Particle filter for object tracking
Part 1: project profile
Project name
Particle filter for object tracking.
Project short description
The aim of this project is to construct a robust particle filter for object tracking, able to follow a moving object given its starting position (in a fixed scene). To obtain this goal, we compare different similarity measures, such as color histograms and joint spatial-color mixtures of gaussians, in different color spaces (RGB, HSV).
Dates
Start date: 01/01/2008
End date: 15/07/2008
Internet site(s)
People involved
Project head(s)
Matteo Matteucci - matteucc (at) elet (dot) polimi (dot) it
Other Politecnico di Milano people
Davide Migliore - migliore (at) elet (dot) polimi (dot) it
Students
Manuel Fossati - manuel (dot) fossati (at) mail (dot) polimi (dot) it
External personnel:
Laboratory work and risk analysis
Since laboratory work for this project is limited to software related activities, there are no potential risks.
Part 2: project description
State of the art
This project is based on the paper Effective Appearance Model and Similarity Measure for Particle Filtering and Visual Tracking, Hanzi Wang, Suter David, Schindler Konrad, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2007.
We compared the SMOG model proposed in the paper with a standard color histogram, using the PETS dataset as our primary test source.
The Problem
The standard color histogram model used in particle filtering suffers from several limitations, and fails under many circumstances.
The Idea
Wang et al. propose a new adaptive model which incorporates additional information about the spatial layout, making the similarity measure more discriminative. This model is known as SMOG (joint spatial-color mixture of gaussians). The main idea is to keep track of both color and spatial distributions, in order to be able to successfully identify the object even in critical conditions (such as changes in foreground appearance, occlusions, presence of similar objects).
The Project
The particle filter algorithm has been implemented in C, using both histogram and SMOG models and allowing the user to choose which one to use.
The algorithm is as follows:
- model the object appearance (color histogram/SMOG)
- initialize random samples
- propagate the samples (according to a motion model)
- assign a weight to each sample according to its similarity measure
- estimate the new object position as the weighted mean of samples
- update the object model
- repeat from step 3
Experimental Results
Some sample videos showing the particle filter in action can be downloaded here.
Conclusions
The performed tests have shown that the SMOG model is indeed more robust than the standard color histogram, allowing us to succesfully track objects under many different circumstances.
