Real-Time nonparametric background subtraction with tracking-based foreground update

Description

This site contains some supplementary material associated to the detection strategy proposed in [*]:

• Downloadable software pack.
• Results.

The work [*] proposes a high-quality nonparametric moving object detection strategy, along with its real-time implementation in a GPU. This strategy features robust spatiotemporal models of both the background and the foreground, the latter augmented with a novel tracking system based on a particle filter capable of dealing with a variable and unknown number of moving regions.The filter updates the positions of reference data and provides prior probability estimations for a Bayesian classifier that is able to combine spatio-temporal models with different spatial distribution of reference data. In addition, the background model is selectively analyzed thanks to the automatic selection of regions of interest in the input images, yielding equivalent results at a fraction of the computational cost.

For any question about the article [*] or about the described test data, please contact Daniel Berjón at This email address is being protected from spambots. You need JavaScript enabled to view it. or Carlos Cuevas at This email address is being protected from spambots. You need JavaScript enabled to view it..

Citation

[*] D. Berjón, C. Cuevas, F. Morán, and N. García, "Real-time nonparametric background subtraction with tracking-based foreground update", Pattern Recognition, vol. 74, pp. 156-170, Feb. 2018. (doi: 10.1016/j.patcog.2017.09.009).

Software

The source code can be downloaded here and a binary can be downloaded here.

The binary has been compiled for 64-bit Linux systems and a GPU with CUDA capabilities >= 2.0; it depends on the following external libraries to be installed:

• CUDA Runtime, version >= 4.x
• FreeImagePlus, version >= 3.x
• Boost.Filesystem and Boost.System, version >= 1.54
• Qt, version >= 5.x

Only two parameters may be indicated in the execution of the binary:

• Number of reference images in the background modeling.
  • It must be high enough to cover the dynamic changes of the background.
  • Default value: 200
• Width (appearance) of the Gaussian kernels in the foreground modeling.
  • The larger this value, the less selective is the foreground model, and therefore it has less weight in the Bayesian classifier.
  • Default value: 0.02 (range [0,1]).

Results:

SABS database [1]:

The sequences and the ground-truth can be downloaded from here.

Our results can be downloaded from here.

STAR database [2]:

These sequences and the ground-truth can be downloaded from here.

Our results can be downloaded from here.

LASIESTA database:

These sequences and the ground-truth can be downloaded from here.

Our results can be downloaded from here.

Software

1. S. Brutzer, B. Höferlin, and G. Heidemann, “Evaluation of background subtraction techniques for video surveillance,” IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1937–1944, 2011.
2. L. Li, W. Huang, I. Y.-H. Gu, and Q. Tian, “Statistical modeling of complex backgrounds for foreground object detection,” IEEE Transactions on Image Processing, vol. 13, no. 11, pp. 1459–1472, 2004.