Efficient crop row detection using transformer-based parameter prediction

Abstract

The detection of crop rows is crucial for achieving visual navigation and is one of the key technologies for enabling autonomous management of maize fields. However, the current mainstream approach to maize crop row detection often involves two steps - feature extraction followed by post-processing. While useful, this method is inefficient, and the heuristic rules designed by humans limit the scalability of these methods. To simplify the solution and enhance its generality, crop row detection is defined as a process of approximating curves. Polynomial parameter learning is adopted to constrain the parameters of crop row shapes, and utilise a model built on the Transformer architecture to learn the elongated structures and global context of crop rows, achieving end-to-end output of crop row shape parameters. The proposed approach has achieved rapid and excellent detection results in complex field environments, even in the presence of curved crop rows.