Mining Spatial-Temporal Similarity for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Correlation Filter (CF) is vital for visual object tracking, improving accuracy and speed.
• Existing CF methods often underutilize the spatial-temporal information present in video sequences.
• Consecutive video frames typically exhibit high similarity, offering potential for predictive tracking.

Purpose of the Study:

• To develop a novel Correlation Filter (CF) tracking framework that exploits spatial-temporal priors in videos.
• To enhance the performance of visual object tracking by predicting target states using previous observations.
• To address the limitations of existing CF methods in leveraging video inherent properties.

Main Methods:

• Proposing a prediction-based CF tracking framework that learns spatial-temporal similarity between consecutive frames.
• Utilizing learned similarities for sample management, template regularization, and training response pre-weighting.
• Developing a novel objective function and effective optimization algorithms for the learning task.

Main Results:

• The proposed scheme significantly improves the accuracy of Correlation Filter (CF) tracking.
• Two implemented CF trackers demonstrate competitive performance against state-of-the-art methods on popular benchmarks.
• Experimental validation confirms the efficacy and efficiency of the proposed prediction-based CF tracking method.

Conclusions:

• The novel prediction-based CF framework effectively utilizes spatial-temporal video information for improved tracking.
• The method offers significant accuracy boosts and competitive performance for real-world visual tracking applications.
• Further analysis confirms the method's efficacy and the trackers' efficiency.