Between Detection and Segmentation: A Row Selection Based Detector for Crop Line Detection in Paddy Field

Visual-based navigation in fields is mainly achieved by following crop lines, so crop line detection is a fundamental task of agricultural automation. The main challenge for crop line detection in paddy fields is to overcome challenges from a semi-constrained environment. It means that the surroundings of the paddy field are variable and exist many distractions. In a typical paddy field, UGVs may encounter brightness variation, surface reflection, and irregular crop seedling placement which can cause row detection inaccuracy. We hope to use the powerful deep learning technique to overcome environmental distractions that can affect the accuracy of crop line detection. Inspired by similar lane detection task for auto driving in semiconstrained environments. This article proposed a crop line detection model based on the row selection method and auxiliary learning technique. Unlike previous works regarding crop line detection as a semantic segmentation task or object detection task, using the row selection method can directly distinguish each line in the field and result in a lower computational cost. As there is no publicly available large-scale dataset for crop line detection in the paddy field, this work also makes a small dataset containing crop line in various growing conditions and propose an offline data argumentation called LineMix. Experiments were conducted using the dataset we create and the results confirmed that the proposed model achieves better accuracy and less computational cost compared with the semantic segmentation method and object detection method which achieve 92.6 in the accuracy metric and 84.4 fps when inference.