RGB-D Detection and Localization for Agricultural Robotics in Low-Resource Settings

Agriculture in low-resource settings faces challenges such as limited access to electricity, low profits, variability in environmental conditions, and high costs. Machine learning (ML) offers solutions to these issues, particularly in tasks like automated fruit picking. This paper focuses on the detection and 3D localization of apples using YOLOv5n and YOLOv10n models, which are lightweight variants suitable for deployment in resource-constrained environments. We provide in-depth details about our fruit dataset and annotation protocols to elucidate the diversity and scalability of our approach. Further, we conduct a broader comparison with other state-of-the-art models to substantiate our claims of superior performance. Our experiments, conducted with an Intel RealSense camera, include detailed depth perception analysis under various conditions, revealing that YOLOv10n outperforms YOLOv5n in both detection accuracy and 3D localization precision. We also address potential deployment challenges such as hardware costs, environmental variability, and energy efficiency, and discuss limitations to guide future research. Our results suggest that YOLOv10n offers a highly effective solution for real-time, reliable fruit detection in low-resource agricultural settings. Our code is available here.