利用卷積神經網路建立雞隻張嘴行為、散佈程度與活動力之預警系統

Abstract

雞肉在全球蛋白質供應鏈中佔有重要地位，隨著市場需求不斷增長，採取集約化飼養方式，將大量雞隻放置於同一環境中已經成為常態。在這種情況下，對於雞群生長狀態的持續監控成為確保產量穩定的關鍵要素。傳統上，雞隻的張嘴行為、散佈程度和活動力都是由人工監控，然而這種方式不僅耗時且耗力，並且難以實現即時異常偵測。為解決上述問題，本研究提出了預警系統，藉由量化雞隻的張嘴行為、散佈程度與活動力，進行監測。預警系統由多台嵌入式設備、Wi-Fi網狀網路、雞隻張嘴偵測模型和雞隻追蹤模型組成。嵌入式系統包含鏡頭分別安裝在雞舍的支柱和橫樑上，從側視和俯視角度捕捉雞隻的影像，再透過Wi-Fi網狀網路傳送到雲端伺服器儲存。雞隻張嘴偵測模型用於偵測側視影像中的雞隻頭部，並將其分為兩類：張嘴與未張嘴。透過雞隻張嘴偵測模型偵測結果，可計算張嘴雞隻占所有偵測到的雞隻比例。雞隻偵測與追蹤模型用於偵測俯視影像中的雞隻位置，並利用最近臨演算法與Bytetrack演算法，分別計算雞隻的散佈程度與活動力。經過量化完成之張嘴雞隻比例、散佈程度與活動力，分別使用平均值與標準差、自回歸整合移動平均（ARIMA）的95%信賴區間、以及季節性自回歸整合移動平均模型含有外生變數（SARIMAX）的95%信賴區間來確定其安全區間。當數值超出該安全區間的數值即被認為是警告。在結果方面，雞隻張嘴偵測模型在雞隻頭部的分類與偵測上，整體平均精度達到91.3%。雞隻偵測與追蹤模型在雞隻偵測上，平均精度達到95.8%，再多目標追蹤準確率達到89.5%。此外，在自回歸整合移動平均模型的平均絕對百分比誤差達到3.44%，而季節性自回歸整合移動平均模型含有外生變數的平均絕對百分比誤差達到13.76%。本研究提供了一個完整且全自動化的預警系統，旨在為雞場管理員提供實時且有效的數據支援，以便他們能更有效地管理雞場。

Chicken is a major source of dietary protein worldwide. To meet the growing demand for chicken meat, chickens are usually raised using intensive farming approach, in which thousands of chickens are housed together. To ensure chicken production, it is essential to monitor the chickens. Typical monitoring indicators include open beak (OB) behavior, spatial dispersion, and movement of chickens. Conventionally, chicken monitoring was achieved in routinely patrol. However, manually monitoring a large flock of chickens is time-consuming and may not detect adverse events in real-time. Thus, this study proposes to monitor OB behavior, spatial dispersion, and movement of chickens on commercial farms using machine vision. The proposed early warning system comprised customized embedded systems, Wi-Fi mesh, an open-beaked behavior detection model (OBDM), and a chicken detection and tracking model (CDTM). The customized embedded systems comprised single board computers and cameras installed on pillars and roof beams to acquire side-view and top-view videos, respectively, of chickens. The acquired videos were transmitted to a cloud server through Wi-Fi mesh and 4G network. Subsequently, OBDM detected chicken heads in the side-view videos and quantified the ratio of the chickens with OB behaviors (also referred to as OB ratio). CDTM localized chickens in the top-view videos, tracked the chickens and quantified spatial dispersion and movement of the chickens using nearest neighbor (NN) and Bytetrack algorithm, respectively. The safe zones of OB ratio, dispersion, and movement, respectively, were determined using mean and standard deviation, 95% confidence intervals of autoregressive integrated moving average (ARIMA), and 95% confidence intervals seasonal autoregressive integrated moving average with exogenous factors (SARIMAX). The values outside the safe zones were considered as warnings. OBDM achieved an overall mAP of 91.3% in chicken head detection. CDTM achieved a mAP of 95.8% in chicken localization. CDTM achieved an overall MOTA of 89.5% in chicken tracking. The ARIMA and SARIMAX models, respectively, achieved a mean absolute percentage error (MAPE) of 3.44% and 13.76%. This research can provide an assistance for chicken farmers to more efficiently manage their farms.