Navigasi Mobil Otonom Menggunakan Segmentasi Visual Dan Data GPS Untuk Manuver Di Jalan Raya ITS

Priyatna, Hernanda Achmad (2026) Navigasi Mobil Otonom Menggunakan Segmentasi Visual Dan Data GPS Untuk Manuver Di Jalan Raya ITS. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5022221114-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Navigasi mobil otonom umumnya bergantung pada GPS, namun akurasinya menurun di lingkungan dengan banyak gedung tinggi dan pepohonan seperti kampus Institut Teknologi Sepuluh Nopember (ITS). Sistem navigasi mobil riset otonom ITS saat ini masih menggunakan GPS sebagai acuan utama pengendali arah sehingga kendaraan tidak dapat mengoreksi arah berdasarkan kondisi jalan yang teramati. Penelitian ini mengintegrasikan hasil segmentasi visual dengan rute titik acuan GPS melalui proses path fusion untuk menghasilkan manuver belok yang adaptif di persimpangan sekaligus mengurangi ketergantungan terhadap GPS. Sistem dibangun dalam kerangka ROS2 dan terdiri atas lima subsistem: persepsi visual, lokalisasi, perencanaan rute global, perencanaan rute lokal, dan kendali kendaraan. Subsistem persepsi visual menggunakan model segmentasi InternImage yang di-fine-tuning pada dataset lokal kampus ITS, dioptimasi dengan TensorRT, lalu ditransformasi menjadi representasi point cloud Bird’s Eye View. Subsistem lokalisasi memfusikan data GPS, IMU, dan odometri roda menggunakan Kalman Filter, sedangkan subsistem navigasi menggabungkan perencanaan rute global berbasis Dijkstra dengan perencanaan rute lokal Dynamic Window Approach. Pengujian dilakukan pada mobil listrik Chery Omoda E5 di lingkungan nyata dan simulasi. Hasil menunjukkan segmentasi mencapai mIoU 0,9381 pada kondisi siang mendung dan 0,9068 pada malam hari dengan laju inferensi 18 FPS. Lokalisasi menghasilkan RMSE 0,173 m dan tetap stabil pada rentang 1,07–1,56 m saat sinyal GPS terdegradasi, jauh lebih andal dibanding GPS murni yang menyimpang hingga 9,43 m. Mode integrasi path fusion mencapai tingkat keberhasilan navigasi 74,4%, melampaui mode hanya citra (63,6%) dan hanya GPS (42,0%), serta memiliki toleransi galat GPS yang lebih besar. Integrasi persepsi visual terbukti meningkatkan keandalan navigasi dan mengurangi ketergantungan pada GPS di lingkungan kampus yang kompleks.
===========================================================================================================================
Autonomous vehicle navigation commonly relies on GPS, yet its accuracy degrades in environments crowded with tall buildings and trees such as the campus of Institut Teknologi Sepuluh Nopember (ITS). The current navigation system of the ITS autonomous research vehicle still uses GPS as its primary steering reference, so the vehicle cannot correct its heading based on the observed road condition. This research integrates visual segmentation results with a GPS waypoint route through a path fusion process to produce adaptive turning maneuvers at intersections while reducing dependence on GPS. The system is built within the ROS2 framework and comprises five subsystems: visual perception, localization, global route planning, local route planning, and vehicle control. The visual perception subsystem uses an InternImage segmentation model fine-tuned on a local ITS campus dataset, optimized with TensorRT, and transformed into a Bird’s Eye View point cloud representation. The localization subsystem fuses GPS, IMU, and wheel odometry data using a Kalman Filter, while the navigation subsystem combines Dijkstra-based global route planning with Dynamic Window Approach local route planning. Testing was conducted on a Chery Omoda E5 electric vehicle in both real-world and simulation environments. The results show that segmentation achieves an mIoU of 0.9381 in overcast daylight and 0.9068 at night, with an inference rate of 18 FPS. Localization yields an RMSE of 0.173 m and remains stable within 1.07–1.56 m under degraded GPS signal, far more reliable than raw GPS, which drifts up to 9.43 m. The path fusion mode attains a 74.4% navigation success rate, surpassing vision-only (63.6%) and GPS-only (42.0%) modes, and tolerates larger GPS errors. Integrating visual perception is shown to improve navigation reliability and reduce GPS dependence in complex campus environments.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Navigasi Mobil Otonom, Segmentasi Visual, Kalman Filter, GPS, DWA, Autonomous Vehicle Navigation, Visual Segmentation
Subjects:	T Technology > TJ Mechanical engineering and machinery > TJ211.415 Mobile robots T Technology > TL Motor vehicles. Aeronautics. Astronautics > TL152.8 Vehicles, Remotely piloted. Autonomous vehicles. T Technology > TL Motor vehicles. Aeronautics. Astronautics > TL589.2.N3 Navigation computer
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User:	Hernanda Achmad Priyatna
Date Deposited:	23 Jul 2026 06:46
Last Modified:	23 Jul 2026 06:46
URI:	http://repository.its.ac.id/id/eprint/136485

Actions (login required)

View Item