A GPU-accelerated particle filter with pixel-level likelihood

We present in this paper a GPU-accelerated particle filter based on pixel-level segmentation and matching, for real-time object tracking. The proposed method achieves real-time perfomance, while computing for each particle the corresponding filled model silhouette through the rendering engine of the graphics card, and comparing it with the underlying binary map of the segmentation preprocess. With the proposed approach, a better precision and generality is obtained with respect to related feature-level likelihoods such as color histograms, while keeping low computational requirements.

• Type: Conference
• Year: 2008
• Where: VMV
• Authors: Claus Lenz, Giorgio Panin, Alois Knoll
• Keywords: Corresponding Filled Model, GPU-accelerated Particle Filter, Real-time Object Tracking, Visualization, VMV 2008