Study On Individual Recognition And Drinking Behavior Analysis Of Topview Group-housed Pigs Based On Machine Vision

At present,pig-breeding industry in China is marching toward the orientation of large scale,professionalization,intelligence and refinement.The application of monitoring devices in farms is one of the principal approaches to realize the professional and intelligent pig breading.In addition,large scale and refinement development require breeders to be able to recognize and analyze the individual behaviors of group-housed pigs to improve their welfares specifically.However,it is a laborious job and inefficient for breeders to monitor multiple screens simultaneously in the long term.Therefore,it is of great theoretical value and great application potential for China,the greatest country of pig breeding and pork consumption in the world,to conduct the in-depth research on foreground detection,object extraction,recognition and behavior analysis on individual topview group-housed pigs in video sequences based on machine vision.The research subject,sourced from National Natural Science Foundation of China,is centered on surveillance video of topview group-housed pigs to conduct the research on foreground detection,object extraction,identity recognition and drinking behavior analysis.The primary innovative research achievements of this paper include:(1)A foreground detection method based on the combination of Mixture of Gaussians(MoG)using prediction mechanism(PM)and threshold segmentation algorithm is proposed.The experiment is on the basis of topview group-housed pigs in video sequences in complex scenes.These complex scenes include light changes,the influence of ground water stains and urine stains,pigs’ slow movement patterns,and varying colors of pigs.First,the “valid area”,namely pigs’ movement area,is manually set to obtain the foreground objects of individual pigs accurately.Second,the foreground objects of topview group-housed pigs are detected using the PM-MoG algorithm.At the parameter updating stage of background model,the algorithm uses the detected binary image of the previous frame to predict the current frame in the valid area for pixels that fulfill background updating conditions.Different update strategies are used to update the background for different circumstances.In addition,a foreground detection result is obtained using the maximum entropy threshold segmentation algorithm and mathematical morphology processing on images in “valid area”.Finally,the results of the two previous steps of foreground detection are fused.The experimental results show that the method is effective and can extract relatively complete foreground objects of pigs in complex scenes.(2)A multi-object extraction method for topview group-housed pig breeding based on adaptive partitioning and multilevel thresholding segmentation is proposed.Given the real-time inefficiency of PM-MOG algorithm and incompetence to process issue of pig adhesion effectively,a single frame image is processed to extract multiple group-housed pig objects standing in a standard position.First,“valid area” is further set to divide pigs’ movement area into drinker area,feeder area and resting area.Each of these areas relates to a specific behavior of pigs.For instance,pigs like to huddle in a corner to sleep in the resting area;pigs in the feeder and drinker area always stand in a standard position;and pigs who are drinking or feeding often do not like to be disturbed.All of these pig behaviors have provided the favorable conditions for this method.Second,a single frame image is enhanced using histogram equalisation,and then it is segmented with a maximum entropy global threshold.The initial segmentation objects are obtained by mathematical morphology processing.Then,each object centroid is calculated from the initially segmented objects,and the original image is adaptively divided into multiple circular sub-blocks whose origin is the centroid and radius is the maximum distance from the centroid to the edge point.Finally,an accurate secondary segmentation result is obtained using multilevel thresholding segmentation in each subblock.The experiment result shows that this method could accurately extract individual pigs from a drinker and feeder area with good real-time performance.This research lays a foundation for the subsequent individual recognition and behavior analysis of group-housed pigs.(3)An individual recognition method for topview group-housed pigs based on(Isometric mapping algorithm)ISOMAP algorithm and support vector machine(SVM)is proposed.First,to obtain each individual pig,foreground detection and object extraction are conducted on a top-view video sequence of the group-housed pigs.Second,the training samples are established,and the color,texture and shape of the individual pig are extracted;through the combination of these features,a feature vector representing an individual pig is then built with feature normalization.Then,the combined features are fused using the ISOMAP algorithm,which reduces the feature dimension on the basis of the maximum retention of the effective recognition information.Finally,the features are trained and recognized using a support vector machine(SVM)classifier with an optimal kernel function.The videos used in the present study are topview group-housed pigs without adhesion;and a higher recognition rate is obtained,showing ISOMAP algorithm’s advantage in feature fusion.SVM based on mixed kernel function has a better classification effect than SVM based on single kernel function.In this paper,a method for recognizing group-housed pigs individually from a top-view video sequence is explored based on the machine vision,which differs from traditional radio frequency identification(RFID)of individual pig in a disturbing way.This study provides a new idea for the recognition of individual pig without stressing the animals.(4)A method to recognize individual pig in drinker area using machine vision technology and to analyze whether the pig is drinking is proposed.Water use could be an important indicator to judge pigs’ health status.First,individual pig is extracted from video sequence of topview group-housed pig breeding.Second,the distance between individual pig and water tap is obtained to determine whether the pig contacts a water tap or not.If it contacts,color moment,area,perimeter and other optimal features are extracted with data normalization.Then,individual pig is recognized via its Euclidean distance with standard sample.Finally,the contact time between individual pig and water tap is used to judge whether the pig is drinking.The experiment results show that this method has its advantages.First,this method analyzes and recognizes individual pig in drinker area with only one fixed water tap,indicating no influence of pig cohesion in other areas on the analysis and recognition to overcome the inefficiency of manually selected test images.Second,the influence of various pig positions on recognition result of individual pig is avoided.Third,feature extraction is further optimized,the recognition process is simplified;and the real-time performance is upgraded.Based on the experiment results,breeders could monitor water use of individual pig with reduced workload;predict disease outbreak related to water use in advance for the reference of further exploration of other behaviors of group-housed individual pigs.The fast-changing IT development and the decreasing price and increasing cost performance of monitoring devices have laid a hardware foundation for the intelligent development of breeding industry.Meanwhile,national policies have been formulated in succession to encourage and guide the intelligent development of breeding industry.The research achievement of this paper is innovative in constructing and completing a computer-based visual surveillance system of topview group-housed pigs,with a great scientific value and reference significance to upgrade the intelligent management level of pig breeding industry and improve pigs’ welfare.