Research On Digital Video Stabilization Algorithm Based On Android Platform

This thesis addresses the problem of digital video stabilization based on mobile platform.Home videos captured by hand-held devices(e.g.,cell-phones or DVs)often suffer from a significant amount of unexpected camera shakiness,which significantly decrease the video quality and bring visual fatigue to users.What's more,shaky videos open have less accuracy on the tasks of objects detection and tracking applied in security monitoring.Digital video stabilization technique aims to transform a shaky video by camera motion estimation and path optimization so that the stabilized video appears as if it were taken from a professional device.We propose a grid-based feature extraction and matching method.The video frames are firstly divided into separate regions and FAST feature detectors are then assigned into each region independently.By adaptively adjusting the feature detecting threshold in each region,high-quality feature points which are uniformly distributed within the frame can be obtained.In order to improve the accuracy of feature matching,we propose a plane-based feature refinement method.By eliminating the outliers in each plane separately,we can obtain as many as possible high-quality feature points.We apply the camera motion estimation and adaptive path optimization method based on projective motion model.By normalizing feature points and applying DLT to compute homograph,we can obtain camera motion between consecutive frames.We also design evaluation methods to verify the motion accuracy.By analyzing the shaky camera motions,we propose an iterative path smooth method to remove the high-frequency jitters along the motion.The method applies a bilateral filter with Gaussian kernel in path smoothing and adaptively adjusts the smooth strength to avoid distortions such as over-cropping and wobble occurring in quick rotation scene.Finally,we synthesize the stabilized video along the smoothed camera path.We design an Android application based on our algorithm to stabilize various consumer-level videos.With limited computational resource in a mobile phone,our application can process full-HD videos in real-time.Experiments on stabilizing challenging video scenes demonstrate the effectiveness and robustness of our technique.