基於深度學習之泌乳牛採食行為影像監測系統

An Image Monitoring System for Feeding Behavior of Dairy Cows Based on Deep Learning

官承譽(Cheng-Yu Kuan)

指導教授 ： 林達德

國立臺灣大學/生物資源暨農學院/生物產業機電工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201903695

摘要

處於熱帶及亞熱帶地區的國家，泌乳牛容易產生熱緊迫的現象，並因此影響牛隻的生長、繁殖和泌乳等行為。在舒適環境中泌乳牛有正常的採食及飲水行為，當泌乳牛受到熱緊迫影響時，其採食量的下降可達30%以上，因此若能了解其採食行為的改變，進而作為偵測泌乳牛熱緊迫現象的警訊，將可提供給管理者作為管理牛隻健康狀況的依據。本研究之目的為開發一泌乳牛採食行為的影像監測系統，影像監測系統為基於Raspberry pi 3B+ 開發板的嵌入式系統，並搭配Pi camera v2模組進行泌乳牛採食影像的擷取。採食行為的偵測使用基於深度學習的物件偵測模型Tiny-YOLO v2來偵測採食中的牛臉影像，並進行偵測模型訓練參數最佳化，最終模型F1-score為0.98。本研究的個別牛臉影像資料庫中共有19頭，並以MobileNets v1進行個別牛臉的辨識，辨識模型的平均F1-score為0.972，以人工計算驗證系統個別牛白天採食時間得到R^2為0.79，而系統辨識一完整採食行為需時27秒。
本研究最終得到個別泌乳牛的白天採食時間作為採食行為，以分娩日期將牛隻分為三類，並分析採食時間與溫濕度指標(Temperature and humidity index, THI)及產乳量的關聯性，將舒適溫濕度(THI<80)的平均採食時間、平均產乳量定為基準進行分析。分析結果得到，處於泌乳前期的牛隻THI對於採食及產乳量的影響趨勢較小，而處於泌乳中後期的牛隻，2天前平均THI對於採食時間有負相關性，對照於產乳量，2天前的平均THI對於產乳量亦有負相關性，然而不同牛隻有不同的耐熱性，因此根據THI對採食時間及產乳量的下降能夠找到對於熱緊迫較無耐受性的牛隻。

關鍵字

熱緊迫 ； 嵌入式系統 ； 牛臉偵測 ； 牛臉辨識 ； 深度學習

並列摘要

In sub-tropical countries, dairy cows tend to experience heat stress. This phenomenon may lead to declines in feed intake, milk production, and fertility. In a comfortable environment, dairy cows have normal feeding and drinking behaviors. Dry matter intake can drop by 30% when dairy cows are affected by heat stress. For this reason, changes in feeding behavior can be a possible indicator to detect heat stress. Alerts on the heat stress of dairy cows can be provided to farm managers as a basis for monitoring the health of dairy cows. In order to monitor and record the feeding behavior of dairy cows, an imaging system is proposed in this study. The imaging system uses Raspberry Pi 3B+ as the embedded system, and a Pi Camera v2 module can acquire images for dairy cow feeding behavior detection. The feeding behavior of dairy cow is detected through the deep learning based object detection model Tiny-YOLO v2. The detection model is optimized, and the final F1-score is 0.98. In this research, there are 19 individual cows in imaging database. Individual cow face recognition is achieved by MobileNet v1 and the recognition average F1-score is 0.972. To validate the feeding time correctness, the manual observed feeding time is compare to feeding time obtain from imaging system. The R^2 is 0.79. It takes 27 seconds to recognize one feeding behavior in imaging system.
Finally, the daytime food intake time of individual dairy cows is taken as the feeding behavior. The dairy cows are divided into 3 groups according to the calving date. The correlation between the feeding time, Temperature and Humidity index (THI) and milk production are analyzed individually. The influence of THI is not obviously in feeding time and milk production when the cow is in early stage of claving. As for the cow in late stage of claving, the 2 day backward shift of THI has negative correlation between their feeding time and milk production. As a result, since different dairy cow has different tolerance to hot weather, the relation between THI, feeding time and milk production can be used to find the cow which has the most tolerance to hot condition.

並列關鍵字

Heat Stress ； Embedded System ； Cow Face Detection ； Cow Face Recognition ； Deep Learning

參考文獻

蔡雨錡。2018。乳牛熱緊迫影像監控系統之建置與資料分析。碩士論文。台北：臺灣大學生物機電工程學研究所。

Allen, J. D., Anderson, S. D., Collier, R. J., and Smith, J. F. 2013. Managing heat stress and its impact on cow behavior. 28th Annual Southwest Nutrition and Management Conference, 68, 150-159.

Armstrong, D. 1994. Heat stress interaction n with shade and cooling. Journal of Dairy Science, 77(7), 2044-2050.

Beatty, D. T., Barnes, A., Taylor, E., Pethick, D., McCarthy, M., and Maloney, S. K. 2006. Physiological responses of Bos taurus and Bos indicus cattle to prolonged, continuous heat and humidity. Journal of Animal Science, 84(4), 972-985.

Bernabucci, U., Lacetera, N., Baumgard, L. H., Rhoads, R. P., Ronchi, B., and Nardone, A. 2010. Metabolic and hormonal acclimation to heat stress in domesticated ruminants. Animal, 4(7), 1167-1183.