Online Weighted Multi-instance Learning For Visual Tracking

Visual tracking is one of the cardinal problems of computer vision. It has been widely used in many applications such as surveillance, driving assistant systems and augmented reality to human-computer interaction. As an active research topic, visual tracking has been extensively studied. However, it is still a challenging problem to track a target in real world environment because there are many influencing factors such as occlusion, illumination change, motion blur and clutter background. Based on multi-instance learning algorithm, two algorithms are proposed for visual tracking.Firstly, an online feature fusion algorithm with weighted multi-instance learning for visual tracking is proposed. Multiple instance learning method for visual tracking can alleviate drift to some extent. However, the MIL tracker does not discriminatively consider the sample importance in its learning procedureI In order to solve this problem, two classifiers with different features (HOG and Haar-like) are trained accordingly and those classifiers are online fused into a weighted multiple instance learning framework.Secondly, a robust visual tracking algorithm based on online discriminant feature selection and weight multiple instance learning is proposed. In this algorithm, an online discriminative feature selection algorithm is adopted, which optimizes the objective function in the steepest ascent direction. The average weak classifier output and the average gradient direction instead of each gradient direction is used for every samples. Then two classifiers with different features (HOG and Haar-like) are trained accordingly and fused into a weighted multiple instance learning framework to achieve a robust visual tracking algorithm.Finally, numerous experimental evaluations with state-of-the-art algorithms on challenging sequences demonstrate that the proposed tracking algorithms can handle the background clutter, occlusion, motion blur and illumination change and work well.