| The behavior of livestock is closely related to their health status and living environment.It is a simple,easy to understand and commonly used method to evaluate the welfare of animals through behavior analysis.The core of precision livestock farming is the intelligent monitoring and automatic analysis of individual information and behavior.Continuous and automatic monitoring of chickens through advanced information technology is a crucial problem to be solved in precision chicken breeding.Towards the goal of the automatic monitoring of chickens' behavior in free-range feeding mode,this paper studied some key issues,such as deep learning for behavior recognition method,multiple object tracking method and the behavior indicators and abnormal detection.(1)The behavior recognition method of free-range chickens in the hen house was studied.To solve the problem of class imbalance in data set,image processing algorithm is used to enhance data.Aiming at the real-time requirement and the large number of hens and occlusion problems,YOLO V4 target detection algorithm is adopted and repulsion loss function is added to the original loss function.Automatic recognition of six basic behaviors of feeding,drinking,standing,stretching,lying and preening and two abnormal behaviors of feather pecking and fighting,the mean average precision reached 93.9%,and the average precision rates of each behavior were 98.6%of feeding,93.7%of drinking,95.6%of standing,88.9%of stretching,89.7%of lying,89.5%of preening,95.6%of feather pecking and 99.8%of fighting.(2)The multiple individual chickens tracking in the hen house is studied.To deal with the problems of motion blur and non-rigid deformation when chickens move fast,expand the bounding box of the Deep SORT algorithm,and using DIoU instead of IoU to match the predict box.It can track well while chickens fast moving,sudden change of direction and severe non-rigid deformation.The multiple object tracking accuracy is reached 89.3%.It is more robust than SORT algorithm and original Deep SORT algorithm,and meets the requirement of real-time processing.(3)Based on behavior recognition,the behavior indicators and the abnormal alarm methods are proposed.YOLO v4 algorithm was used to identify chickens' behaviors from video data of 44 days.The number of 8 kinds of behaviors was analyzed statistically by the frame,hour and day respectively.According to the feeding behavior characteristics of chickens before and after feeding,the abnormal feeding situation was alerted.According to the statistics of drinking behavior,it alarms the abnormal water supply.Eating index was established based on the historical data of feeding and drinking to alarm the abnormal eating of chickens.According to the time duration characteristics of feather pecking behavior,algorithm was developed to improve the recognition accuracy,the number of feather pecking behavior over 30s can be recorded.The distribution index was established to estimate the clustering condition in the he house and the activities of chickens inside and outside the hen house.(4)Behavior monitoring indicators based on individual chicken movement speed and group average speed were put forward,and abnormal alarms were carried out.By expanding prediction box and DIoU matching,the improved Deep SORT algorithm was used to track multiple chickens,and the individual movement speed and the average movement speed of the group were calculated.Based on the analysis of individual and group movement speed,an alarm can be made for sick or dead chickens.The scare alarm can be done by the analysis of the average movement velocity of the group.Based on two fast-moving two individuals,the distance between them and the identification of fighting behavior,the fight alarm can be made.The activity index is established by moving speed.In this paper,the object detection and tracking technology based on deep learning was used to study the behavior of free-range chickens in the hen house.Through a large number of data analysis,the behavior characteristics and rules of chicken s were found out,and the indicators of behavior and activity monitoring were put forward.A chicken behavior monitoring platform has been developed to monitor chicken' behaviors automatically,continuously and in real-time.It can alarm for abnormal conditions in time,manage and control the feeding process continuously.Use deep learning method to identify and monitor chicken behaviors,which provides a new technical development direction for precision livestock farming. |
PDF Full Text Request