0422106801 - Dewi Rosmala , S.Si, M.IT. Dosen Pembimbing 1 DIMAS SANTOSO / 15-2021-218 Penulis

Peningkatan jumlah kendaraan bermotor di Indonesia telah menyebabkan kemacetan lalu lintas yang semakin parah, sehingga diperlukan pengukuran volume lalu lintas yang akurat menggunakan teknologi computer vision. Namun, proses pelacakan kendaraan sering mengalami ID switching akibat occlusion dan kemiripan objek. Penelitian ini dilakukan untuk mereduksi ID switching dengan mengintegrasikan algoritma DeepSORT pada model deteksi YOLOv11 dalam pengukuran volume lalu lintas. Dataset yang digunakan terdiri dari rekaman video CCTV pada waktu sibuk di Jalan Raya Imbanagara, Ciamis, dan Jalan Siliwangi, Sukabumi, dengan kelas kendaraan meliputi motor, mobil, truk, dan bus. Proses pelatihan YOLOv11 menghasilkan precision 0,910, recall 0,887, mAP@0.5 0,951, dan mAP@[0.5:0.95] 0,700. Pengujian dengan DeepSORT menunjukkan akurasi pengukuran volume lalu lintas sebesar 96,3% dan reduksi ID switching hingga 197 kejadian, lebih baik dibandingkan metode Euclidean Distance dan YOLOv10n. Metode counting line diterapkan untuk menghitung kendaraan berdasarkan perlintasan garis virtual. The increase in motorized vehicles in Indonesia has led to severe traffic congestion, necessitating accurate traffic volume measurement using computer vision technology. However, vehicle tracking processes often encounter ID switching due to occlusion and object similarity. This study was conducted to reduce ID switching by integrating the DeepSORT algorithm with the YOLOv11 detection model in traffic volume measurement. The dataset utilized consists of CCTV video recordings during peak hours on Jalan Raya Imbanagara, Ciamis, and Jalan Siliwangi, Sukabumi, with vehicle classes including motorcycles, cars, trucks, and buses. The YOLOv11 training process yielded a precision of 0.910, recall of 0.887, mAP@0.5 of 0.951, and mAP@[0.5:0.95] of 0.700. Testing with DeepSORT demonstrated a traffic volume measurement accuracy of 96.3% and a reduction in ID switching to 197 occurrences, outperforming the Euclidean Distance method and YOLOv10n. The counting line method was applied to count vehicles based on virtual line crossings.