Research

Publications & thesis

Peer-reviewed work at the intersection of computer vision, deep learning, and intelligent transportation systems.

IEEE · PUBLISHED 2025 First Author

Optimizing YOLOv11 for Vehicle Detection in Low-Visibility CCTV Footage

ICSINTESA 2025 — 5th International Conference of Science and Information Technology in Smart Administration

0.896
mAP@0.5: 0.678
mAP@0.5-0.95: 122
FPS: 8.2 ms/img
latency

Abstract

This study optimizes YOLOv11 for vehicle detection in low-visibility environments (rain, fog, low-light) using CCTV data. Systematic evaluation of batch sizes (8, 16, 32), image resolutions (228, 448, 640), and optimizers (AdamW, SGD, Adam, etc.) identified the optimal configuration: batch size 16, resolution 640x640, AdamW optimizer. The model achieved mAP@0.5 of 0.896 and mAP@0.5-0.95 of 0.678, with real-time inference at 122 FPS (8.2 ms/img) on an NVIDIA RTX 4090 GPU.

Co-authors: William Eka Chandra, Rafi Andhika Putra Pratama, Felix Corputty · GPU: NVIDIA RTX 4090

IEEE Xplore DOI

Cite

R. Firdaus, W. E. Chandra, R. A. P. Pratama, and F. Corputty, “Optimizing YOLOv11 for Vehicle Detection in Low-Visibility CCTV Footage,” in Proc. 2025 5th Int. Conf. of Science and Information Technology in Smart Administration (ICSINTESA), IEEE, 2025.

IN PROGRESS Undergraduate thesis

Traffic Flow Estimation under Low-Visibility Conditions

Extending the IEEE paper's object detection work to traffic flow estimation using multi-object tracking (ByteTrack) and speed measurement per Indonesian Highway Capacity Manual (PKJI) standards. Focus on low-visibility robustness from CCTV footage.

University

Telkom University, Bandung — School of Informatics, B.Sc. Data Science

Technologies

YOLOv11YOLOv26ByteTrackPyTorchOpenCV

Co-authored

Other publications

Contributions spanning medical imaging and wireless communications.

2026 Co-author (4th of 6)

Comparative Study of Enhanced Deep Learning Models for Cervical Pap Smear Cell Segmentation

A strictly fair comparative benchmark of four deep learning architectures (U-Net, U-Net++, Swin-UNet, SegNet) for cervical cell semantic segmentation on the Herlev dataset. Swin-UNet achieved the highest performance (F1: 0.8868 ± 0.0027; IoU: 0.7974 ± 0.0043), with statistical significance confirmed via one-way ANOVA and Bonferroni-corrected pairwise t-tests.

Co-authors: Fadly Huwaiza Khalid, Muhammad Daffa Izzati, Akmal Yaasir Fauzaan, William Eka Chandra, Naufal Hanan Lutfianto

2026 Co-author (4th of 4)

Deep Learning-Based Resource Allocation for Visible Light Communication Networks

VLC-Net: A supervised deep learning framework for real-time resource allocation in multi-cell visible light communication networks employing NOMA. A feedforward DNN with Bayesian-optimized architecture achieves 388× speedup over iterative solvers and 66% improvement in Jain's fairness index.

Co-authors: Marcelia Chintya Hartakaadi, Aminah Indahsari Marsuki, Intan Nisa Bani

Focus

Research interests

Computer Vision for Traffic Systems

Object detection, multi-object tracking, and traffic-parameter estimation from CCTV footage under real-world, low-visibility conditions.

NLP for Underrepresented Languages

Indonesian NLP with transformer models — emotion classification, NER-based anonymization, and retrieval-augmented systems.

Explainable AI & Fairness

SHAP-based interpretability, evidence-grounded scoring, and transparent decision pipelines in high-stakes applications.

Production ML Systems

Bridging research and deployment: containerization, self-hosted APIs, and scalable RAG architectures on commodity infrastructure.