Deep Learning for Robust Foreground Detection

GLOBAL RESEARCH INSTITUTES

  • Deep Learning for Robust Foreground Detection

Deep Learning for Robust Foreground Detection

Project Summary:

Foreground detection in video is a key processing step in video analytics, figure 1.  Whilst, there have been recent advances that improve their robustness for real world applications, there is much still to be done.  Concurrently, deep learning methods have been applied very successfully to a range of supervised computer vision problems.  However, to date there has been very little progress made with respect to unsupervised learning, such as required for background modelling – a key component of foreground detection techniques.  Furthermore, there has been little effort in video processing.  We hope to address this shortfall by investigating the use of unsupervised deep learning to improve the robustness of foreground detection in video.

In this project we propose to investigate the use of deep unsupervised learning and to investigate the relationship between neural networks and generative modelling in a deep learning framework.  Our hypothesis is that there is some link between generative learning and neural-network based learning that if found could help provide a theoretical explanation for the latter. In our previous work on unsupervised generative learning for foreground-background separation, we imposed a user designated fixed region-size over which the Gaussian mixture model (GMM) could learn.  In this work we propose to investigative the use of deep generative Gaussian mixture models (DGMMs) for learning the region size automatically.  In addition, we propose to investigate the use of regularisation for stochastic gradient descent by incorporating momentum and higher order dynamics as a means to improve the convergence rate of the learning algorithm.  A key aspect of this task will be to develop a convergence proof and rate-of-convergence proof for the new online deep learning algorithm.  Once again, this will build on previous work we have performed on regularised spatial-region GMMs.  We will then implement deep adversarial auto encoders for figure-background separation and compare the hierarchical feature representations with those produced by the DGMMs.  This will help give us insights into how well, or not, we can explain neural network performance using generative models. Finally, we will seek to extend the work by developing a unified framework for human detection.

 

Objectives:

To investigate:

  1. Deep generative models for learning region size
  2. Deep adversarial autoencoders
  3. Longer memory with momentum
  4. Extend to unified framework for human detection

 

Academic Requirements: 

A minimum 2.1 honours degree or equivalent in Computer Science or Electrical and Electronic Engineering or relevant degree is required.

 

GENERAL INFORMATION

This 3 year PhD studentship, potentially funded by the Department for Employment and Learning (DEL), commences on 1 October 2017. 

Eligibility for both fees and maintenance (approximately £14,000) depends on the applicants being either an ordinary UK resident or those EU residents who have lived permanently in the UK for the 3 years immediately preceding the start of the studentship. Non UK residents who hold EU residency may also apply but if successful may receive fees only.

Applicants should apply electronically through the Queen’s online application portal at: https://dap.qub.ac.uk/portal/ 

Further information available at:  http://www.qub.ac.uk/schools/eeecs/StudyattheSchool/PhDProgrammes

 

How to Apply

Applicants should apply electronically through the Queen’s online application portal at: https://dap.qub.ac.uk/portal/

 

Contact Details:

Dr. Paul Miller

Email: p.miller@qub.ac.uk

Telephone: +44 (0)28 9097 1809