3D data driven prediction for active contour models based on geometric bounding volumes

摘要: Abstract Active contour models have proven to be a promising approach for data driven object tracking without knowledge about the problem domain and object. Problems arise in case of fast moving objects natural scenes with heterogeneous background. In these cases, prediction step is an essential part mechanism. this paper we describe new modelling 3D world. The key point description by simplified geometric model, so-called bounding volume. predicted estimating movement shape volume world projecting its image plane. Stochastic optimization algorithms are used estimate parameters 2D initialize active contour, which then extracts Thus, update motion model possible. No task specific no priori necessary. We will show that convex polyhedral volumes, method can applied real-time closed-loop on standard Unix workstations. Furthermore, present experiments prove robustness front background improved.