合成与实况视频评估的背景减除算法全面综述
需积分: 9 9 浏览量
更新于2024-07-18
收藏 3.9MB PDF 举报
《计算机视觉与图像理解：背景减除算法的全面评估》是一篇发表于2014年5月的计算机视觉领域的重要论文，其研究重点是针对合成和真实视频对背景减除算法进行全面的评价。背景减除是一种在计算机视觉中常用的技术，它通过识别和去除图像中的静态背景元素，以便突出动态对象或变化，从而增强监控、行为分析等应用的效果。
作者Andrews Sobral和Antoine Vacavant在文中系统地探讨了各种背景减除算法的优缺点，这些算法包括但不限于基于统计的方法、光流技术、帧间差异、混合模型等。他们通过精心设计的实验，利用合成视频来验证算法在理想条件下的性能，并进一步用实际拍摄的视频数据来测试算法在复杂场景中的鲁棒性和准确性。
这篇论文提供了详尽的算法比较，评估了它们在处理不同光照条件、运动模糊、噪声以及视频质量变化等情况下的表现。这使得读者能够更好地理解哪种类型的算法在特定应用场景下更为有效。论文的引用次数高达181次，显示了其在学术界的影响和认可度，阅读量超过10,533次，体现了其广泛的读者基础。
论文的两位作者不仅在背景减除算法的研究上有所贡献，还在矩阵、张量、深度学习等领域有深入工作。他们的其他项目如“Matrices, Tensors, Deep Learning and Beyond”和“Image Vectorization”展示了他们在多方面的专业知识。Antoine Vacavant来自法国奥弗涅大学，拥有61篇出版物和565次引用，表明他在计算机视觉领域的深厚学术积累。
此外，论文的后期更新由Andrews Sobral于2015年3月上传，可能包含了后续的研究成果或对原有内容的改进。对于希望深入了解背景减除算法的科研人员和工程师来说，这篇文章是宝贵的学习资源和参考文献，提供了一个系统性的理解和实践指导。如果你正在寻找优化背景处理或开发新的实时视频分析系统的线索，这篇论文无疑是一个值得深入研读的起点。
image. This approach, called further Frame Difference, works with
some background changes but fails if the moving object stops sud-
denly. In [27], the authors suggest the initialization and mainte-
nance of the background model by the arithmetic mean (or
weighted mean) of the pixels between successive images. So, given
a video V with length l containing gray scale images defined by
V ¼fZ
1
; ...; Z
l
g, the background model B can be defined by:
B ¼
1
l
X
l
t¼1
Z
t
: ð1Þ
Typically Eq. (1) is used to initialize the background model.
1
How-
ever, after the initialization, to perform the background model main-
tenance, it is also common to use Eq. (1) recursively by:
B
t
¼ð1
a
ÞB
t1
þ
a
Z
t
; ð2Þ
where B
t
is the background model at time t 2f1; lgZ and
a 2½0; 1R is the learning rate. The main advantage of this meth-
od is the adaptive maintenance of the background model while
changes occur in the scene (see Fig. 1). Afterwards, [34] clarify that
some foreground pixels are included in the background model up-
date. To solve this issue, an adaptive-selective method is proposed.
In this approach, only the regions with no moving object are
updated.
After building the background model, the next step is the fore-
ground detection. The first and most common way is to compute
the absolute difference between the current frame and the back-
ground model, similarly to Static Frame Difference method. How-
ever, in this case the background model is continuously adapted
instead of a static image. The foreground detection can be per-
formed in other ways. More recent methods, such as
[1,24,23,26,46] suggest the use of color, texture and edges features
to improve the foreground detection. In [46], the authors present a
Table 2
Possible combinations of synthetic data generated for BMC, and their respective number.
Parameter Value Description Number
Scenes 1 Rotary 10
2 Street 10
Event types 1 Cloudy, without acquisition noise, as normal mode 4
2 Cloudy, with salt and pepper noise during the whole sequence 4
3 Sunny, with noise, which generates moving cast shadows 4
4 Foggy, with noise, making both background and foreground hard to analyze 4
5 Wind, with noise, to produce a moving background 4
Use cases 1 10 s without objects, then moving objects during 50 s 10
2 20 s without event, then event (e.g. sun uprising or fog) during 20 s, finally 20 s without event 10
Fig. 3. Examples of synthetic (top) and real (bottom) videos and their associated ground truth in the BMC benchmark.
Table 3
Parameter settings of each BS algorithm.
Method ID Settings
Basic methods, mean and variance over time
StaticFrameDifferenceBGS T ¼ 15
FrameDifferenceBGS T ¼ 15
WeightedMovingMeanBGS T ¼ 10
WeightedMovingVarianceBGS T ¼ 15
AdaptiveBackgroundLearning T ¼ 15;
a ¼ 0:5
DPMeanBGS
T ¼ 2700;
a ¼ 10
7
; LF ¼ 30
DPAdaptiveMedianBGS T ¼ 20; LF ¼ 30; SR ¼ 10
DPPratiMediodBGS T ¼ 30; SR ¼ 5; HS ¼ 16;
c ¼ 5
Fuzzy based methods
FuzzySugenoIntegral T ¼ 0:67; LF ¼ 10;
a
learn
¼ 0:5;
a
update
¼ 0:05; RGB þ LBP
FuzzyChoquetIntegral T ¼ 0:67; LF ¼ 10;
a
learn
¼ 0:5;
a
update
¼ 0:05; RGB þ LBP
LBFuzzyGaussian T ¼ 160; LR ¼ 150;
q ¼ 100;
r
¼ 195
Statistical methods using one Gaussian
DPWrenGABGS T ¼ 12:15; LF ¼ 30;
a ¼ 0:05
LBSimpleGaussian LR ¼ 50;
q ¼ 255;
r
¼ 150
Statistical methods using multiple gaussians
DPGrimsonGMMBGS T ¼ 9;
a ¼ 0:05; n ¼ 3
MixtureOfGaussianV1BGS T ¼ 10;
a ¼ 0:01
MixtureOfGaussianV2BGS T ¼ 5;
a ¼ 0:01
DPZivkovicAGMMBGS T ¼ 20;
a ¼ 0:01; n ¼ 3
LBMixtureOfGaussians T ¼ 80;
a ¼ 60; q ¼ 120;
r
¼ 210
Type-2 Fuzzy based methods
T2FGMM_UM T ¼ 1; K
m
¼ 2:5; n ¼ 3; a ¼ 0:01
T2FGMM_UV T ¼ 1; K
v
¼ 0:6; n ¼ 3; a ¼ 0:01
T2FMRF_UM T ¼ 1; K
m
¼ 2:0; n ¼ 3; a ¼ 0:01
T2FMRF_UV T ¼ 1; K
v
¼ 0:9; n ¼ 3; a ¼ 0:01
Statistical methods using color and texture features
MultiLayerBGS Original default parameters from [45]
Non-parametric methods
PixelBasedAdaptiveSegmenter Original default parameters from [22]
GMG T ¼ 0: 7; LF ¼ 20
1
Other authors such as McFarlane and Schofield [31], Prati et al. [33] and Calderara
et al. [9] suggests to use the median filter instead of the mean or average filter.
6 A. Sobral, A. Vacavant / Computer Vision and Image Understanding 122 (2014) 4–21
剩余18页未读，继续阅读
349 浏览量
2019-03-12 上传
2023-09-19 上传
2023-06-06 上传
点击了解资源详情
点击了解资源详情
点击了解资源详情
2018-03-25 上传
2010-05-02 上传
zhouzmtco
- 粉丝: 0
- 资源: 2
我的内容管理
展开
最新资源
- Angular程序高效加载与展示海量Excel数据技巧
- Argos客户端开发流程及Vue配置指南
- 基于源码的PHP Webshell审查工具介绍
- Mina任务部署Rpush教程与实践指南
- 密歇根大学主题新标签页壁纸与多功能扩展
- Golang编程入门：基础代码学习教程
- Aplysia吸引子分析MATLAB代码套件解读
- 程序性竞争问题解决实践指南
- lyra: Rust语言实现的特征提取POC功能
- Chrome扩展：NBA全明星新标签壁纸
- 探索通用Lisp用户空间文件系统clufs_0.7
- dheap: Haxe实现的高效D-ary堆算法
- 利用BladeRF实现简易VNA频率响应分析工具
- 深度解析Amazon SQS在C#中的应用实践
- 正义联盟计划管理系统：udemy-heroes-demo-09
- JavaScript语法jsonpointer替代实现介绍
