Redefined target sample-based background-aware correlation filters for object tracking

Wanli Xing, Hong Zhang, Yujie Wu, Yawei Li, Ding Yuan

Abstract: Object tracking has always been an important research task in computer vision. Since the bounding box is rectangular, there are inevitably a large number of background contents in target samples. When the background changes drastically, it is easy to cause template drift. Traditional methods alleviate template drift by discarding samples and restricting tracker updates when the background interference is severe. However, it is difficult to assess the degree of interference, and these methods also waste part of the undisturbed target information. To bridge these gaps, we proposed a Redefined Target Sample-Based Background-Aware Correlation Filters (RTSBACF), which effectively reduces background interference in target samples and makes full use of undisturbed target information. Specifically, we first proposed a tracking model with the redefined target samples which are calculated by sample suppression items and compensation samples. Then, based on this model, a strategy to generate sample suppression items and compensation samples was designed to reduce background interference in target samples by redefining target sample areas without modifying background sample areas. Finally, we adopted the combined features which are more suitable for redefined target samples, and designed the scale penalties for the combined features to further enhance tracking performance. The experimental results demonstrated that the performance of our proposed tracker exceeds several advanced trackers. In addition, compared with the baseline tracker, the overall precision and success rate of the proposed RTSBACF increased by 6.1% and 5.2% in the OTB2013 dataset and by 9.5% and 5.3% in the TC128 dataset, respectively.