Abrupt motion tracking via intensively adaptive Markov-chain Monte Carlo sampling

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Robust tracking of abrupt motion is a significant challenge in computer vision due to high motion uncertainty.
• Existing methods like particle filters and Markov-chain Monte Carlo (MCMC) often struggle with the local-trap problem and slow convergence.

Purpose of the Study:

• To propose a novel sampling-based tracking scheme within the Bayesian filtering framework specifically for abrupt motion.
• To enhance the accuracy and efficiency of visual tracking systems dealing with sudden, unpredictable movements.

Main Methods:

• Introduced Stochastic Approximation Monte Carlo (SAMC) sampling into the Bayesian filter to adaptively estimate the filtering distribution.
• Developed a new MCMC sampler with intensive adaptation, combining a density-grid-based predictive model with SAMC for proposal adaptation.
• Implemented a sampling-based tracking scheme to address the local-trap problem and improve convergence rates.

Main Results:

• The proposed method effectively handles the local-trap problem inherent in abrupt motion tracking.
• Achieved a good approximation to the target distribution through adaptive estimation during sampling.
• Demonstrated superior effectiveness and computational efficiency compared to alternative tracking algorithms in experiments.

Conclusions:

• The novel sampling-based tracking scheme provides an effective and efficient solution for robustly tracking abrupt motion in computer vision.
• The integration of SAMC and an adaptive MCMC sampler significantly improves sampling efficiency and tracking performance.
• Extensive experiments validate the method's capability in handling diverse abrupt motion scenarios.