Author: Wang, Liwen
Title: Video object detection/tracking with learning approach
Advisors: Siu, Wan-chi (EIE)
Degree: M.Sc.
Year: 2018
Subject: Hong Kong Polytechnic University -- Dissertations
Digital video
Pattern recognition systems
Image processing -- Digital techniques
Department: Department of Electronic and Information Engineering
Pages: xv, 121 pages : color illustrations
Language: English
Abstract: A Close-up Monitoring System (CMS) has been designed in the group with a consistently project, which aims at avoiding any potential collision risk by detecting the frontal train's distance from the captured video. To achieve the goal, several modules were designed to support the whole system, like railway detection module, train detection module, distance estimation module and so on. This thesis firstly gives a brief introduction to the whole structure of CMS. Then it gives focus on introducing train recognition process that is the key part in CMS system. This recognition process adopts several orthogonal or semi-orthogonal features and shows an outstanding performance in various conditions. However, in some tough cases, there were still a small number of false detections. Thanks to the feedback of CMS system, those false detections could be solved through the collected data. It is well known to us all that video object detection/tracking is a very fundamental task in computer version filed, and machine learning is a powerful tool that can handle tough conditions at the expense of larger computational power. There are many new videos captured every day, the traditional manual data analysis approach cannot make full use of these data. Under the circumstance, the machine learning approach shows great advantages that it can automatically process the obtained data by a suitable framework. In other words, the object detector with learning approach can improve its performance gradually by using the increasing video data. In this thesis, we try to combine object detection technology with learning approach to achieve vehicle (train) detection process in moving scene challenge. After analyzing the feedbacks of CMS, some false alarms were released caused by color-rich background. Considering the limitation of local color features, a template matching method is presented to verify the detection result based on the global features. These templates are designed using K-means algorithm that is a common unsupervised learning approach. Besides, a simple verification process that uses the position relationship between the train and bumper is proposed based on the statistical data. We adopt histogram of orientated gradient (HOG) as a feature descriptor, because it shows a robust performance in various illumination conditions. Random forest algorithm is a popular machine learning approach, but it is new in the driving assistant application. Besides, the predicting process of this classifier is very fast because it only depends on a limit number of simple tests in each random-trained decision tree. Based on the HOG features and random forest algorithm, a close-range red train detector was designed. This proposed detector worked as one detection module in CMS, and the correct detection rate of close-range red train was nearly 100%, which means there was no missing detection. Compared with original CMS system, our approach could release more reliable warning signals with a lower false-alarm rate of 0.09% in our control experiment.
Rights: All rights reserved
Access: restricted access

Files in This Item:
File Description SizeFormat 
991022144625103411.pdfFor All Users (off-campus access for PolyU Staff & Students only)3.67 MBAdobe PDFView/Open

Copyright Undertaking

As a bona fide Library user, I declare that:

  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.

By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

Show full item record

Please use this identifier to cite or link to this item: