Seedling Area Recognition Algorithm for Autonomous Rice Planter Based on Deep Learning and Image Processing

When the global positioning system (GPS) signal is poor, GPS-based autonomous driving rice transplanters often drive over areas where seedlings have already been planted. To solve this problem, in this paper, an algorithm was proposed to distinguish planted fields by using deep learning (DL) and red-green-blue (RGB) images. The differences between the learning data and the test data, referred to as the domain gap, must be reduced. To reduce the domain gap, three methods were used in this study: domain randomization, domain normalization, and style blend. The DL model provided information regarding locations where seedlings have been planted. Next, to control the rice transplanter autonomously, a linear boundary between the cultivated and un cultivated areas was established using the RANSAC algorithm. Finally, inverse-perspective mapping was performed to obtain the bird’s eye view, which was then used to obtain the desired steering angle command of the rice transplanter. © ICROS 2023.