Estimasi Traversability Berbasis Visi Monokuler Dan Penerapannya Pada Kontrol Kecepatan Adaptif Kendaraan Tak Berawak Off-Road Menggunakan Model Predictive Control

Daniyah, Daniyah (2026) Estimasi Traversability Berbasis Visi Monokuler Dan Penerapannya Pada Kontrol Kecepatan Adaptif Kendaraan Tak Berawak Off-Road Menggunakan Model Predictive Control. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022211056-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Navigasi otonom di medan off-road menuntut kemampuan menilai traversability (keterlintasan) terrain secara cepat dan andal sebelum kendaraan melintasinya. Pendekatan konvensional mengandalkan sensor jarak aktif seperti LiDAR atau kamera stereo yang akurat namun mahal, berat, dan berkonsumsi daya tinggi. Penelitian ini mengembangkan pipeline estimasi traversability berbasis kamera tunggal yang mengintegrasikan segmentasi semantik UperNet-ConvNeXt-Small dan estimasi kedalaman monokuler MiDaS-small untuk menghasilkan skor traversability kontinu melalui formula heuristik, untuk lima kelas terrain yaitu grass, sand, puddle, rubble, dan mud. Skor traversability dimodulasi oleh kapabilitas penggerak kendaraan, membedakan konfigurasi dua roda (2WD) dan empat roda (4WD). Skor tersebut kemudian dimanfaatkan sebagai batasan kecepatan pada tiga strategi kendali yang dibandingkan, yaitu P-Controller, Model Predictive Control (MPC) tanpa batasan traversability, dan MPC dengan batasan traversability. Kestabilan pelacakan dievaluasi melalui analisis empiris fungsi Lyapunov. Pengujian subsistem persepsi pada rekaman off-road menunjukkan bahwa model dengan bobot pra-latih ADE20K tanpa penyetelan-halus hanya mengklasifikasikan dengan benar dua dari lima kelas terrain akibat kesenjangan domain, sehingga subsistem persepsi masih bersifat proof-of-concept. Hasil simulasi kendali menunjukkan bahwa MPC dengan batasan traversability menekan pelanggaran batas kecepatan terrain sebesar 98,6% dibandingkan MPC tanpa batasan, dihitung sebagai rata-rata reduksi per konfigurasi penggerak, sambil mempertahankan kecepatan rata-rata yang sebanding dengan P-Controller. Penelitian ini menunjukkan bahwa skor traversability berbasis kamera tunggal berpotensi menjadi mekanisme yang transparan dan hemat biaya untuk navigasi off-road, meskipun implementasi operasional masih memerlukan optimasi inferensi karena pipeline saat ini berjalan sekitar 0,26 frame per detik pada CPU.
=================================================================================================================================
Autonomous navigation in off-road terrain requires the ability to assess terrain traversability quickly and reliably before a vehicle traverses it. Conventional approaches rely on active range sensors such as LiDAR or stereo cameras, which are accurate but expensive, heavy, and power-intensive. This research develops a single-camera traversability estimation pipeline that integrates UperNet-ConvNeXt-Small semantic segmentation and MiDaS-small monocular depth estimation to produce a continuous traversability score through a heuristic formula for five terrain classes, namely grass, sand, puddle, rubble, and mud. The traversability score is modulated by the vehicle drivetrain capability, distinguishing two-wheel-drive (2WD) and four-wheel-drive (4WD) configurations. The score is then used as a speed constraint in three compared control strategies, namely a P-Controller, Model Predictive Control (MPC) without traversability constraint, and MPC with traversability constraint. Tracking stability is evaluated through empirical analysis of a Lyapunov function. Testing the perception subsystem on off-road footage shows that the model with ADE20K pre-trained weights without fine-tuning correctly classifies only two of the five terrain classes due to the domain gap, so the perception subsystem remains a proof-of-concept. Control simulation results show that MPC with traversability constraint reduces terrain speed-limit violations by 98.6% compared to MPC without constraint, computed as the average reduction per drivetrain configuration, while maintaining an average speed comparable to the P-Controller. This research demonstrates that single-camera traversability scores can potentially serve as a transparent and cost-effective mechanism for off-road navigation, although operational implementation still requires inference optimization since the current pipeline runs at approximately 0.26 frames per second on a CPU.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	traversability, kendali kecepatan adaptif, Model Predictive Control, visi monokuler, segmentasi semantik, kendaraan tak berawak, off-road, traversability, adaptive speed control, Model Predictive Control, monocular vision, semantic segmentation, unmanned ground vehicle, off-road
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7888.3 Digital computers
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Daniyah Daniyah
Date Deposited:	05 Aug 2026 04:34
Last Modified:	05 Aug 2026 04:34
URI:	http://repository.its.ac.id/id/eprint/143928

Actions (login required)

View Item