利用卷積神經網路自動偵測保育舍斷奶仔豬行為

Abstract

豬肉佔全球畜牧生產的重要部分，根據聯合國糧食及農業組織（FAO）的預測，到2031 年豬肉消費量將增長 17%。2022 年，FAO 估計全球豬肉生產量約為 1.246 億噸。豬的生產主要包括三個階段，即分娩、保育和肥育。在保育的第一週，不同分娩欄的仔豬會被分配到不同的保育欄中。剛斷奶的仔豬因剛剛與母豬分離而特別脆弱。新環境和陌生的同伴經常會引起仔豬的壓力和敏感，導致明顯的攻擊性行為。這些對仔豬的負面互動經常導致壓力、食慾減退和生長遲緩。因此，在傳統的豬場管理中，農民需要經常巡視保育舍，以人工方式觀察仔豬的狀況，並處理不正常情況。然而，人工方式的觀察過於耗時且可能無法及時發現異常情況。因此，本研究提出應用卷積神經網絡（CNNs）監控仔豬在保育階段第一週的關鍵行為和運動，關鍵行為包括進食、飲水和攻擊性行為。此研究一共使用兩種模型。Yolov7 被訓練為豬檢測模型（PDM）以定位保育欄中的單個和有互動之仔豬。PDM 與 SORT 跟踪算法（PDMS）之結合應用於跟踪單個仔豬。EfficientNet 與 LSTM 之結合被訓練為仔豬行為識別模型（PBRM）用於識別仔豬行為。PDM 在仔豬檢測中達到了 94.5%的平均精度。跟踪性能達到 83.1%的 MOTA。PBRMs 在仔豬行為識別中達到了 93.0%的準確度。PBRMs 和 PDMS 能夠識別上述仔豬行為的頻率和仔豬運動的長期分析。本研究進行了為期 30 天的長期分析。所提出的研究提供了一種自動化的方法來監控保育階段仔豬的行為和運動。本研究可以幫助農民發現仔豬異常情況，改善農場管理。此外，它可能通過降低仔豬死亡率和人工成本來提高豬產業的利潤。

關鍵詞：卷積神經網路、深度學習、豬隻行為偵測系統、畜舍管理

Pork constitutes a significant segment of global livestock production. Pig production primarily comprises three stages, namely farrowing, nursery, and fattening. During the first weeks of nursery, piglets from different farrowing crates are grouped into the same nursery pens. New environments and unfamiliar companions often induce stress to piglets, leading to notable aggressive behavior. The stress may also result in loss of appetite and growth retardation. Thus, in conventional pig farming practices, farmers need to frequently patrol the nursery houses, manually observe piglet conditions and troubleshoot irregular situations. However, manual observation is time-consuming and may not detect the irregular situations early enough. Therefore, this research proposed to monitor key behaviors and movement of the piglets in a pen from grouping to 35 days of the nursery stage using computer vision. The key behaviors included feeding, drinking, and aggression. Two models were used. A Yolov7 was trained as the pig detection model (PDM) to localize individual piglet and interactive piglets in nursery. The combination of PDM and simple online realtime tracking (SORT) algorithm was applied to quantify piglet movement. An EfficientNet combined with long-short term memory was trained as the pig behavior recognition models (PBRMs) to recognize piglet behaviors. PDM achieved a mean average precision of 94.5% in piglet detection. SORT achieved a multiple object tracking accuracy of 83.1%. PBRMs achieved an overall accuracy of 93.0% in piglet behavior recognition. The proposed research provides an automatic approach to monitor piglet behaviors and movement in nursery. This research can help farmers to discover irregular piglet conditions and improve the farm management. Moreover, it may raise the profit of pig industry by reducing the mortality rate of piglets and staffing costs.