基于点云旋转归一化的猪体体尺测点提取方法

家畜体尺测量是现代畜牧业研究的重要课题,而体尺测点的提取是体尺测量过程中的关键。为了解决猪体点云的三维体尺测点的自动提取问题,提出了一种基于旋转归一化的猪体体尺测点提取方法。对于采集的猪体点云数据,首先利用随机采样一致性算法删除地面区域并提取目标猪体,同时存储地面法向量;然后采用基于地面法向量矫正的主成分分析方法对猪体点云进行旋转归一化,将其统一到标准测量坐标系;最后利用体尺测点的几何形态特征和测点间的结构关系提取猪体体尺测点。试验结果表明,基于地面法向量矫正的主成分分析方法可以得到较好的旋转归一化结果,使猪体的体长,体高和体宽方向基本与坐标系的x轴,y轴,z轴相一致,体长、体宽、体高、臀宽和臀高测点提取结果与实际的体尺测点测量结果位置平均误差在16 mm以内,与其他体尺测点检测方法相比,该方法可以对输入猪体的点云方向进行校正,输入的猪体点云体高方向不需要和相机坐标系的任意一个坐标轴平行,降低了原始数据获取难度。该方法可为猪体的自动化体尺测量提供参考。

Extraction method of pig body size measurement points based on rotation normalization of point cloud

Abstract: The body size of the livestock is an important indicator for breeding and animal production, however, the traditional measurement of body size of the livestock is a hard manual operation and it is inaccurate, expensive and time-consuming. In the field of computer vision measurement, the location of the pig body size measurement points is an essential work. To solve the automatic extraction of pig body size measurement points based on point clouds, a method of automatic extraction of pig body size measurement points was developed based on rotation normalization approach. In order to acquire the measurement points of the livestock, several processing steps were applied, and the steps were as follows: (1) Random sample consensus algorithm was used to acquire the plane parameter of the ground and the ground calculated by the parameter was then removed. The target pig and the ground normal vector were acquired. (2) Since the livestock acquired was from different coordinate systems, we proposed a rotation normalization method. First, the method of principal component analysis was used to get the coordinate axis of data cloud of the pig body and the initial measuring coordinate system was obtained. Secondly, the direction of the y axis was replaced by the direction of the ground normal vector acquired. The normal direction had 2 possibilities, that was, the normal vector was either pointing straight to the pig or going back away from the pig. The direction of the normal vector was determined by whether there was intersection between the constructed cuboid region and the pig. The cuboid region was constructed from the point on the ground along the normal vector. Finally, by using the geometrical relationship among the coordinate axis, standard measuring coordinate system defined in this paper was acquired. (3) Some of the measurement points of the livestock were usually the extreme points on the livestock body contour. By cutting the cloud points of the pig along the x axis direction with a step length, the slice data of the pig were acquired and the point cloud of contour line was obtained. By using the structural relation between the measurements points and the geometric feature of the measurement points either at the contour points or on the whole pig, the measurements points were calculated. Landrace sow specimens with 100 days old were selected as the experimental samples. To validate the rotation normalization performance of the proposed method, the scene point cloud including the specimen was used to evaluate the effects objectively. The point cloud was acquired by the 3D (three-dimensional) camera Xtion sensor using KinectFusion technology. Xtion is a sensor based on technology of structured light and consists of an infrared emitter, an infrared camera and an RGB (red, green, blue) camera. KinectFusion provides a function of 3D object scanning using a sensor. The results indicated that our approach could produce a reasonable rotation normalization result, and the directions of the length, height and width of the pig were basically consistent to the directions of the x, y and z axis. In order to verify the precision of measurement point''s position, the proposed method was applied to locate the neck midpoint, body length measuring point, body width measuring point, body height measuring point and hip height measuring point. Average error of the results between the automatic extraction and artificial measurement of the selected points was less than 16 mm. The method in this paper can provide a reference for the automatic body measurement of the point cloud of the pig. This work is expected to be a useful contribution for animal production and breeding genetics.