AI-Enabled Autonomous Visual Navigation in Complex and Data-Constrained Agricultural Environments by Leveraging Image Augmentation and Translation Methods
Semester of Graduation
Spring 2026
Degree Type
Thesis
Degree Name
Masters of Science in Computer Science
Department
College of Computing & Software Engineering
Committee Chair/First Advisor
Taeyeong Choi
Second Advisor
Maria Valero de Clemente
Third Advisor
Jian Zhang
Abstract
The field of agricultural robotics remains one of the most underexplored domains within the broader robotics community. As a result, there is a scarcity of robust, scalable, and adaptable robotic systems capable of operating under the highly variable conditions of real-world farms. Visual navigation is essential for field autonomy, yet state-of-the-art perception models rely heavily on large annotated datasets and consistent environmental conditions. This thesis addresses these limitations by proposing two novel frameworks to enhance model robustness against structural occlusions, environmental variability, and low-visibility conditions under nighttime settings. First, I introduce Crop-Aligned Cutout (CA-Cut), a spatially guided augmentation method that simulates realistic under-canopy disturbances by placing masks along crop rows. Experiments demonstrate that CA-Cut impoves keypoint prediction accuracy by up to 36.9%, significantly enhancing model performance under severe occlusions. Second, to enable 24-hour operations, I propose an unsupervised day-to-night cross-modal image translation framework. By leveraging a CLIP-enhanced CycleGAN and a visibility masking strategy, this framework converts daytime RGB images into realistic nighttime near-infrared (NIR) counterparts. This allows for the reuse of daytime labels to train nighttime perception models, a process validated through our newly created dataset, AgriNight dataset and a physical robot deployment. Together, these contributions provide a comprehensive pathway for reliable autonomous navigation in diverse and data-constrained agricultural environments.
This is the signed approval of my thesis committee.
Previous Versions
Incorrect file (withdrawn)