Long-Term Object Tracking Method Based on Dimensionality Reduction

Abstract

Long-term object tracking is challenging as target objects often undergo drastic appearance changes over time. Recently, the FDSST algorithm has performed very well which reduces number of the FFT by dimensionality reduction and operates at real-time. But in case of long-term tracking the performance of FDSST degrades, and the existing long-term tracking methods cannot guarantee the accuracy and real-time performance simultaneously. To solve the above problems, we input a set of sample patches of the target appearance to a multi-channel correlation filter to locate the position of the target in a new frame. At the same time, the number of FFTs is reduced by dimensionality reduction, and an online SVM is trained as the detector to ensure the accuracy of target tracking. Finally, we get a method to track long-term object accurately and in real time. To evaluate the method, we did extensive experiments on a benchmark with 100 sequences. The results show that the proposed method performs well both in accuracy and real-time performance and outperforms than the state-of-the-art methods.