数据图像处理程序代写

Background Subtraction Using Low Rank
and Group Sparsity Constraints
Xinyi Cui1, Junzhou Huang2, Shaoting Zhang1, and Dimitris N. Metaxas1
1 CS Dept., Rutgers University
Piscataway, NJ 08854, USA
2 CSE Dept., Univ. of Texas at Arlington
Arlington, TX, 76019, USA
{xycui,shaoting,dnm}@cs.rutgers.edu,
jzhuang@uta.edu
Abstract. Background subtraction has been widely investigated in recent
years. Most previous work has focused on stationary cameras. Recently,
moving cameras have also been studied since videos from mobile
devices have increased significantly. In this paper, we propose a unified
and robust framework to effectively handle diverse types of videos,
e.g., videos from stationary or moving cameras. Our model is inspired
by two observations: 1) background motion caused by orthographic cameras
lies in a low rank subspace, and 2) pixels belonging to one trajectory
tend to group together. Based on these two observations, we introduce
a new model using both low rank and group sparsity constraints. It is
able to robustly decompose a motion trajectory matrix into foreground
and background ones. After obtaining foreground and background trajectories,
the information gathered on them is used to build a statistical
model to further label frames at the pixel level. Extensive experiments
demonstrate very competitive performance on both synthetic data and
real videos.
1 Introduction
Background subtraction is an important preprocessing step in video surveillance
systems. It aims to find independent moving objects in a scene. Many algorithms
have been proposed for background subtraction under stationary. In recent years,
videos from moving cameras have also been studied [1] because the number of
moving cameras (such as smart phones and digital videos) has increased significantly.
However, handling diverse types of videos robustly is still a challenging
problem (see the related work in Sec. 2). In this paper, we propose a unified
framework for background subtraction, which can robustly deal with videos from
stationary or moving cameras with various number of righd/non-rigid objects.
The proposed method for background subtraction is based on two sparsity
constraints applied on foreground and background levels, i.e., low rank [2] and
group sparsity constraints [3]. It is inspired by recently proposed sparsity theories
[4, 5]. There are two “sparsity” observations behind our method. First, when
A. Fitzgibbon et al. (Eds.): ECCV 2012, Part I, LNCS 7572, pp. 612–625, 2012.
c
Springer-Verlag Berlin Heidelberg 2012
Background Subtraction Using Low Rank and Group Sparsity Constraints 613
the scene in a video does not have any foreground moving objects, video motion
has a low rank constraint for orthographic cameras [6]. Thus the motion of
background points forms a low rank matrix. Second, foreground moving objects
usually occupy a small portion of the scene. In addition, when a foreground
object is projected to pixels on multiple frames, these pixels are not randomly
distributed. They tend to group together as a continuous trajectory. Thus these
foreground trajectories usually satisfy the group sparsity constraint.
These two observations provide important information to differentiate independent
objects from the scene. Based on them, the video background subtraction
problem is formulated as a matrix decomposition problem. First, the video
motion is represented as a matrix on trajectory level (i.e. each row in the motion
matrix is a trajectory of a point). Then it is decomposed into a background
matrix and a foreground matrix, where the background matrix is low rank,
and the foreground matrix is group sparse. This low rank constraint is able to
automatically model background from both stationary and moving cameras, and
the group sparsity constraint improves the robustness to noise.
The trajectories recognized by the above model can be further used to label
a frame into foreground and background at the pixel level. Motion segments on
a video sequence are generated using fairly standard techniques. Then the color
and motion information gathered from the trajectories is employed to classify
the motion segments as foreground or background. Our approach is validated on
various types of data, i.e., synthetic data, real-world video sequences recorded
by stationary cameras or moving cameras and/or nonrigid foreground objects.
Extensive experiments also show that our method compares favorably to the
recent state-of-the-art methods.
The main contribution of the proposed approach is a new model using low
rank and group sparsity constraints to differentiate foreground and background
motions. This approach has three merits: 1) The low rank constraint is able
to handle both static and moving cameras; 2)The group sparsity constraint
leverages the information of neighboring pixels, which makes the algorithm
robust to random noise; 3) It is relatively insensitive to parameter settings.
2 Related Work
Background Subtraction. A considerable amount of work has studied the
problem of background subtraction in videos. Here we review a few related
work, and please refer to [7] for comprehensive surveys. The mainstream in
the research area of background subtraction focuses on stationary cameras.
The earliest background subtraction methods use frame difference to detect
foreground [8]. Many subsequent approaches have been proposed to model the
uncertainty in background appearance, such as, but not limited to,W4 [9], single
gaussian model [10], mixture of gaussian [11], non-parametric kernel density [12]
and joint spatial-color model [13]. One important variation in stationary camera
based research is the background dynamism. When the camera is stationary, the
background scene may change over time due to many factors (e.g. illumination
614 X. Cui et al.
a) Input video
b) Tracked
trajectories
c) Trajectory
matrix
d) Decomposed
trajectories
e) Decomposed trajectories
on original video
f) Optical flow g) Motion segments
h) Output
Trajectory level separation
Pixel level labeling
Fig. 1. The framework. Our method takes a raw video sequence as input, and produces
a binary labeling as the output. Two major steps are trajectory level separation and
pixel level labeling.
changes, waves in water bodies, shadows, etc). Several algorithms have been
proposed to handle dynamic background [14–16].
The research for moving cameras has recently attracted people’s attention.
Motion segmentation approaches [17, 18] segment point trajectories based on
subspace analysis. These algorithms provide interesting analysis on sparse trajectories,
though do not output a binary mask as many background subtraction
methods do. Another popular way to handle camera motion is to have strong
priors of the scene, e.g., approximating background by a 2D plane or assuming
that camera center does not translate [19, 20], assuming a dominant plane [21],
etc. [22] propose a method to use belief propagation and Bayesian filtering
to handle moving cameras. Different from their work, long-term trajectories
we use encode more information for background subtraction. Recently, [1] has
been proposed to build a background model using RANSAC to estimate the
background trajectory basis. This approach assumes that the background motion
spans a three dimensional subspace. Then sets of three trajectories are randomly
selected to construct the background motion space until a consensus set is
discovered, by measuring the projection error on the subspace spanned by the
trajectory set. However, RANSAC based methods are generally sensitive to
parameter selection, which makes it less robust when handling different videos.
Group sparsity [23, 24] and low rank constraint [2] have also been applied to
background subtraction problem. However, these methods only focus on stationary
camera and their constraints are at spatial pixel level. Different from their
work, our method is based on constraints in temporal domain and analyzing the
trajectory properties.
3 Methodology
Our background subtraction algorithm takes a raw video sequence as input,
and generates a binary labeling at the pixel level. Fig. 1 shows our framework.
It has two major steps: trajectory level separation and pixel level labeling. In
Background Subtraction Using Low Rank and Group Sparsity Constraints 615
the first step, a dense set of points is tracked over all frames. We use an offthe-
shelf dense point tracker [25] to produce the trajectories. With the dense
point trajectories, a low rank and group sparsity based model is proposed to
decompose trajectories into foreground and background. In the second step,
motion segments are generated using optical flow [26] and graph cuts [27].
Then the color and motion information gathered from the recognized trajectories
builds statistics to label motion segments as foreground or background.
3.1 Low Rank and Group Sparsity Based Model
Notations: Given a video sequence, k points are tracked over l frames. Each
trajectory is represented as pi = [x1i, y1i, x2i, y2i, ...xli, yli] ∈ R1×2l, where x and
y denote the 2D coordinates in each frame. The collection of k trajectories is
represented as a k × 2l matrix, φ = [pT1
, pT2
, ..., pTl
]T, φ∈ Rk×2l.
In a video with moving foreground objects, a subset of k trajectories comes
from the foreground, and the rest belongs to the background. Our goal is to
decompose tracked k trajectories into two parts: m background trajectories and
n foreground trajectories. If we already know exactly which trajectories belong
to the background, then foreground objects can be easily obtained by subtracting
them from k trajectories, and vice versa. In other words, φ can be decomposed
as:
φ = B + F, (1)
where B ∈ Rk×2l and F ∈ Rk×2l denote matrices of background and foreground
trajectories, respectively. In the ideal case, the decomposed foreground matrix
F consists of n rows of foreground trajectories and m rows of flat zeros, while
B has m rows of background trajectories and n rows of zeros.
Eq. 1 is a severely under-constrained problem. It is difficult to find B and F
without any prior information. In our method, we incorporate two effective priors
to robustly solve this problem, i.e., the low rank constraint for the background
trajectories and the group sparsity constraint for the foreground trajectories.
Low Rank Constraint for the Background. In a 3D structured scene
without any moving foreground object, video motion solely depends on the scene
and the motion of the camera. Our background modeling is inspired from the
fact that B can be factored as a k×3 structure matrix of 3D points and a 3×2l
orthogonal matrix [6]. Thus the background matrix is a low rank matrix with
rank value at most 3. This leads us to build a low rank constraint model for the
background matrix B:
rank(B) ≤ 3, (2)
Another constraint has been used in the previous research work using RANSAC
based method [1]. This work assumes that the background matrix is of rank
three: rank(B) = 3. This is a very strict constraint for the problem. We refer
the above two types of constraints as the General Rank model (GR) and the
Fixed Rank model (FR). Our GR model is more general and handles more
situations. A rank-3 matrix models 3D scenes under moving cameras; a rank-2
616 X. Cui et al.
matrix models a 2D scene or 3D scene under stationary cameras; a rank-1 matrix
is a degenerated case when scene only has one point. The usage of GR model
allows us to develop a unified framework to handle both stationary cameras and
moving cameras. The experiment section (Sec. 4.2) provides more analysis on
the effectiveness of the GR model when handling diverse types videos.
Group Sparsity Constraint for the Foreground. Foreground moving objects,
in general, occupy a small portion of the scene. This observation motivates
us to use another important prior, i.e., the number of foreground trajectories
should be smaller than a certain ratio of all trajectories: m ≤ αk, where α
controls the sparsity of foreground trajectories.
Another important observation is that each row in φ represents one trajectory.
Thus the entries in φ are not randomly distributed. They are spatially clustered
within each row. If one entry of the ith row φi belongs to the foreground,
the whole φi is also in the foreground. This observation makes the foreground
trajectory matrix F satisfy the group sparsity constraint:
F2,0 ≤ αk, (3)
where  ·2,0 is the mixture of both L2 and L0 norm. The L2 norm constraint is
applied to each group separately (i.e., each row of F). It ensures that all elements
in the same row are either zero or nonzero at the same time. The L0 norm
constraint is applied to count the nonzero groups/rows of F. It guarantees that
only a sparse number of rows are nonzero. Thus this group sparsity constraint not
only ensures that the foreground objects are spatially sparse, but also guarantees
that each trajectory is treated as one unit. Group sparsity is a powerful tool in
computer vision problems [28, 23]. The standard sparsity constraint, L0 norm,
(we refer this as Std. sparse method) has been intensively studied in recent
years. However, it does not work well for this problem compared to the group
sparsity one. Std. sparse method treats each element of F independently. It does
not consider any neighborhood information. Thus it is possible that points from
the same trajectory are classified into two classes. In the experiment section
(Sec. 4.1), we discuss the advantage of group sparsity constraint over sparsity
constraint through synthetic data analysis, and also show that this constraint
improves the robustness of our model.
Based on the low rank and group sparsity constraints, we formulate our
objective function as:

ˆ B, ˆ F

= argmin
B,F

 φ − B − F 2
F

,
s.t. rank(B) ≤ 3,  F 
2,0 < αk, (4)
where  · F is the Frobenius norm. This model leads to a good separation of
foreground and background trajectories. Fig. 2 illustrates our model.
Eq. 4 only has one parameter α, which controls the sparsity of the foreground
trajectories. In general, user-tuning parameter is a key issue for a good model. It
is preferable that the parameters are easy to tune and not sensitive to different
Background Subtraction Using Low Rank and Group Sparsity Constraints 617
= +
Low rank matrix
B=UΣVT
Group sparse
matrix F
…
…
Trajectory
matrix􀁉
Fig. 2. Illustration of our model. Trajectory matrix φ is decomposed into a background
matrix B and a foreground matrix F. B is a low rank matrix, which only has a
few nonzero eigenvalues (i.e. the diagonal elements of Σ in SVD); F is a group
sparse matrix. Elements in one row belongs to the same group (either foreground or
background), since they lie on one trajectory. The foreground rows are sparse comparing
to all rows. White color denotes zero values, while blue color denotes nonzero values.
datasets. Our method is relatively insensitive to parameter selection. Using α =
0.3 works for all tested videos.
Low rank constraint is a powerful method in computer vision and machine
learning area [29]. Low rank constraints and Robust PCA have been recently
used to solve vision problems [30, 2, 31], including background subtraction at the
pixel level [2]. It assumes that the stationary scenes satisfy a low rank constraint.
However, this assumption does not hold when camera moves. Furthermore, that
formulation does not consider any group information, which is an important
constraint to make sure neighbor elements are considered together.
3.2 Optimization Framework
This subsection discusses how to effectively solve Eq. 4. The first challenge is that
it is not a convex problem, because of the nonconvexity of the low rank constraint
and the group sparsity constraint. Furthermore, we also need to simultaneously
recover matrix B and F, which is generally a Chicken-and-Egg problem.
In our framework, alternating optimization and greedy methods are employed
to solve this problem. We first focus on the fixed rank problem (i.e., rank equals
to 3), and then will discuss how to deal with the more general constraint of
rank ≤ 3.
Eq. 4 is divided into two subproblems with unknown B or F, and solved by
using two steps iteratively:
Step 1: Fix B, and update F. The subproblem is:

ˆ F

= argmin
F

 φ
 − F 2
F

, s.t. F
2,0 < αk, (5)
where φ = φ − B.
Step 2: Fix F, and update B. The subproblem is:

ˆB

= argmin
B

 φ
 − B 2
F

, s.t. rank(B) = 3, (6)
where φ = φ − F.
618 X. Cui et al.
To initialize this optimization framework, we simply choose Binit = φ, and
Finit = 0. Greedy methods are used to solve both subproblems. To solve Eq. 5,
we compute Fi2, i ∈ 1, 2, ..., k, which represents the L2 norm of each row. Then
the αk rows with largest values are preserved, while the rest rows are set to zero.
This is the estimated F in the first step. In the second step, φ is computed as per
newly-updated F. To solve Eq. 6. Singular value decomposition (SVD) is applied
on φ. Then three eigenvectors with largest eigenvalues are used to reconstruct
B. Two steps are alternatively employed until a stable solution of ˆB is found.
Then ˆ F is computed as φ − ˆB. The reason of updating ˆ F after all iterations is
that the greedy method of solving Eq. 5 discovers exact αk number of foreground
trajectories, which may not be the real foreground number. On the contrary, B
can be always well estimated, since a subset of unknown number of background
trajectories is able to have a good estimation of background subspace. Thus we
finalize ˆ F by φ − ˆB . Since the whole framework is based on greedy algorithms,
it does not guarantee a global minimum. In our experiments, however, it is
able to generate reliable and stable results. The above-mentioned method solves
the fixed rank problem, but the rank value in the background problem usually
cannot be pre-determined. To handle this undetermined rank issue, we propose
a multiple rank iteration method. First, B and F are initialized as B(0)
init = φ
and F(0)
init = 0k×2l . Then the fixed rank optimization procedure is performed on
each specific rank starting from 1 to 3. The output of the current fixed rank
procedure is fed to the next rank as its initialization. We obtain the final result
B(3) and F(3) in the rank-3 iteration. Given a data matrix of K × 2L with K
trajectories over L frames, the major calculation is O(KL2) for SVD on each
iteration. Convergence of each fixed rank problem is achieved 6.7 iterations on
average. The overall time complexity is O(KL2).
To explain why our framework works for the general rank problem, we discuss
two examples. First, if the rank of B is 3 (i.e., moving cameras), then this
framework discovers an optimal solution in the third iteration, i.e., using rank-3
model. The reason is that the first two iterations, i.e. the rank-1 and rank-
2 models, cannot find the correct solution as they are using the wrong rank
constraints. Second, if the rank of the matrix is 2 (i.e., stationary cameras),
then this framework obtains stable solution in the second iteration. This solution
will not be affected in the rank-3 iteration. The reason is that the greedy
method is used to solve Eq. 6. When selecting the eigenvectors with three largest
eigenvalues, one of them is simply flat zero. Thus B does not change, and the
solution is the same in this iteration. Note that low rank problems can also
be solved using convex relaxation on the constraint problem [2]. However, our
greedy method on unconstrained problem is better than convex relaxation in
this application. Convex relaxation is not able to make use of the specific rank
value constraint (≤ 3 in our case). The convex relaxation uses λ to implicitly
constrain the rank level, which is hard to constrain a matrix to be lower than a
specific rank value.
Background Subtraction Using Low Rank and Group Sparsity Constraints 619
3.3 Pixel Level Labeling
The labeled trajectories from the previous step are then used to label each frame
at the pixel level (i.e. return a binary mask for a frame). In this step, each frame
is treated as an individual labeling task. First, the optical flow [26] is calculated
between two consecutive frames. Then motion segments are computed using
graph cuts [27] on optical flow. After collecting the motion segments, we want
to label each motion segment s as f or b, where f and b denotes the label of
foreground and background. There are two steps to label the segments. First,
segments with high confidence belonging to f and b are selected. Second, a
statistical model is built based on those segments. This model is used to label
segments with low confidence. The confidence of a segment is determined by
the number of labeled f and b trajectories. The low confidence segments are
those ambiguous areas. To predict the labels of these low confidence segments,
a statistical model is built for f and b based on high confidence ones. First, 20%
pixels are uniformly sampled on segments with high confidence. Each sampled
pixel w is represented by color in hue-saturation space (h, s), optical flow (u, v)
and position on the frame (x, y). The reason we use sampled points to build
the model instead of the original trajectories is that the sparse trajectories
may not cover enough information (i.e., color and positions) on the motion
unit. Uniform sampling covering the whole segment can build a richer model.
The segments are then evaluated using a kernel density function:P(si|c) =
1
N·|si|

N
i=1

j∈si
κ(ej − wi), c ∈ {f, b}, where κ(·) is the Normal kernel, N
is the total number of sampled pixels, and |si| is the pixel number of si. For
every pixel j lying on si, ej denotes the vector containing color, optical flow and
position.
4 Experiments
To evaluate the performance of our algorithm, we conduct experiments on different
data sources: synthetic data, real-world videos from both moving and
stationary cameras. The performance is evaluated by F-Measure, the harmonic
mean of recall and precision. This is a standard measurement for background
subtraction [7, 16]: F = 2·recall ·precision/(recall+precision). Since our major
contribution is to separate background/foreground motions at trajectory level,
thus our comparisons and analysis mainly focus on the first step: trajectory level
labeling.
4.1 Evaluations on Synthetic Data
Experimental Settings. A grid of background points and four shapes are
generated to simulate objects moving under camera movements. The homogeneous
representation of each point is (X, Y,Z, 1) as its position in the 3D world.
Then the projected points (x, y) are obtained in a 2D image by (x, y, 1)T =
C · (X, Y,Z, 1)T, where C is a 3 × 4 camera projection matrix. The depth value
620 X. Cui et al.
Z in the 3D world for foreground shapes and background grid is 10±5 and 20±10,
respectively. The foreground shapes move and the background grid stays still.
Changing the camera projection matrix C simulates the camera movement and
generates projected images.
0 50 100
0
0.2
0.4
0.6
0.8
1
Variance
F-Measure
(a) Stationary Camera
0 20 40
0
0.2
0.4
0.6
0.8
1
Variance
F-Measure
(b) Moving Camera
L20
L0
Fig. 3. Comparison between group sparsity constraint
(L2,0 norm) and Std. sparse constraint. The
left is synthetic data simulated with a stationary
camera, and the right is simulated with a moving
camera.
Group Sparsity Constraint
versus Standard Sparsity
Constraint. We compare the
performance between the group
sparsity constraint (L2,0 norm)
and Std. sparse constraint
(L0 norm). The sparsity constraint
aims to find a sparse
set of nonzero elements, which
is F0 < αk × 2l in our
problem. k × 2l denotes the
total number of nonzero elements.
It is equivalent to the
total number of nonzero elements
in the group sparsity
constraint. Note that the formulation with this Std. sparse method is similar
to Robust PCA method [2]. The difference is that we use it at the trajectory
level instead of pixel level.
Two sets of data are generated to evaluate the performance. One is simulated
with a stationary camera, and the other is from a moving one. Foreground keeps
moving in the whole video sequence. Random noise with variance v is added to
both the foreground moving trajectories and camera projection matrix C. The
performance is shown in Fig. 3. In the noiseless case (i.e. v = 0), the motion
pattern from the foreground is distinct from the background in the whole video
sequence. Thus each element on the foreground trajectories is different from
the background element. Sparsity constraint produces the same perfect result as
group sparsity constraint. When v goes up, the distinction of elements between
foreground and background goes down. Thus some elements from the foreground
may be recognized as background. On the contrary, the group sparsity constraint
connects the elements in the neighboring frames. It treats the elements on one
trajectory as one unit. Even some elements on this trajectories are similar to
the background, the distinction along the whole trajectory is still large from the
background. As shown in Fig. 3, using the group sparsity constraint is more
robust than using the sparsity constraint when variance increases.
4.2 Evaluations on Videos
Experimental Settings. We test our algorithm on publicly available videos
from various sourcese. One video source is provided by Sand and Teller [32]
(refer this as ST sequences). ST sequences are recorded with hand held cameras,
both indoors and outdoors, containing a variety of non-rigidly deforming objects
Background Subtraction Using Low Rank and Group Sparsity Constraints 621
(hands, faces and bodies). They are high resolution images with large frame-toframe
motion and significant parallax. Another source of videos is provided from
Hopkins 155 dataset [33], witch has two or three motions from indoor and outdoor
scenes. These sequences contain degenerate and non-degenerate motions,
independent and partially dependent motions, articulated and non-rigidmotions.
We also test our algorithm on a typical video for stationary camera: “truck”.
The trajectories in these sequences were created using an off-the-shelf dense
particle tracker [25]. We manually label the pixels into foreground/background
(binary maps). If a trajectory falls into the foreground area, it is considered as
foreground, and verse versa.
0 0.5 1
0
1000
2000
3000
GR Rank 1
0 0.5 1
0
50
100
150
GR Rank 2
0 0.5 1
0
20
40
60
GR Rank 3
0 0.5 1
0
1000
2000
3000
0 0.5 1
0
50
100
150
0 0.5 1
0
50
100
150
0 0.5 1
0
20
40
60
FR Rank 3
0 0.5 1
0
10
20
30
Fig. 4.  ˆ Fi2 distribution of the GR and the FR
model. Green means foreground and blue means
background. Separation means good result. Top
row is from moving camera, and bottom row is
from stationary camera. The four columns are
GR−1, GR−2, GR−3 and FR.
Handling Stationary andMoving
Cameras. We first demonstrate
that our approach handles
both stationary cameras and
moving cameras automatically in
a unified framework, by using the
General Rank constraint (GR) instead
of the Fixed Rank constraint
(FR). We use two videos
to show the difference. One is
“VHand” from a moving camera
(rank(B) = 3), and the other
is “truck” captured by stationary
camera (rank(B) = 2). We use
the distribution of L2 norms of
estimated foreground trajectories
(i.e.,  ˆ Fi2, i ∈ 1, 2, ..., k) to show how well background and foreground is
separated in our model. For a good separation result, F should be well estimated.
Thus  ˆ Fi2 is large for foreground trajectories and small for background ones.
In other words, its distribution has an obvious difference between the foreground
region and the background region (see examples in Fig. 4).
We use GR-i, i ∈ 1, 2, 3 to denote the optimization iteration on each rank
value.  ˆ Fi2 of each specific rank iteration is plotted in Fig. 4. The GR method
works for both cases. When the rank of B is 3 (the first row of Fig. 4), the
FR model also finds a good solution, since rank-3 perfectly fits the FR model.
However, the FR constraint fails when the rank of B is 2, where the distribution
of  ˆ Fi2 between B and F are mixed together. On the other hand, GR−2 can
handle this well, since the data perfectly fits the constraint. On GR−3 stage, it
uses the result from GR−2 as the initialization, thus the result on GR−3 still
holds. The figure shows that the distribution of  ˆ Fi2 from the two parts has
been clearly separated in the third column of the bottom row. This experiment
demonstrates that the GR model can handle more situations than the FR model.
Since in real applications it is hard to know the specific rank value in advance,
the GR model provides a more flexible way to find the right solution.
622 X. Cui et al.
Table 1. Quantitative evaluation at trajectory level labeling
RANS GPCA LSA RANS Std. Ours
AC-b AC-m Sparse
VPerson 0.786 0.648 0.912 0.656 0.616 0.981
VHand 0.952 0.932 0.909 0.930 0.132 0.987
VCars 0.867 0.316 0.145 0.276 0.706 0.993
cars2 0.750 0.773 0.568 0.958 0.625 0.976
cars5 0.985 0.376 0.054 0.637 0.779 0.990
people1 0.932 0.564 0.087 0.743 0.662 0.955
truck 0.351 0.368 0.140 0.363 0.794 0.975
Input
video
RANSAC-b GPCA LSA RANSAC-m Standard sparse Our method
Fig. 5. Results at trajectory level labeling. Green: foreground; purple: background.
From top to bottom, five videos are “VCars”, “cars2 ”, “cars5 ”, “people1” and
“VHand”.
Performance Evaluation on Trajectory Labeling.We compare our method
with four state-of-art algorithms: RANSAC-based background subtraction (referred
as RANSAC-b here) [1], Generalized GPCA (GPCA) [34], Local Subspace
Affinity (LSA) [18] and motion segmentation using RANSAC (RANSAC-m)
[33]. GPCA, LSA and RANSAC-m are motion segmentation algorithms using
subspace analysis for trajectories. The code of these algorithms are available
online. When testing these methods, we use the same trajectories as for our own
method. Since LSA method runs very slow when using trajectories more than
5000, we randomly sample 5000 trajectories for each test video. The three motion
Background Subtraction Using Low Rank and Group Sparsity Constraints 623
segmentation algorithms ask for the number of regions to be given in advance.
We provide the correct number of segments n, whereas our method does not need
that. Motion segmentation methods separate trajectories into n segments. Here
we treat the segment with the largest trajectory number as the background and
rest as the foreground. For RANSAC-b method, two major parameters influence
the performance: projection error threshold th and consensus percentage p.
Inappropriate selection of parameters may result in failure of finding the correct
result. In addition, as RANSAC-b randomly selects three trajectories in each
round, it may end up with finding a subspace spanned by part of foreground
and background. The result it generates is not stable. Running the algorithm
multiple times may give different separation of background and foreground,
which is undesirable. In order to have a fair comparison with it, we grid search
the best parameter set over all teste videos and report the performance under
the optimal parameters.
The quantitative and qualitative results on the trajectory level separation
is shown in Fig. 5 and Tab. 1, respectively. Our method works well for the test
videos. Take “cars5 ” for example. GPCA and LSA misclassify some trajectories.
RANSAC-b randomly selects three trajectories to build the scene motion. On this
frame, the three random trajectories all lie in the middle region. The background
model built from these 3 trajectories do not cover the left and right region of the
scene, thus the left and right regions are misclassified as foreground. RANSAC-m
produces similar behavior to RANSAC-b. Std. sparse method does not have any
group constraint in the consecutive frames, thus some trajectories are classified
as foreground in one frame, and classified as background in the next frame. Note
that the quantitative results are obtained by averaging on all frames over 50
iterations. Fig. 5 only shows performance on one frame, which may not reflect
the overall performance shown in Tab. 1.
Performance Evaluation at Pixel Level Labeling.We also evaluate the performance
at the pixel level using four methods: RANSAC-b [1], MoG [11], Std.
sparse method and the proposed method. GPCA, LSA, RANSAC-m are not evaluated
in this part, since these three algorithms do not provide pixel level labeling.
Due to space limitations, one typical video “VPerson” is shown here in Fig. 6.
MoG does not perform well, because a statistical model of MoG is built on
pixels of fixed positions over multiple frames, but background objects do not stay
in the fixed positions under moving cameras. RANSAC-b can accurately label
Fig. 6. Quantitative results at pixel level labeling on “VPerson” sequence
624 X. Cui et al.
pixels if the trajectories are well classified. However, it is also possible to build
a wrong background subspace because RANSAC-b is not stable and sensitive to
parameter selection. Our method can robustly handle these diverse types of data,
due to our generalized low rank constraint and group sparsity constraint. One
limitation of our method is that it classifies the shadow as part of the foreground
(e.g., on the left side of the person in “VPerson” video). This could be further
refined by using shadow detection/removal techniques [16].
5 Conclusions and Future Work
In this paper, we propose an effective approach to do background subtraction
for complex videos by decomposing the motion trajectory matrix into a low rank
one and a group sparsity one. Then the information from these trajectories is
used to further label foreground at the pixel level. Extensive experiments are
conducted on both synthetic data and real videos to show the benefits of our
model. The low rank and group sparsity constraints make the model robust to
noise and handle diverse types of videos.
Our method depends on trajectory-tracking technique, which is also an active
research area in computer vision. When the tracking technique fails, our
method may not work well. A robust way is to build the tracking errors into the
optimization formulation. We will investigate it in our future work.
References
1. Sheikh, Y., Javed, O., Kanade, T.: Background subtraction for freely moving
cameras. In: ICCV (2009)
2. Candes, E.J., Li, X., Ma, Y., Wright, J.: Robust principal component analysis?
Journal of ACM (2011)
3. Yuan, M., Lin, Y.: Model selection and estimation in regression with grouped
variables. Journals of the Royal Statistical Society (2006)
4. Candes, E., Tao, T.: Near-optimal signal recovery from random projections:
Universal encoding strategies? TIT (2006)
5. Starck, J.L., Elad, M., Donoho, D.: Image decomposition via the combination of
sparse representations and a variational approach. TIP (2005)
6. Tomasi, C., Kanade, T.: Shape and motion from image streams under orthography:
a factorization method. IJCV (1992)
7. Brutzer, S., Hoferlin, B., Heidemann, G.: Evaluation of background subtraction
techniques for video surveillance. In: CVPR (2011)
8. Jain, R., Nagel, H.: On the analysis of accumulative difference pictures from image
sequences of real world scenes. TPAMI (2009)
9. Haritaoglu, I., Harwood, D., Davis, L.: W4: real-time surveillance of people and
their activities. TPAMI (2000)
10. Wren, C., Azarbayejani, A., Darrell, T., Pentland, A.: Pfinder: Real-time tracking
of the human body. TPAMI (2002)
11. Stauffer, C., Grimson, W.: Learning patterns of activity using real-time tracking.
TPAMI (2000)
12. Elgammal, A., Duraiswami, R., Harwood, D., Davis, L.: Background and foreground
modeling using nonparametric kernel density estimation for visual surveillance.
Proceedings of the IEEE (2002)
Background Subtraction Using Low Rank and Group Sparsity Constraints 625
13. Sheikh, Y., Shah, M.: Bayesian object detection in dynamic scenes. In: CVPR
(2005)
14. Monnet, A., Mittal, A., Paragios, N., Ramesh, V.: Background modeling and
subtraction of dynamic scenes. In: ICCV (2003)
15. Zhong, J., Sclaroff, S.: Segmenting foreground objects from a dynamic textured
background via a robust kalman filter. In: ICCV (2003)
16. Liao, V.S., Zhao, G., Kellokumpu, Pietikainen, M., Li, S.Z.: Modeling pixel process
with scale invariant local patterns for background subtraction in complex scenes.
In: CVPR (2010)
17. Rao, S., Tron, R., Vidal, R., Ma, Y.: Motion segmentation in the presence of
outlying, incomplete, or corrupted trajectories. TPAMI (2010)
18. Yan, J., Pollefeys, M.: A General Framework for Motion Segmentation: Independent,
Articulated, Rigid, Non-rigid, Degenerate and Non-degenerate. In: Leonardis,
A., Bischof, H., Pinz, A. (eds.) ECCV 2006, Part IV. LNCS, vol. 3954, pp. 94–106.
Springer, Heidelberg (2006)
19. Hayman, E., Eklundh, J.-O.: Statistical background subtraction for a mobile
observer. In: ICCV (2003)
20. Ren, Y., Chua, C., Ho, Y.: Statistical background modeling for non-stationary
camera. PR Letters (2003)
21. Yuan, C., Medioni, G., Kang, J., Cohen, I.: Detecting motion regions in the
presence of a strong parallax from a moving camera by multiview geometric
constraints. TPAMI (2007)
22. Kwak, S., Lim, T., Nam, W., Han, B., Han, J.H.: Generalized background
subtraction based on hybrid inference by belief propagation and bayesian filtering.
In: ICCV (2011)
23. Huang, J., Zhang, T.: The benefit of group sparsity. The Annals of Statistics (2010)
24. Huang, J., Huang, X., Metaxas, D.: Learning with dynamic group sparsity. In:
ICCV, pp. 64–71 (2009)
25. Sundaram, N., Brox, T., Keutzer, K.: Dense Point Trajectories by GPUAccelerated
Large Displacement Optical Flow. In: Daniilidis, K., Maragos, P.,
Paragios, N. (eds.) ECCV 2010, Part I. LNCS, vol. 6311, pp. 438–451. Springer,
Heidelberg (2010)
26. Liu, C.: Beyond pixels: Exploring new representations and applications for motion
analysis. Doctoral thesis, Massachusetts Institute of Technology (2009)
27. Boykov, Y., Veksler, O., Zabih, R.: Fast approximate energy minimization via
graph cuts. TPAMI (2001)
28. Zhang, S., Huang, J., Huang, Y., Yu, Y., Li, H., Metaxas, D.: Automatic image
annotation using group sparsity. In: CVPR (2010)
29. Liu, G., Lin, Z., Yu, Y.: Robust subspace segmentation by low-rank representation.
In: ICML (2010)
30. Zhang, Z., Liang, X., Ma, Y.: Unwrapping low-rank textures on generalized
cylindrical surfaces. In: ICCV (2011)
31. Mu, Y., Dong, J., Yuan, X., Yan, S.: Accelerated low-rank visual recovery by
random projection. In: CVPR (2011)
32. Sand, P., Teller, S.: Particle video: Long-range motion estimation using point
trajectories. In: CVPR (2006)
33. Tron, R., Vidal, R.: A benchmark for the comparison of 3-D motion segmentation
algorithms. In: CVPR (2007)
34. R., Vidal, Y.M., Sastry, S.: Generalized principal component analysis (GPCA). In:
CVPR (2003)