Particle Filter Based Object Tracking with Discriminative Feature Extraction and Fusion

This paper presents an object tracking algorithm based on the unscented particle filtering (UPF) approach. In this algorithm, occlusion tolerant features are first obtained for the images of the object in the consecutive frames based on the color, texture and shape (edge) information, and then a variant of the Fisher's linear discriminant function approach is applied for reducing the dimensionality of the feature space. Similarities of the two images in each feature dimension are computed by matching the histograms of the quantized feature values, and finally these similarity values are aggregated into an over all similarity measure by a novel feature fusion technique embedded in the UPF framework. Results of experimentation with two different data sets indicate that our algorithm is both efficacious in handling severe occlusions (almost as high as 80%) and efficient with respect to tracking accuracy...