Sistem Deteksi dan Penghitung Kendaraan pada Video Survei Lalu Lintas Harian Rata-Rata

Bihanda, Yusuf Gladiensyah (2024) Sistem Deteksi dan Penghitung Kendaraan pada Video Survei Lalu Lintas Harian Rata-Rata. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025222011-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (8MB) | Request a copy

Abstract

Survei Lalu Lintas Harian Rata-Rata menjadi salah satu acuan penting dalam pengambilan keputusan terhadap suatu ruas jalan setiap tahunnya. Metode yang digunakan umumnya bersifat semi-manual yakni merekam ruas jalan tertentu menggunakan CCTV serta menghitung secara manual kendaraan yang ada pada ruas jalan tersebut. Namun pendekatan ini tidak efektif serta tidak efisien dikarenakan berpotensi terjadi human error. Disamping itu penelitian-penelitian terkait penghitung kendaraan dan deteksi kendaraan masih belum mampu menyelesaikan permasalahan pada survei karena perbedaan posisi kamera dan dataset yang digunakan. Pendekatan deep learning Real Time Detection Transformer (RT-DETR) memiliki kemampuan dalam mengidentifikasi objek secara akurat dengan frame per second (FPS) yang tinggi dibandingkan arsitektur DETR lainnya serta algoritma tracking Bytetrack mampu untuk melakukan penelusuran objek dengan memperhatikan semua bounding box yang ada sehingga kombinasi dari RT-DETR dengan ByteTrack digunakan dalam penelitian ini.
Penelitian ini menggunakan dataset dari CCTV yang dipasang pada samping ruas jalan dengan latar siang maupun malam hari. Untuk membandingkan hasil penerapan RT-DETR dengan ByteTrack, dipilih YOLOX dan YOLOv9 serta algoritma tracking SORT lalu dihasilkan 6 kombinasi berbeda. Tahapan dimulai dari mendeteksi objek dalam satu frame, menelusuri objek hingga melewati garis bantu perhitungan, terakhir objek dihitung setelah melewati garis bantu perhitungan.
Hasil evaluasi model deteksi menunjukkan bahwa model RT-DETR Resnet101 menghasilkan mAP tertinggi dengan nilai 0.891. Hasil perhitungan menunjukkan bahwa akurasi perhitungan terbaik menggunakan kombinasi RT- DETR dengan ByteTrack dengan nilai akurasi sebesar 82.48% dengan FPS keseluruhan sebesar 1.93. Sedangkan nilai FPS tertinggi untuk keseluruhan video uji diraih YOLOX dengan ByteTrack dengan nilai 4.95 dan nilai akurasi sebesar 77.46%. Hasil pengujian juga membenarkan klaim bahwa RT-DETR menghasilkan akurasi deteksi objek yang tinggi namun FPS yang dihasilkan belum mampu menyaingi model deteksi yang dibandingkan. Selain itu hasil perbandingan penggunaan algoritma tracking menunjukkan bahwa ByteTrack mampu menanggulangi pergantian ID objek pada saat objek terhalang objek lain selama kurang dari 30 frame. Selain itu, sistem berbasis desktop telah dilakukan proses rancang bangun untuk membantu surveyor dalam menghitung kendaraan.
===========================================================================================================================
The Average Daily Traffic Survey is an important reference for making annual road section decisions. The method used is generally semi-manual, recording certain road sections using CCTV and manually counting the number of vehicles on those road sections. However, it is ineffective and inefficient due to the potential for human error. In addition, research related to vehicle counters and vehicle detection is still unable to solve the survey problems due to differences in camera positions and datasets. The deep learning Real Time Detection Transformer (RT-DETR) approach has the ability to identify objects accurately with high frames per second (FPS) compared to other DETR architectures, and the Bytetrack tracking algorithm is able to search for objects by paying attention to all existing bounding boxes; thus, the combination of RT-DETR with ByteTrack is used in this study.
This study uses datasets from CCTV installed side by side of the road in both daytime and nighttime settings. To compare the results of applying RT-DETR with ByteTrack, YOLOX and YOLOv9 and the SORT tracking algorithm were selected and 6 different combinations were produced. The stages begin by detecting the object in one frame and tracking it until it passes the counting line. Finally, the object is counted after passing the counting line.
The results of the detection model evaluation indicated that the RT-DETR Resnet101 model produced the highest mAP (0.891). The results demonstrate that the best calculation accuracy was achieved using the combination of RT-DETR and ByteTrack, with an accuracy value of 82.48% and an overall FPS of 1.93. The highest FPS value for the entire test video was achieved by YOLOX with ByteTrack, with a value of 4.95 and an accuracy value of 77.46%. The test results also justified the claim that RT-DETR produces high object detection accuracy; however, the resulting FPS could not rival the detection models being compared. In addition, the results of the comparison of the use of tracking algorithms demonstrate that ByteTrack can overcome the change in object ID when an object is blocked by other objects in less than 30 frames. In addition, a desktop-based system was designed to help surveyors count vehicles.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Survei Lalu Lintas Harian Rata-Rata, Sistem Deteksi dan Perhitungan Kendaraan, RT-DETR, ByteTrack, Average Daily Traffic Survey, Vehicle Detection and Counting System, RT-DETR, ByteTrack
Subjects:	T Technology > T Technology (General) T Technology > TE Highway engineering. Roads and pavements > TE7 Transportation--Planning
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Yusuf Gladiensyah Bihanda
Date Deposited:	07 Aug 2024 21:40
Last Modified:	07 Aug 2024 21:40
URI:	http://repository.its.ac.id/id/eprint/112094

Actions (login required)

View Item