本研究以Derek Hugger設計之Colibri木製仿生蜂鳥機構為對象，建立一套以單視角公開影片之外部週期位移為輸入，逆向重建等效凸輪輪廓的方法。公開影片經解析度重取樣與影格插值後，以Tracker完成尺度標定、座標設定與特徵點追蹤，取得撲翼垂向位移、尾羽展收與胸腔起伏三個子系統之時間－位移資料。尾羽與胸腔資料先扣除撲翼相關構件之共同位移。其後以Mathematica進行單週期切分、端點閉合、Hampel異常點抑制及 Fourier 週期擬合，建立角度域位移函數。再依等效基圓與滾子幾何生成節曲線及滾子補償後凸輪輪廓。本文建立10條主要理論關係式，並以閉合、曲率／欠切、自交／非鄰近間距及法向一致性四類條件進行建模前幾何檢核。各凸輪輪廓於單週期輸出2，000點。驗證階段依購得之原廠設計資料建立兩組同架構 CAD／ADAMS模型。原廠凸輪工作輪廓不參與影像量測、訊號處理與重建參數選定，僅於重建完成後作為獨立基準，以避免重建與驗證資料互相回饋。除凸輪幾何外，兩組模型其餘條件一致。三子系統之均方根誤差分別為0.307、0.294與0.362 mm，升距幅值比分別為0.978、1.003與0.965；對應總升距差異為−2.2%、+0.3%與−3.5%。結果顯示，重建模型可再現原廠基準模型之週期位移特徵與總升距。本方法僅需單視角公開影片與少數固定構件尺寸，無須拆解實體或接觸式量測，即可取得可供CAD與多體動力學分析之等效凸輪輪廓。研究成果定位為運動功能等效重建，不主張唯一復原原廠凸輪幾何。;This study develops an image-tracking-based method for reverse reconstruction of equivalent cam profiles from observable periodic motion in Derek Hugger’s Colibri wooden biomimetic hummingbird mechanism. After spatial-resolution resampling and frame interpolation， Tracker is used for scale calibration， coordinate definition， and feature-point tracking. Time-displacement data are obtained for wing vertical displacement， tail deployment and retraction， and chest oscillation; common wing-related motion is removed from the tail and chest data to obtain relative displacement. Mathematica is then used for single-cycle extraction， endpoint closure， Hampel outlier suppression， and Fourier periodic fitting to establish angular-domain displacement functions. The formulation comprises ten principal theoretical relations. Pitch curves and roller-compensated cam profiles are generated and checked against four pre-modeling geometric criteria: closure， curvature/undercut， self-intersection/non-adjacent spacing， and normal consistency; each profile is exported with 2，000 points per cycle. For validation， purchased design data are used to construct two CAD/ADAMS models with the same mechanism architecture. The original cam working profiles are excluded from image measurement， signal processing， and reconstruction-parameter selection and are introduced only after reconstruction as an independent validation reference， thereby preventing feedback between reconstruction and validation data. Except for the cam profiles， the two models share identical non-cam geometry， assembly relationships， and simulation conditions. Root-mean-square errors are 0.307， 0.294， and 0.362 mm， while lift-amplitude ratios are 0.978， 1.003， and 0.965， corresponding to total-lift differences of −2.2%， +0.3%， and −3.5%. The reconstructed models therefore reproduce the principal periodic displacement characteristics and total lift of the reference models. The method requires only a single-view publicly available video and a few fixed-component dimensions， without physical disassembly or contact measurement， and yields equivalent cam profiles suitable for CAD and multibody-dynamics analysis. The results represent functional-equivalence reconstruction rather than unique recovery of the original cam geometry.
