A quick review of background subtraction techniques
The current advancement of Intelligent Transportation Systems (ITS) has accelerated the era of a smart city and its component, including smart vehicle, intelligent surveillance, pedestrian detection, vehicle detection, etc. As a result, the foreground detection technology is becoming increasingly important, because it is inevitably applied as pre-processing for higher level vision application. To date, deep learning is now becoming a hot research topic due to its high precision and very fast computation coupled with GPU architecture. Recently, a supervised-based deep learning method called FgSegNet [1] outperformed other state-of-the-art approaches in ChangeDetection.net [2]. As mentioned in the abstract of their work, they adopt a pre-trained convolutional network, i.e. VGG-16 Net, under a triplet framework in the encoder part to embed an image in multiple scales into the feature space and use a transposed convolutional network in the decoder part to learn a mapping from feature space to image space. FgSegNet is clearly addressing the essential issues of the real world, e.g., abrupt illumination changes, background or camera motion, shadow interference, camouflage, etc. Here are the top 10 methods in the overall evaluation of ChangeDetection.net (taken on April 18, 2018):
In this article, we aim to quickly review several important types of object detection methods from the conventional to the newest one. It then allows the reader to know the history of low-level vision processing, especially background subtraction that has been widely discussed.
Some literature has been comprehensively reviewed the classification of background subtraction models. In [3], Bouwmans divided it into two main categories; traditional models and recent models. As he explained in the paper:
Traditional models which present the following characteristics:
(1) These models are the first models used in the field and are generally basics. (2) These models allow us to handle some specific challenges, and are generally easy to implement. (3) Numerous improvements concern these models and their limitations seem to be reached.Recent models which present on the other hand the following characteristics:
(1) These models are more sophisticated to robustly handle a lot of challenges. (2) Most of them need improvements to achieve incremental and real-time requirements.
Most of the traditional models only focus on figuring out one challenge (such as shadow, illumination changes or water ripple) which is not really capable of a real-world problem. Conversely, recent models aim to cope multiple challenges by introducing advanced approach such as statistical, fuzzy, and tracking methods. While the variant of smart and/or cyber surveillance applications requiring robust-efficient-accurate results hence the traditional models have been abandoned. Moreover, every application has its specific requirement, that encourages the scholars to devise a specific-purpose workflow to achieve the goal. For example, an insect surveillance application will need to build a model in order to distinguish the background with high-velocity objects, while river surveillance needs to handle slow solid waste in the river stream.
It is important to note that the recent models can also be further divided into two sub-categories; supervised and unsupervised approach. This is because of quickly-growing of deep learning [1] and universal change detection techniques [4,5,6]. By defining the requirement of an application, either supervised or unsupervised, we can easily localize, analyze and examine the state-of-the-art methods accordingly.
References
[1] Lim, Long Ang, and Hacer Yalim Keles. “Foreground Segmentation Using a Triplet Convolutional Neural Network for Multiscale Feature Encoding.” arXiv preprint arXiv:1801.02225 (2018).
[2] Goyette, Nil, et al. “Changedetection. net: A new change detection benchmark dataset.” Computer Vision and Pattern Recognition Workshops (CVPRW), 2012 IEEE Computer Society Conference on. IEEE, 2012.
[3] Bouwmans, Thierry. “Traditional and recent approaches in background modeling for foreground detection: An overview.” Computer Science Review 11 (2014): 31–66.
[4] Barnich, Olivier, and Marc Van Droogenbroeck. “ViBe: A universal background subtraction algorithm for video sequences.” IEEE Transactions on Image processing 20.6 (2011): 1709–1724.
[5] St-Charles, Pierre-Luc, Guillaume-Alexandre Bilodeau, and Robert Bergevin. “Subsense: A universal change detection method with local adaptive sensitivity.” IEEE Transactions on Image Processing 24.1 (2015): 359–373.
[6] Hofmann, Martin, Philipp Tiefenbacher, and Gerhard Rigoll. “Background segmentation with feedback: The pixel-based adaptive segmenter.” Computer Vision and Pattern Recognition Workshops (CVPRW), 2012 IEEE Computer Society Conference on. IEEE, 2012.
