A Digital Video Stabilization System Based on Reliable SIFT Feature Matching and Adaptive Low-Pass Filtering

A real-time digital video stabilization system is proposed to remove unwanted camera shakes and jitters. Firstly, SIFT algorithm is improved to extract and match features between the reference frame and current frame reliably, and then global motion param

  • PDF / 306,085 Bytes
  • 10 Pages / 439.37 x 666.14 pts Page_size
  • 28 Downloads / 234 Views

DOWNLOAD

REPORT


Department of Automation, University of Science and Technology of China, Hefei 230026, China {harrtjun,zfwang}@ustc.edu.cn 2 Institute of Intelligent Machines, Chinese Academy of Sciences, Hefei 230031, China

Abstract. A real-time digital video stabilization system is proposed to remove unwanted camera shakes and jitters. Firstly, SIFT algorithm is improved to extract and match features between the reference frame and current frame reliably, and then global motion parameters are obtained based on the geometric constraint consistency between feature matches through random sample consensus algorithm. Secondly, multiple evaluation criteria are fused by an adaptive lowpass filter to smooth global motion for obtaining correction vector, which is used to compensate the current frame. Finally, stabilized video is obtained after each frame is completed by combining the texture synthesis method and the spatio-temporal information of video. The objective experiments demonstrate the system can increase the average peak signal-to-noise ratio of jittered videos around 6.12 dB, The subjective experiments demonstrate the system can increase the identification ability and perceptive comfort on video content. Keywords: Global motion estimation · Motion filtering and compensation · Video completion

1

Introduction

Videos retrieved from hand-held video cameras are affected by unwanted camera shakes and jitters, resulting in video quality loss [1]. Digital video stabilization techniques have gained consensus, for they permit to obtain high quality and stable video footages by making use only of information drawn from footage images and do not need any additional knowledge about camera physical motion [2][3]. There are three stages for digital video stabilization: global motion estimation [4], motion filtering and compensation [5], video completion [6]. Global motion estimation can be performed by global intensity alignment approaches [7-10] or feature-based approaches [11-13]. Feature-based methods are generally faster than global intensity alignment approaches, while they are more prone to local effects. A good survey on global motion estimation can be found in [4]. After estimating the global motion, motion filtering is removing the annoying irregular jitter to recognize intentional movement. It can be performed by DFT filtered © Springer-Verlag Berlin Heidelberg 2015 H. Zha et al. (Eds.): CCCV 2015, Part II, CCIS 547, pp. 180–189, 2015. DOI: 10.1007/978-3-662-48570-5_18

A Digital Video Stabilization System Based on Reliable SIFT Feature Matching

181

frame position smoothing [10], Kalman filtering [14] and motion vector integration [15] according to real system constraints [16][17]. After motion filtering, motion compensation is applied to spatially displace image frames by correction vector from the filtering result. The goal of video completion is filling in missing image areas in a video [18]. It can be performed by mosaicing [19], sampling spatio-temporal volume patches [20], multi-layers segmenting [21][22] and local motio