| Forage intake of sheep can reflect their health status to some extent,and real-time and continuous monitoring of feed intake of sheep is one of the effective measures to improve the breeding efficiency.Previous studies have shown that there is a high correlation between the short-time chewing behavior of ruminants and their forage intake.Here,the short-time chewing behavior refers to the behavior corresponding to a single maxor-mandible tension for the purpose of grinding grass/forage.In the grazing or house feeding environment,the short-time mastication behavior recognition model based on the data of jaw pressure and acoustic signal generated by chewing and grinding of sheep with wearable sensing equipment has been experimented and initially verified,and has achieved good recognition effect,but this contact data acquisition method has some shortcomings,such as difficult battery replacement,easy device damage and low data stability.Automatic recognition and analysis of individual sheep behavior by using computer vision technology is one of the important methods of animal behavior recognition at present.However,the existing research on automatic monitoring of sheep diet-related behaviors based on computer vision technology mainly focuses on identifying animals’ long-term behaviors(such as feeding,drinking and ruminating).Taking sheep as the research object,this paper puts forward a kind of automatic identification from sheep feeding video short chewing behavior of the algorithm,based on a short video data identification problem into a sheep mouth chewing behavior status label coding sequence fragment classification problem,realize the sheep automatic analysis of short-time chewing frequency and duration,and sheep intake estimate model is established based on this,the main work is as follows:(1)Video data collection of feeding of Hu sheep.The experiment was carried out under the condition of house feeding with individual sheep in a sheep pen and the feed was fermented corn straw silage.A camera with a fixed height and Angle of view was used to collect the feeding video of Hu sheep without contact,there is a camera corresponding to its number in front of each sheepfold.Before and after each video capture,the experimenters weighed and recorded the weight of the silage.The original video collected was saved to SD card in MP4 format,and all video data was transferred to computer hard disk to establish a video database after the collection.The video library consisted of 32 videos of singing behavior and 104 videos of feeding behavior,and the size of the video library was183 G.screening,framing,labeling and other processing were carried out on the feeding video,and the sheep mouth status recognition data set and the short-time chewing behavior recognition database were constructed respectively.(2)A model of sheep’s mouth state recognition is based on Efficient Det network with target box filter module.Analyses the chewing characteristics of Hu Sheep during feeding,and adds a target box selection module in the Efficient Det network,detected the three states of open,staggered and closed beak in the feeding video frame and combined with three sheep head relative to the camera angle relationship(side face on the camera,sheep head is for camera,sheep head are to camera)identified seven breeding sheep mouth state.The dataset was divided into training set and test set in an 8:2 ratio at first,the training set was input into the target detection model of Yolov5,SSD,and Efficent Det-D0 ~D4respectively to establish the sheep mouth state recognition model;Then,the performance of each model was tested by using the test set data,and precision,recall rate and mean average precision were introduced as the evaluation criteria of the model.The comparison and verification results show that the improved Efficient Det-D1 model can obtain 95.64% and98.84% precision of the sheep mouth state detection at 28.18 frames per second,which is superior to YOLO-V5 and SSD networks.(3)The classification rules were constructed to realize the automatic recognition of short-term chewing and singing behaviors.Each feeding video detected by the improved Efficient Det-D1 mouth status recognition model corresponds to a mouth status digital coded label sequence,the regular expression is used to extract the label coding sequence fragment of the mouth state corresponding to a maxillary tension in a continuous video frame.The mark at the beginning of short-term chewing behavior corresponds to the first frame of the first mouth state of each action,and the mark at the end corresponds to the first frame of the last action.In order to realize automatic recognition of short-time chewing behavior,classification rules were constructed for the snippet text of the label coding sequence corresponding to the mandibular and upper mandibular opening movements,such as chewing on the camera with the side of the sheep face,chewing on the camera with the head up,chewing on the camera with the head down and chewing on the camera,and singing.The results showed that the rule could recognize the sheep’s chewing behaviors in a100% accuracy rate,including side chewing,front head chewing,front head chewing and song,the results showed that the sheep’s chewing behaviors could be recognized by the rule with 100% accuracy.(4)A model for estimating Hu sheep forage intake was established.Efficent Det-D1 and regular expression based automatic recognition method of short-time chewing behavior extracted the chew_num and chew_time variables of short-time chewing behavior in all feeding videos from the database,combined with the actual forage intake,a forage intake estimation model based on univariate and multivariate was established.This method avoids the misidentification of singing behavior as short-time chewing behavior and makes the extraction results of the two variables more accurate.For the univariate forage intake model,the least square method was used to predict forage intake,of which,the parameter chew_num showed the best performance,and the R2 of the test set reached 83.65%.For multivariate forage intake model,the positive and negative correlations between each variable in the variable set and forage intake were determined according to Pearson correlation coefficient,and six multivariate forage intake estimation models were built with Bagging and Boosting integration algorithms respectively.The verification results show that the Random Forest model based on multivariate has the best effect,and the R2 reaches90.73%. |
PDF Full Text Request