利用卷積神經網路自動辨識母豬哺乳行為及追蹤仔豬

Abstract

豬肉是台灣和世界許多國家的重要蛋白質來源。維持仔豬的斷奶率對於滿足豬肉不斷增長的需求至關重要。新生仔豬比較脆弱，因此需要更多照顧。然而，人工觀察是費時且勞動密集的，因此，本研究的目的為開發一種自動方法在影片中識別母豬的泌乳行為、定位仔豬、追蹤個別仔豬並量化仔豬的移動狀況。在研究中，視頻以一秒五張的速率轉換為圖像。圖像經過了預處理。開發了手機網路第二板（MobileNetV2）模型來識別泌乳行為。然後開發了精細旋轉視網膜網路（R3Det）模型來定位仔豬。隨後，將簡單的即時跟踪（SORT）演算法應用追蹤個別仔豬。最後，通過頻率和距離閾值過濾量化的仔豬運動誤差。本研究利用圖像顯示卡 (Graphics Processing Unit, GPU)計算的MobileNetV2在哺乳辨識確度達到95.45%的準確率及13.9毫秒∕影像的辨識速度。本研究利用GPU運算的R3Det在仔豬定位上的的平均精度達到87.08%、精準度為93.52%、召回率為88.52%及10.2 影像∕秒的辨識速度。本研究利用CPU運算的SORT演算法在仔豬追蹤上的多物體追蹤準確率為97.53%、多物體追蹤精準率為96.97%及IDF1為97.89%及171.6影像∕秒的辨識速度。本研究所提出的方法可以達到即時的性能表現。

Pork is an essential source of protein in Taiwan and many countries around the world. Maintaining the weaning rate of piglets is essential to meet the increasing demand of pork. Newborn piglets are relatively fragile and need more attention. Manual observation is, however, time consuming, and labor intensive. This study aimed to develop an automatic approach for recognize lactating behavior of sows, localizing piglets, tracking individual piglets and quantifying piglet movements in videos. In the proposed approach, a video was converted to images using a rate of 5 fps. Images were preprocessed. A MobileNetV2 model was developed to recognize lactating behavior. A Refined rotation RetinaNet (R3Det) model was then developed to localize piglets. Subsequently, simple online and realtime tracking (SORT) algorithm was applied for individual piglets tracking. Finally, the quantified movement errors of the piglets were filtered by frequency and distance thresholding. The developed MobileNetV2 reached an overall accuracy of 95.45% and a test time of 13.9 ms per image using a GPU. The developed R3Det reached an overall mAP of 87.08%, a precision of 93.52%, a recall of 88.52% and a processing speed of 10.2 fps using a GPU. The SORT algorithm reached an overall MOTA of 97.53%, MOTP of 96.97%, IDF1 of 97.89% and a processing speed of 171.6 fps using a CPU. The proposed approaches could reach a realtime performance.