Improved Background-Aware Correlation Filters based on Content-Related Spatial Regularization

摘要

Discriminative correlation filters(DCF) have caught much attention in visual tracking. However, DCFs suffer from the boundary effects. To alleviate the boundary effects, the background-aware correlation filter (BACF) approach has been proposed to resolve this issue by enforcing the filter coefficients outside the target bounding box to zeros, which can improve the tracking accuracy and robustness. However, when occlusion occurs, the target may drift and even leave the target bounding box and the correlation coefficients of the true target could be falsely forced to zeros. In that situation, the tracker drifts towards distracters and loses the target. To tackle this limitation, we put forward a content-related spatial regularization map, which can enhance the ability of picking the target from background. By incorporating both content-related and spatial regularization map into the BACF objective function, our tracker improves the PR score and the AUC score by 4.5% and 4.1% respectively over BACF on OTB2013 datasets.