Date of Publication
8-15-2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Electronics and Communications Engineering
Subject Categories
Electrical and Computer Engineering
College
Gokongwei College of Engineering
Department/Unit
Electronics And Communications Engg
Thesis Advisor
Dr. Elmer P. Dadios
Defense Panel Chair
Dr. Argel A. Bandala
Defense Panel Member
Prof. Raouf N. G. Naguib
Dr. Edwin Sybingco
Dr. Laurence A. Gan Lim
Dr. Ryan Rhay P. Vicerra
Abstract (English)
This study presents the development and evaluation of a Universal Robust Vehicle Quantification and Classification System capable of detecting, classifying, counting, and identifying the color of vehicles from CCTV footage using deep learning. The system integrates synthetic dataset generation, model benchmarking, multi feature analytics, and a graphical user interface to support real time traffic monitoring. A large synthetic dataset was produced through the semi automated compositing of cropped vehicle images onto road backgrounds, enabling rapid creation of annotated training data without manual labeling. Three candidate models—Faster R CNN, YOLOv11, and YOLOv12—were evaluated using identical datasets, with YOLOv11 demonstrating superior accuracy and becoming the primary detection model. The complete system combines YOLOv11 for detection and classification, DeepSORT for tracking and counting, and an HSV based clustering method for color identification. All detection outputs, including class, type, color, bounding box dimensions, coordinates, and confidence scores, were stored for further analysis. Model performance was assessed using confusion matrices and standard metrics such as accuracy, precision, recall, and F1 score. Training on 10,000 synthetic images followed by iterative fine tuning using real CCTV frames from three traffic scenarios resulted in substantial improvements across all features. Additional experiments using 20,000 domain specific synthetic images confirmed the importance of camera aligned data in achieving high detection accuracy. Final results showed strong performance, including 97% class accuracy, 90% type accuracy, 82% color accuracy, and 96.67% counting accuracy. Overall, the study validates synthetic dataset generation and iterative fine tuning as effective strategies for building robust vehicle analytics systems. The proposed framework offers a scalable, adaptable solution for intelligent transportation applications across diverse road networks and camera configurations.
Abstract Format
html
Abstract (Filipino)
Ipinapakita ng pag‑aaral na ito ang pagbuo at pagsusuri ng isang Universal Robust Vehicle Quantification and Classification System na may kakayahang mag‑detect, mag‑classify, magbilang, at tukuyin ang kulay ng mga sasakyan mula sa CCTV footage gamit ang deep learning. Pinagsasama ng sistema ang synthetic dataset generation, model benchmarking, multi‑feature analytics, at graphical user interface upang suportahan ang real‑time na pagmamanman ng trapiko.
Nabuo ang isang malaking synthetic dataset sa pamamagitan ng semi‑automated na pag‑composite ng mga cropped na larawan ng sasakyan sa mga background ng kalsada, na nagpapabilis sa paggawa ng annotated training data nang hindi nangangailangan ng manual labeling. Tatlong modelong kandidato—Faster R‑CNN, YOLOv11, at YOLOv12—ang sinuri gamit ang magkaparehong dataset, kung saan nagpakita ng mas mataas na accuracy ang YOLOv11 at naging pangunahing detection model. Pinagsama ng kumpletong sistema ang YOLOv11 para sa detection at classification, DeepSORT para sa tracking at counting, at HSV‑based clustering para sa color identification. Lahat ng output ng detection—kabilang ang class, type, color, bounding box dimensions, coordinates, at confidence scores—ay iniimbak para sa karagdagang pagsusuri.
Sinuri ang performance ng modelo gamit ang confusion matrices at mga karaniwang metric tulad ng accuracy, precision, recall, at F1 score. Ang training gamit ang 10,000 synthetic images, na sinundan ng iterative fine‑tuning gamit ang totoong CCTV frames mula sa tatlong traffic scenarios, ay nagresulta sa malalaking pag‑angat sa lahat ng feature. Dagdag na eksperimento gamit ang 20,000 domain‑specific synthetic images ang nagpatunay sa kahalagahan ng camera‑aligned data upang makamit ang mataas na detection accuracy. Ipinakita ng huling resulta ang matatag na performance, kabilang ang 97% class accuracy, 90% type accuracy, 82% color accuracy, at 96.67% counting accuracy.
Sa kabuuan, pinatutunayan ng pag‑aaral na ang synthetic dataset generation at iterative fine‑tuning ay epektibong mga estratehiya sa pagbuo ng matatag na vehicle analytics systems. Nagbibigay ang iminungkahing framework ng isang scalable at adaptable na solusyon para sa intelligent transportation applications sa iba’t ibang road networks at camera configurations.
Abstract Format
html
Language
English
Keywords
Computer vision
Recommended Citation
Ambata, L. U. (2026). Universal robust vehicle quantification and classification using deep learning. Retrieved from https://animorepository.dlsu.edu.ph/etdd_ece/15
Upload Full Text
wf_yes
Embargo Period
8-14-2026
