Refaldi, Darrell Athaya (2026) Pengembangan Model Terintegrasi Berbasis Deep Learning Untuk Autentikasi Identitas Kendaraan Dan Pengendara Sebagai Dasar Keputusan Sistem Gerbang Akses Otomatis. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5026221094-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5026221094-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (33MB) | Request a copy |
Abstract
Sistem kontrol akses kawasan terbatas seperti kompleks perumahan dan areaparkir umumnya masih bergantung pada verifikasi manual yang rentan terhadapkelelahan petugas, antrean panjang, serta modus identity spoofing, vehicle cloning, dan tailgating. Penelitian terdahulu umumnya memperlakukan pengenalan wajah pengendara dan pengenalan pelat nomor sebagai dua persoalan terpisah, sehingga masih terdapat celah dalam memverifikasi kesamaan identitas pengendara-kendaraan antara sesi masuk dan keluar. Penelitian ini mengusulkan kaskade multimodal yang mengintegrasikan deteksi objek, pengenalan wajah, dan pengenalan pelat nomor: wajah diverifikasi melalui embedding ArcFace, sedangkan pelat dilokalisasi oleh model berbasis YOLO dan dibaca melalui kesepakatan beberapa engine OCR, lalu dicocokkan secara berjenjang dengan teks pelat sebagai bukti utama dan wajah sebagai bukti pendukung. Diuji atas 71 citra pasangan masuk-keluar dari tangkapan CCTV nyata, sistem mencocokkan 66,2% pasangan tanpa satu pun pelanggaran format maupun digit, dengan deteksi pelat mencapai mAP@50 sebesar 0,811. Sistem juga diperluas dengan jalur inferensi berbasis klip video singkat yangmengumpulkan bukti Lintas-frame sebelum pencocokan, dan tetap menjagaprinsip nol penerimaan salah.
=================================================================================================================================
Access control systems for restricted areas such as residential complexes and parking facilities still largely rely on manual verification that is vulnerable to guard fatigue, long queues, and threats such as identity spoofing, vehicle cloning, and tailgating. Prior research generally treats rider face recognition and license-plate recognition as two separate problems, leaving a gap in verifying that the rider & vehicle identity is the same on entry and exit. This study proposes a multimodal cascade integrating object detection, face recognition, and license-plate recognition: the face is verified through ArcFace embeddings, while the plate is localized by a YOLO-based detector and read through a consensus of several OCR engines, then matched in a tiered manner with the plate text as primary evidence and the face as supporting evidence. Evaluated on 71 entry-exit image pairs from real CCTV captures, the system matched 66.2% of pairs with no format or digit violations, and plate detection reached a mAP@50 of 0.811. The system was also extended with a short-video-clip inference path that pools evidence across frames before matching, while preserving the zero-false-accept principle.
Actions (login required)
 |
View Item |