Experimental Study on Navigation for Wheat Seedling Root Cutting Based on Deep Learning

Abstract

In response to the automatic extraction of navigation lines for wheat root cutting, this paper conducted field experiments and analyses on the navigation line extraction algorithm, based on the improved YOLOv5 algorithm. Firstly, based on the characteristics of wheat seedling rows during the wheat rejuvenation period, the YOLOv5 algorithm was improved by using rotation detection box labels, and navigation lines were extracted by fitting the detection boxes using clustering methods. Then, an experimental system was established to conduct field experiments on the algorithm: (1) Tests were conducted at three speeds of 0.5 m/s, 1.0 m/s and 1.5 m/s respectively, and the position error of the root cutter was measured and analyzed, indicating that the actual navigation path position error increased with the speed. The best navigation performance was observed at 1 m/s, with an average positional error of 18.56 mm, meeting the requirements for wheat root cutting. (2) Robustness analysis of the algorithm was conducted using data collected from 2019 to 2022. Comparative tests were conducted from four aspects: different years, different time periods, different environments, and different yaw angles. The results showed that the algorithm proposed in this paper has stronger robustness and higher accuracy.