Related Experiment Videos
MCMC-based particle filtering for tracking a variable number of interacting targets.
Zia Khan1, Tucker Balch, Frank Dellaert
1GVU Center, College of Computing, Georgia Institute of Technology, Atlanta 30332-0760, USA. zkhan@cc.gatech.edu
IEEE Transactions on Pattern Analysis and Machine Intelligence
|November 16, 2005
Summary
This study introduces an improved particle filter using Markov random fields (MRF) and Markov chain Monte Carlo (MCMC) sampling. This approach enhances multitarget tracking, especially for interacting targets, by reducing failures and improving efficiency.
Area of Science:
- Computer Vision
- Robotics
- Machine Learning
Background:
- Particle filters are widely used for tracking but struggle with interacting targets.
- Maintaining target identity during interactions is a significant challenge in multitarget tracking.
Purpose of the Study:
- To develop an efficient particle filter capable of handling interacting targets.
- To reduce tracker failures and improve computational efficiency in multitarget tracking scenarios.
Main Methods:
- Incorporated a Markov random field (MRF) motion prior into the particle filter's importance weights.
- Replaced traditional importance sampling with Markov chain Monte Carlo (MCMC) sampling for improved efficiency.
- Extended the MCMC-based filter to accommodate a variable number of targets.
Main Results:
- The MRF prior effectively maintains target identity during interactions.
- The MCMC-based filter significantly reduces computational requirements compared to traditional methods.
- The proposed filters demonstrate efficient and effective handling of complex target interactions.
Conclusions:
- The novel MCMC-based particle filter offers an efficient solution for multitarget tracking with interacting targets.
- This method improves robustness and scalability for complex tracking problems.