Simultaneous detection of reference lines in paddy fields using a machine vision-based framework

Accurate and robust detection of reference lines in the field is essential for formulating linear tracking and steering strategies for agricultural machinery. The identification of planted and ridge areas poses challenges due to their similar appearance to unplanted areas, along with complex field conditions such as uneven illumination. In this study, we proposed a machine vision-based framework for the simultaneous detection of auxiliary navigation lines and ridge boundary lines. Firstly, we constructed a multi-area Paddy Area Segmentation dataset named PASeg, which contained ridge areas, planted areas, and unplanted areas. Additionally, a deep learning network called G-STDC that integrated the Ghost module into the STDC network was developed for efficient and robust area segmentation. Finally, a multi-line detection method was applied based on a central axis-based point clustering algorithm and random sample consensus (RANSAC) algorithm to extract reference lines. According to the results on PASeg, the proposed G-STDC model obtained a mean intersection over union (mIoU) of 95.23 %, outperforming the baseline model (with the mIoU of 93.18 %). The attitude error and distance error of line extraction on 640 × 512 resolution images were within 0.776° and 4.687 pixels, respectively. The overall detection speed reached 13.9 frames per second (FPS), while the faster G-STDC model (G-STDC-t) achieved 16.7 FPS. The proposed method could provide real-time reference lines for turning path planning and automatic navigation in agro-machinery.