Efficient minimum error bounded particle resampling L1 tracker with occlusion detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Sparse representation is used in visual tracking for target identification via minimum reconstruction error.
• Existing L1 trackers incur high computational costs and have uncharacterized occlusion insensitivity.

Purpose of the Study:

• To propose an efficient L1 tracker, the bounded particle resampling (BPR)-L1 tracker.
• To incorporate a minimum error bound and occlusion detection for enhanced tracking performance.

Main Methods:

• Calculates a minimum error bound from a linear least squares equation to guide particle resampling within a particle filter (PF) framework.
• Employs two-step testing (τ testing and max testing) to efficiently remove insignificant samples before computationally expensive L1 minimization.
• Performs occlusion detection by analyzing trivial coefficients in L1 minimization to improve template updating.

Main Results:

• The BPR-L1 tracker achieves significant speed-up through efficient sample removal techniques.
• Occlusion detection enhances template updating, improving robustness in challenging scenarios.
• Experimental evaluation on diverse video applications shows superior performance compared to nine state-of-the-art trackers on eleven benchmark sequences.

Conclusions:

• The BPR-L1 tracker offers an efficient and robust solution for visual tracking, particularly in scenarios with occlusions.
• The proposed minimum error bound and occlusion detection mechanisms are beneficial for particle filter-based trackers, especially those using sparse representations.