Two-Space Spatial-Temporal Regularized Correlation Filters for UAV-Based Aerial Tracking

Abstract

Discriminative Correlation Filters (DCF) are proved to be effective and efficient tracking methods for unmanned aerial vehicle (UAV) because of their high speed and accuracy. However, DCF suffers from unwanted boundary effects. Spatially regularized DCF (SRDCF) tackles this problem by adding spatial penalty on DCF coefficients. Afterwards, spatial-temporal regularized DCF (STRCF) introduces temporal regularization in the DCF coefficient space to SRDCF for more robust appearance learning and faster computation. In this work, we present two-space spatial-temporal regularized DCF (TSSTCF) by introducing spatial-temporal regularization in the regression space to STRCF. The spatial-temporal regularization in the regression space is complement with it in the DCF coefficient space, leading to more robust target appearance learning. More importantly, in the regression space, we can easily control the learning of every single circular shifted sample. This property helps the model suppress the responses of hard negative samples while still keep temporal regularization of other samples. Furthermore, an adaptive factor is introduced to dynamically control the penalty quantity in the regression space. The TSSTCF model can be efficiently solved via the Alternating Direction Method of Multipliers (ADMM). Extensive experiments have been done to prove the superiority of our method.