APR-Net Tracker: Attention Pyramidal Residual Network for Visual Object Tracking
Bing Liu, Di Yuan, Xiaofang Li · IEEE Transactions on Emerging Topics in Computational Intelligence · 2024
Visual object tracking has attracted much attention thanks to its remarkable capability to identify a moving target accurately in real-world video scenarios. Recently, tracking performance has improved significantly. However, there is still a lot of progress space to achieve consummate tracking performance because of the complex and varied target and its surrounding background. How to trade-off between accuracy, robustness, and approximately real-time performance is crucial for visual object tracking. We find that when attention mechanisms are integrated into the pyramidal residual network, they can subtly describe the appearance of the target object, which provides a strong guarantee for accurate, robust, and high-performance visual object tracking. Based on this find, we propose a simple yet powerful visual object tracking algorithm, APR-Net, that automatically searches various scales and significant regions. The comprehensive experimental results on multiple benchmarks demonstrate that the proposed APR-Net tracker attains the desired performance improvement in accuracy and robustness while achieving near real-time tracking performance in various real-world applications.