Image Analysis Method and System for Honey Bee Movement Behavior Monitoring in Beehive

蜜蜂為開花植物的重要授粉者，糧食與經濟作物的生產皆依賴蜜蜂授粉。在2006~2007年期間北美洲發生蜜蜂族群崩潰症候群，並導致嚴重的農業與經濟損失，因此該現象的原因亟需加以探討。基於此，本研究以前人的蜜蜂巢內行為影像系統為基礎，進行改良並重新開發系統，提高巢內影像系統性能以提供更大的觀察範圍與更加準確的蜜蜂巢內運動軌跡資訊，以提供一套科學工具探究蜜蜂族群崩潰症候群的原因。為偵測個別蜜蜂的軌跡，首先將欲觀察的蜜蜂貼上防水文字標籤，並使用霍爾圓轉換演算法 (Hough circle transform)偵測個別蜜蜂身上之標籤位置，然後將標籤影像進行方向梯度直方圖特徵提取演算法 (Histogram of oriented gradient, HOG)提取影像邊緣特徵，並使用主成分分析法降維與去除冗餘資訊，最後使用支持向量機演算法 (Support Vector Machine, SVM)作為字元辨識方法。獲得標籤位置與編號之後，使用軌跡追蹤演算法串接蜜蜂位置產生行為軌跡資訊。所擷取的個別蜜蜂軌跡進一步藉由切割片段軌跡並轉換成為個別步態、軌跡運動指標以及蜜蜂互動頻率，並搭配蜜蜂巢外覓食行為影像監測系統所記錄的蜜蜂進出巢資訊，分析蜜蜂在不同分工與日齡的行為差異。本研究設計三項實驗以驗證影像系統效能及蜜蜂行為分析方法，第一項實驗為蜜蜂使用不同標記方法後對於活動力的影響，由實驗得知負壓吸取法在標記後能夠維持較高的活動力。第二項實驗為箱內蜂與外勤蜂的運動行為比較，並由實驗結果得知外勤蜂於巢內的平均移動速度較快且具有週期性，不同分工的蜜蜂會因天氣狀況而動態調整於巢內的行為。第三項實驗則是不同日齡的蜜蜂行為模式比較，並由實驗得知隨日齡不同而導致軌跡步態比例與移動速度差異。所建立的蜂巢內影像監測系統與量化分析方法未來將可應用於各種蜜蜂行為之實驗與分析，為蜜蜂行為相關研究領域提供一個有用的研究工具。Honey bees (Apis mellifera) are important pollinators in nature. In 2006~2007, North America occurred serious CCD (colony collapse disorder) problem. Abnormal and inadequate pollination cause enormous economic and agricultural loss, the reason of CCD problem is necessary to study. Base on the problem, we redevelop an image monitoring system in the beehive, it can provide larger FOV (field of view) and more accurate trajectory information in the beehive than former design. Provide a scientific tool to study the reason of CCD problem. In order to capture individual honey bee information in beehive, we label a water-proof text tag on each bee. After labeling, Hough circle transform technique is applied to detect tag position and to segment out the tag image. In order to extract the tag character feature, tag images are transformed into feature vectors by histogram of oriented gradient (HOG) algorithm. Extracting feature by HOG leads to feature dimension increase which slows down the computational speed. Hence, principle component of analysis (PCA) is applied to reduce the feature dimensionality first. Then, support vector machine (SVM) algorithm is applied to train and predict the tag text. After retrieving the position and text of tag, trajectory tracking algorithm is applied to generate trajectories. We establish a trajectory analysis method and activity indices are built up for evaluating honey bee’s activity in the beehive. Analyze the behavior differences of honey bee with different division of labor or different age. We design 3 experiments to verify our image system and analysis method, first part of experiment is honey bee activity comparison experiment by different labeling method. According to the experiment result, it is clear that the labeling by vacuum fixation method can keep honey bee higher activity ability than freezing method. The second part of experiment is the foraging bee and in-hive bee behavior comparison experiment. We can compare that different task of honey bee have different moving speed and activity cycle from the experiment result. The last part of experiment is different age of honey bee behavior comparison experiment, we can find out the motion pattern ratio is different from the age of honey bee and the activity indices are also different with different age of honey bee