Visual Tracking via Weighted Local Cosine Similarity

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Visual tracking is crucial for analyzing dynamic scenes.
• Existing methods often struggle with target appearance variations and complex backgrounds.

Purpose of the Study:

• To introduce a novel weighted local cosine similarity (WLCS) for improved visual tracking.
• To develop a discriminative objective function for obtaining optimal feature weights.
• To create an effective and efficient particle filter-based tracker.

Main Methods:

• Proposed weighted local cosine similarity (WLCS) to measure target-candidate similarities.
• Developed an objective function to model discriminative component abilities.
• Utilized quadratic programming to derive discriminative weights.
• Integrated WLCS into a particle filter framework with a simple update strategy.

Main Results:

• The proposed WLCS method effectively measures similarities between target templates and candidates.
• Discriminative weights were obtained by solving an objective function using quadratic programming.
• The WLCS-based tracker demonstrated superior performance on challenging image sequences compared to existing methods.

Conclusions:

• The novel WLCS method offers a robust approach for visual tracking.
• The proposed tracker is both effective and efficient, outperforming competing techniques.
• WLCS provides a valuable tool for enhancing the accuracy of visual tracking systems.