Development of Dynamic Background Subtraction Algorithms for Visual Surveillance Applications

Abstract

Recent developments in the field of computer vision and the easy availability of low-cost digital cameras have spurred demands for video surveillance systems. These advancements have helped manufacturers to market surveillance cameras as a consumer electronic product for both personal and public use. Surveillance cameras have a wide range of applications such as security, anomaly detection, traffic control, congestion analysis, target tracking, and daycare/ nanny monitoring.
Background subtraction (BS) is a fundamental step for moving object detection in various video surveillance applications. It detects a moving object by comparing every incoming frame with the up-to-date (learned) statistical background model. Based on the deviation from the background model, it classifies pixels into foreground or background pixels. Dynamic environments such as swaying trees, ripples in water, fast and gradual changes in illumination, relocation of background objects, initialization with moving objects, and crowded scenes make background modelling a challenging task. Moreover, problems like shadows, camera jitters, and noise make it even more difficult. Therefore, it is important to develop efficient BS algorithms that can address these challenges and offer effective performance in real-time scenarios.
In this thesis, robust BS techniques are developed to handle dynamic background conditions. Initially, algorithms relating to background modelling are proposed. Here, new features which are robust to dynamic backgrounds are developed using Gabor filter and the concept of difference domain. In the later part of the thesis, new foreground detection algorithms are formulated using Wronskian change detection model (WM).
A novel fuzzy colour difference histogram (FCDH) based BS algorithm has been proposed by employing fuzzy c-means (FCM) clustering and colour difference histogram (CDH). This is done by measuring the colour difference between a pixel and its neighbourhood. The use of CDH reduces the number of false errors due to the non-stationary background, illumination variation, and camouflage. The use of the FCM clustering algorithm in CDH reduces the large dimensionality of the histogram bins in the computation.
The length of the feature vector in FCDH based BS algorithm depends on the number of clusters used in the FCM. If the number of clusters taken is large, then the length of the feature vector becomes large. Moreover, FCDH does not completely remove the effect of dynamic background. In order to overcome the limitations posed by the FCDH based BS, a multi-modal BS algorithm based on difference of intensity between the original and the quantised image is computed. Further, the difference of intensity values between the original and the Gaussian filtered image is also calculated. To reduce the effect of noise and small variations, the difference of the original intensity values with the local mean of the reduced quantised value is performed. Further, to reduce the effect of a dynamic background such as swaying trees, spouting fountains and flowing rivers, the difference of original intensity values with the local mean of the Gaussian filtered image is determined. The effect of cluttered environment is reduced due to the blurring introduced by the Gaussian filtered image.
The use of Gabor feature in BS is the next contribution. A hierarchical BS algorithm using Gabor filter is proposed where both the block-based and the pixel-based approaches are combined. Both coarse and fine level background modelling is done using the magnitude feature obtained from the Gabor filter. First, the coarse level background modelling is accomplished for identifying blocks which are fully or partially occupied by the foreground objects. In the second stage of pixel-level analysis, a pixel in a foreground block is further analysed and then classified using the Gabor feature for improving the precision of the detected moving object. The number of features extracted from the Gabor filter for the hierarchical BS is large. In order to reduce the length of the feature vector generated from the Gabor filter, a novel Gabor Hu moment is proposed from the orientations of the Gabor filter at different scales and is used as a feature in the codebook BS algorithm for efficient moving object detection. The use of Gabor features makes the algorithm less susceptible to background variations and models the background in an efficient manner for accurate silhouette detection of the foreground object.
Foreground detection algorithms are proposed using the Wronskian change detection model (WM). An adaptive spatio-temporal BS technique using improved WM in Gaussian mixture model (GMM) framework is developed for moving object detection. The GMM does not support the spatial relationship among neighbouring pixels and uses a fixed learning rate for every pixel during the parameter update. On the other hand, WM is a spatial-domain BS technique which solves the misclassification of pixels but fails in the presence of a dynamic background. To address this, a novel spatio-temporal BS technique is proposed, which exploits spatial relation of Wronskian function and employs it with a new fuzzy adaptive learning rate in GMM framework. Instead of using WM directly, an improved WM (IWM) is proposed by adaptively finding out the ratio of the current pixel to the background pixel or its reciprocal. Further, a weighted Wronskian function is developed to mitigate the effect of dynamic background pixels. A new multi-channel and multi-resolution Wronskian change detection model (MCMRWM) based codebook BS algorithm is proposed for moving object detection in the presence of dynamic background conditions. In the proposed MCMRWM, the multi-channel information helps to reduce the false negative of the foreground object; and the multi-resolution data suppresses the background noise resulting in reduced false positives. The proposed algorithm considers the ratio between feature vectors of current frame and the background model or its reciprocal in an adaptive manner, depending on the l2 norm of the feature vector, which helps to detect the foreground object completely without any false negatives.
From experimental results, it is observed that the proposed algorithms exhibit a considerable improvement in moving object detection and hence are expected to be quite useful in applications like traffic monitoring, target tracking, security, and surveillance.