One-shot classification-based tilled soil region segmentation for boundary guidance in autonomous tillage

Abstract

In recent years, machine vision has been frequently used to estimate local paths for autonomous tractors. The incorporation of deep learning-based techniques to its methodology is expected to induce robust performance under practical field conditions. Typical tillage of autonomous tractors requires the tilled soil region detection to provide the local paths, which are usually determined based on the boundaries between tilled and non-tilled soil. However, tilled soil regions usually exhibit different feature distributions by field (domain) in terms of factors like soil characteristics, crop type, and method of tillage. This makes the development of deep learning-based systems with highly generalizable performance difficult. In this study, this issue is addressed by proposing a few-shot learning (FSL)-based effective tilled soil region segmentation, which is utilized to detect tilled soil region in 2D perspective scenes to provide tillage path guidance to autonomous tractors. The proposed approach used an image patch classification-based field region segmentation pipeline as the backbone and a one-shot classification model was designed to classify each image patch as tilled or non-tilled soil. The tilled soil region is then segmented by clustering the image patches belonging to the corresponding classes. The results demonstrate that one-shot classification exhibits a classification accuracy exceeding 0.9 over the entire dataset obtained from various fields, including new data, while the typical classification method exhibits a classification accuracy of approximately 0.74 on new data. Further, segmentation is successfully performed on the detected tilled soil regions belonging to various domains. In conclusion, the proposed method can provide the relative feature difference between tilled and non-tilled soils, allowing for expansion to new data without fine-tuning. This robust field region segmentation can be utilized practically as an essential skill to support the tillage path decision for an autonomous tractor.