Camera Array Based Bat Flight Motion Capture

Research in flapping flight has gained new attention in recent years.As one of the most unique flying creatures,bats are particularly agile,stable,and efficient compared to other flapping-wing fliers,which makes them excellent models for the design of flapping-wing air vehicles.With the development of camera imaging and computer vision technology,it is possible to study the bat flight quantitatively through visual 3D reconstruction.In order to obtain accurate 3D reconstruction,it is often necessary to place a large number of fiducial landmarks on bat bones,joints,and wing membranes.Also,a well calibrated multi-camera system is essential.Given the large data quantity,solving the correspondences manually is intractable.Thus,automating the landmark identification and tracking process will be significant for speeding up the study of bat flight mechanism.Commonly used feature identification schemes such as corner detection,Blob extraction and scale invariant descriptor typically extracts a large number of features firstly,then removes outliers according to some certain contraints for matching.Such methods often come without any assumption of experiment scenarios,thus they are suitable for lack of prior knowledge.However,when extracting specific objects such as landmarks,these general methods are subject to noise and produce low extraction precision.Based on the characteristics of bat landmarks,this paper proposes a method of landmark identification method based on image segmentation.In order to suppress the similar features from background,we propose an adaptive threshold ViBe to obtain accurate foreground segmentation,and then perform image enhancement using Lapcian of Gaussian operator.Otsu's method is employed to further obtain the segmentation of landmarks.The precise location of landmarks are finally determined by computing the centroids of each contected components.In the landmark tracking phase,large non-rigid motion of bat wing has made tracking a very difficult problem.Through computing the correlation of local image,this paper analyses the reason why landmarks on bat membrane are difficult to track.To solve this problem,we define some key points on bat wings,such as wing tips and joints,where features are rich for tracking.We employ the pyramidal Lucas-Kanade sparse optical flow to track those key points,and polygons are further formed based on these vertices.Markers inside the the polygon between two frames will be matched through Coherent Point Drift point set registration method.We have experimentally shown that,compared to feature descriptors including SIFT and BRISK,and traditional optical flow tracking method,our proposed partition based point registration method has improved the tracking accuracy by about 50%and 30%,respectively.Therefore it is suitable for dense marker tracking application.At last,we discuss the drawback of using classic camera calibration method with a plannar chessboard in calibrating a multi-camera array,and introduce a new Svoboda self-calibration method and obtain the extrinc parameters for 28 cameras.3D reconstrion of bat landmarks are further conducted with the landmark coordinates and correspondences obtained by our proposed algorithm.