基于目标特征与决策树支持向量机的生猪体态识别

The Recognition of Pig Posture Based on Target Features and Decision Tree Support Vector Machine

为实现生猪异常行为的自动化监测,提出了一种数字化表示生猪目标特征的体态识别研究。首先对猪场采集到的视频图像采用改进的Grabcut分割算法进行生猪目标提取;然后基于生猪轮廓图像建立生猪目标特征集,包括圆形度、矩形度和Hu不变矩等12个特征;并利用类内类间距离判据对样本数据建立的特征集进行特征优选;最后构建决策树支持向量机(DT-SVM)对生猪体态进行分层识别。实验结果表明,选择的最优特征集可以有效地表征生猪体态信息,DT-SVM对单只猪的站立、躺卧和扎堆猪的适度扎堆、过度扎堆都有较高的识别率,为进一步探索生猪异常行为分析奠定了基础。

In order to realize the automatic monitoring of pigs＇ abnormal behavior,the research on the recognition of pigs＇ posture is proposed based on digitized target features.Firstly,the improved Grabcut segmentation algorithm is used to obtain the target of pigs from the video images collected in the farm.Then,the target features set of pigs is established based on the contour images of pigs,including twelve features,that is,circularity,rectanglarity and Hu invariant moments.The feature selection of the feature set based on samples is judged by within-class and among-class distance criteria.Finally,a decision tree support vector machine（DT-SVM） is constructed to identify the posture of pigs hierarchically.The experimental results show that the optimal features selected can effectively describe the posture of pigs and DT-SVM has higher recognition rate for four postures,which are respectively standing and lying of single pig,the proper stacking and over-stacking of pigs.This study lays a foundation for further analysis of the abnormal behavior of pigs.