Recursive-Batch Estimation of Motion and Structure from Monocular Image Sequences

Structure from motion causally integrated over time

“…The first attempts to prove stability of the schemes proposed are recent [26]. However, few of the schemes cited address occlusions [8], which makes them prone to the tradeoffs just described and, therefore, hardly usable in realistic scenes where occlusions are the norm. The first attempts to handle occlusions in a causal scheme 6 came only a few years ago: McLauchlan et al [23] proposed a filter with variable states, that however requires a batch initialization, while Soatto and Perona [35] proposed several schemes in which the problem of occlusions was bypassed by eliminating structure from the model.…”

“…The first attempts to handle occlusions in a causal scheme 6 came only a few years ago: McLauchlan et al [23] proposed a filter with variable states, that however requires a batch initialization, while Soatto and Perona [35] proposed several schemes in which the problem of occlusions was bypassed by eliminating structure from the model. Cui et al [8] have also proposed a method to handle occlusions, although that relies on a batch step to estimate interframe motion and, therefore, suffers the tradeoffs described in Section 1.2. Our approach is similar in spirit to the work of Azarbayejani and Pentland [3], extended to handle occlusions.…”

“…This corresponds to reducing the number of the states of the model (2), with the states y i H substituted for the measurements y i H, which causes the model noise nt to be nonzero-mean. 8 When the zero-mean assumption, implicit in the use of the Kalman filter, is violated, the filter settles at a biased estimate. When adding new points, such biases add up to catastrophic consequences, as we show in Fig.…”

Robot Hand-Eye Calibration Using Structure-from-Motion

“…A relevant class of structure-from-motion methods is known as Euclidean reconstruction (Cui, Weng, and Cohen 1994;Taylor, Kriegman, and Anandan 1991;Poelman and Kanade 1994;Tomasi and Kanade 1991;Koenderink and van Doorn 1991;Christy and Horaud 1996). It assumes that the camera is calibrated (i.e., the camera intrinsic parameters are known).…”

“…Hence, only approximations of the Euclidean camera motions can be obtained. Besides, solutions exist (Cui, Weng, and Cohen 1994;Taylor, Kriegman, and Anandan 1991), based on the perspective model, that offer some Euclidean information on the camera motions but are nonlinear. Our choice lies, in fact, between these two classes of methods: we propose a method for Euclidean reconstruction by successive affine approximations of the perspective model (Christy and Horaud 1996), which combines the simplicity of affine methods and the accuracy of nonlinear methods.…”