Home


Research


Publications


Downloads


Teaching


Links

Video Background Subtraction Based on Dynamic Group Sparsity

.Introduction

Background subtracted images are typical DGS data. They generally correspond to the interested foreground objects. Compared with the whole scene, they tend to be not only spatially sparse but also cluster into dynamic groups, although the sparsity number and group structures are not known. As we know, the sparsity number must be provided in most of current recovery algorithms, which make them impractical for this problem. In contrast, the proposed AdaDGS can apply well to this task since it not only can automatically learn the sparsity number and group structures but also is a fast enough greedy algorithm.


Reference

Junzhou Huang, Xiaolei Huang, Dimitris Metaxas, ”Learning with Dynamic Group Sparsity”, The 12th International Conference on Computer Vision, Kyoto, Japan, October 2009. [SLIDES] [CODE]

Notice: The codes was tested on Windows and MATLAB 2008. If you have any suggestions or you have found a bug, please contact us via email at jzhuang@uta.edu


Video Datasets for Evaluations

We provide datasets used in evaluating background substraction. If you use them in your work, please cite our paper and related rescources correctly.

Original Video

1) "WaterObject" dataset

2) "RainCar" dataset

3) "OceanPerson" dataset

Resulted Video

1) "WaterObject" dataset

2) "RainCar" dataset

3) "OceanPerson" dataset



Figure 3.1 Results on the "WaterObject" dataset (a) original frame, (b) background subtracted image with AdaDGS, (c) the binary maskwith AdaDGS and d) with robust Kalman model

 

Figure 3.2 Results on "RainCar" dataset. (a) original frame, (b) with AdaDGS, (c) with KDE model, (d) with MoG

 

Figure 3.3 Results on "OceanPerson" dataset. (a) original frame, (b) with AdaDGS, (c) with Monnet’s method (PCA) and (d) with MoG


Related Sources

Tracking With Gaussian Mixture Model

Non-Parametric Model for Background Subtraction

Tracking with a Dynamic Model in a PCA Subspace

Dynamic Background Sequences