Sistem Cerdas Penghitungan Volume Lalu Lintas Berbasis CCTV Menggunakan YOLO

Wardhani, Muhammad Azzikri Putra (2026) Sistem Cerdas Penghitungan Volume Lalu Lintas Berbasis CCTV Menggunakan YOLO. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5024221029-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (13MB) | Request a copy

Abstract

Ketergantungan analitik CCTV lalu lintas pada komputasi terpusat dapat menimbulkan latensi jaringan, biaya server tinggi, dan ketidakstabilan saat kualitas koneksi CCTV berubah. Penelitian ini mengajukan purwarupa penghitung volume lalu lintas berbasis edge computing yang mengeksekusi seluruh proses akuisisi, deteksi, pelacakan, penghitungan, penyimpanan, dan visualisasi secara lokal pada perangkat NVIDIA Jetson Orin Nano Super. Sistem menggunakan model YOLO26s yang dilatih pada 9.135 citra kendaraan dari CCTV ATCS siang dan malam, dataset MIO-TCD, serta cuplikan video daring, kemudian dikompilasi ke format TensorRT FP32, FP16, dan 8-bit integer (INT8). Alur operasional memanfaatkan OpenCV untuk akuisisi RTSP/HLS, ByteTrack untuk mempertahankan identitas kendaraan, Redis untuk komunikasi asinkron, SQLite untuk penyimpanan historis, dan FastAPI untuk dashboard pemantauan. Hasil komparasi arsitektur menunjukkan YOLO26s TensorRT FP16 memproses 4.500 bingkai dalam 86,90 detik dengan 51,78 FPS dan latensi P95 16,23 ms. Pada evaluasi presisi, konfigurasi TensorRT INT8 menjadi yang paling efisien dengan ukuran model 12,60 MB, kecepatan 52,79 FPS, latensi rerata 15,65 ms, latensi P95 15,76 ms, mAP50 0,8548, dan mAP50-95 0,6587. Meskipun membutuhkan RAM maksimum 2.040,88 MB, uji beban kontinu sekitar tiga jam mempertahankan 24,96 FPS tanpa failed frame dan suhu puncak 52,28◦C. Hasil tersebut menunjukkan bahwa arsitektur ini mampu mereduksi latensi analitik sekaligus menjaga stabilitas perangkat keras untuk operasional pemantauan lalu lintas berbasis CCTV secara real-time.
=====================================================================================================================================
The dependence of traffic CCTV analytics on centralized computing may introduce network latency, high server cost, and unstable operation when CCTV connection quality changes.
This study proposes an edge-computing traffic volume counting prototype that executes video acquisition, detection, tracking, counting, storage, and visualization locally on an NVIDIA Jetson Orin Nano Super. The system uses a YOLO26s model trained on 9,135 vehicle images collected from daytime and nighttime ATCS CCTV, the MIO-TCD dataset, and online traffic video clips, then compiled into TensorRT FP32, FP16, and 8-bit integer (INT8) formats. The operational pipeline employs OpenCV for RTSP/HLS acquisition, ByteTrack to preserve vehicle identities, Redis for asynchronous communication, SQLite for historical storage, and FastAPI for the monitoring dashboard. The architectural comparison shows that YOLO26s TensorRT
FP16 processes 4,500 frames in 86.90 seconds with 51.78 FPS and 16.23 ms P95 latency. In the precision evaluation, TensorRT INT8 is the most efficient configuration, producing a 12.60 MB model size, 52.79 FPS, 15.65 ms average latency, 15.76 ms P95 latency, 0.8548 mAP50, and 0.6587 mAP50-95. Although it requires a maximum RAM allocation of 2,040.88 MB, the approximately three-hour continuous stress test maintains 24.96 FPS with zero failed frames and a 52.28◦C peak temperature. These results demonstrate that the proposed architecture reduces analytics latency while preserving hardware stability for real-time CCTV-based traffic monitoring operations.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	CCTV Lalu Lintas, Edge Computing, TensorRT, Traffic Counting, YOLO26s, Traffic CCTV, Edge Computing, TensorRT, Traffic Counting, YOLO26s
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.6 Computer programming. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) R Medicine > R Medicine (General) > R858 Deep Learning T Technology > T Technology (General) > T57.5 Data Processing T Technology > TA Engineering (General). Civil engineering (General) > TA1573 Detectors. Sensors
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Muhammad Azzikri Putra Wardhani
Date Deposited:	27 Jul 2026 01:20
Last Modified:	27 Jul 2026 01:20
URI:	http://repository.its.ac.id/id/eprint/137381

Actions (login required)

View Item