A novel multi-feature fusion method for tracking based on discriminative power of feature

Visual object tracking essentially deals with nonstationary data, both the target and background that change over time, and no single feature can remain reliable in various situations. Most existing multiple feature fusion trackers simply used fixed weights to combine the features. In this paper, we propose a novel multiple features fusion approach which can adaptively evaluate and adjust the effect of each feature online. The framework is embedded in particle filter, different feature extraction mechanisms are applied to train and update different Incremental Fisher Linear Discriminant Analysis (IFLD) classifiers online independently. The IFLD classifiers label the particles, target or background, and determine the weights to generate likelihood maps. The fusion of the likelihood maps is accomplished with a linear fusion method and the confidence score is adaptively determined by measuring the separability of foreground and background, as we believe that the feature which best distinguishes between object and background is the best feature for tracking. Experimental results demonstrate the robustness of our algorithm in handling appearance changes, low contrast image and cluttered background. Compared to other state-of-the-art algorithms, our method is more accurate.