Research On Behavior Recognition Approach Of Group-Housed Goats Based On Deep Learning

Daily behavior is one important manifestation for health and welfare status of livestock.Machine vision-based analysis of animal behavior has the inherent characteristics,in the efficient and noninvasive perspective,compared with manual observation and touched sensing.This paper proposes one efficient behavior recognition approach of group-housed goats based on deep learning,where the daily behavior including eating,drinking,active and inactive behaviors of group-housed goats are investigated by analyzing the frame images collected by the camera located on the upper side of the pen.The main research content covers the following three aspects: data acquisition of goat activities and experimental platform construction,research on goat target detection algorithm based on deep learning and recognition method of goat behavior based on YOLOv4 and spatiotemporal features.The research results show that this approach will effectively overcome the deficiency of the traditional methods that heavily depend on head detection,depth cameras and other auxiliary methods for identifying animal behavior,and providing reference for further exploring the other behavior analysis methods of group-housed animals.The main research contents are as follows,1)Data acquisition of goat activities and experimental platform construction.According to the on-site conditions of the sheepfold,the camera is installed above the side of the captive area to collect goat videos in real time,and the experimental platform is built according to the characteristics of the target detection algorithm.In order to carry out the experiment more conveniently and accurately,data enhancement operations such as rotation and contrast adjustment are implemented for image data,and the dataset is annotated.According to the proportion of 8:2,the labeled dataset is divided into training dataset and test dataset,which provides the data basis for target detection and behavior recognition of group-housed goats.2)Research on goat target detection algorithm based on deep learning.Typical deep learning target detection algorithms of Faster R-CNN、YOLOv3 and YOLOv4 are adopted.By analyzing the model structure and working principle of the different detection algorithms,comparative experiments are carried out from three aspects: thresholds for bounding boxes appearance,detection accuracy of goats and detection speed of frame images.The experimental results show that YOLOv4 is superior to the other two algorithms both in detection accuracy of group-housed goats speed and detection speed of frame images,in real-time manner with the average analysis speed of 17 FPS on a conventional hardware configuration.3)Behavior recognition method based on the relationship between captive area and goat bounding box.A general behavior recognition framework for group-housed goats is proposed to analyze the characteristics goat behavior changes between continuous frame images,and the relationship between feeding/drinking zones and goat bounding box.The experimental results show that the average recognition accuracies of this approach for eating,drinking,active and inactive behaviors of goats are 97.87%,98.27%,96.86% and96.92%,respectively.The cause of goat behavior misrecognition and the effect of goat behavior recognition under different frame image intervals are analyzed in order to provide reference for the real-time recognition of group-housed goats.