Real-time level set based tracking with appearance model using Rao-Blackwellized particle filter

In this paper, a computationally efficient algorithm for level set based tracking is suggested for near real-time implementation. The problem of computational complexity in level set based tracking is tackled by combining a sparse field level set method (SFLSM) with a Rao-Blackwellized particle filter (RBPF). Under the RBPF framework, affine motion is estimated using an appearance-based particle filtering (PF) to provide the initial curves for SFLSM and the local deformation of contours is analytically estimated through SFLSM. SFLSM is adopted to significantly reduce the computational complexity of the level set method (LSM) implementation. For the initial curve estimation in SFLSM, the estimated position and object scale are provided by the appearance-based PF in order to achieve the desired efficiency. Furthermore, the appearance-based PF alleviates inaccurate segmentation incurred by an incorrect initial curve. Experimental results with a real-video confirm the promising performance of this method.