Robust Multitask Multiview Tracking in Videos.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Sparse representation methods are common for visual tracking.
• Least squares (LSs) criteria in traditional methods are sensitive to heavy-tailed noise.
• Robustness is crucial for accurate tracking in challenging environments.

Purpose of the Study:

• To develop a robust visual tracking approach using approximate least absolute deviation (LAD).
• To leverage multitask and multiview learning for enhanced tracking performance.
• To integrate multiple visual features (intensity, color, texture) for improved accuracy.

Main Methods:

• An approximate least absolute deviation (LAD)-based multitask multiview sparse learning method is proposed.
• The method is integrated into a particle filter framework, treating each particle's view as a task.
• A unified robust multitask formulation based on LAD exploits relationships between tasks.
• Representation matrix decomposition enhances approximation robustness and accuracy.
• Nesterov's smoothing and accelerated proximal gradient methods are used for efficient solving.

Main Results:

• The proposed tracker, using four feature types, demonstrates superior performance on synthetic and real-world sequences.
• Evaluated on the CVPR2013 tracking benchmark and ALOV++ dataset.
• Qualitative and quantitative results show significant improvements over state-of-the-art trackers.

Conclusions:

• The proposed approximate LAD-based multitask multiview sparse learning tracker offers robust and accurate visual tracking.
• The method effectively handles heavy-tailed noise and integrates diverse visual features.
• This approach represents a significant advancement in robust visual object tracking.