基于机器视觉的母猪哺乳行为监测方法研究

Research on monitoring method of sow’s nursing behaviour based on machine vision

[目的]针对母猪哺乳行为及哺乳时长的自动化监测水平低、人工观测费时费力等问题,提出一种基于YOLOv5结合db4小波的方法,实现非接触式母猪哺乳行为的判定和哺乳时长的监测。[方法]利用YOLOv5对目标母猪和仔猪进行识别并输出母猪姿态,选择姿态为侧卧的母猪获取其预测框面积,根据哺乳特征的预测框面积变化与哺乳行为建立对应关系,综合判定母猪的哺乳行为和哺乳时长;利用高、低通滤波器对母猪预测框面积进行下采样卷积,判定母猪哺乳行为和哺乳时长;对比加入db4小波前、后的识别准确率。[结果]母猪哺乳行为监测模型的精度均值和召回率分别为94.62%和93.70%,在加入db4小波前、后对哺乳行为时长判定的平均准确率为93.52%和96.04%,对清晰度为720P的视频平均监测速度分别为23.89和19.35 f·s^(-1)。[结论]深层卷积神经网络模型结合db4小波为判定母猪的哺乳行为和哺乳时长提供技术支撑,识别准确率和监测速度均可满足猪场实际需求。

[Objectives]Aiming at the low level of automatic monitoring of sows’nursing behaviour and nursing duration,time-consuming and laborious manual observation,a method based on YOLOv5 combined with db4 wavelet was proposed to realize the judgment of non-contact sows’nursing behavior and nursing duration monitoring.[Methods]YOLOv5 was used to identify the target sows and piglets and output the sows posture.The sows with lateral posture were selected to obtain the predictive frame area,and the corresponding relationship was established according to the changes in the predictive frame area of lactation characteristics and nursing behavior,and the nursing behavior and duration of sows were comprehensively determined.The sows’nursing behavior and duration were determined by down-sampling convolution of the sows’predictive frame area with high and low pass filters.The recognition accuracy before and after adding db4 wavelet was compared.[Results]The average accuracy and recall rate of the sows’nursing behavior monitoring model were 94.62%and 93.70%,respectively.The average accuracy of the determination of nursing behavior duration before and after db4 wavelet was added was 93.52%and 96.04%,respectively.The average monitoring speed of 720P video was 23.89 and 19.35 f·s^(-1),respectively.[Conclusions]The deep convolutional neural network model combined with db4 wavelet provides technical support for judging the nursing behavior and duration of sows,and the recognition accuracy and monitoring speed can meet the actual needs of pig farms.