Basic Motion Behavior Recognition Of Dairy Cows Based On Skeleton

The health of dairy cows is closely related to their exercise behavior.Monitoring the basic movement behaviors of dairy cows such as walking,standing and lying is of great significance for disease prevention,treatment and improving the production efficiency of dairy farms.As for the problems that most of the existing basic motion behavior recognition algorithms of dairy cows study specific behavior,do not consider the time dimension information,and recognize with the help of reference objects in the environment,the model has poor antiinterference,this study carried out the research on non-contact dairy cow skeleton extraction and basic movement behavior recognition method based on video intelligent analysis,so as to provide a new method for dairy cow health monitoring in unstructured dairy cow breeding environment.The main research contents and conclusions of this paper are as follows:(1)A multi-objective cow skeleton extraction network based on partial affinity field was proposed.The feed-forward network used to extract image features was the improved VGG-19.The multi-stage two branch network was used to get the hot spot map and partial affinity field of dairy cows,and then the skeleton of dairy cows was extracted.By introducing the feature fusion method,the shallow details and deep semantic features of cow image can be extracted at the same time,which further improves the accuracy of cow skeleton extraction.The results showed that the overall confidence of single dairy cow was 78.90%.The overall confidence of double dairy cows was 67.94%,and it had good robustness to light change and dairy cow scale change.(2)Based on the analysis of the characteristics and advantages of multi-resolution network,a cow skeleton extraction method based on cascaded pyramid network was proposed.In order to describe the model objectively,the evaluation indexes to measure the prediction ability of the model were introduced and compared with the extraction results of cascaded pyramid network.The experimental results showed that the accuracy of the optimal model trained by High-Resolution Net(HRNet)was 86.90%,the m AP was 0.637 and the m AR was0.678.The accuracy of Cascaded Pyramid Network(CPN)on the same test set was 63.00%,the m AP was 0.519 and the m AR was 0.593.The results showed that the extraction of cow skeleton using HRNet could provide data basis for the subsequent recognition of cow basic movement behavior.(3)A multi feature recognition method of dairy cow basic motion behavior based on dairy cow image,optical flow and skeleton was proposed.As for the problem that the existing 2D basic motion behavior recognition could not effectively use the time dimension information,and was excessively dependent on the environmental profile characteristics and weak antiinterference,a multi feature branch prediction network was constructed by combining the cow skeleton,RGB image and global optical flow.The network was divided into three branches according to different characteristics,branch 1,3 used the Convolutional Neural Networks(CNN)to extract and predict the RGB image features and global optical flow features respectively,and the branch 2 used the Spatial Temporal Graph Convolutional Networks(STGCN)to extract and predict the bone frame features,adjust the weight and perform score fusion to complete the basic running behavior recognition of dairy cows.The experimental results showed that ACC of the model was 94.00%.This study could provide a new idea for the recognition of basic motor behavior of dairy cows.(4)A recognition method of cow basic motion behavior based on cow skeleton and mixed convolution was proposed.In view of the problem that the parameters of the existing 3D basic motion behavior recognition methods were too large,the network depth was insufficient,and the network anti-interference was weak when only RGB images were used for recognition,the depth of 3D convolution neural network was increased by concatenating a depth 2D convolution after each 3D convolution,and then adding parallel 2D convolution on the basis of concatenation.Using the correlation between 3D and 2D feature images,3D and 2D convolution shared spatial information.At the same time,the key point information of the cow skeleton corresponding to the frame was added in the form of hot spot map in the parallel 2D convolution feature.While improving the depth of 3D convolution network,the model parameters and robustness were effectively controlled.The experimental results showed that the recognition accuracy of this research method was 91.80%.Compared with the improved hybrid convolution network,the accuracy was improved by 3.40%.This method was effective for the classification of walking,standing and lying of dairy cows.