應用深度學習於蜂巢內蜜蜂軌跡追蹤與運動行為分析

蜜蜂為世界上最重要的授粉昆蟲，在農業與生態上皆扮演舉足輕重的腳色。但因其族群個體數量龐大，且為高度社會性動物，具有複雜的行為多態性，若欲對其進行觀察，往往耗時且易流於主觀。因此本研究建立巢內及巢口影像監測系統，對蜂群進行自動化觀察與數據化分析。巢內影像系統改良前人硬體設置，提高解析度並擴大43%相機視野，有效擷取巢片各功能區域的軌跡資訊；巢口系統則進行蜜蜂個體偵測與整體交通量之計數。實驗設計防水文字標籤，黏貼於蜜蜂胸背板上作為標記，以影像系統拍攝，並透過霍夫圓轉換進行標籤偵測；辨識方面選用深度學習MobileNet V2模型取代傳統影像處理方法，將文字標籤辨識準確率由77%提高至90%，再將結果串接成完整運動軌跡。本研究建構之蜜蜂行為分析方法，係將巢內軌跡切割成固定長度的小段副軌跡，計算六種特徵值，利用PCA主成分分析進行降維後，再以高斯混合模型將副軌跡分成靜止、徘徊、移動三種運動模式，結合巢片上卵及幼蟲、封蓋、儲蜜等功能分區，再搭配巢口系統紀錄的進出資訊進行分析，並設計三部分實驗驗證前述之分析方法。第一部分為內外勤蜂行為比較實驗，結果顯示內勤蜂在巢內出現頻率明顯高於外勤蜂，各項指標皆較為穩定，移動比例高且軌跡遍布巢片各功能區域，外勤蜂則較常在空巢房區靜止休息。第二部分為失王實驗，失王群比起正常群更為躁動，內外勤蜂皆有較高的移動比例與較低的靜止比例；內勤蜂造訪蜜區的比例提高，且試飛情形踴躍，外勤蜂則幾乎不在幼蟲區滯留。第三部分觀察不同日齡幼蜂的行為轉變，結果顯示幼蜂在幼蟲區停留比例明顯較高，育幼行為與其日齡分工相呼應。

關鍵字

蜜蜂標記；蜜蜂行為分析；卷積神經網路；軌跡分段；高斯混合模型

並列摘要

Western honey bee (Apis mellifera), one the most important pollinators in the world, plays a pivotal role in both agriculture and ecosystem. The sophisticated social behaviors and highly flexible division of labor have driven considerable amount of previous research studies. However, each colony usually contains tens of thousands of individuals, making it laborious and time-consuming to quantitatively assess and understand their behavior. This work aims to establish in-hive and in-and-out imaging monitoring system to automatically collect bee behavioral data and quantitatively analyze bee behavior. In-hive imaging system was adapted from previous works, and several improvements have been achieved. The field of view (FOV) of the cameras were enlarged by 43%, allowing them to capture a complete bee comb, which is essential for monitoring bee behaviors in different comb areas. A waterproof round-shaped paper tag with characters was designed to label individual bees by gluing them on the scutum of bees. After collecting in-hive images, Hough circle transform was applied to detect the location of the paper tag. A MobileNetV2 deep learning model was trained to classify the characters on the tags, and reached a 90% classification accuracy. The positions of the tags were further combined to trajectories with tracking algorithm. After sampling trajectories to fixed-length segments, six features of the segment were calculated. Principle components analysis (PCA) was implemented to reduce the dimensionality of the features, and the principle components were further fitted by Gaussian mixture model (GMM) with three clusters. Judging from the characteristic of the clusters, the segments were further classified into three motion patterns, namely stationary, loitering, and moving. As for in-and-out activity monitoring system, besides tag detection, unlabeled bee detection was also performed to monitor the traffic of the bee entrance using tiny-YoloV4 model and SORT tracking algorithm, which achieved an mAP of 98.83% and MOTA of 94%, respectively. Three experiments were conducted with both systems with comb area information to validate the proposed system as well as analyze honey bee behaviors. The first experiment was designed to compare the behaviors of house bees and field bees. The results indicated that house bees were more frequently captured and more vigorous in each areas of bee comb, while field bees were often idling or resting in empty areas. The objective of the second experiment was to identify the behavioral differences in queenright and queenless colonies. The results showed that the bees of queenless colony tend to have a ratio of moving pattern. House bees in queenless colony appear in honey areas more frequently, while field bees rarely stayed in brood area. Moreover, more occasions of orientation flights in queenless colony were observed. The third experiment was designed to observe the behavioral changes with respect to the aging of bees. The results showed that young bees were frequently detected in brood areas, and their behavioral changes corresponded to their ages.