Segmentation Tracking and Recognition Based on Foreground-Background Absolute Features, Simplified SIFT, and Particle Filters

We propose an approach to tracking and recognition based on segmentation by scanning foreground-background absolute difference (FBAD) features, simplified scale-invariant feature transform (s-SIFT), and evolutionary particle filter. Particle filter is shown to be efficient in visual tracking due to its sequential propagation ability of the conditional posterior density of the states, i.e., the tracking parameters. First, we obtain FBAD features and perform segmentation tracking of moving objects by 4-directional scanning. Second, the segmentation mask is applied to the SIFT key-points to obtain the key-points of moving objects. Third, those reduced key-points and the associated key-descriptors are found by our simplified technique. Once the reference SIFT key-descriptors are registered, two different matching procedures, a full-search technique and an evolutionary particle filter approach, are applied. The experiments show that both schemes are robust and efficient in visual tracking and recognition even if a target object is occluded in a cluttered background.