Depth Assisted Background Modeling and Super-resolution of Depth Map

Sun, B
(2019) Depth Assisted Background Modeling and Super-resolution of Depth Map. Master of Philosophy thesis, University of Liverpool.

[img] Text

Download (8MB)


Background modeling is one of the fundamental tasks in the computer vision, which detects the foreground objects from the images. This is used in many applications such as object tracking, traffic analysis, scene understanding and other video applications. The easiest way to model the background is to obtain background image that does not include any moving objects. However, in some environment, the background may not be available and can be changed by the surrounding conditions like illumination changes (light switch on/off), object removed from the scene and objects with constant moving pattern (waving trees). The robustness and adaptation of the background are essential to this problem. Mixture of Gaussians (MOG) is one of the most widely used methods for background modeling using color information, whereas the depth map provides one more dimensional information of the images that is independent of the color. In this thesis, the color only based methods such as Gaussian Mixture Models (GMM), Hidden Markov Models (HMM), Kernel Density Estimation (KDE) are thoroughly reviewed firstly. Then the algorithm that jointly uses color and depth information is proposed, which uses MOG and single Gaussian model (SGM) to represent recent observations of the color and depth respectively. And the color-depth consistency check mechanism is also incorporated into the algorithm to improve the accuracy of the extracted background. The spatial resolution of the depth images captured from consumer depth camera is generally limited due to the element size of the senor. To overcome this limitation, depth image super-resolution is proposed to obtain the high resolution depth image from the low resolution depth image by making the inference on high frequency components. Deep convolution neural network has been widely successfully used in various computer vision tasks like image segmentation, classification and recognitions with remarkable performance. Recently, the residual network configuration has been proposed to further improve the performance. Inspired by this residual network, we redesign the popular deep model Super-Resolution Convolution Neural Network (SRCNN) for depth image super-resolution. Based on the idea of residual network and SRCNN structure, we proposed three neural network based approaches to address the problem of depth image super-resolution. In these approaches, we introduce the deconvolution layer into the network which enables the learning directly from original low resolution image to the desired high resolution image, instead of using conventional method like bicubic to interpolate the image before entering the network. Then in order to minimize the sharpness loss near the boundary regions, we add layers at the end of network to learn the residuals. The main contributions of this thesis are investigating the utilization of the depth information for background modeling and proposing three approaches on depth image super-resolution. For the first part, the property of depth image is exploited and added into the commonly used background models. By doing so, the background model can be constructed more efficiently and accurately because the depth information is not affected by the color information. During the investigation, we found that the depth image usually has two problems, which are spatial resolution and accuracy, which need to be addressed. Most of the depth images either have small resolution or the accuracy is very bad. In the second part of this thesis, we investigate three methods to obtain the accurate high resolution depth image from the low resolution one.

Item Type: Thesis (Master of Philosophy)
Uncontrolled Keywords: Background Subtraction, Depth Map, Convolutional Neural Network, Image Super-resolution
Divisions: Fac of Science & Engineering > School of Electrical Engineering, Electronics and Computer Science
Depositing User: Symplectic Admin
Date Deposited: 25 Jun 2019 15:13
Last Modified: 09 Jan 2021 04:05
DOI: 10.17638/03037405
  • Al-Nuaimy, Waleed