Improved Correlation Filter Visual Tracker By Using Scale Estimation Network

Object tracking is to accurately track target information in continuous video sequences, and the bounding box of target is also indispensable as an important metric for evaluating tracker algorithms. In recent years, correlation filter trackers have been successfully achieved powerful robustness. To estimate object scale, most correlation filter trackers use a simple multi-scale search that has limited the development of correlation filters in precision tracking. We propose a scale estimation network to solve the problem, which uses the experience of bounding box estimation in object detection. Through extensive offline learning, high-level knowledge is incorporated into target estimation. Our scale estimation network is trained to optimize object scale. We further introduce the fusion method between correlation filter and scale estimation network coordinated operation. Our improved tracker evaluated on three benchmarks: OTB2015, VOT2018, and VOT2019. The evaluation experiments on these benchmarks demonstrate that our tracker has better performance.