Video Stabilization Based on Speeded-Up Robust Features

With the advancement of science and technology, portable devices using video cameras are becoming more popular. Videos captured by ordinary commercial cameras always suffer from undesired motion which results from human hand shaking and mobile platform vibration. The undesired motion would not only blur the image degrading the image quality leading to inaccurate results in automatic object recognition and tracking, but also make it difficult for people to focus on specific object regions. It is also a possibility that one may feel dizzy while watching shaky video for a long time. Many hardware and software methodologies have been developed by several researchers for stabilization of shaky video, but an accurate methodology for real time stabilization of video streams by software means is still an important need. An innovative and effective video stabilization method based on feature point extraction and tracking is presented in this thesis. We employ speeded-up robust features (SURF) as a feature detector to extract the feature points from each frame in a video. To improve the performance of video stabilization, we adopt Random Sample Consensus (RANSAC) which eliminates wrong matching interest points and feature points belonging to moving objects that cannot reflect the real camera motion. These refined interest points would be used to estimate the inter-frame motion parameters. The estimated camera motion is eventually filtered through Adaptive Motion Vector Integration to separate the undesired camera motion from intentional shaky motion. In addition, to improve the quality of video streams captured under poor illumination, we use the Adaptive Integrated Neighborhood Dependent Approach for Nonlinear Enhancement (AINDANE) technique to boost image brightness and to stretch the image contrast. Experimental results conducted on several video streams confirm the effectiveness of the proposed video stabilization method in terms of accuracy and speed performance when compared with other state of the art feature based stabilization techniques. Research work is in progress to apply the proposed video stabilization technique as a preprocessing stage for automatic object detection, tracking and identification specifically from videos captured in a maritime environment.