Lecture-14 Motion detection

1、12022-3-6Lecture 1: Introduction 14.1 Introduction 14.2 Motion detection methods 14.3 Video processingLecture 14: Motion D22022-3-614.1 Introduction First, the world is dynamic, and motion information is one of the keys to many practical applications such as video mining, Robert navigation, intellig

2、ent traffic system, etc. Motion Detection is also a basic step to machine learning, pattern recognition in dynamic context.Why Motion Detection? 32022-3-614.1 IntroductionExtract the moving foreground from image sequence, which is to separate the moving foreground from background.將運(yùn)動前景從圖像序列中提取出來，也就是

3、說將背景與運(yùn)動前景分離開。42022-3-6 Static Camera Static Objects Static Camera Moving Objects Moving Camera Static Objects Moving CameraMoving Objects14.1 I52022-3-614.2 Motion detectionTwo approaches : 1. Motion detection based on foreground information directly.Foreground information直接利用前景信息檢測前景。62022-3-614.2

4、Motion detectionTwo approaches : 2. The input image subtract the background image and get a foreground image.先得到背景圖象，然后將輸入圖象減去背景圖像從而得到前景圖象。Input imageBackground 72022-3-6一些場景的背景相對固定，而前景變化較大。Two approaches : 1. Motion detection based on foreground information directly.2. The input image subtract the

5、background image and get a foreground image.According to the specific application decide which approach is adopted. 兩種思路，視具體環(huán)境選擇.一些場景的背景有較大變化，而前景的一些特征變化不大。14.2 Motion 82022-3-6幀間差分法(Frame differencing)背景差法(Background subtraction)光流法(Optical flow)前景建模方法14.2 Motion 92022-3-614.2 Motion detection A sim

6、ple way to detect the variation between frames is to do subtraction on gray-scale values of corresponding pixels. Double frame difference（雙幀差分）1( , ) |( , )( , )|nnnD i jI i jIi jDifference Image：1,nnIIwhere are two images of adjacent frame.0( , )1nMi j( , )( , )nnD i jTD i jTMoving foregroundT10202

7、2-3-614.2 Motion detection Three-frame difference （三幀差分）11( , ) |( , )( , )|( , )( , )|nnnnnD i jIi jI i jI i jIi jDifference Image：11,nnnIIIwhere are three images of adjacent frame.0( , )1nMi j( , )( , )nnD i jTD i jTMoving foregroundT112022-3-6 優(yōu)點(diǎn)：適用于動態(tài)變化的背景環(huán)境。 不足：較難準(zhǔn)確檢測運(yùn)動速度過快的物體；較難準(zhǔn)確檢測運(yùn)動速度過慢的物體；較

8、難準(zhǔn)確檢測場景中同時(shí)存在的多個(gè)運(yùn)動物體。14.2 Motion 122022-3-614.2 Motion detectionPrinciple：Construction the background image of scene , Calculate the difference on gray-scale value between current image and background image pixel-by-pixel.構(gòu)建背景圖像，計(jì)算當(dāng)前圖像與背景圖像的逐象素的灰度差。( , ) |( , )( , )|nnD i jI i jB i jCurrent imageBack

9、ground image0( , )1nMi j( , )( , )nnD i jTD i jTMoving foregroundT132022-3-6Input imageBackground image14.2 Motion detection?Moving foreground How to construct background image？- -求其均值圖像；求其均值圖像；背景未知：背景已知：- -圖像訓(xùn)練集的中值/均值圖像；142022-3-6 1 2 3 4 5 6a.均值圖像：若干背景圖像求逐點(diǎn)的灰度均值。14.2 Motion detection簡單，易于實(shí)現(xiàn)；不適用于背景變化的情況152022-3-614.3 Video processingCommands for Image ProcessingRead a video file:avi