Sistem Navigasi Berbasis ORB-SLAM3 Dengan Penghapusan Titik Fitur Dinamis Menggunakan YOLOV11 Pada Single Board Computer

Ahmad Taufiq, Ika Candradewi, S.Si., M.Cs.

2025 | Skripsi | ELEKTRONIKA DAN INSTRUMENTASI

Lingkungan dinamis masih menjadi tantangan besar dalam pengembangan sistem SLAM (Simultaneous Localization and Mapping), khususnya yang berbasis kamera. Ketika terdapat objek bergerak sering menyebabkan kesalahan estimasi posisi karena fitur visual dari objek dinamis turut digunakan dalam pelacakan. Untuk mengatasi permasalahan tersebut, penelitian ini akan menganalisis sebuah sistem bernama YDM-SLAM (YOLO Dynamic Mapping SLAM) yang mengintegrasikan metode ORB-SLAM3 dengan model segmentasi objek dinamis berbasis YOLOv8 dan YOLOv11 segmentasi. Evaluasi dilakukan menggunakan dataset TUM RGB-D dan Bonn RGB-D pada perangkat Jetson AGX Xavier untuk mengukur akurasi serta performa komputasinya secara real-time.

Penilaian akurasi dilakukan secara kuantitatif menggunakan metrik Absolute Trajectory Error (ATE) dan Relative Pose Error (RPE), serta secara kualitatif melalui visualisasi trajektori dan fitur keypoint. Hasil penelitian menunjukkan bahwa YDM-SLAM dengan YOLO baik YOLOv8 dan YOLOv11 berhasil meningkatkan akurasi estimasi posisi secara signifikan dibandingkan ORB-SLAM3. Rata-rata penurunan ATE mencapai 89.55%, dan penurunan RPE sebesar 41.85% pada seluruh sequence yang diuji. Sistem juga terbukti mampu melakukan relokalisasi ketika terjadi kehilangan pelacakan (tracking lost).

Dari sisi performa komputasi, sistem menunjukkan waktu inisialisasi 21.4 detik, mean waktu tracking 80.5 ms, FPS rata-rata 12.5, serta penggunaan daya CPU dan GPU masing-masing sebesar 5.9 W dan 2.6 W, dengan konsumsi memori rata-rata 7.7 GB.

Dengan demikian, YDM-SLAM menunjukkan keunggulan dalam akurasi dan ketahanan pada lingkungan dinamis, walaupun performa real-time masih dapat ditingkatkan melalui optimasi model dan runtime.

Dynamic environments remain a major challenge in the development of SLAM (Simultaneous Localization and Mapping) systems, particularly those based on visual sensors. The presence of moving objects often leads to pose estimation errors because dynamic features are mistakenly used for tracking. To address this issue, this study analyzes a system called YDM-SLAM (YOLO Dynamic Mapping SLAM), which integrates the ORB-SLAM3 method with dynamic object segmentation using YOLOv8 and YOLOv11 segmentation models. The evaluation was conducted using the TUM RGB-D and Bonn RGB-D datasets on a Jetson AGX Xavier device to assess both accuracy and real-time computational performance.

Accuracy evaluation was carried out quantitatively using the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) metrics, as well as qualitatively through trajectory and keypoint visualization. The results demonstrate that YDM-SLAM, using either YOLOv8 or YOLOv11, significantly improves pose estimation accuracy compared to ORB-SLAM3. The system achieved an average ATE reduction of 89.55% and an RPE reduction of 41.85% over all tested sequences. Furthermore, the system proved capable of relocalizing after tracking loss events.

In terms of computational performance, the system recorded an initialization time of 21.4 seconds, an average tracking time of 80.5 ms, an average frame rate of 12.5 FPS, CPU and GPU power consumption of 5.9 W and 2.6 W respectively, and an average memory usage of 7.7 GB.

In conclusion, YDM-SLAM offers improved robustness and accuracy in dynamic environments, although its real-time performance could still be enhanced through model and runtime optimization.

Kata Kunci : ORB-SLAM3, SLAM Dinamis, Computer Vision, Robotika