Particle filter based scale adaptive compressive tracking

Compressive Tracking is a very popular vision tracking method based on Compressive Sensing theory. In the Compressive Tracking, the measurement matrix is used to transform the image patch to a feature vector and plays a fundamental role in the tracking procedure. However, based on our analysis the traditional way of constructing the measurement matrix has intrinsic problems. In this paper, we propose a loop-blocked matrix which can extract more complete and discriminative information than the original one. In order to make our method robust to scale variation, a scale adaptive window model is also developed and its parameters are estimated by the particle filter. Regarding to the issue of occlusion, a forgetting model is proposed to improve the tracking robustness, especially when complete occlusion happens or the occlusion lasts too long. Experiments show that our algorithm has good adaption to the scale changes of the target in the image and good robustness to occlusion.