Author: Rezka Bunaiya Prayudha
Prayudha, Rezka Bunaiya, 2020 AddInsight-Vision: An intelligent video system for object detection in traffic scenes and vehicle re-identification, Flinders University, College of Science and Engineering
Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.
This thesis project involved the development of the AddInsight-Vision - an intelligent video system designed to extract information from traffic scenes which is part of SAGE Automation's Intelligent Transport System. The primary aim of the project was to design and develop such a stream-processing pipeline which functioned to detect traffic objects, including vehicle and pedestrian accurately and send the information to the management system for further analysis. The initial stage of AddInsight-Vision development covered the state-of-art analysis and determining evaluation metrics—the state-of-art analysis including proposed hardware requirements, tools and software framework. This project used Average Precision (AP) and Cumulative Matching Characteristic (CMC) to evaluate AddInsight-Vision. Following the development of AddInsight-Vision, the stream processing pipeline was constructed and tested. Early testing result explicitly showed that AddInsight-Vision were not able to detect object such as bus and pedestrian in certain conditions such as night times. After initial testing, AddInsight-Vision was modified to improve object detection accuracy and overall performance. The object detection and vehicle re-identification were models re-trained using transfer learning based on DPTI and Ve-RI datasets. In the final evaluation, the average precision of object detection model achieved over 90% for car and bus detection while maintaining high Frame Rate per Second (FPS). In addition, the CMC result of vehicle re-identification model achieved second-best in contrast with other baselines with a similar task. As such, the overall performance of AddInsight-Vision during testing and evaluation shows the project development was successful.
Keywords: computer vision, deep learning, intelligent transport system, object detection, intelligent video analytics
Subject: Engineering thesis
Thesis type: Masters
Completed: 2020
School: College of Science and Engineering
Supervisor: Dr. Nasser Asgari