학술논문
基于BRISK的实时视频抖动检测算法 / Real-time video shaking detection algorithm based on BRISK
Document Type
Academic Journal
Author
Source
计算机工程与设计. 37(8):2132-2137
Subject
Language
Chinese
ISSN
1000-7024
Abstract
针对目前视频抖动检测算法只能处理离线视频、简单判断视频是否抖动，不能识别抖动类型及具体帧序号的问题，提出一种实时视频抖动检测算法。采用BRISK算法进行视频图像特征检测，采用 OPQANN 算法进行特征点匹配，对匹配的特征点采用RANSAC算法进行运动参数估计并构建其运动轨迹，采用高斯低通滤波器对运动轨迹进行滤波处理，据此进行视频水平、垂直及旋转抖动检测。大量实验结果表明，在复杂环境下，与ORB+前向-后向 LK算法相比，该算法抖动检测正确率提高18.3%，误报率、漏报率分别降低13.8%、34%，在抖动类型识别等方面该算法也表现出良好的性能。
Aiming at the problem that current algorithms for video shaking detection can only process off-line videos and simply detect shakiness,which are incapable of figuring out the type of shakiness and which video frame has the shakiness,a real-time video shaking detection algorithm was proposed.The relationships of the successive frames were formulated by extracting the bi-nary robust invariant scalable keypoints (BRISK)feature points that matched using optimized product quantization for approxi-mate nearest neighbor search (OPQANN)algorithm,which then applied to RANdom SAmple consensus (RANSACE)algo-rithm to estimate the motion parameters and curves.The noises in motion curves were smoothed out using Gaussian filter,which were then adopted to detect the horizontal,vertical and rotated shakiness in video.Extensive experimental results show that the proposed algorithm performs favorably against oriented FAST and rotated BRIEF (ORB)+forward-backward Lucas-Kanade (LK)algorithm,the accuracy is increased by 1 8.3%,while the false positive and false negative are decreased by 1 3.8% and 34% respectively.Furthermore,the proposed approach also shows good performance in recognizing the shakiness type.
Aiming at the problem that current algorithms for video shaking detection can only process off-line videos and simply detect shakiness,which are incapable of figuring out the type of shakiness and which video frame has the shakiness,a real-time video shaking detection algorithm was proposed.The relationships of the successive frames were formulated by extracting the bi-nary robust invariant scalable keypoints (BRISK)feature points that matched using optimized product quantization for approxi-mate nearest neighbor search (OPQANN)algorithm,which then applied to RANdom SAmple consensus (RANSACE)algo-rithm to estimate the motion parameters and curves.The noises in motion curves were smoothed out using Gaussian filter,which were then adopted to detect the horizontal,vertical and rotated shakiness in video.Extensive experimental results show that the proposed algorithm performs favorably against oriented FAST and rotated BRIEF (ORB)+forward-backward Lucas-Kanade (LK)algorithm,the accuracy is increased by 1 8.3%,while the false positive and false negative are decreased by 1 3.8% and 34% respectively.Furthermore,the proposed approach also shows good performance in recognizing the shakiness type.