Pengembangan Sistem Deteksi Kerusakan Kendaraan Berbasis Video Menggunakan Yolov8

Nitilaksito, Hario Kuncoro Nitilaksito (2026) Pengembangan Sistem Deteksi Kerusakan Kendaraan Berbasis Video Menggunakan Yolov8. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5024201022-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Proses penanganan klaim asuransi kendaraan bermotor secara konvensional memakan waktu yang lama (10 hingga 30 hari) dan rentan terhadap penilaian yang subjektif serta risiko penipuan (fraud). Penelitian ini bertujuan untuk merancang bangun purwarupa sistem deteksi jenis kerusakan kendaraan berbasis video dengan mengintegrasikan algoritma You Only Look Once (YOLO), platform orkestrasi alur kerja n8n, dan mesin inferensi FastAPI. Arsitektur sistem dibangun secara terpisah (decoupled architecture) di dalam lingkungan Docker, yang terdiri dari antarmuka pengguna Streamlit, n8n, FastAPI, dan basis data PostgreSQL. Model kecerdasan buatan dilatih menggunakan varian arsitektur YOLOv8m yang diakselerasi melalui komputasi perangkat keras GPU lokal NVIDIA RTX 5060 Ti dan teknik Transfer Learning, secara native untuk mendeteksi 6 kelas kerusakan asli dari dataset publik CarDD (penyok, goresan, retak, kaca pecah, lampu pecah, ban kempes) tanpa remapping label. Kontribusi ilmiah dan evaluasi kuantitatif penelitian ini difokuskan murni pada akurasi deteksi jenis kerusakan; penentuan tingkat keparahan maupun prioritas penanganan tetap sepenuhnya menjadi wewenang petugas/surveyor manusia. Guna memproses masukan berupa rekaman video, sistem mengimplementasikan logika deduplikasi berbasis dictionary dan pemilihan best frame untuk mencegah tumpang tindih deteksi. Hasil pengujian menunjukkan bahwa model terbaik (v7) mencapai nilai Mean Average Precision (mAP@0.5) sebesar 0,721, tingkat Presisi 0,773, dan Recall 0,678 pada data pengujian akhir (Test Set). Sistem ini berhasil memangkas waktu inspeksi dengan latensi pemrosesan gambar 0,126–0,181 detik dan latensi pemrosesan video (durasi ±30 detik) 0,602–0,731 detik berdasarkan pengukuran langsung ke backend, meskipun saat ini inferensi masih dijalankan menggunakan CPU-only pada kontainer Docker. Secara keseluruhan, integrasi arsitektur sistem ini menghasilkan ringkasan temuan kerusakan per jenis serta menawarkan otomasi inspeksi yang cepat dan efisien, yang bermanfaat bagi industri asuransi kendaraan bermotor sebagai salah satu pemangku kepentingan.
==========================================================================================================================================
The conventional motor vehicle insurance claim handling process is time-consuming (10 to 30 days) and vulnerable to subjective assessments as well as fraud risks. This study aims to develop a prototype for a video-based vehicle damage detection and severity classification system by integrating the You Only Look Once (YOLO) algorithm, the n8n workflow orchestration platform, and the FastAPI inference engine. The system architecture is built in a decoupled architecture within a Docker environment, consisting of the Streamlit user interface, n8n, FastAPI, and a PostgreSQL database. The artificial intelligence model was trained using the YOLOv8m architecture variant, accelerated by NVIDIA RTX 5060 Ti local GPU hardware computation and Transfer Learning techniques. The model was trained to detect all 6 original CarDD damage classes (dent, scratch, crack, glass shatter, lamp broken, tire flat) without label remapping; the severity level (Light, Medium, Heavy) is then derived by the system as an additional triage heuristic layer based on the detected damage type. To process video recording inputs, the system implements dictionary-based deduplication logic and best-frame selection to prevent overlapping detections. The test results showed that the best model (v7) achieved a Mean Average Precision (mAP@0.5) of 0.721, a Precision rate of 0.773, and a Recall of 0.678 on the Test Set. This system successfully reduces inspection time with a processing latency of 0.126–0.181 seconds per image and 0.602–0.731 seconds per video (±30 seconds duration) based on direct backend measurement, even though the inference is currently still run using CPU-only on the Docker container. Overall, the integration of this system architecture produces a summary of damage findings by type and severity level, offering fast and efficient operational automation that benefits the vehicle insurance industry as one of its stakeholders.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Kerusakan Kendaraan, YOLOv8, Orkestrasi n8n, Mesin Inferensi FastAPI, Vehicle Insurance, YOLOv8, n8n Orchestration, FastAPI Inference Engine.
Subjects:	T Technology > T Technology (General) > T174.5 Technology--Risk assessment.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Hario Kuncoro Nitilaksito
Date Deposited:	29 Jul 2026 01:34
Last Modified:	29 Jul 2026 01:34
URI:	http://repository.its.ac.id/id/eprint/139138

Actions (login required)

View Item