An Improved Adaptive Background Mixture Model for Real-time Tracking with Shadow Detection
Summary (2 min read)
1 Introduction
- Background subtraction involves calculating a reference image, subtracting each new frame from this image and thresholding the result.
- They applied a selective update scheme to include only the probable background values into the estimate of the background.
2 Background Modelling
- The authors discuss the work of Grimson and Stauffer [2,3] and its shortcomings.
- The probable background colours are the ones which stay longer and more static.
- Static single-colour objects trend to form tight clusters in the colour space while moving ones form widen clusters due to different reflecting surfaces during the movement.
- The measure of this was called the fitness value in their papers.
- Every new pixel value is checked against existing model components in order of fitness.
2.1 Adaptive Gaussian Mixture Model
- The threshold T is the minimum fraction of the background model.
- If none of the K distributions match that pixel value, the least probable component is replaced by a distribution with the current value as its mean, an initially high variance, and a low weight parameter.
- The details of its robustness were explained in their papers [1,2,3]; however, with a simple discussion, the authors can see its incapability.
- If the authors assume that at least 60% of the time the background is present and α is 0.002 (500 recent frames), it would take 255 frames and 346 frames for the component to be included as part of the background and the dominant background component, respectively.
- The situation can be worse in busy environments where a clean background is rare.
2.2 Online EM Algorithms
- The authors begin their estimating of the Gaussian mixture model by expected sufficient statistics update equations then switch to L-recent window version when the first L samples are processed.
- The expected sufficient statistics update equations provide a good estimate at the beginning before all L samples can be collected.
- This initial estimate improves the accuracy of the estimate and also the performance of the tracker allowing fast convergence on a stable background model.
- The L-recent window update equations gives priority over recent data therefore the tracker can adapt to changes in the environment.
- The online EM algorithms by expected sufficient statistics are shown in the left column while the by L-recent window version in the right.
2.3 Shadow Detection and Colour Model
- As it is evidence in their papers [1,2,3], Grimson et al’s tracker can not identify moving shadows from the objects casting them.
- As many colour spaces can separate chromatic and illumination components, maintaining a chromatic model regardless of the brightness can lead to an unstable model especially for very bright or dark objects.
- As the requirement to identify moving shadows, the authors need to consider a colour model that can separate chromatic and brightness components.
- It should be compatible and make use of their mixture model.
- The calculation of a and c are trivial using vector dot product.
3 Experiment
- This section demonstrates the performance of the Grimson model [2,3] and their proposed algorithms on an image sequence.
- In the shadow detection module, the brightness threshold, τ of 0.7 was used.
- To show the performance of the background models, higher level processes such as noise cleaning or connected component analysis algorithms were not introduced to the results of background subtractions.
- The sequence includes strong sunshine, large shaded area, tree, reflections from windows and long moving shadows.
4 Conclusion
- The authors have presented new update algorithms for learning adaptive mixture models of background scene for the real-time tracking of moving objects.
- The algorithm run under the framework of the real-time robust tracker proposed by Grimson et al.
- Shadow detection need only be performed upon pixels labelled as foreground and therefore with negligible computational overheads the moving shadows can be detected successfully.
Did you find this useful? Give us your feedback
Citations
1,777 citations
Cites methods from "An Improved Adaptive Background Mix..."
...The GMM of [70] and the Bayesian algorithm of [47] were tested using their implementations available in Intel’s IPP image processing library....
[...]
...We now compare the results of ViBe with those of five stateof-the-art background subtraction algorithms and two basi c methods: (1) the gaussian mixture model proposed in [47] (hereafter referred to as GMM); (2) the gaussian mixture model of [50] (referred to as EGMM); (3) the Bayesian algorithm based on histograms introduced in [70]; (4) the codebook algorithm [62]; (5) the zipfianΣ − ∆ estimator of [37]; (6) a single gaussian model with an adaptive varianc e (named “gaussian model” hereafter); and (7) the first-order low-pass filter (that isBt = αIt +(1−α)Bt−1, whereIt and Bt are respectively the input and background images at time t), which is used as a baseline....
[...]
1,483 citations
Cites background or methods from "An Improved Adaptive Background Mix..."
...Keywords: Background subtraction; On-line density estimation; Gaussian mixture model; Non-parametric density estimation...
[...]
...The standard GMM update equations are extended in (KaewTraKulPong and Bowden, 2001; Lee, 2005) to improve the speed of adaptation of the model....
[...]
1,462 citations
Cites methods from "An Improved Adaptive Background Mix..."
...[18] and used the implementation in the OpenCV library....
[...]
1,355 citations
Cites background from "An Improved Adaptive Background Mix..."
...To the authors’ knowledge, none of the earlier studies have utilized discriminative texture features in dealing with the problem....
[...]
1,222 citations
References
7,660 citations
"An Improved Adaptive Background Mix..." refers methods in this paper
...When incorporate with the shadow detection, our method results in far better segmentation than that of Grimson et al....
[...]
3,631 citations
"An Improved Adaptive Background Mix..." refers background or methods in this paper
...According to their papers [1,2,3], only two parameters, α and T, needed to be set for the system....
[...]
...Because of no clean images at the beginning, an artefact of the initial image left in Grimson et al’s tracker lasted for over a hundred frames....
[...]
...The results show the speed of learning and the accuracy of the model using our update algorithm over the Grimson et al’s tracker....
[...]
...The algorithm run under the framework of the real-time robust tracker proposed by Grimson et al. A comparison has been made between the two algorithms....
[...]
...Our method is based on Grimson et al’s framework [1,2,3], the differences lie in the update equations, initialisation method and the introduction of a shadow detection algorithm....
[...]
2,432 citations
"An Improved Adaptive Background Mix..." refers methods in this paper
...Elgammal et al used a kernel estimator for each pixel [6]....
[...]
2,093 citations
1,971 citations
"An Improved Adaptive Background Mix..." refers methods in this paper
...Other techniques using high level processing to assist the background modelling have been proposed; for instance, the Wallflower tracker [7] which circumvents some of these problems using high level processing rather than tackling the inadequacies of the background model....
[...]