Convolution operators for visual tracking based on spatial-temporal regularization

摘要

In recent years, the method based on discriminative correlation filter has been shown excellent performance in short-term visual tracking. However, discriminative correlation filter-based method heavily suffers from the problem of the multiple peaks and model drift in responds maps incurred by occlusion and rotation. To solve the above problem, we proposed convolution operators for visual tracking based on spatial-temporal regularization. Firstly, we add spatial-temporal regularization in loss function, which will guarantee continuity of the model in time. And we use preconditioned conjugate gradient algorithm to obtain filter coefficients. Secondly, we proposed channel reliability to estimate quality of the learned filter and fuse the different reliability coefficients to weight response map in location. We set a threshold to reduce the number of iteration in location and accelerate the compute speed of algorithm. Finally, we use two different correlation filters to estimate location and scale of target, respectively. Extensively experiment in five video sequences show that our tracker has been significantly improved performance in case of occlusion and rotation. The AUC in success plot improves 33.2% than ECO-HC and 41.5% than STRCF, respectively.