基于两阶段关键点定位算法的奶牛体型评定指标自动测量

Automatic measurement of cow linear appraisal indicators based on two-stage key point locating algorithm

为解决传统奶牛体型评定指标测量方法受主观影响大、自动化程度以及体型关键点定位存在误差等问题,提出一种基于两阶段关键点定位算法的奶牛体型评定指标自动测量方法。对采集的奶牛背部深度图像序列,首先基于滤波方法进行边缘平滑与缺失区域修补;之后基于YOLO v5体型关键区域检测算法确定体型关键区域并重建相关区域三维点云;进而计算区域点云曲率与z轴最值定位体型关键点;最后依据关键点间相对位置自动测得体型评定指标。结果表明,该方法可完成俯视视角下奶牛体长、肩宽、胸宽、腹宽和腰宽指标的精准测量。对15头奶牛5个体型评定指标,算法测量值与实测值平均绝对误差为1.55 cm,均方根误差为1.78 cm,决定系数R2最大为0.9394。该方法可在实际养殖环境下实现奶牛体型评定指标的精准测量,对生产实际具有一定现实意义。

Aiming at solving the problems of the traditional measurement methods of cow linear appraisal indicators, such as subjective influence, low degree of automation, and errors in locating the key points of linear appraisal, an automatic measurement method of cow linear appraisal indicators based on two-stage key point locating algorithm was proposed. For the depth image sequence of the cow back, the edge smoothing and missing area repairment were performed based on filtering method.Then, the key area detection algorithm based on YOLO v5 was used to determine the key areas of linear appraisal and reconstruct the 3D point cloud model of relevant area. Next, the key points of linear appraisal were located by calculating the regional point cloud curvature and the maximum value of z-axis.Finally, the linear appraisal indicators were obtained according to the relative position of key points. The experimental results showed that the method could accurately measure the body length, shoulder width, chest width, belly width and waist width indicators of dairy cows from an overhead perspective.For the five linear appraisal indicators of 15 dairy cows, the average absolute error was 1.55 cm, the root mean square error was 1.78 cm between the algorithm measurement value and the actual measurement value, and the maximum coefficient of determination R~2 was 0.9394. The experiment proved that the method could realize the accurate measure-ment of cow linear appraisal indicators in the actual breeding environment, which had certain practical significance for production.