基于DeepLabCut算法的猪只体尺快速测量方法研究

Rapid Measurements of Pig Body Size Based on DeepLabCut Algorithm

为解决基于计算机视觉猪只体尺测量过程中存在的对猪只姿态依赖度高、测定效率低等问题,提出了一种基于DeepLabCut算法的非接触式猪只体尺快速测量方法。本研究以长白猪为研究对象,使用RealSense L515深感相机作为图像数据采集单元获取猪只背部RGB-D数据,通过分析对比ResNet、MobileNet-V2、EfficientNet系列的10个主干网络训练效果,选取EfficientNet-b6模型作为DeepLabCut算法最优主干网络进行猪只体尺特征点检测;为实现猪只体尺数据的精准计算,本文采用SVM模型识别猪只站立姿态,筛选猪只自然站立状态;在此基础上,采用深度数据临近区域替换算法对离群特征点进行优化,并计算猪只体长、体宽、体高、臀宽和臀高5项体尺指标。经对140组猪只图像进行测试发现,本研究提出的算法可实现猪只自然站立姿态下体尺的实时、精准测量,体尺最大均方根误差为1.79 cm,计算耗时为每帧0.27 s。

At present,the computer vision-based pig body measurement shows a high dependence on pig posture and low measurement efficiency.To solve these problems,a rapid and non-contact pig body size measurement method based on DeepLabCut was proposed.The top view RGB-D images of landrace pigs were captured by a RealSense L515 camera.The training effects of 10 backbone networks of ReNet,MobileNet-V2,and EfficientNet series were compared and analyzed,and then the EfficientNet-b6 model was selected as the optimal backbone network of DeepLabCut algorithm for feature point detection of pig body size.In order to achieve accurate calculation of pig body size data,SVM model was used to identify the standing stance of pigs and screen the natural standing stance of pigs.Based on this,the depth-valued proximity region replacement algorithm was used to optimize the outlier feature points and calculate the five body size indexes of pig body length,body width,body height,rump width and rump height by Euclidean distance.This method was tested on 140 groups of standing images of pigs,and it was found that the algorithm could achieve real-time and accurate measurement of body size in the natural standing posture of pigs,with maximum root mean square error of 1.79 cm and computation time of 0.27 s per frame.