Abstract:
Abstract: Digital plant is the key of technology for agricultural digitization and informatization, and it has become a hot research field both at home and abroad. This technology can analyze the morphological structure and growth process of plants with the aid of the digital and visual three-dimensional (3D) model. The technology can analyze growth and development status for plants and evaluate the type characteristics of plants in real time, and hence guide agricultural production. It can provide powerful conditions for agricultural mechanization and automation production. Obtaining 3D point cloud data of plants is the difficulty and key for the plant digital and visual modeling. Domestic and foreign research focused on 3D laser scanning to obtain 3D point cloud data of plant, and achieved some results. High-precision 3D laser scanning can obtain accurate 3D point cloud data for the reconstruction of plants and realize the digitization of plants, but this method needs expensive equipment. This processing algorithm is complex and has poor real-time performance, because of a large amount of data. This method is unfavorable to the application and popularization. To design a low cost and good practicability method for obtaining plant 3D point cloud data had important research value and significance. TOF (time of flight) depth sensor was a new method for depth information acquisition based on the principle of time of flight. The method could quickly acquire the depth information of the image, the depth calculated was not affected by surface gray level and characteristics, and the calculation accuracy did not change with distance. There were few researches on the application of TOF depth sensor in plant digitalization and visualization, and the research field abroad was just in its infancy. In order to improve the accuracy of 3D reconstruction for plants and realize the digitization of plants, the obtaining and denoising method of plants 3D point cloud data was proposed based on TOF depth sensor. This method possessed the advantages of low cost and good practicability. Firstly, the plant point cloud data were obtained through TOF depth sensor, and then pretreated by direct pass filter in order to reduce the background noise. Secondly, the improved outlier denoising algorithm based on density analysis was combined with neighboring average distance and neighborhood point number, and these were 2 characteristic parameters. The algorithm of outlier could detect and denoise the outliers of point cloud data. Finally, small-size noise in the point cloud was detected and denoised by bilateral filtering algorithm. Tomato plants were selected to carry out experiment. Compared with the traditional bilateral filtering algorithm, experiments showed that the maximum error was reduced by 11.2%, and the average error was reduced by 23.2%. Compared with the Laplace filtering algorithm, experiments showed that the maximum error was reduced by 20.6%, and the average error was reduced by 39.2%. It showed that the proposed algorithm could simply and effectively denoise noise with different scales in 3D point cloud data under the condition of keeping the characteristics of point cloud. The point cloud data are large for 3D reconstruction process, and therefore the artificial intelligence method can analyze the point cloud data in the further study.