應用深度學習與軌道式監控系統於雞舍之監測

Abstract

雞隻是全球重要的動物性蛋白來源。在臺灣，台灣土雞因其肉質佳而備受青睞，但其長達十二週的飼養期，造成如果在養殖期間發生疾病肆虐或環境問題，可能造成業者重大經濟損失。傳統上，養雞業者依賴人工的方式，觀察雞群散佈情形以及活動狀況，來評估雞隻健康。然而，此方式不僅耗時費力，亦因人員頻繁進出雞舍而增加疾病傳播風險。為解決上述問題，本研究所開發的自動化軌道式監測系統，包含軌道、智慧監視模組、雞隻偵測模型、散佈程度量化演算法、活動力量化演算法以及低溫雞隻偵測演算法。智慧監視模組包含雙通道攝影機，架設於40公尺的軌道上，在每日進行來回移動的過程中，於九個固定停靠點蒐集可見光及熱影像資料。為確保資料穩定傳輸與模組即時通訊，雞舍內建置了Wi-Fi 網狀網路，透過快速且穩定的網路連線，將資料傳輸到雲端伺服器進行儲存。雞隻偵測模型用於偵測可見光影像中的雞隻位置；散佈程度量化演算法整合雞隻偵測結果以及最近鄰演算法量化雞隻散佈程度；活動力量化演算法使用simple online and realtime tracking (SORT)演算法根據影像中雞隻位置量化活動力；低溫雞隻偵測演算法結合同時間拍攝的可見光影像以及熱影像和雞隻偵測結果與動態溫度閾值進行低溫雞隻檢測。結果顯示，軌道式監測系統在雞隻偵測上準確率達到 95.8%，平均精度達94.3%。於多目標追蹤準確率上93.9%，並透過分析量化的散佈程度以及活動力量值發現雞隻行為、環境溫度以及養殖的區域具有明顯相關性。此外，本研究亦成功檢測出體溫異常低下之死亡個體。本研究所提出的軌道式監控系統可於實際飼養環境中進行雞隻以及環境的監控，有助於提升飼養管理效率、促進雞隻福祉，於降低人力需求的同時，可以更有效的管理雞舍。

Chickens are a vital source of animal protein worldwide. In Taiwan, Taiwan Native Chickens (TNCs) are particularly valued for their meat quality but require a long rearing period of up to 12 weeks. Any health or environmental disturbances during this period can result in significant economic losses. Traditionally, chicken farmers conduct manual inspections to monitor flock health through observation of dispersion and activity levels. However, this approach is labor-intensive, inefficient, and increases the risk of disease introduction due to frequent human entry into the chicken houses. To overcome these challenges, this study developed an automated rail-based monitoring system that integrates a smart rail surveillance module with chicken detection model (CDM), dispersion quantification algorithm (DQA), movement quantification algorithm (MQA), and hypothermal detection algorithm (HDA) to monitor chickens in a commercial farming environment. The system operated along a 40-meter monorail and collected RGB and thermal images at nine fixed locations. To ensure data transmission and system communication, a 5G Wi-Fi Mesh network was deployed, providing stable, high-speed wireless connectivity throughout the chicken house. The results showed the CDM could accurately detect chickens with a precision of 95.8% and an average precision (AP) of 94.3%. The MQA achieved a Multiple Object Tracking Accuracy (MOTA) of 93.9%. Additionally, the HDA successfully identified hypothermic chickens by estimating body surface temperature from thermal images, using spatial alignment of RGB and thermal images and dynamic thresholding techniques. A clear relationship was observed between temperature and chicken behavior through the analyzation of quantification results. This automated and objective system demonstrated reliable performance under real-world conditions, offering a practical tool to enhance farm management, improve animal welfare, and enable early detection of potential health issues in commercial poultry operations.