Deteksi Kendaraan Dengan Kondisi Oklusi Menggunakan YOLO-OVD

Al-Farisi, Muhammad Hanif (2025) Deteksi Kendaraan Dengan Kondisi Oklusi Menggunakan YOLO-OVD. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002211090-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (37MB) | Request a copy

Abstract

Pada umumnya, mobilisasi manusia memerlukan kendaraan, apabila luasnya jalan raya tidak sebanding dengan peningkatan kendaraan maka tentu akan menyebabkan berbagai masalah, salah satunya adalah kemacetan yang akan menghambat mobilisasi manusia. Salah satu cara mengatasinya adalah dengan melakukan rekayasa lalu lintas. Sebagai langkah awal, diperlukan suatu sistem yang dapat mendeteksi kendaraan dalam pengambilan keputusan rekayasa lalu lintas. Akan tetapi kondisi nyata di lapangan adalah sulitnya melakukan deteksi kendaraan yang tepat dan akurat, karena faktor oklusi, yaitu kondisi terhalangnya objek oleh objek lain yang menyebabkan visual dari objek tersebut tidak terlihat secara utuh. Pada penelitian ini dilakukan deteteksi kendaraan dengan YOLO (You Only Look Once) yang telah dilakukan beberapa perubahan pada arsitekturnya sehingga dapat menangani masalah oklusi parsial, yaitu dengan metode YOLO-OVD (YOLO-Occluded Vehicle Detection). Melalui perubahan arsitektur YOLO, proses ekstrasi fitur menjadi lebih baik dan dengan bantuan attention mechanism model dapat lebih fokus pada bagian kendaraan yang tidak terhalang. Proses pelatihan dilakukan dengan berbagai skenario hyperparameter dan menggunakan metrik precision, recall, mAP50 dan mAP50:95 sebagai metrik evaluasi selama proses validasi untuk didapatkan suatu model terbaik. Proses pengujian model dihitung berdasarkan persentase antara banyak frame objek tersebut berhasil diprediksi selama mengalami oklusi dan jumlah frame objek tersebut mengalami oklusi. Secara keseluruhan, persentase keberhasilan prediksi objek teroklusi pada berbagai skenario data uji menunjukkan hasil yang lebih tinggi pada model YOLO-OVD daripada YOLOv5s dengan selisih rata-rata persentase keberhasilan prediksi antara kedua model dari seluruh objek yang diamati sebesar 17.45% pada seluruh skenario uji coba. Sementara itu, hasil prediksi lebih tinggi pada model YOLO-OVD daripada YOLOv5s didapatkan pada 90% objek diamati dari seluruh skenario uji coba.
==============================================================================================================================================
Human mobility requires vehicles for moving from one place to another. However, if the width of the roads does not keep up with the increase in the number of vehicles, it will inevitably cause various problems, one of which is traffic congestion that hampers human mobility. One way to address this issue is through traffic engineering. As an initial step, a vehicle detection system is needed to support decision making in traffic engineering. However, in real-world scenarios, accurately detecting vehicles is challenging due to occlusion. Occlusion occurs when an object is partially or fully blocked by another object, making it visually incomplete. This study addresses on vehicle detection using YOLO (You Only Look Once), with several architectural modifications to handle partial occlusion. The proposed method is called YOLO-OVD (YOLO-Occluded Vehicle Detection), that build by modifying the YOLO architecture, then the feature extraction becomes more effective, and with the help of an attention mechanism, the model can focus more precisely on the unoccluded parts of vehicles. Training were conducted using some hyperparameter tuning scenarios and evaluated using precision, recall, mAP50, and mAP50:95 as evaluation metrics to get the best model during validation step. The model evaluation was conducted by calculating the percentage of frames in which occluded objects were successfully detected compared to the total number of frames where the objects were occluded. Overall, the model’s success rate in detecting occluded objects across different test scenarios showed better results for YOLO-OVD compared to YOLOv5s , with an average difference in prediction success rate between the two models of 17.45% across all observed objects in every test scenario. Furthermore, YOLO-OVD outperformed YOLOv5s in 90% of the observed objects throughout all test scenarios.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	YOLO-OVD, Deteksi Objek, Oklusi, Lalu Lintas, YOLOv5, YOLO-OVD, Object Detection, Occlusion, Traffic, YOLOv5
Subjects:	Q Science > QA Mathematics > QA76.9 Computer algorithms. Virtual Reality. Computer simulation.
Divisions:	Faculty of Mathematics, Computation, and Data Science > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Hanif Al-farisi
Date Deposited:	01 Aug 2025 06:43
Last Modified:	01 Aug 2025 06:43
URI:	http://repository.its.ac.id/id/eprint/125893

Actions (login required)

View Item