Hierarchical Spatiotemporal Context-Aware Correlation Filters for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Discriminative Correlation Filter (DCF)-based trackers offer high precision and speed in visual tracking.
• Existing DCF methods often lack sufficient descriptive power for complex scenarios like occlusion and rapid target variation due to reliance on spatial information only.

Purpose of the Study:

• To introduce a novel tracking framework that leverages spatiotemporal context for improved accuracy and robustness.
• To develop a hierarchical spatiotemporal context model and an optimization fusion approach for adaptive learning of temporal instances.

Main Methods:

• A hierarchical spatiotemporal context model is proposed, with each layer being a spatial correlation filter learned from different temporal instances.
• An optimization fusion approach adaptively learns the effect of each hierarchical layer.
• An adaptive model update strategy dynamically selects hierarchical layers to boost appearance diversity and reduce parameters.

Main Results:

• The proposed tracker achieves the best success rates among state-of-the-art trackers using handcrafted features.
• Performance is comparable to deep-learning-based trackers on standard benchmarks (OTB-2013, OTB-2015, VOT-2016, UAV-20L).
• The method significantly outperforms deep trackers in terms of speed.

Conclusions:

• The novel framework effectively exploits spatiotemporal context for robust and accurate visual tracking.
• The proposed methods offer a significant advancement over existing DCF trackers, achieving competitive performance with real-time efficiency.