Adaptive multiple-feature fusion for moving-object tracking using particle filter

This paper presents an adaptive mechanism for fusing multiple features for moving-object tracking in video sequences using a particle filter. The multiple features are fused to improve the representability of the tracking target. However, tracking success or failure also depends primarily on how distinguishable an object is from its adjacent background. Thus, to further improve the tracking performance, the proposed mechanism not only fuses multiple features to represent the tracking target, but dynamically balances the effects of feature similarity between the target object and the candidate with the feature discriminability between the target object and its adjacent background. We demonstrate the algorithm using color, edges, and local binary pattern (LBP) texture. Color is described with part-wise normalized histograms. Edges are described with histograms of gradient directions that represent the shape and the internal edges of a target. The improved LBP texture is adopted to describe the texture histograms. We use the Bhattacharyya coefficient to define feature similarity, and we adopt the variance ratio of the log-likelihood function to describe feature discriminability. The performance of the proposed method was evaluated on numerous sequences involving different types of challenges, including scale changes, occlusion, and rotation. The experimental results show that the proposed method was more efficient and robust than classical approaches.