Video foreground extraction for mobile camera platforms

Abstract

Foreground object detection is a fundamental task in computer vision with many applications in areas such as object tracking, event identification, and behavior analysis. Most conventional foreground object detection methods work only in a stable illumination environments using fixed cameras. In real-world applications, however, it is often the case that the algorithm needs to operate under the following challenging conditions: drastic lighting changes, object shape complexity, moving cameras, low frame capture rates, and low resolution images. This thesis presents four novel approaches for foreground object detection on real-world datasets using cameras deployed on moving vehicles.The first problem addresses passenger detection and tracking tasks for public transport buses investigating the problem of changing illumination conditions and low frame capture rates. Our approach integrates a stable SIFT (Scale Invariant Feature Transform) background seat modelling method with a human shape model into a weighted Bayesian framework to detect passengers. To deal with the problem of tracking multiple targets, we employ the Reversible Jump Monte Carlo Markov Chain tracking algorithm. Using the SVM classifier, the appearance transformation models capture changes in the appearance of the foreground objects across two consecutives frames under low frame rate conditions. In the second problem, we present a system for pedestrian detection involving scenes captured by a mobile bus surveillance system. It integrates scene localization, foreground-background separation, and pedestrian detection modules into a unified detection framework. The scene localization module performs a two stage clustering of the video data.In the first stage, SIFT Homography is applied to cluster frames in terms of their structural similarity, and the second stage further clusters these aligned frames according to consistency in illumination. This produces clusters of images that are differential in viewpoint and lighting. A kernel density estimation (KDE) technique for colour and gradient is then used to construct background models for each image cluster, which is further used to detect candidate foreground pixels. Finally, using a hierarchical template matching approach, pedestrians can be detected.In addition to the second problem, we present three direct pedestrian detection methods that extend the HOG (Histogram of Oriented Gradient) techniques (Dalal and Triggs, 2005) and provide a comparative evaluation of these approaches. The three approaches include: a) a new histogram feature, that is formed by the weighted sum of both the gradient magnitude and the filter responses from a set of elongated Gaussian filters (Leung and Malik, 2001) corresponding to the quantised orientation, which we refer to as the Histogram of Oriented Gradient Banks (HOGB) approach; b) the codebook based HOG feature with branch-and-bound (efficient subwindow search) algorithm (Lampert et al., 2008) and; c) the codebook based HOGB approach.In the third problem, a unified framework that combines 3D and 2D background modelling is proposed to detect scene changes using a camera mounted on a moving vehicle. The 3D scene is first reconstructed from a set of videos taken at different times. The 3D background modelling identifies inconsistent scene structures as foreground objects. For the 2D approach, foreground objects are detected using the spatio-temporal MRF algorithm. Finally, the 3D and 2D results are combined using morphological operations.The significance of these research is that it provides basic frameworks for automatic large-scale mobile surveillance applications and facilitates many higher-level applications such as object tracking and behaviour analysis.