Electronic Theses and Dissertation

Peningkatan jumlah kendaraan di kawasan perkotaan menuntut adanya sistem pemantauan lalu lintas yang akurat dan mampu bekerja secara real-time. Penelitian ini bertujuan untuk menerapkan serta mengevaluasi kinerja algoritma YOLOv12 dalam mendeteksi dan menghitung jumlah kendaraan berbasis video. Penelitian ini menggunakan tiga varian model YOLOv12, yaitu YOLOv12n, YOLOv12s, dan YOLOv12m, yang dilatih menggunakan dataset citra kendaraan hasil anotasi manual. Model yang telah dilatih kemudian diuji pada video lalu lintas nyata yang direkam di kawasan JPO Masjid Oman, Banda Aceh. Evaluasi kinerja sistem dilakukan menggunakan parameter precision, recall, mAP@0.5, dan F1-score. Data ground truth diperoleh melalui perhitungan manual jumlah kendaraan pada video uji dengan durasi 1 menit 10 detik. Hasil pengujian menunjukkan bahwa ketiga model YOLOv12 mampu mendeteksi kendaraan dengan tingkat performa yang berbeda. Model YOLOv12m menunjukkan kinerja paling stabil dan hasil penghitungan yang paling mendekati ground truth, dengan nilai precision sebesar 80,1%, recall sebesar 67,3%, mAP@0.5 sebesar 78,7%, serta F1-score pada kisaran 0,70–0,75. Berdasarkan hasil tersebut, YOLOv12m dinilai sebagai model terbaik dalam penelitian ini karena mampu memberikan keseimbangan antara kinerja deteksi dan kestabilan penghitungan kendaraan. Model ini berpotensi untuk diterapkan pada sistem pemantauan lalu lintas otomatis berbasis video sebagai pendukung pengelolaan lalu lintas perkotaan.

The increase in vehicle volume in urban areas requires an accurate traffic monitoring system that is capable of working in real time. This study aims to apply and evaluate the performance of the YOLOv12 algorithm in detecting and counting the number of vehicles based on video. This study uses three variants of the YOLOv12 model, namely YOLOv12n, YOLOv12s, and YOLOv12m, which were trained using a dataset of manually annotated vehicle images. The trained models were then tested on real traffic videos recorded in the JPO Masjid Oman area, Banda Aceh. System performance evaluation was conducted using the parameters of precision, recall, the mAP@0.5s, and F1-score. Ground truth data was obtained through manual counting of the number of vehicles in the test video with a duration of 1 minute and 10 seconds. The test results showed that all three YOLOv12 models were able to detect vehicles with varying levels of performance. The YOLOv12m model showed the most stable performance and calculation results closest to the ground truth, with a precision value of 80.1%, recall of 67.3%, mAP@0.5 of 78.7%, and an F1-score in the range of 0.70–0.75. Based on these results, YOLOv12m is considered the best model in this study because it provides a balance between detection performance and vehicle calculation stability. This model has the potential to be applied to video-based automatic traffic monitoring systems as a support for urban traffic management.