Automated detection of boundary line in paddy field using MobileV2-UNet and RANSAC

The detection of boundary lines in farmlands is critical for precision agriculture and automatic navigation. This study proposed a novel method for automatic detection of boundary lines in paddy fields based on images acquired from a vision system. These images were collected from different environmental conditions with wheel marks, shadows, weeds, and uneven illumination in paddy fields. To alleviate the effect of this environmental noise on the detection of boundary lines, the proposed method was designed with the required robustness, which included two sequentially linked phases: farmland area segmentation and boundary line detection. For the segmentation of the farmland area, this study proposed an effective deep learning model, called MobileV2-UNet, which used modified inverted residual blocks and the dilated convolution to achieve the accurate segmentation of the farmland area and nonfarmland area. For the detection of farmland boundary lines, a multiboundary detection method based on the frame correlation and random sample consensus (RANSAC) algorithm was applied to detect the side boundary and end boundary, which could provide critical information for agricultural machinery steering. Results showed that the mean intersection over union (mIoU) for area segmentation reached 0.908, and the average angular and vertical errors for boundary line detection were 0.865° and 0.021, respectively. Moreover, the processing speed reached 8 frames per second, which could meet the real-time work demands of agricultural machinery.