Real-Time Tracking Based on Keypoints and Discriminative Correlation Filters

Recently, tracking methods based on discriminative correlation filters (DCF) and CNN features have seen a great improvement in accuracy and performance. However, increasingly complex models and heavy computation burdens can reduce their speed and real-time capability. In this paper, we analyze the key factors that increased computation in the state-of-the-art DCF trackers and provide some solutions for solving these problems. We propose a novel method for real-time tracking based on keypoint consensus clustering and improved DCFs. First, we use the consensus-based keypoint clustering scheme to coarsely locate the bounding box, which employs the geometric compatibility of keypoints and separates correct correspondences from erroneous ones by voting. Next, we propose regions around the estimated location and exploit an improved DCF to track the object. In the new DCF, we modify the core formulation and reduce the number of parameters. In our tracking scheme, we also propose a strategy to collect a training sample set based on keypoints, which contributes to a clear acceleration in training. The experiments are carried out based on two well-known benchmarks: VOT2016 and OTB2015. Compared to the state-of-the-art tracking systems, our method can provide a competitive accuracy with a significant improvement in tracking speed.