Uncertainty Optimization for Robust Dynamic Optical Flow Estimation.pdfVIP

Uncertainty Optimization for Robust Dynamic Optical Flow Estimation ∗ Volker Willert, Marc Toussaint , Julian Eggert, Edgar K¨orner HRI Europe GmbH TU Berlin∗ Carl-Legien-Str. 30 Franklinstr. 28/29 D-63073 Offenbach, Germany D-10587 Berlin, Germany volker.willert@honda-ri.de mtoussai@cs.tu-berlin.de Abstract Recent and very accurate optical flow methods [3] use local spatiotemporal information by means of a structure We develop an optical flow estimation framework that tensor and combine it with global spatiotemporal smooth- focuses on motion estimation over time formulated in a Dy- ness constraints on the flow field. Further on, robust penalty namic Bayesian Network. It realizes a spatiotemporal in- functions or robust statistics [2] improve the estimations tegration of motion information using a dynamic and ro- especially at motion discontinuities because they are less bust prior that incorporates spatial and temporal coher- sensitive to outliers. In addition, there are some new at- ence constraints on the flow field. The main contribution tempts to extract detailed scene-dependent prior knowledge is the embedding of these particular assumptions on opti- about the underlying spatial statistics of optical flow that cal flow evolution into the Bayesian propagation approach hold specific spatial relations between the velocity vectors that leads to a compu