Kinect-based 3D Reconstruction of Leaf Lettuce

Abstract. Existing 3D reconstruction methods of plants require manual adjustment of parameters, which cannot meet the needs of rapid and high-throughput reconstruction. This paper takes leaf lettuce as the object, captures leaf lettuce point cloud data from multiple perspectives by the Kinect data acquisition platform, and conducts research on automatic 3D reconstruction of leaf lettuce. Firstly, the platform was used to capture RGB images and depth images of leaf lettuce from multiple perspectives, and the original 3D point cloud data from multi-view were obtained by fusing RGB information and depth information. Secondly, the process of registration by using PCL tools was analyzed, through numerous repetitions of registration experiment between two pieces of point clouds, the paper thoroughly studied the effects of relevant algorithms and their parameters on two point clouds‘ registration accuracy. Accordingly, an evaluation system of registration effect was proposed to carry out automatic registration of two point clouds, and a corresponding algorithm was developed. Then, based on the automatic registration algorithm of two point clouds, a global reconstruction algorithm with multi-view point clouds was investigated and developed. System tests showed that the automatic registration algorithm could greatly reduce time consumption and improve the accuracy and stability of registration results. Based on pairwise registrations, the global reconstruction was carried out according to certain point clouds sequence. The results showed that error accumulation was reduced by distributing evenly to the multi-perspective point clouds registration, and the registration success rate reached 93.3% within the error range of 0.0015 m.