| In order to realize the fast judgment of sheep’s attitude and behavior in the environment of breeding farms,and to solve the problems of poor recognition accuracy and low efficiency of existing models,this paper proposes sheep’s attitude recognition algorithm based on improved YOLOv3 and sheep’s behavior recognition algorithm based on improved Convolutional neural network based on self built data sets.The main work of this article is as follows:(1)Constructing a sheep pose recognition model based on YOLOv3-SE-RE.Sheep pose recognition is studied based on surveillance videos from a certain breeding farm.The 3047 surveillance videos obtained were preprocessed through filtering,deletion,keyframe extraction,classification,etc.The dataset images were annotated,resulting in 2000 images of different sheep poses,forming the dataset for sheep pose recognition experiments.The videos used in the sheep behavior recognition data set were collected in the farm on the spot,and the different behavior videos of sheep were screened.The classification,key frame extraction,data enhancement and other operations were carried out.Finally,9400 different behavior images of sheep were obtained,which constituted the data set used in the sheep behavior recognition experiment.(2)Construct a sheep pose recognition model based on YOLOv3-SE-RE.Introduce Squeeze and Excitation Networks(SENet)in the backbone network of YOLOv3;Secondly,replace some residual modules in the network with Recurrent Feature Shift Aggregator(RESA);The cosine annealing dynamic Learning rate is used to replace the original Learning rate for dynamic fine tuning in the training process;Finally,the multi-scale training method is used to improve the robustness of the network to input images of different sizes.The experimental results show that the mAP of YOLOv3-SE-RE algorithm has increased by 9.98% compared to the original YOLOv3 algorithm,and the detection speed has also been improved.(3)A sheep behavior recognition model based on improved Convolutional neural network was constructed.Construct a convolutional kernel with a size of 3 × 3’s Convolutional neural network;Use the scaled exponential linear units(SeLU)as the Activation function;Using maximum pooling as downsampling;In the full connection layer,the discarding operation is used,and the cosine annealing dynamic Learning rate is used for dynamic tuning;Finally,a softmax classifier was used as the network output to construct a sheep behavior recognition network model.The experimental results show that the model can monitor different behaviors of sheep,with recognition accuracy of over 90%. |
