Perancangan Sistem Navigasi Visual Real-Time Berbasis Segmentasi Otsu dan Proyeksi Bird's-Eye View pada Jalur Greenhouse Terkontrol

Abstract

Penelitian ini merancang sistem panduan navigasi visual real-time pada Unmanned Ground Vehicle (UGV) untuk jalur greenhouse terkontrol. Pipeline klasik yang diterapkan mencakup CLAHE, segmentasi Otsu sub-region 2×2 dengan confidence gate, proyeksi Bird’s-Eye View (BEV), deteksi centroid, perhitungan error lateral, serta pembatasan laju dan perataan EMA. Arsitektur dual-controller Jetson Nano–ESP32 memisahkan visi dari aktuasi ber-shared-control. Pengujian pada Jetson Nano memakai dua rekaman jalur greenhouse nyata: Data Uji A (bayangan sebagian, 283 frame) dan Data Uji B (tanpa bayangan, 256 frame), mencapai 62,4 dan 62,1 FPS dengan latensi rata-rata di bawah 17 ms per frame, melampaui target rancangan =15 FPS. Pada perbandingan A/B di rekaman identik, deviasi terhadap pusat citra turun 49% di zona bayangan dan 35% pada Data Uji B. Akurasi sebenarnya diuji melalui anotasi manual garis tengah jalur pada 60 frame Data Uji A: di zona bayangan Otsu sub-region lebih akurat 15,97 piksel daripada Otsu tunggal (MAE 44,02 berbanding 59,98 piksel) dan tetap unggul 6,03 piksel setelah bias masing-masing dihilangkan, sehingga perbaikan murni segmentasi. Di zona tanpa bayangan keunggulan itu murni pergeseran bias, dan sisa error didominasi bias pemasangan kamera. Validasi hardware-in-the-loop menunjukkan sinyal panduan menggerakkan kemudi secara end-to-end. Hasil berbasis video dan HIL, bukan navigasi lapangan otonom.

This research designs a real-time visual guidance system for an Unmanned Ground Vehicle (UGV) on controlled greenhouse pathways. The classical pipeline comprises CLAHE, 2 × 2 sub-region Otsu segmentation with a confidence gate, Bird’s-Eye View (BEV) projection, centroid detection, lateral error computation, rate limiting, and EMA smoothing. A dual-controller Jetson Nano–ESP32 architecture separates vision from shared-control actuation. Testing on a Jetson Nano used two real greenhouse path recordings, Test Data A (partial shadow, 283 frames) and Test Data B (no shadow, 256 frames), reaching 62.4 and 62.1 FPS with mean latency below 17 ms per frame, exceeding the =15 FPS design target. In an A/B comparison on identical footage, image-center deviation fell 49% in shadow and 35% on Test Data B. Manual centerline annotation on 60 Test Data A frames tested true accuracy: in shadow, sub-region Otsu is 15.97 pixels more accurate than single Otsu (MAE 44.02 versus 59.98 pixels) and stays 6.03 pixels better once each method’s bias is removed, so the gain comes from segmentation. Without shadow, that advantage is purely a bias shift, with residual error dominated by camera-mounting bias. Hardware-in-the-loop validation showed the signal driving steering end-to-end. Results are video- and HIL-based, not autonomous field navigation.