High-speed tracking based on multi-CF filters and attention mechanism

Recently, correlation filters and deep convolutional network show good performance for visual tracking. Many real-time and high accuracy tracking algorithms are realized; however, there are still some challenges to build a robust tracker. In this paper, we present a novel tracking framework named multi-attention filter (MAF) to solve some challenges for tracking like object drift in a long time, lack of training samples and fast motion. Our framework consists of two components, a basic CNN network to extract feature maps and a set of classifiers to distinguish between the target and the background. First, to solve the problem of object drift in a long time, a simple but effective evaluation mechanism is proposed to the framework, the evaluation mechanism checks the tracking results and corrects it when needed. In addition, the results from different classifiers are fused to predict the object location according to intersection over union. Second, to overcome the lack of training samples, MAF stores some positive and negative samples in two queues, one named long-term queue and another named short-term queue. Third, to deal with fast motion of the target, attention mechanisms including channel attention and location attention are added to the tracker. In our experiments on the popular benchmarks including OTB-2013 and OTB-2015. MFA achieves state of the art among trackers, and as a correlation filter framework, MAF is very flexible and has great rooms for improvement and generalization.

Automated summary

This paper presents a novel tracking framework named multi-attention filter (MAF) to address challenges such as object drift over time, lack of training samples, and fast motion in tracking. The framework achieves good performance in tracking tasks by incorporating attention mechanisms and evaluation mechanisms, and outperforms existing trackers on popular benchmarks.