基於牛臉辨識與熱影像技術之自動化乳牛眼溫監測系統

Abstract

本研究提出了一個基於牛臉辨識與熱影像技術的自動化乳牛體溫監測系統，旨在提升乳牛疾病的早期檢測與管理效率。系統分為三個主要部分：首先，使用熱成像網絡攝像機結合YOLOv4 牛眼偵測模型實時測量牛眼溫度，該模型有著0.9的準確度；其次，通過Resnet50 CNN 骨架結合Arcface 損失函數採用度量學習的策略進行牛臉辨識；最後，將溫度數據與牛的身份結合起來。該系統能自動記錄每頭牛的每日體溫數據，並分析這些長期數據以估計每頭牛的日常體溫是否在合理範圍內，從而實現異常體溫的早期預警 。系統使用了YOLOv8模型進行牛臉偵測(mAP@50 = 0.987)，並引入了YOLOv8實例分割技術 (mAP@50 = 0.992) 以及SORT (Simple Online and Realtime Tracking) 物件追蹤演算法來提高圖像質量和辨識精度，以克服系統實際部署至場域中遇到的辨識問題。通過這些技術的結合，系統能夠在現實場景中有效應對圖像遮擋和質量不佳的問題，最終在現實場景的應用中達到了0.82的準確度。我們分別在台大實驗牧場及豐樂牧場進行為期兩個月的溫度監測，台大實驗牧場的平均眼溫為35.35 ± 0.75°C，而豐樂牧場的平均眼溫為35.25 ± 1.19°C 。這些結果表明，系統能夠有效監測乳牛的眼溫，紀錄個體牛隻每日眼溫並且根據統計資料找出當日眼溫異常的牛隻。同時我們也觀察到使用熱影像攝影機為乳牛測溫時THI指數與量測到的溫度有著高達0.84的相關係數。此外，系統也分析了牛眼移動速度和不同觀測區域對溫度測量結果的影響，發現這些因素對溫度分佈的影響並不顯著。結果顯示本研究所開發的系統能夠為成功監測個體乳牛眼溫為牧場經營提供一個可靠的工具，有助於提高乳牛健康管理的效率和準確性。

This study presents an automated dairy cow temperature monitoring system based on cow face recognition and thermal imaging technology, aiming to enhance the efficiency of early disease detection and management in dairy cows. The system comprises three main components: first, it uses a thermal imaging network camera combined with the YOLOv4 cow eye detection model to measure cow eye temperatures in real-time, achieving an accuracy of 0.9; second, cow face recognition is performed using a ResNet50 CNN architecture combined with the ArcFace loss function, employing a metric learning strategy; finally, it integrates temperature data with cow identities. The system can automatically record each cow's daily temperature data and analyze these long-term data to estimate whether the daily temperatures of each cow fall within a reasonable range, thereby enabling early warning of abnormal temperatures. The system employs the YOLOv8 model for cow face detection (mAP@50 = 0.987) and incorporates YOLOv8 instance segmentation (mAP@50 = 0.992) along with the SORT (Simple Online and Realtime Tracking) object tracking algorithm to enhance image quality and recognition accuracy, addressing the identification challenges encountered during real-world deployment. Through the combination of these technologies, the system effectively handles image occlusion and poor quality issues, ultimately achieving an accuracy of 0.82 in real-world applications. We conducted two-month eye temperature monitoring at the NTU Experimental Farm and Home Love Farm. The average eye temperatures recorded were 35.35 ± 0.75°C at the NTU farm and 35.25 ± 1.19°C at Home Love farm. These results indicate that the system can effectively monitor cow eye temperatures, record daily eye temperatures for individual cows, and identify cows with abnormal eye temperatures based on statistical data. Additionally, we observed a high correlation coefficient of 0.84 between the Temperature-Humidity Index (THI) and the measured eye temperatures when using thermal imaging cameras for cow temperature monitoring. Moreover, the system analyzed the impact of cow eye position movement speed and different observation areas on eye temperature measurement results, finding that these factors did not significantly affect eye temperature distribution. The results demonstrate that the developed system provides a reliable tool for successfully monitoring individual cow eye temperatures, contributing to improved efficiency and accuracy in dairy cow health management.