本研究為了解決智慧牧場中自動化乳牛發情辨識及身分識別的問題。我們透過MIAT方法論提出了一個結合影像追蹤、發情辨識、身分識別的系統。使用DeepSORT追蹤，觀察並追蹤牧場中乳牛位置及動作，並計算乳牛行動或動作變化的Hu Moments動量。取連續影像幀透過LSTM根據它行為時的動量變化來預測其發情狀態。在發情辨識方面我們在資料集中獲得了總體79.8%的精確率、79.3%的召回率，說明我們總共有79.3%的發情樣本被正確預測；在身分識別上透過Siamese孿生神經網路比對影像中發情牛隻切割與資料庫牛隻影像，以歐式距離計算兩張影像在特徵向量上的相差程度，並透過歸一化轉成一個相似度值，我們以自建的資料庫測試，獲得Rank-1為91.73%的準確率，最後在整個系統的呈現上我們成功將三項功能整合成一個系統。;The purpose of this research was to resolve the problem caused by automatic estrus recognition and identity recognition of dairy cows in smart ranch.We have proposed the system that combined with image tracking, estrus recognition and identification though the method of MIAT. Using DeepSORT to track and observe cows’ movement and position in Smart Ranch and then calculate their changes of Hu Moments. Then, by taking continues image through LSTM to predict dairy cows’ estrus. In terms of estrus recognition, we have collected samples with total 79.8% Precision rate, 79.3% Recall rate, and it means there were 79.3% of samples that have been correct predicted.With regard to identity recognition, we used Siamese Networks comparing with the images of Cows under estrus and in database. Euclidean distance calculation helped to figure out the difference of eigenvector and by normalization, which transfer them to a similarity value.After testing with our self-built database, we have gained 91.73% accuracy, and finally, we succeeded to combine three functions into one system.
