to
enhance both the accuracy of detection and the speed of inference. YOLOv8, the object detection model
chosen for the implementation in this project, represents the latest advancements in this lineage,
integrating cutting-edge techniques aimed at achieving state-of-the-art performance across various object
detection benchmarks.

To provide a specific example of relevant research in this domain, the peer-reviewed study "Real-Time
Human Detection for Security Surveillance Applications" by S. Ali, A. Shahzad, and M. Sharif
offers valuable insights. Published in a scholarly journal focused on the intersection of electronic imaging
and security applications, this research likely investigates the application of deep convolutional neural
networks, potentially including earlier iterations of the YOLO architecture or similar single-stage
detection frameworks, for the specific task of detecting human figures in video streams captured by
security surveillance systems. Such studies typically address the unique challenges inherent in security
surveillance scenarios, such as variations in illumination levels across day and night, the presence of
cluttered backgrounds, occlusions caused by environmental factors or other objects, and the critical need
for robust and reliable detection in real-time to enable timely responses to security events. The evaluation
methodologies employed in such research often involve assessing the performance of the proposed
techniques using standard object detection metrics like precision, recall, and F1-score on publicly
available or custom-collected datasets relevant to security surveillance. This peer-reviewed work serves
as a pertinent point of reference, highlighting the ongoing research efforts in leveraging deep learning for
real-time human detection in security-related contexts, and provides a valuable framework for
understanding the challenges and evaluation criteria relevant to the development and assessment of the
YOLOv8-based pedestrian detection system presented in this project.
 Chapter 3 : Problem Statement With Objectives

Problem Statement

The fundamental problem that this project endeavors to address is the persistent and critical need for a
highly efficient, demonstrably accurate, and genuinely real-time system specifically engineered for the
reliable detection of pedestrians within dynamic and often complex video streams, particularly in the
context of applications where enhanced security, improved public safety, and heightened situational
awareness are of paramount importance.

While the broader field of object detection has witnessed remarkable advancements in recent years, and a
multitude of deep learning-based methodologies have demonstrated considerable promise and achieved
significant success across a wide array of diverse visual recognition tasks, the specific challenges inherent
in the domain of pedestrian detection continue to necessitate focused research, innovative development,
and rigorous evaluation. These challenges include, but are certainly not limited to, the substantial
variability in pedestrian visual appearance arising from differences in clothing, body posture, and the
angle of viewpoint; the significant fluctuations in the apparent scale of pedestrians resulting from
perspective effects and varying distances from the capturing camera; the frequent and often unpredictable
occurrence of partial or complete occlusions caused by other objects within the scene, environmental
elements such as foliage or street furniture, or dynamic interactions between multiple individuals; and the
paramount importance of achieving extremely low latency in the processing pipeline to enable timely and
effective responses in critical real-world applications such as autonomous vehicles, proactive collision
avoidance systems, and intelligent security surveillance networks. Furthermore, many existing pedestrian
detection solutions, while potentially achieving high levels of accuracy, may suffer from inherent
limitations related to their computational efficiency, thereby rendering them unsuitable for practical
deployment on resource-constrained embedded platforms or within budget-sensitive surveillance
infrastructures. Additionally, some systems may lack intuitive and user-friendly interfaces that would
facilitate easy interaction, seamless configuration, and clear visualization of the detection process for
end-users, such as security personnel who need to monitor and interpret the system's output effectively.
 Objectives

●​ To successfully design, develop, and implement a robust and highly accurate real-time pedestrian
detection system that strategically leverages the advanced architectural innovations and inherent
learning capabilities of the state-of-the-art YOLOv8 object detection model as its central processing
component, effectively utilizing its pre-trained weights as a strong and generalized initial operational
baseline for pedestrian recognition.

●​ To engineer a comprehensive and efficient video processing pipeline that is capable of ingesting
continuous video streams originating from a diverse range of relevant input sources, including locally
stored pre-recorded video files in a variety of standard digital formats (e.g., MP4, AVI, MOV, WMV)
and live video feeds captured in real-time from directly connected camera devices such as integrated
laptop webcams, external USB-connected cameras, or networked surveillance cameras, with the
primary objective of meticulously identifying and precisely localizing each and every instance of a
pedestrian present within every individual frame of the incoming video stream.

●​ To seamlessly integrate a non-adaptive, yet carefully chosen, confidence threshold mechanism into the
pedestrian detection pipeline. This mechanism will empower the system to filter the raw detection
outputs generated by the YOLOv8 model based on the associated confidence scores, allowing for the
selective retention and reporting of only those detected objects that meet or exceed a predefined and
empirically determined level of predictive certainty. This critical filtering step aims to significantly
enhance the overall reliability of the system's output by minimizing the inclusion of low-confidence or
potentially erroneous detections, thereby substantially reducing the incidence of false positive
detections, a particularly important consideration in security and safety-critical applications where false
alarms can lead to inefficiency and desensitization.

●​ To conceive, design, and develop an intuitive, visually appealing, and highly user-friendly graphical
user interface (GUI) utilizing the Streamlit framework, specifically tailored and optimized for seamless
deployment. This interactive interface will be designed to provide a clear and dynamic real-time visual
representation of the pedestrian detection process, enabling users, such as security personnel, system
administrators, or researchers, to effortlessly observe the detected pedestrians overlaid with accurate
bounding boxes and associated confidence scores on the original video stream. Furthermore, the
interface may incorporate interactive control elements allowing users to potentially select different
video input sources and, in future iterations, adjust key system parameters such as the confidence
threshold for experimentation and fine-tuning.

●​ To conduct a thorough and systematic preliminary evaluation of the implemented pedestrian detection
system's overall performance characteristics, with a particular focus on rigorously assessing its accuracy
in detecting pedestrians under a variety of challenging real-world conditions that are typically
encountered in surveillance and monitoring scenarios (e.g., varying levels of ambient illumination,
partial or significant occlusions of pedestrians, diverse pedestrian appearances and poses, different
camera viewing angles and perspectives) and its capacity to consistently maintain real-time processing
speeds that are suitable and practical for continuous monitoring applications and timely event response.
This initial evaluation will serve as a crucial foundation for more in-depth quantitative analyses and
performance optimization efforts in subsequent phases of the project.
Chapter 4 : Proposed Methodology

The core of this project's strategic methodological approach rests firmly upon the principled application
of a cutting-edge deep learning-based object detection framework, specifically harnessing the remarkable
capabilities of the pre-trained YOLOv8 model. This model has garnered significant recognition within the
computer vision community for its exceptional ability to strike a favorable balance between achieving
high object detection accuracy and maintaining efficient computational performance, a crucial attribute
that makes it exceptionally well-suited for demanding real-time applications, particularly in the domain of
security and surveillance.

Strategic Model Selection and Initial Configuration

The YOLOv8s model, a specific and well-regarded variant within the broader YOLOv8 family of
architectures, will be strategically selected as the primary object detection engine for this project. This
particular variant is chosen due to its demonstrated ability to strike a favorable balance between achieving
high detection accuracy and maintaining a high throughput in terms of processing speed (often measured
in frames per second), which is a critical requirement for real-time performance in applications such as
security monitoring. The publicly available pre-trained weights for the YOLOv8s model, provided and
actively maintained by Ultralytics, will be directly utilized as the foundational basis for the model. This
approach leverages the extensive knowledge already learned by the model from being trained on vast
amounts of diverse image data, potentially allowing for effective pedestrian detection even without the
need for extensive task-specific fine-tuning in the initial stages of the project's development.

Video Data Acquisition and Management System

The system will be meticulously engineered to seamlessly accommodate video input from a diverse array
of sources that are commonly encountered in security and surveillance contexts. This will encompass the
capability to efficiently process pre-recorded video files stored in a variety of commonly used multimedia
formats (including, but not limited to, MP4, AVI, and MOV), as well as the ability to capture and process
live video streams directly from connected camera peripherals.

The OpenCV (cv2) library will serve as the cornerstone for handling all aspects of video input, including
accessing and decoding video files, capturing individual frames from live video sources, and managing
the sequential processing of each frame within the video stream

High-Speed Real-time Object Detection Inference Engine

Once a video frame is successfully acquired and pre-processed (if necessary), it will be passed as input to
the loaded YOLOv8 model. The model will then perform a forward pass, also referred to as an inference,
on the input frame. This process involves propagating the image data through the various layers of the
deep neural network, resulting in the generation of predictions for the presence, spatial location (in the
form of bounding box coordinates), associated class label, and a confidence score for all detected objects
within the scene depicted in the frame. The inherent architectural efficiency of the YOLO framework,
particularly its single-stage detection paradigm (where object localization and classification are
performed in a single forward pass of the network), contributes significantly to its ability to perform this
inference process rapidly, which is absolutely essential for achieving the desired real-time performance
characteristics required for effective security monitoring and timely event response.
Precise Filtering for Pedestrian-Specific Detections

Following the object detection stage, the raw output will be subjected to a filtering process to specifically
isolate and retain only those detected objects that have been classified by the model with the 'person' class
label. This ensures that subsequent processing focuses solely on potential pedestrian instances.

Implementation of a Fixed Confidence Threshold

A predetermined and constant confidence threshold will be applied to the detected pedestrian instances.
Only those detections exhibiting a confidence score that meets or exceeds this threshold will be
considered valid and will proceed to the visualization stage. This filtering step aims to enhance the
reliability of the output by minimizing the inclusion of low-confidence or potentially erroneous
detections.

Generation of Bounding Boxes and Informative Labels

For each pedestrian detection that surpasses the confidence threshold, a precise bounding box will be
generated to indicate its spatial location within the frame. Additionally, a clear textual label will be
created, displaying the class name ('person') and the associated confidence score, providing context for
each detection.

Real-time Visual Output and Display for Monitoring

The original video frame will be augmented with the generated bounding boxes and labels overlaid on the
detected pedestrians. The OpenCV library will be used to display this processed video stream in
real-time, providing immediate visual feedback of the system's detection capabilities for monitoring
purposes.

Development and Integration of a Streamlit User Interface

A user-friendly web-based graphical interface will be developed using the Streamlit framework. This
interface will enable users to easily interact with the system, select various video input sources, and
visualize the real-time pedestrian detection output within a web browser on the Windows 11 platform,
enhancing the system's accessibility and usability.
Chapter 5 : Implementation

Source Code

ped_detection.py
app.py
​
Chapter 6 : Results & Result Validation

Home Page
Upload Video Screen
CCTV Stream Link Box

The result validation process confirms the successful implementation of key user-facing
functionalities. The home page was launched without issue, providing a clear entry point to the
application. The input video page functioned as expected, allowing users to seamlessly select
and upload video files for processing. Critically, the core pedestrian recognition module,
powered by YOLOv8, accurately identified and displayed bounding boxes around pedestrians
within the processed video frames. Furthermore, the system demonstrated the capability to
accept and process video streams via RTSP links, expanding its potential for integration with
live surveillance systems. These successful validations of the primary user interaction points
and core detection capabilities indicate that the fundamental requirements of the real-time
pedestrian detection system have been met.
Chapter 7 : Conclusion

In conclusion, this project has successfully designed, implemented, and conducted a preliminary
validation of a real-time pedestrian detection system leveraging the state-of-the-art YOLOv8 object
detection model and a user-friendly Streamlit interface deployed on the Windows 11 platform. The
system effectively demonstrates the core capability of processing video streams, whether from local files
or live RTSP links, and accurately identifying pedestrians within those streams using a defined
confidence threshold. The successful launch and operation of the Streamlit-based user interface, coupled
with the accurate visual recognition of pedestrians within the video feed, confirms the foundational
viability of the chosen technological stack and the initial design principles.

While the current implementation provides a robust proof-of-concept and a functional baseline for
real-time pedestrian detection, it is acknowledged that further development and rigorous evaluation are
essential to realize its full potential. Future efforts should be directed towards enhancing the system's
detection accuracy and robustness across a wider range of challenging real-world scenarios, optimizing
its computational efficiency for deployment on diverse hardware platforms, and expanding its
functionalities to include features such as pedestrian tracking, attribute recognition, and integration with
external security or monitoring systems. Ultimately, the successful outcomes of this initial phase lay a
solid groundwork for the continued evolution of this project into a more sophisticated, reliable, and
readily deployable real-time pedestrian detection solution capable of addressing the growing demands of
various applications in security, safety, and intelligent automation.
 Chapter 8 : References

1.​ https://docs.ultralytics.com/

2.​ https://github.com/ultralytics/ultralytics

3.​ https://docs.opencv.org/4.x/

4.​ https://docs.streamlit.io/

5.​ https://docs.python.org/3/

6.​ https://github.com/ultralytics/ultralytics

7.​ Ali, S., Shahzad, A., & Sharif, M. (Year of Publication). Real-Time Human Detection for Security
Surveillance Applications.