Simple Median-Based Method for Stationary Background Generation - - PowerPoint PPT Presentation

▶

Jun 01, 2023 17 likes •179 views

Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms B. Laugraud, S. Pirard, M. Braham, M. Van Droogenbroeck INTELSIG Laboratory, University of Lige, Belgium September 8th 2015 Scene

SLIDE 1

Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms

B. Laugraud, S. Piérard, M. Braham, M. Van Droogenbroeck

INTELSIG Laboratory, University of Liège, Belgium

September 8th 2015

Scene Background Modeling and Initialization Workshop (SBMI2015) Genova, Italy

SLIDE 2

Introduction

We want to estimate an image containing the background (BG) of a scene taken from a static viewpoint. This problem can be solved by applying a temporal median filter per pixel if the foreground (FG) is visible less than half of the time (MED method). For most of the sequences of the SBI dataset, it is not the case!

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 2 / 15

SLIDE 3

Ideas

In order to get the MED method work:

1 Select a set of relevant frames to have the BG visible for half of the time. 2 Such a selection could be based on motion, discarding the frames containing a

large “amount of motion”. Observation Background subtraction (BGS) algorithms can detect motion!

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 3 / 15

SLIDE 4

Ideas: Graphical Summary

Background Subtraction Probabilistic Selection MED Method

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 4 / 15

SLIDE 5

Proposed Method: Buffers

Median computed on a per-pixel subset of selected values. The size of the subsets is fixed by the S parameter. h w S

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 5 / 15

SLIDE 6

Proposed Method: Patches

The image plane is divided in N × N (parameter) non-overlapping patches. An observed value is selected according to a probability denoted p∗

+ of FG ele-

ments in the patch of the considered pixel. Pixels are classified as FG or BG using a BGS algorithm (denoted by A). h N w N p∗

+ = # pixels classified as foreground in the patch

# pixels in the patch

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 6 / 15

SLIDE 7

Proposed Method: Passes

BGS algorithms need an initialization period. The number of frames needed for this period is algorithm dependent. It can be larger than the number of frames in the sequences of the SBI dataset. We process the sequences with several passes. The number of passes (γ parameter) is chosen to be odd. forwards (odd passes) backwards (even passes)

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 7 / 15

SLIDE 8

Proposed Method: Processing

For each pixel: Our method selects the S values, encountered during the γ passes, with the low- est probability p∗

When S is too small to select all the values with equal p∗

+ probabilities, the last

encountered ones are selected. At the end of the γ passes, the BG color is estimated by the median of the S selected values.

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 8 / 15

SLIDE 9

Classification (According to Maddalena and Petrosino)

Hybrid Non-recursive Selective Our method

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 9 / 15

SLIDE 10

Experiments

The proposed method has been tested with all combinations of: A ∈ { F. Diff, Pfinder, MoG G., MoG Z., S-D, KDE, ViBe, PBAS, SuBS., SOBS } S ∈ {5,11,21,51,101,201} N ∈ {1,3,5,10,25,50}

γ ∈ {1,3,5,...,19}

Note that, for our results, we arbitrarily chose to work with the pEPs metric.

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 10 / 15

SLIDE 11

Results

CaVignal Foliage Hall&Monitor HighwayI HighwayII People&Foliage Snellen

Figure : The 7 video sequences of the SBI dataset (50th frame on the 1st row), the result obtained by the MED method (2nd row), the result obtained by minimizing the pEPs score averaged over all the sequences (A = F . Diff., S = 21, N = 3, γ = 11) (3rd row), and the corresponding ground-truth (4th row).

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 11 / 15

SLIDE 12

Results: Metrics POV

Sequence AGE pEPs pCEPs MS-SSIM PSNR CQM CaVignal 9.2286 0.0003 0.0000 0.9933 27.5385 39.7264 Foliage 11.9949 0.0198 0.0000 0.9916 26.0057 34.0230 HallAndMonitor 2.5051 0.0025 0.0004 0.9880 35.0603 43.1707 HighwayI 2.1235 0.0083 0.0018 0.9833 35.8519 52.9773 HighwayII 2.2706 0.0019 0.0000 0.9927 37.2908 45.7950 PeopleAndFoliage 12.2607 0.0345 0.0014 0.9902 25.6114 33.5995 Snellen 17.0920 0.1159 0.1000 0.9646 21.2595 41.0757 Average 8.2108 0.0262 0.0148 0.9863 29.8026 41.4811 Table : Scores computed on the results obtained by minimizing the pEPs score averaged over all the sequences (A = F . Diff., S = 21, N = 3, γ = 11).

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 12 / 15

SLIDE 13

Optimal Set of Parameters Per Sequence

Sequence

CaVignal Foliage Hall&Monitor HighwayI HighwayII People&Foliage Snellen Average

pEPs (lower is better) - Log. Scale

10-4 10-3 10-2 10-1

Optim. per sequence
Optim. average

Improvement from 0.0001 (CaVignal) to ≃ 0.06 (Snellen). Mean improvement ≃ 0.015. Note that the optimal BGS algorithm is not the same for all the sequences.

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 13 / 15

SLIDE 14

Optimal Set of Parameters Per BGS Algorithm

Algorithm

Frame Diff. PBAS MoG S&G SuBSENSE KDE MoG Zivkovic ViBe Sigma-Delta SOBS Pfinder

Averaged pEPs (lower is better)

0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 14 / 15

SLIDE 15

Conclusion

New method to estimate the background image of a cluttered static scene. The method is simple and easy to implement. It allows to embed any BGS algorithm without any modification as we only use the segmentation maps. Surprisingly, the frame difference outperforms more advanced BGS algorithms in this particular context. The obtained results are almost free of foreground objects. Source code is available: http://hdl.handle.net/2268/182893.

Benjamin Laugraud (University of Liège) Simple Median-Based Method for Stationary Background Generation Using Background Subtraction Algorithms 15 / 15

SLIDE 16