Date of Publication

2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Electronics and Communications Engineering

Subject Categories

Electrical and Computer Engineering | Engineering

College

Gokongwei College of Engineering

Department/Unit

Electronics And Communications Engg

Thesis Advisor

Edwin Sybingco

Defense Panel Chair

Argel Bandala

Defense Panel Member

Robert Kerwin Billones
Alexis Fillone
Elmer Dadios
Raouf Naguib

Abstract (English)

Video-based traffic monitoring systems provide essential data for transportation planning, urban management, and infrastructure maintenance. Recently, there has been a rapid growth in video surveillance studies. However, video image processors for traffic data collection still have weaknesses, such as low accuracy during nighttime and missing counts caused by occlusion. Also, many deep learning-based systems need a massive amount of data to be collected for training and creating an effective model. This study aims to propose a video-based traffic data collection method that can collect traffic data during daytime and nighttime and count vehicles even with occlusions using deep learning algorithms that do not require costly annotations. The proposed method comprises an enhancement network, instance segmentation network, tracking module, and automatic camera calibration module. The enhancement network employs a self-supervised strategy for image enhancement training. The instance segmentation network trains the model by predictions of Cascade MaskRCNN to generate annotations for unlabeled training images. The tracking module, Mask- OCSORT, uses mask features and bounding box coordinates of the instance segmentation network to link formerly tracked objects with new detections. Then, the automatic camera calibration module, for converting 2D coordinates of tracked objects viito 3D world coordinates, uses the tracks and segmentation masks from previous modules to generate vanishing points for the camera calibration problem. By implementing the proposed algorithms, the study obtained traffic data like vehicle classification, speed, volume, and density in traffic videos. The study was able to enhance nighttime images without boosting the light effects from vehicles and streetlights. The study used unlabeled real-video traffic scenes and a small traffic dataset comprising more than 1300 annotations from over 600 traffic images. The results show that the proposed traffic data collection can track and count vehicles with occlusions on daytime and nighttime traffic videos.

Abstract Format

html

Language

English

Format

Electronic

Keywords

Traffic flow; Traffic patterns; Deep learning (Machine learning); Intelligent transportation systems

Upload Full Text

wf_yes

Embargo Period

12-11-2023

Download

Share

COinS