Video foreground segmentation with deep learning

Abstract

This thesis tackles the problem of foreground segmentation in videos, even under extremely challenging conditions. This task comes with a plethora of hurdles, as the model needs to distinguish the difference between moving objects and irrelevant background motion which can be caused by the weather, illumination, camera movement etc. As foreground segmentation is often the first step of various highly important applications (video surveillance for security, patient/infant monitoring etc.), it is crucial to develop a model capable of producing excellent results in all kinds of conditions. In order to tackle this problem, we follow the recent trend in other computer vision areas and harness the power of deep learning. We design architectures of convolutional neural networks specifically targeted to counter the aforementioned challenges. We first propose a 3D CNN that models the spatial and temporal information of the scene simultaneously. The network is deep enough to successfully cover more than 50 different scenes of various conditions with no need for any fine-tuning. These conditions include illumination (day or night), weather (sunny, rainy or snowing), background movements (trees moving from the wind, fountains etc) and others. Next, we propose a data augmentation method specifically targeted to illumination changes. We show that artificially augmenting the data set with this method significantly improves the segmentation results, even when tested under sudden illumination changes. We also present a post-processing method that exploits the temporal information of the input video. Finally, we propose a complex deep learning model which learns the illumination of the scene and performs foreground segmentation simultaneously.

Date of Award	4 May 2020
Original language	English
Awarding Institution	Northumbria University
Supervisor	Graham Sexton (Supervisor), Edmond Ho (Supervisor) & Jungong Han (Supervisor)