Full
Full
PROJECT SYNOPSIS
ON
“Using deep learning for Real-time object detection”
Submitted in partial fulfilment of the requirement of the Seventh Semester
Bachelor of Engineering In
INFORMATION SCIENCE AND ENGINEERING
Submitted By
AIM 2
OBJECTIVES 2
2 ABSTRACT 3
3 INTRODUCTION 4
4 EXISTING SYSTEM 5
5 PROPOSED SYSTEM 6
6 REFERENCES 8
LIST OF FIGURES
1
AIM & OBJECTIVES
AIM:
The aim of the proposed system is to develop an Android application equipped with robust
real-time object detection capabilities, leveraging the SSD (Single Shot Multibox Detector)
method and OpenCV. The primary goal is to seamlessly integrate OpenCV into the Android
app, utilizing its powerful computer vision functionalities. The system will capture frames in
real-time from the device's camera using either the Android Camera API or CameraX, ensuring
adaptability across different Android devices. A key component of the system is the Object
Detection Module, which relies on OpenCV functions to perform inference on camera frames,
drawing bounding boxes around detected objects.
OBJECTIVES:
• Develop a real-time object detection system.
• Utilize state-of-the-art deep learning techniques for the implementation.
• Emphasize achieving high accuracy in object detection.
• Prioritize efficiency in the detection and localization of objects within a video stream.
• Extend the system's applicability to sequences of images for comprehensive object
tracking.
• Explore and integrate cutting-edge technologies in deep learning for improved
performance.
• Assess and compare the proposed system against existing object detection methods.
• Optimize the system for real-time processing, ensuring minimal latency in object
detection.
• Investigate techniques to enhance the robustness of object detection across various
environmental conditions.
• Provide a comprehensive evaluation of the developed system's accuracy and efficiency in
real-world scenarios.
2
ABSTRACT
In the realm of technological advancements, the burgeoning field of object detection, integral
to video and image analysis, has become a focal point for researchers. The early stages of object
recognition primarily relied on hand-crafted features, imprecise architectures, and trainable
algorithms. A notable drawback in many existing object detection systems is their dependence
on auxiliary computer vision methods to bolster their deep learning-based methodologies,
resulting in sluggish and suboptimal performance.
This Project introduces a comprehensive end-to-end solution to the object detection challenge,
employing a deep learning-based methodology. A key innovation lies in the adoption of the
Single Shot Detector (SSD) technique, recognized for its unparalleled speed in object detection
from images, achieved through a singular layer of a convolutional network. Unlike
conventional systems, SSD eliminates reliance on additional computer vision methods,
streamlining the process and significantly enhancing overall performance.
The primary research objective is to elevate the accuracy of the SSD method. By presenting a
novel approach that mitigates the shortcomings of previous techniques, this work contributes
to the advancement of real-time object detection. The proposed solution not only addresses the
need for swiftness in detection but also emphasizes precision, offering potential applications in
diverse fields where rapid and accurate object detection is imperative.
3
INTRODUCTION
In the rapidly evolving landscape of computer vision, object detection emerges as a pivotal task,
streamlining the identification and localization of multiple objects within an image or video
frame. The integration of deep learning techniques has notably propelled this process, offering
unmatched accuracy and efficiency. A standout method in this realm is the Single Shot
Multibox Detector (SSD), strategically developed to meet the demand for real-time object
detection. SSD harnesses the combined strengths of convolutional neural networks (CNNs)
and multibox object detection, enabling the swift and precise identification of objects of
interest.
This introduction delves into the intricate intersection of object detection, the SSD method, and
the application of deep learning techniques. It underscores the significance of OpenCV, a
versatile computer vision library, in implementing and optimizing these cutting-edge methods
for real-world applications.
4
EXISTING SYSTEM
Paul Viola and Michael Jones pioneered object detection with an effective facial detection
algorithm, demonstrating real-time face detection on a webcam. Traditional methods involve
inspecting the entire image, feature extraction, and classification, using steps like generating
candidate boxes with sliding windows, feature extraction with methods like Haar-Like features
and HOG, and classification with algorithms such as SVM and Adaboost. The framework's
effectiveness relies on the feature set, classifier, learning method, and training set.
Despite traditional methods' shortcomings, detectors using sliding windows prevailed until
2012, with marginal improvements. The reliance on manual feature design and model
understanding prompted a shift towards computer- driven learning.
This shift intensified after the 2010 ImageNet competition, where the convolutional neural
network (CNN) AlexNet significantly reduced the error rate. This marked a critical opening for
CNNs, leading to their dominance and a drastic reduction in the ImageNet competition's
classification error rate since 2012.
5
PROPOSED SYSTEM
The proposed system envisions an Android application with real-time object detection
capabilities using the SSD (Single Shot Multibox Detector) method and OpenCV. The system
integrates OpenCV into the Android app to harness its computer vision functionalities and
seamlessly captures frames from the device's camera using the Android Camera API or
CameraX. A pre-trained SSD model, optimized for real-time object detection, is employed and
converted to a format compatible with OpenCV. The heart of the system lies in the Object
Detection Module, a dedicated component leveraging OpenCV functions to perform inference
on camera frames, drawing bounding boxes around detected objects. The user interface is
designed to be intuitive, featuring a live camera feed and a section displaying real-time object
detection results. Users have the flexibility to customize detection parameters, such as
confidence thresholds and input resolutions. The system incorporates optimization strategies,
including model quantization and multi-threading, to ensure efficient performance across a
range of Android devices. Thorough testing and validation are essential, encompassing
different devices, environmental conditions, and error scenarios. The system also includes
comprehensive documentation for future development and maintenance, ensuring clarity on
codebase structure and model update procedures.
6
Fig 3. Detection of human from background subtraction
7
REFERENCES
1. Subhani Shaik, Ida Fann. Performance indicator using machine learning techniques,
Dickensian Journal. 2022;22(6).
2. Vijaya Kumar Reddy R, Subhani Shaik B, Srinivasa Rao. Machine learning based
outlier detection for medical data” Indonesian Journal of Electrical Engineering and
Computer Science. 2021;24(1)
3. Towards Data Science. Available: https://towardsdatascience.com/ssd-single- shot-
detector-for-object detection-using- multibox1818603644ca?gi=f02e06e2d636
4. Subhani Shaik, Ganesh. Taming an autonomous surface vehicle for path following and
collision avoidance using deep reinforcement learning, Dickensian Journal.
2022;22(6).
5. Vijaya Kumar Reddy R, Shaik Subhani, Rajesh Chandra G, Srinivasa Rao B. Breast
Cancer Prediction using Classification Techniques, International Journal of Emerging
Trends in Engineering Research. 2020;8(9).