Adaptive Features Fusion Correlation Filter for Real-time Object Tracking

Abstract

Recently, visual object tracking has encountered many challenges. The response maps of multiple templates are linearly combined with a fixed weight. Meanwhile, the model updating strategy is realized by linear interpolation of continuous frames. It is easy to introduce too much background information and cause model pollution. Herein, we propose a novel tracking approach. The four features including fHOG, local binary pattern (LBP), color, and gray features are integrated into the tracker, which can make the model have strong expression ability. The fusion weight has been adaptively assigned according to the response peak value and smooth constraint of confidence maps (SCCM) indicator. The experimental results show that the overlap precision score (OPS) and distance precision score (DPS) are improved compared with some existing tracking algorithms. The average running frame rate can meet real-time requirements.