Cattle Behavior Model Classification And Monitoring System Research Based On An Improved Fully Convolutional Network

Cattle behavior classification is a critical technology for achieving intelligent and refined cattle breeding.By classifying cattle behavior patterns,farmers can better understand bovine behavior changes and improve management approaches to increase cattle welfare.In this paper,we studied the cattle behavior pattern classification approach and developed a cattle behavior pattern classification monitoring system based on an IMU sensor and an improved full convolutional network.The main research work and results are as follows:(1)Cattle behavior data acquisition and behavior definition.Using a combination of video and sensor recording,the cattle behavioral data collection was completed and the target behaviors for this study were determined to include: ruminating(lying),lying,feeding,rub(legs),self-grooming,rub(neck),and social licking.Tagging of the data set was completed by corresponding the sensor data to the timestamps of the surveillance videos.Finally,the cattle behavior dataset with slice sizes of 64 and 128 is produced.(2)Classification of cattle behavioral patterns based on time series data.By comparing three feature extraction algorithms on a one-dimensional full convolutional network(FCN1D),the principal component analysis was selected as the optimal solution for feature extraction of the cattle behavior dataset.Then the FCN1 D was optimized using long-short memory network(LSTM)with channel attention module(SENet),and the results showed that the cattle behavior pattern classification model based on channel attention module optimized long-short memory full convolutional network(SENet-LSTM-FCN1 D,SLFCN1D)on the data sets with window size of 64 and 128,and the classification accuracies were 87.10% and 89.92%,achieving the classification of behavioral patterns based on time series data.(3)Classification of cattle behavior patterns based on time-frequency images.In order to improve the classification accuracy of cattle behavior,this paper transforms the time series data into time-frequency feature images by CWT.Then a two-dimensional full convolutional network(FCN2D)based cattle behavior pattern classification network is built,and the network is optimized by external attention mechanism(EANet).The results show that the cattle behavior pattern classification model based on the external attention module optimized full convolutional network(EA-FCN2D)has the highest classification accuracy of 93.83%on the dataset with a window size of 128,and the classification accuracy of each cattle behavior pattern reaches 90%.Also,comparing the classification results of EA-FCN2 D and SL-FCN1 D,it was found that EA-FCN2 D had the most significant improvement in the classification accuracy of two behavior patterns,namely,cattle ruminating(lying)and social licking,both by 10%.In summary,the time-frequency image-based cattle behavior pattern classification scheme can meet the cattle behavior classification needs and provide a new way for cattle behavior classification.(4)The visualization interface of cattle behavior pattern classification monitoring is completed based on pyQt5 framework.Combining the above research of cattle behavior pattern classification model with intelligent collar,we build a cattle behavior pattern classification monitoring system to realize cattle information management and cattle 7behavior pattern classification monitoring.This study combined IMU sensor and improved FCN,built a classification model of cattle behavior using time series data and time-frequency feature images,and accurately classified 7 cattle behavior patterns.A behavior monitoring system was constructed based on the classification network of cattle behavior to accomplish cattle information management as well as the classification and monitoring of cattle behavior patterns.