EEE3032 – Computer Vision and Pattern Recognition (handout)
Simple Blob Tracker
Lecture 13 – Tracking (I)
close all;
clear all;
F=aviread('c:\visiondemo\ball_uncomp.avi');
%F=aviread('c:\horse_ucomp.avi');
NUMFRAMES=length(F);
%% 1) Get a colour model from Frame 1
img=double(F(1).cdata)./255;
imshow(img);
[u v]=ginput(2);
topleft=round([u(1);v(1)]);
botright=round([u(2);v(2)]);
samps=[]; % will store RGB values of user-selected pixels
for x=topleft(1):botright(1)
for y=topleft(2):botright(2)
r=img(y,x,1);
g=img(y,x,2);
b=img(y,x,3);
samps=[ samps [ r ; g ;b ] ] % one column per rgb
end
end
% build eigenmodel
e=Eigen_Build(samps);
lastpos=round(topleft+(botright-topleft)./2);
%% 2) Track from frame 2 onwards
history=[];
for i=2:NUMFRAMES
% reshape the pixels in the image to be 1 column per pixel (r,g,b)
img=double(F(i).cdata)./255;
samps=[ reshape(img(:,:,1),1,[]) ; reshape(img(:,:,2),1,[]) ;
reshape(img(:,:,3),1,[]) ];
% get Mahalanobis distance of each pixel from eigenmodel
dst=Eigen_Mahalanobis(samps,e);
% reshape the Mahalanobis distances into a rectangular image
dst=reshape(dst,size(img,1),size(img,2));
% threshold the Mahalanobis distances
thresholded=dst<3;
% dilate the mask a bit
structuringelement = strel('square',3);
thresholded=imdilate(thresholded,structuringelement);
1 of 5
EEE3032 – Computer Vision and Pattern Recognition (handout)
% find all the centroids of the next frame's connected components
map=bwlabel(thresholded);
reglabels=setdiff(unique(map),0);
fprintf('There are %d connected components\n',max(reglabels));
for cc=1:max(reglabels)
mask=(map==cc);
[y x]=find(mask);
possiblecentroid{cc}=[mean(x);mean(y)];
end
bestdist=inf;
bestidx=-1;
for cc=1:max(reglabels)
thisdist=sqrt(sum((possiblecentroid{cc}-lastpos).^2));
if (thisdist<bestdist)
bestdist=thisdist;
bestidx=cc;
end
end
currentpos=possiblecentroid{bestidx};
history=[history currentpos];
% display
cla; hold on;
thresholded(:,:,2)=thresholded(:,:,1);
thresholded(:,:,3)=thresholded(:,:,1);
imgshow([img thresholded]);
plot(history(1,:),history(2,:),'y-');
plot(currentpos(1),currentpos(2),'m*');
drawnow;
title(i);
lastpos=currentpos;
end
2 of 5
EEE3032 – Computer Vision and Pattern Recognition (handout)
Descriptor based tracking (using RGB Histogram)
Lecture 13 – Tracking (I)
close all;
clear all;
%F=aviread('c:\visiondemo\ball_uncomp.avi');
F=aviread('c:\horse_ucomp.avi');
NUMFRAMES=length(F);
%% 1) Cut a template from Frame 1
img=double(F(1).cdata)./255;
imshow(img);
[u v]=ginput(2);
topleft=round([u(1);v(1)]);
botright=round([u(2);v(2)]);
targetwnd =img(topleft(2):botright(2),topleft(1):botright(1),:);
targetdesc=getTrackDescriptor(targetwnd);
[winh winw winc]=size(targetwnd);
lastpos=topleft+(botright-topleft)./2;
%% 2) Track from frame 2 onwards
history=[];
for i=2:NUMFRAMES
% reshape the pixels in the image to be 1 column per pixel (r,g,b)
img=double(F(i).cdata)./255;
STEP=2;
halfwinw=floor(winw/2);
halfwinh=floor(winh/2);
scores=ones(size(img,1),size(img,2)).*inf;
for x=(1+halfwinw):STEP:(size(img,2)-(halfwinw+1))
for y=(1+halfwinh):STEP:(size(img,1)-(halfwinh+1))
wnd=img((y-halfwinh):(y+halfwinh),(x-halfwinw):(x+halfwinw),:);
thisdesc=getTrackDescriptor(wnd);
eucdst=sqrt(sum((targetdesc-thisdesc).^2));
scores(y,x)=eucdst;
end
end
% threshold the descriptor distances
thresholded=scores<0.25;
% dilate the mask a bit
structuringelement = strel('square',3);
thresholded=imdilate(thresholded,structuringelement);
% find all the centroids of the next frame's connected components
3 of 5
EEE3032 – Computer Vision and Pattern Recognition (handout)
map=bwlabel(thresholded);
reglabels=setdiff(unique(map),0);
if isempty(reglabels)
fprintf('No connected components\n');
continue;
end
fprintf('There are %d connected components\n',max(reglabels));
for cc=1:max(reglabels)
mask=(map==cc);
[y x]=find(mask);
possiblecentroid{cc}=round([mean(x);mean(y)]);
end
bestdist=inf;
bestidx=-1;
for cc=1:max(reglabels)
thisdist=sqrt(sum((possiblecentroid{cc}-lastpos).^2));
if (thisdist<bestdist)
bestdist=thisdist;
bestidx=cc;
end
end
if (bestidx==-1)
fprintf('Not found! skipping frame\n');
continue;
end
currentpos=possiblecentroid{bestidx};
history=[history currentpos];
% display
cla; hold on;
thresholded(:,:,2)=thresholded(:,:,1);
thresholded(:,:,3)=thresholded(:,:,1);
imgshow([img thresholded]);
plot(history(1,:),history(2,:),'y-');
plot(currentpos(1),currentpos(2),'m*');
drawnow;
title(i);
lastpos=currentpos;
end
4 of 5
EEE3032 – Computer Vision and Pattern Recognition (handout)
Compute an RGB Histogram from an image
(used in Descriptor based tracking)
Lecture 13 – Tracking (I)
function F=getTrackDescriptor(img)
red=img(:,:,1);
green=img(:,:,2);
blue=img(:,:,3);
red=reshape(red,1,[]);
green=reshape(green,1,[]);
blue=reshape(blue,1,[]);
%F=[mean(red) ; mean(green); mean(blue)];
%return;
Q=4; % Split RGB space into 4x4x4 cubes
sf=1./Q;
redindex=floor(red./sf); % when Q=4 each pixel is in set {0,1,2,3}
greenindex=floor(green./sf);
blueindex=floor(blue./sf);
N=redindex*(Q^2) + greenindex*Q + blueindex;
bins=Q^3;
% Make histogram from N
F=hist(N,1:bins);
% Normalise the histogram
F=F./sum(F);
5 of 5