Crop Row Detection Algorithm Based on 3-D LiDAR: Suitable for Crop Row Detection in Different Periods

The 2D visual inspection algorithm improves the performance of crop row detection in downstream tasks, but it is easily affected by weather, shadow and other factors, and it is difficult to detect crop rows with severe occlusion. Considering the problem that crop pixel features are difficult to separate under complex field conditions, it is still a very challenging problem to ensure the real-time operation of agricultural machinery. In this paper, a crop row detection method based on 3D LiDAR is proposed. It is divided into three procedures: point cloud preprocessing, feature point extraction, and crop canopy row centerline detection. Aiming at the influence of mutual occlusion between crop rows on crop row detection, a double thresholds method is proposed for extracting crop canopy point clouds, i.e., height threshold for the near area and density threshold for the far area. Aiming at the problem that crops are not planted in a straight line and the growth direction of crops is uncontrollable, a method of dynamically dividing clustering areas in horizontal strips is proposed to determine the feature points of crop canopy rows. Finally, the least square method is used to fit the centerline of crop canopy row. The results of the experiments demonstrate this algorithm is capable of detecting the centerlines of maize rows at the V4, V7, VT, and R4 growth periods, and the average correct detection rate of the four periods is greater than 86%, and the average processing time is less than 120ms, which meets the requirements of precision and real-time field assisted driving of agricultural machinery. The superiority of the proposed algorithm is proven by comparative experiments, and it is less sensitive to the density of the crop. The proposed algorithm provides technical support for alignment-assisted driving of field management machinery.