ODSC - Batch to Stream workshop - integration of Apache Spark, Cassandra, Pos...
Major PRC-1 ppt.pptx
1. OBJECT DETECTION IN SELF DRIVING
CARS USING DEEP LEARNING AND IOT
UNDER THE GUIDANCE OF
Dr.D.V.Lalitha Parameswari
Associate Professor
Department Of CSE
BY
20251A6611-K.Shravani
20251A6638-D.Jagruthi
20251A6653-S.Sai Richa
BATCH- M10
2. TABLE OF CONTENTS
• Introduction
• Literature survey
• Drawbacks
• Problem statement
• Objectives
• Proposed system
• conclusion
3. INTRODUCTION
•Every year, approximately 1.5 lakh people dies on India roads, which
translate, on an average, into 1130 accidents and 422 deaths every day or 47
accidents and 18 deaths every hour.
•In the face of this alarming reality, there is an increasing call for
transformative technologies to revolutionise road safety.
•Through enhanced object detection capabilities, autonomous vehicles
empowered by deep learning and IOT can bring about a paradigm shift in
accident prevention and road safety.
4. S.No Title Year Methods used
Performance
Metrics
1.
Object Detection System in Self-
Driving Cars by Calibrated Camera and
LiDAR Sensor data-Implementation of
the Hybrid PointPillar Method
2023
Pillar Feature Encoder processes pillar data, extracting
high-level features. Region Proposal Network (RPN)
identifies regions, generating anchor boxes and
proposals. Non-Maximum Suppression (NMS) selects
top-scoring proposals, removing redundant ones by
confidence scores.
The system achieved an average precision
of 75.8% on the KITTI validation dataset.
The mean average precision reached
70.8%, indicating the overall accuracy of
object detection across multiple classes
(Cars, Pedestrians, Cyclists).
2.
An Improved Deep Network-Based
Scene Classification Method for Self-
Driving Cars
2022
This dataset contains five categories: crosswalk, gas
station, parking lot, highway, and street, with 15,000
images per category from the KITTI and Place365
datasets. The proposed deep network uses an enhanced
Faster RCNN for local feature extraction and an
Inception_V1 network for global features, creating the
scene classification architecture. The network is trained
and evaluated on the provided dataset.
The accuracy of the deep network on the
validation set reaches 95.04% after 19,000
training iterations.
LITERATURE SURVEY
5. S.No Title Year Methods used
Performance
Metrics
3.
An Improved Deep Learning Solution
for Object Detection in Self-Driving
Cars
2022
The method in this utilizes a deep neural network
architecture, leveraging both a base network (backbone)
and a pyramid feature structure for feature extraction. It
then employs dense bounding box generation and
classification scoring to detect objects in images,
followed by non-maximum suppression for producing
final results.
With input size 320x320 the average
precision is 58.4 and with input size
512x512 average precision is 63.8.
3.
Object Detection in Self Driving Cars
Using Deep Learning 2021
The object detection process involves three key steps:
loading a pre-trained object detection network using
OpenCV's DNN module, inputting data to the network,
performing a forward pass to get detections, and
iterating through the results to identify objects, their
labels, and confidence scores. The chosen model for
object detection is MobileNet SSD in the OpenCV
library.
The system achieved a mean average
precision of up to 72.8% on the PASCAL
VOC dataset.
6. Existing System
• Traditional computer vision techniques, such as Haar cascades (HOG), rely on handcrafted
features and rule-based methods for object detection, lacking the ability to learn from data
like modern deep learning approaches.
• R-CNN (Region-based Convolutional Neural Network) is a conventional object detection
system used in self-driving cars that divides the task into multiple stages, including region
proposal generation, feature extraction, and object classification.
• Single Shot MultiBox Detector (SSD) system combines the benefits of high accuracy, real-
time performance, and versatility by utilizing a set of default bounding boxes at multiple
scales and aspect ratios to predict object classes and refine bounding box locations
simultaneously.
7. Drawbacks of Existing System
• R-CNN, which relies on multi-stage processing and region proposals, an advantage of the
alternative algorithm lies in its real-time processing capability. Processing the entire image in a
single pass significantly enhances its suitability for time-sensitive applications such as self-
driving cars, where rapid decision-making is imperative for safety and navigation.
• SSD cannot accurately detect small or closely spaced objects.
• Haar cascades have a drawback compared to some modern methods in their limited
adaptability to complex and changing scenarios. Haar cascades rely on handcrafted features
and rule-based approaches, which can hinder their performance in dynamically evolving self-
driving environments, unlike more flexible deep learning techniques like YOLO.
8. Problem Statement
This project aims to create an Self driving car system that can
safely and intelligently navigate through its surroundings, obey
traffic signals, and avoid collisions. This involves the integration
of hardware components, the development of sophisticated
software for object detection and rigorous testing to ensure the
system's reliability and accuracy.
9. OBJECTIVES
The objective of this project are:
1. To integrate camera module and ultrasonic sensor with raspberry pi to
collect real time input data.
2. To establish a wireless data communication system between the Raspberry Pi
and a computer to transmit sensor data and receive control instructions.
3. To implement object detection algorithm to recognize and locate stop signs,
traffic signals and obstacles on road.
4. To create a control system that sends steering instructions to the Arduino to
precise RC car control.
10. PROPOSED SYSTEM
The goal of this project is to make a self-driving RC car that integrates a Raspberry Pi,
Arduino, camera module, and ultrasonic sensor. Using YOLO (You Only Look Once) as
the object detection algorithm, the system collects real-time data from the camera and
sensor, wirelessly transmitting it to a computer for processing. The computer runs a neural
network model to generate steering predictions based on the YOLO-detected objects,
enabling precise control of the RC car. The system also includes collision avoidance
algorithms, ensuring the car's safety while navigating autonomously and adhering to traffic
rules.