Improved Correlation Filter Tracking with Enhanced Features and Adaptive Kalman Filter

Abstract

In the field of visual tracking, discriminative correlation filter (DCF)-based trackers have made remarkable achievements with their high computational efficiency. The crucial challenge that still remains is how to construct qualified samples without boundary effects and redetect occluded targets. In this paper a feature-enhanced discriminative correlation filter (FEDCF) tracker is proposed, which utilizes the color statistical model to strengthen the texture features (like the histograms of oriented gradient of HOG) and uses the spatial-prior function to suppress the boundary effects. Then, improved correlation filters using the enhanced features are built, the optimal functions of which can be effectively solved by Gauss–Seidel iteration. In addition, the average peak-response difference (APRD) is proposed to reflect the degree of target-occlusion according to the target response, and an adaptive Kalman filter is established to support the target redetection. The proposed tracker achieved a success plot performance of 67.8% with 5.1 fps on the standard datasets OTB2013.

Fast and Robust Object Tracking Using Tracking Failure Detection in Kernelized Correlation Filter

“…Bertinetto et al [16] presented sum of template and pixel-wise learners (STAPLE) method using HOG and color histograms for target representation. More recently, Yijin et al [17] proposed Parallel Correlation Filters, Zhaohui et al [18] presented a tracker using correlation filter fused with color histogram and Hao et al [19] proposed a correlation filter-based tracker using enhanced features and adaptive Kalman filter.…”