Particle filter based on real-time Compressive Tracking

It remains to be a challenging task to develop effective and efficient models for robust object tracking due to occlusion, motion blur, pose variation, illumination change and other factors. Compressive Tracking (CT) method proposed by Zhang performs perfectly in real-time detecting due to its simple computational load. However, the accuracy will decline due to the imperfection of prediction model as slight inaccuracies lead to cumulative faults in training examples selection, then the classifier degrades and the object is lost in tracking process. Particle filtering (PF), a framework widely used in object tracking, is highly extensible and is able to handle non-linearity and non-normality to some extent. We integrate compressive sensing into particle filtering, so that the strengths of the both methodologies are incorporated into the algorithm. Features are compressively sensed for every particle, and the result is regarded as the weight. Meanwhile, a second-order auto regressive model is introduced for particle transition model in order to predict motions of object efficiently. Therefore, our algorithm can surmount overlap and ambiguities and handle drifting problem flexibly. In terms of efficiency, accuracy and robustness, the combined tracking algorithm runs in real-time and performs favorably against state-of-the-art methods on some challenging sequences.