基于机器学习的奶牛深度图像身体区域精细分割方法

摘要

The recognition of cows' body parts is essential for providing accurate details of the cows' shape,which is the fundamental prerequisite for locomotion scoring,posture detection and behavioral quantifications.The objective was to develop a robust depth feature in order to reduce the difficulty in building the classifier and detect cows' body parts with higher accuracy.Therefore,a method for segmenting cows' body parts was proposed,including the head,neck,body,forelimbs,hind limbs and tail,with high accuracy on the basis of depth image processing and machine learning.The local binary patterns of each pixel under several sampling radii were used as the features with which the filtering rules were designed,and a decision forest was trained and tested to classify the pixels into six groups.Furthermore,totally 288 depth images were captured from 30 cows;150 images were randomly selected to build three decision trees,and the rest images were used for testing.The results showed that when the number of sampling radii and training layers were 30 and 20,respectively,the recognition rate reached 95.15％.Among the cows' body parts,the recognition rate of tail was 54.97％,and the minimum recognition rate of other parts was 89.22％.In some cases that tail was too close to trunk to segment tail from trunk by human marker,the decision trees recognized the tail successfully.The average recognition time for pixel were 0.38 ms and 0.25 ms,and the recognition time for cow target were 20.30 s and 15.25 s for the conventional method and new method,respectively.This LBP-based depth image feature was translation-invariant and rotation-invariant and had fewer parameters.The results showed that the new method proposed was more effective in recognizing small and complex structures of the cow target with higher accuracy.Compared with the typical depth image features,the new feature employed was capable of extracting the details of cows' body and recognizing complex parts more accurately with fewer parameters and simple model.%奶牛目标各区域的精细分割和识别能够提供精确的奶牛形体细节信息,是奶牛体形评价、姿态检测、行为分析和理解的前提和基础.为实现深度图像中奶牛头、颈、躯干和四肢等身体区域的精确分割,提出一种基于深度图像特征和机器学习的奶牛目标各区域精细分割方法.该方法以每个像素点在不同采样半径下的带阈值LBP序列为深度特征值,设置分类约束条件,用决策树森林机器学习方法实现奶牛各区域的精细分类.对10头奶牛的288幅侧视深度图像进行试验,结果表明,当采样半径分段数为30,决策树训练至20层时,奶牛整体各像素点的平均识别率为95.15％,较传统深度图像特征值有更强的细节信息提取能力,可以用较少参数实现对复杂结构的精确识别.