Background Subtraction via Generalized Fused Lasso Foreground Modeling
PEKING UNIVERSITY
Bo Xin, Yuan Tian, Yizhou Wang and Wen Gao
Abstract
Optimization
IEEE 2015 Conference on Computer Vision and Pattern Recognition
Experiments
Background Subtraction (BS) is one of the key steps in video analysis. Due to challenges such as illumination change, dynamic background etc. the resulted foreground segmentation often consists of holes as well as background noise. In this regard, we consider generalized fused lasso (GFL) regularization to quest for intact structured foregrounds. Together with certain assumptions about the background, we formulate BS as a matrix decomposition problem using regularization terms for both the foreground and background matrices. The optimization was carried out via applying the augmented Lagrange multiplier (ALM) method in such a way that a fast parametric-flow algorithm is used for updating the foreground matrix. Experimental results on several popular BS data sets demonstrate better than state-of-the-arts performance.
Formulation Given a sequence of images D β π
πΓπ , we assume it is the summation of a low-rank background matrix B and a structured sparse foreground matrix πΉ characterd by generalized fused lasso. (π (π) denotes the π-th column of πΉ and ππ is the π-th entry of vector π). min ππππ π΅ + πΉ
{π΅,πΉ}
πΉ
πππ
πππ
π
(π) π€ππ |ππ
β1 + π
π=1
β
(π) ππ
|
π,π βπ
When pure background frames are labelled, we separate π· = [π·1 , π·2 ], where π·1 is pure background. We separate π΅ and πΉ accordingly. min ππππ [π΅1 , π΅2 ] + πΉ
We assume rank π΅1 , π΅2 min π
{π,πΉ2 }
1
πΏ π΅, πΉ; π, π = π΅
β
+ πΉ
π + π, π· β π΅ β πΉ + π·βπ΅βπΉ πππ 2
2 πΉ
Algorithm βπ
{π΅,πΉ}
Optimization was carried out via ALM. The Lagrange function is
π . π‘. π· = π΅ + πΉ
π
=
Figure. Alternated updating of the background and the foreground. In each iteration (iter) either the background model or the foreground is updated and the objective value (the green plots) keeps decreasing until convergence.
πππ
π . π‘. π·2 = π΅2 + πΉ2 πππ π·1 = π΅1
= ππππ π΅1 and derive to + πΉ2
πππ
π . π‘. π·2 = π·1 π + πΉ2
1. Input: π· β π
πΓπ , π β₯ 0. 2. Output: B, F β π
πΓπ . 3. Initialization: Set π0 = 0, π΅0 = 0, F0 = 0, π0 = 0, π½ > 1 and ππππ₯ 4. while not converged do 5. π΅π+1 = πππππππ΅ πΏ(π΅, πΉπ , ππ , ππ ) 6. πΉπ+1 = πππππππΉ πΏ(π΅π+1 , πΉ, ππ , ππ ) 7. ππ+1 = ππ + ππ (π· β π΅π+1 β πΉπ+1 ) 8. ππ+1 = min{π½ππ , ππππ₯ } 9. Return π΅π and πΉπ .
Code available at: http://idm.pku.edu.cn/staff/boxin/
Figure. Results on the SABS data set. F-scores are shown Table. Results on the Li data set. F-scores are shown