Research And Implementation On Point Cloud Acquisition Of Beef Cattle Based On Stereo Vision

Computer vision technology can get a lot of information from images under the condition of no contact with the object being measured, particularly suit for the product detection and quality comprehensive evaluation of plant and animal. The research of relevance between domestic beef cattle meat and body shape characteristic parameters based on computer vision technology plays a significant role in promoting domestic beef meat automation and improving the prediction precision and efficiency. The premise to study the relevance is point cloud acquisition. This paper takes the QinChuan cattle as the research object and proposes a method of obtaining the three-dimensional point cloud of a beef cattle in a complex situation based on stereo vision theory. In order to improve the quantity and precision of point cloud, three cameras are adopted to compose three binocular vision systems. The mainly research results contains the following three parts.(1) The camera systematic calibration: In this paper, epipolar geometry is used to remove the false matching points in the matching process and camera imaging model is used to get three-dimension point cloud. Before these operations, camera systematic calibration is required to get some necessary parameters. During the systematic calibration, single camera calibration is conducted firstly to get the intrinsic parameters, and then extrinsic parameters and the fundamental matrix of each binocular vision are obtained through calibrating the three cameras uniformly. Due to the reprojection error of each camera calibration is within sub-pixels which proved by experiments and the fundamental matrix of each binocular vision is solved based on camera parameters, the precision of the fundamental matrix and camera calibration can be ensured.(2) Beef cattle detection: To facilitate the later period of image feature point detection and matching, the cattle must be separated from complex background image. Cattle image obtained is influenced by highlight. The first step is to remove highlight. Then a skin detection algorithm based on naive Bayesian classification is implemented to get the cattle from the complex situation. Denoising and repairing operations are conducted on cattle image through image morphology. Experiments prove that highlight removal algorithm adopted can effectively remove the highlight in the cattle image. The detection algorithm based on naive Bayesian classification has relatively good results in RGB color channels. Finally beef cattle can be separated from the complex background effectively.(3) Feature points matching and point cloud acquisition: The ultimate purpose of this paper is to obtain beef cattle point cloud data. Feature points extraction and matching method based on SIFT algorithm is employed, and epipolar geometric principle is utilized to eradicate the false matching points. Then the three-dimensional coordinate can be gotten by means of camera imaging model. Experiments show that the feature point matching precision is above 70% of five groups of experimental data in three binocular vision system. Through contrast analysis, the error precision between measurements of point cloud and actual manual measured values of cattle body height and slanting length can be controlled within 4% basically.Feature points extraction and matching operation is conducted on cattle images after cattle detection in each binocular vision system. Then the three-dimensional coordinate can be gotten through matching points by means of camera imaging model. Three binocular vision systems adopted can get more matching points in some degree, and improve the density and precision of the cattle point cloud obtained consequently. The experiments show that the proposed method is feasible under the complex situation, and can achieve good point cloud data.