Please use this identifier to cite or link to this item:
Title: Traffic Light Detection using Deep Learning based Feature Fusion
Authors: Yasser Khalil 
Supervisor: Prof. Imtiaz Ahmad
Keywords: Traffic Light : Fusion
Issue Date: 2018
Publisher:  Kuwait university - college of graduate studies
Abstract: Traffic lights are used worldwide for regulating traffic and pedestrian flow on streets. Even with the existence of traffic lights, drivers often miss spotting them because of their positioning and occlusion, and color codes used. Accurate detection of traffic lights is a must requirement for smooth and error-free performance of intelligent systems such as autonomous vehicles, and advanced driver assistance systems. Recently, deep learning has revolutionized intelligent image processing for object detection tasks. The success of deep learning techniques lies in robust feature extraction for the task from a sequence of convolution and sampling operations through a neural network training. In general, these features are widely referred to as deep features. In the traditional machine learning, handcrafted features designed based on the domain knowledge have shown excellent performance in many problems. In this thesis, we present a deep learning based object detection framework by the fusion of deep features with handcrafted features learned in a Convolutional Neural Network (CNN) training. Our object detection framework is based on the state-of-the-art You Only Look Once (YOLO) architecture, where we learn the fusion of deep features for detection with the Integral Channel Features (ICF) while training the YOLO detector. ICF features are defined as multidimensional concatenation of different color channels including RGB and LUV, and edge information encoded as gradient histograms. We demonstrate the application of proposed detection framework based on feature fusion for traffic light detection task where experimental evaluation have been presented on Bosch Small Traffic Lights dataset. Our solution improves the state-of-the-art for this dataset by 4.8% increase in mean average precision (mAP) measure. We also demonstrate the efficiency of the proposed detection framework for common object task where our evaluation on COCO dataset have yielded 0.8% increase in the mAP measure in comparison with our set benchmark. The proposed solution has been implemented on GeForce GTX 1080 GPU desktop. In addition with the implementation details, we have also discussed the challenges and problems that we encountered.
Appears in Programs:0612 Computer Engineering

Files in This Item:
File Description SizeFormat 
Traffic Light Detection using Deep Learning based Feature Fusion.pdf24,73 MBAdobe PDFView/Open    Request a copy
Show full item record

Page view(s)

Last Week
Last month
checked on Nov 17, 2019

Download(s) 50

checked on Nov 17, 2019

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.