| Precision livestock farming is a modern way for accurate and continuous monitoring the groups and individual livestock of large-scale livestock farms.Precision livestock farming systems are applied to model and manage the efficient production of livestock,and they are the current and future development trend of livestock farming.The production of sows in commercial pig farms,the maternal ability of feeding,and the survival rate of piglets are directly related to the economic benefits of farms.The behavior of the sows’ postures and posture change are the important indicators of their health measurement,maternal ability and animal welfare.Traditional manual observations on sows and piglets have problems such as low labor,high cost,high intensity,poor real-time,subjectivity and fatigue,and contact sensor technology can cause sow welfare problems.The sensors worn on animals are easy to cause their stress reaction,and the noise generated by collision also affects the accuracy of behavior detection.The monitoring of computer vision technology is a safer,non-invasive way of harming the sow than sensor technology,and it is also a hot research area for the monitoring of livestock behavior in the precision livestock farming.In this study,the following scientific problems were extracted from the remote,noninvasive behavioral monitoring of sows and piglets in free pens: 1)accurately detecting sow posture to learn posture distribution,behavioral habits,and activity levels in a whole day;2)capturing the active and behavioral association between a sow and piglets;and 3)accurately parsing several action clips from an untrimmed long video segment,so as to learn the level of maternal ability and to further understand the animal behavior mechanism of sow posture transformation process.With these scientific problems,using Microsoft Kinect captured RGB-D datasets,this work started from the actual production application of the sow breeding,and studied posture detection of sows on frames,the behavioral association between sow and piglets,and the temporal and spatial action detection from a long video.This study was based on computer vision technology,deep learning theory and convolutional neural network.Specifically,it includes the following three aspects:(1)The study designed a posture detector based on the deep learning Faster R-CNN framework to detect the five sow postures of standing,sitting,sternal recumbency,ventral recumbency and lateral recumbency in an open free pen.This provided a research model to detect the sow posture and established a research basis for the subsequent action detection of posture change.The depth image datasets were used as a source for posture detection,and they had strong anti-interference ability to the environment of light change.It can be used for automatic monitoring of sows at night and under the hot lamp environment for 24 hours,so as to obtain their whole-day posture distribution,habits,movements,and posture changes.The classification performance of the five postures of the sow reached Accuracy of 93.6%and m AP of 87.1%.From the speed performance of the detection,the posture detector detected about 5 frames per second,which can basically meet the real-time application requirements in livestock farming.(2)The study provided a feasible and concise method for behavioral association between sows and piglets,and made a deeper analysis and further understanding of maternal behavior in large-scale breeding environments.A total of 540 observation scans were performed for 9time periods selected in the experiment.During each scan,the sow postures were detected by the Faster R-CNN detector,and the sow’s activity index was estimated by optical flow field images.Meanwhile,the features of piglets’ activity were acquired by the optical flow field images,and the activity index of piglets was estimated by the linear regression model.Finally,the correlation analysis between the sows’ and the piglets’ activity index was carried out.The results showed that the system could capture the activity association between the sow and the piglets.When the sow was in the posture of lateral recumbency,the activity index of the piglets dropped to a very low level,and when the sow was standing or walking,the activity of the piglets was greatly enhanced.These results were consistent with the viewpoints of animal behavior studies.(3)The study proposed a method for detecting the posture change of sows,which was suitable for automatic analysis of sow maternal ability in large-scale breeding environment.It adopted new ideas combining with frame-level posture detection,temporal action localization and spatio-temporal action tube to achieve the spatio-temporal action detection of sow posture change.For the untrimmed long segment of depth video after acquisition,the improved Faster R-CNN posture detector was used for frame-level posture detection,and the action clips with posture change were analyzed by the coarse-to-fine temporal action localization.The action clips were generally within 1～2 minutes.Then the action tube optimization method was used to obtain the spatio-temporal action tube in the localized clip,which could obtain further video-level interesting action clips and spatio-temporal tubes.By comparing the results of the two sows in two-day posture change,the sow action information with better maternal ability could be obtained.The innovations of the study include the following aspects: a)innovatively proposed and implemented a deep learning detector from depth images for sow posture detection;b)innovatively presented and implemented an automatical analysis method on behavior association between sows and piglets;and c)innovatively proposed and implemented a research method for detecting the behavior of sow posture change at the video level. |
PDF Full Text Request