MCMC data association and sparse factorization updating for real time multitarget tracking with merged and multiple...

Area of Science:

• Computer Vision
• Robotics
• Signal Processing

Background:

• Existing multitarget tracking algorithms struggle with multiple measurements per target and merged measurements between interacting targets.
• Current methods, often derived from radar tracking, are insufficient for these complex scenarios.

Purpose of the Study:

• To develop a novel probabilistic model for tracking interacting targets with complex measurement types.
• To create an efficient algorithm for approximate inference in this model.

Main Methods:

• Introduced a probabilistic model for interacting targets capable of handling multiple and merged measurements.
• Developed a Markov chain Monte Carlo (MCMC)-based auxiliary variable particle filter for approximate inference.
• Employed Rao-Blackwellization to eliminate sampling over continuous state spaces and utilized sparse least squares for efficient updating/downdating.

Main Results:

• The proposed algorithm effectively addresses multiple and merged measurements from interacting targets simultaneously.
• Sparse least squares techniques significantly reduced computational cost per Markov chain iteration.
• The algorithm demonstrated real-time performance on a conventional PC using video and laser range data.

Conclusions:

• The novel probabilistic model and MCMC-based algorithm provide an effective solution for complex multitarget tracking scenarios.
• Computational efficiency was achieved through sparse least squares and a heuristic for focusing on interacting targets.
• The algorithm's accuracy and real-time capabilities were validated using challenging simulations and dual-sensor modalities.