int sobel_argb32_3x3_partial(vbx_uword_t *sobel_out, vbx_uword_t *argb_in, const short image_width,
const short image_height, const short image_pitch, const short renorm)
{
VBX::Prefetcher<vbx_uword_t> input(1,image_width,argb_in,argb_in+image_height*image_pitch,image_pitch);
VBX::Vector<vbx_uword_t> output(image_width);
VBX::Vector<vbx_uhalf_t>* luma[3];
luma[0]=new VBX::Vector<vbx_uhalf_t>(image_width);
luma[1]=new VBX::Vector<vbx_uhalf_t>(image_width);
luma[2]=new VBX::Vector<vbx_uhalf_t>(image_width);
VBX::Vector<vbx_uhalf_t> gradient_x(image_width);
VBX::Vector<vbx_uhalf_t> gradient_y(image_width);
VBX::Vector<vbx_uhalf_t>* sobel_rows[3];
sobel_rows[0]=new VBX::Vector<vbx_uhalf_t>(image_width);
sobel_rows[1]=new VBX::Vector<vbx_uhalf_t>(image_width);
sobel_rows[2]=new VBX::Vector<vbx_uhalf_t>(image_width);
input.fetch();
int rowmod3=0;
for(int row=0;row<image_height;row++){
input.fetch();
VBX::Vector<vbx_uhalf_t>& sobel_top= *sobel_rows[rowmod3];
VBX::Vector<vbx_uhalf_t>& sobel_bot= *sobel_rows[mod3(rowmod3+2)];
VBX::Vector<vbx_uhalf_t>& luma_top= *luma[rowmod3];
VBX::Vector<vbx_uhalf_t>& luma_mid= *luma[mod3(rowmod3+1)];
VBX::Vector<vbx_uhalf_t>& luma_bot= *luma[mod3(rowmod3+2)];
rowmod3=mod3(rowmod3+1);
argb_to_luma8(luma_bot,input[0]);
sobel_row( sobel_bot ,luma_bot);
if(row<2){
continue;
}
gradient_y=absdiff(sobel_top,sobel_bot);
gradient_x=luma_top +luma_bot + luma_mid*2;
gradient_x[1 upto image_width-1]=absdiff(gradient_x[0 upto image_width-2],gradient_x[2 upto image_width]);
output=((gradient_x + gradient_y) >> renorm);
output.cond_move(output>0xFF,0xFF);
output*=0x010101;
//write to output buffer, skipping first and last elements
output[1 upto (image_width -1)].dma_write(sobel_out+(row-1)*image_pitch +1);
}
output=0;
output.dma_write(sobel_out);
output.dma_write(sobel_out+ (image_height-1)*image_pitch);
delete luma[0];
delete luma[1];
delete luma[2];
delete sobel_rows[0];
delete sobel_rows[1];
delete sobel_rows[2];
return 0;
}
int sobel_luma8_3x3(vbx_uword_t *output,
unsigned char *input,
const short image_width,
const short image_height,
const short image_pitch,
const short renorm)
{
VBX::Prefetcher<vbx_ubyte_t> luma_in(3,image_width,input,input+image_height*image_pitch,image_pitch);
VBX::Vector<vbx_uhalf_t> *sobel[3];
sobel[0]=new VBX::Vector<vbx_uhalf_t>(image_width-2);
sobel[1]=new VBX::Vector<vbx_uhalf_t>(image_width-2);
sobel[2]=new VBX::Vector<vbx_uhalf_t>(image_width-2);
VBX::Vector<vbx_uhalf_t> gradient_x(image_width-2);
VBX::Vector<vbx_uhalf_t> gradient_y(image_width-2);
VBX::Vector<vbx_uword_t> row_out(image_width);
VBX::Vector<vbx_uhalf_t> tmp(image_width);
if(!tmp.data){
printf("Out of scratchpad memory\n");
return -1;
}
luma_in.fetch();
luma_in.fetch();
sobel_row(*sobel[0],luma_in[0]);
luma_in.fetch();
sobel_row(*sobel[1],luma_in[1]);
//set first row black
row_out=0;
row_out.dma_write(output);
for(int row=0,s_t=0,s_b=2;row<image_height-(3-1);row++,s_t++,s_b++){
luma_in.fetch();
//use reference to refer to vectors without copying
VBX::Vector<vbx_ubyte_t>& luma_top=luma_in[0];
VBX::Vector<vbx_ubyte_t>& luma_mid=luma_in[1];
VBX::Vector<vbx_ubyte_t>& luma_bot=luma_in[2];
if(s_t>=3)s_t=0;//mod 3
if(s_b>=3)s_b=0;//mod 3
VBX::Vector<vbx_uhalf_t>& sobel_top=*sobel[s_t];
VBX::Vector<vbx_uhalf_t>& sobel_bot=*sobel[s_b];
sobel_row(sobel_bot,luma_bot);
gradient_y=absdiff(sobel_top,sobel_bot);
tmp=luma_top+luma_bot+(luma_mid<<1);
gradient_x=absdiff(tmp,tmp[2 upto image_width]);
//sum the gradients, normalize and saturate at 255
row_out[1 upto image_width-1]=(gradient_x+gradient_y)>>renorm;
row_out.cond_move(row_out>255,255);
//copy lowest byte into the 2nd and 3rd;
row_out*=0x10101;
row_out.dma_write(output+(row+1)*image_pitch);
}
//set last row black
row_out=0;
row_out.dma_write(output+(image_height-1)*image_width);
return 0;
}
// Per OpenVX
// Implements the Sobel Image Filter k.
// This k produces two output planes (one can be omitted) in the x and y plane.
// The Sobel Operators Gx,Gy are defined as:
//
// -1 0 1 -1 -2 -1
// Gx = -2 0 2 Gy = 0 0 0
// 1 0 1 1 2 1
//
// https://www.khronos.org/registry/vx/specs/1.0/html/da/d4b/group__group__vision__function__sobel3x3.html
//
std::pair<Halide::Func, Halide::Func> sobel_3x3(Halide::Func input, bool grayscale) {
Halide::Func kx("kx"), ky("ky");
Halide::Func gradient_x("gradient_x"), gradient_y("gradient_y");
Halide::RDom r(-1,3,-1,3);
Halide::Var x,y,c;
kx(x,y) = 0;
kx(-1,-1) = -1;
kx(0,-1) = 0;
kx(1,-1) = 1;
kx(-1, 0) = -2;
kx(0, 0) = 0;
kx(1, 0) = 2;
kx(-1, 1) = -1;
kx(0, 1) = 0;
kx(1, 1) = 1;
if (grayscale)
gradient_x(x,y) = sum(input(x+r.x, y+r.y) * kx(r.x, r.y));
else
gradient_x(x,y,c) = sum(input(x+r.x, y+r.y, c) * kx(r.x, r.y));
ky(x,y) = 0;
ky(-1,-1) = -1;
ky(0,-1) = -2;
ky(1,-1) = -1;
ky(-1, 0) = 0;
ky(0, 0) = 0;
ky(1, 0) = 0;
ky(-1, 1) = 1;
ky(0, 1) = 2;
ky(1, 1) = 1;
if (grayscale)
gradient_y(x,y) = sum(input(x+r.x, y+r.y) * ky(r.x, r.y));
else
gradient_y(x,y,c) = sum(input(x+r.x, y+r.y, c) * ky(r.x, r.y));
std::pair<Halide::Func, Halide::Func> ret = std::make_pair(gradient_x, gradient_y);
return ret;
}
// Main
int main(int argc, char ** argv) {
// Choose the best GPU in case there are multiple available
choose_GPU();
// Keep track of the start time of the program
long long program_start_time = get_time();
if (argc !=3){
fprintf(stderr, "usage: %s <input file> <number of frames to process>", argv[0]);
exit(1);
}
// Let the user specify the number of frames to process
int num_frames = atoi(argv[2]);
// Open video file
char *video_file_name = argv[1];
avi_t *cell_file = AVI_open_input_file(video_file_name, 1);
if (cell_file == NULL) {
AVI_print_error("Error with AVI_open_input_file");
return -1;
}
int i, j, *crow, *ccol, pair_counter = 0, x_result_len = 0, Iter = 20, ns = 4, k_count = 0, n;
MAT *cellx, *celly, *A;
double *GICOV_spots, *t, *G, *x_result, *y_result, *V, *QAX_CENTERS, *QAY_CENTERS;
double threshold = 1.8, radius = 10.0, delta = 3.0, dt = 0.01, b = 5.0;
// Extract a cropped version of the first frame from the video file
MAT *image_chopped = get_frame(cell_file, 0, 1, 0);
printf("Detecting cells in frame 0\n");
// Get gradient matrices in x and y directions
MAT *grad_x = gradient_x(image_chopped);
MAT *grad_y = gradient_y(image_chopped);
// Allocate for gicov_mem and strel
gicov_mem = (float*) malloc(sizeof(float) * grad_x->m * grad_y->n);
strel = (float*) malloc(sizeof(float) * strel_m * strel_n);
m_free(image_chopped);
int grad_m = grad_x->m;
int grad_n = grad_y->n;
#pragma acc data create(sin_angle,cos_angle,theta,tX,tY) \
create(gicov_mem[0:grad_x->m*grad_y->n])
{
// Precomputed constants on GPU
compute_constants();
// Get GICOV matrices corresponding to image gradients
long long GICOV_start_time = get_time();
MAT *gicov = GICOV(grad_x, grad_y);
long long GICOV_end_time = get_time();
// Dilate the GICOV matrices
long long dilate_start_time = get_time();
MAT *img_dilated = dilate(gicov);
long long dilate_end_time = get_time();
} /* end acc data */
// Find possible matches for cell centers based on GICOV and record the rows/columns in which they are found
pair_counter = 0;
crow = (int *) malloc(gicov->m * gicov->n * sizeof(int));
ccol = (int *) malloc(gicov->m * gicov->n * sizeof(int));
for(i = 0; i < gicov->m; i++) {
for(j = 0; j < gicov->n; j++) {
if(!double_eq(m_get_val(gicov,i,j), 0.0) && double_eq(m_get_val(img_dilated,i,j), m_get_val(gicov,i,j)))
{
crow[pair_counter]=i;
ccol[pair_counter]=j;
pair_counter++;
}
}
}
GICOV_spots = (double *) malloc(sizeof(double) * pair_counter);
for(i = 0; i < pair_counter; i++)
GICOV_spots[i] = sqrt(m_get_val(gicov, crow[i], ccol[i]));
G = (double *) calloc(pair_counter, sizeof(double));
x_result = (double *) calloc(pair_counter, sizeof(double));
y_result = (double *) calloc(pair_counter, sizeof(double));
x_result_len = 0;
for (i = 0; i < pair_counter; i++) {
if ((crow[i] > 29) && (crow[i] < BOTTOM - TOP + 39)) {
x_result[x_result_len] = ccol[i];
y_result[x_result_len] = crow[i] - 40;
G[x_result_len] = GICOV_spots[i];
x_result_len++;
}
}
// Make an array t which holds each "time step" for the possible cells
t = (double *) malloc(sizeof(double) * 36);
for (i = 0; i < 36; i++) {
t[i] = (double)i * 2.0 * PI / 36.0;
}
// Store cell boundaries (as simple circles) for all cells
cellx = m_get(x_result_len, 36);
celly = m_get(x_result_len, 36);
for(i = 0; i < x_result_len; i++) {
for(j = 0; j < 36; j++) {
m_set_val(cellx, i, j, x_result[i] + radius * cos(t[j]));
m_set_val(celly, i, j, y_result[i] + radius * sin(t[j]));
}
}
A = TMatrix(9,4);
V = (double *) malloc(sizeof(double) * pair_counter);
QAX_CENTERS = (double * )malloc(sizeof(double) * pair_counter);
QAY_CENTERS = (double *) malloc(sizeof(double) * pair_counter);
memset(V, 0, sizeof(double) * pair_counter);
memset(QAX_CENTERS, 0, sizeof(double) * pair_counter);
memset(QAY_CENTERS, 0, sizeof(double) * pair_counter);
// For all possible results, find the ones that are feasibly leukocytes and store their centers
k_count = 0;
for (n = 0; n < x_result_len; n++) {
if ((G[n] < -1 * threshold) || G[n] > threshold) {
MAT * x, *y;
VEC * x_row, * y_row;
x = m_get(1, 36);
y = m_get(1, 36);
x_row = v_get(36);
y_row = v_get(36);
// Get current values of possible cells from cellx/celly matrices
x_row = get_row(cellx, n, x_row);
y_row = get_row(celly, n, y_row);
uniformseg(x_row, y_row, x, y);
// Make sure that the possible leukocytes are not too close to the edge of the frame
if ((m_min(x) > b) && (m_min(y) > b) && (m_max(x) < cell_file->width - b) && (m_max(y) < cell_file->height - b)) {
MAT * Cx, * Cy, *Cy_temp, * Ix1, * Iy1;
VEC *Xs, *Ys, *W, *Nx, *Ny, *X, *Y;
Cx = m_get(1, 36);
Cy = m_get(1, 36);
Cx = mmtr_mlt(A, x, Cx);
Cy = mmtr_mlt(A, y, Cy);
Cy_temp = m_get(Cy->m, Cy->n);
for (i = 0; i < 9; i++)
m_set_val(Cy, i, 0, m_get_val(Cy, i, 0) + 40.0);
// Iteratively refine the snake/spline
for (i = 0; i < Iter; i++) {
int typeofcell;
if(G[n] > 0.0) typeofcell = 0;
else typeofcell = 1;
splineenergyform01(Cx, Cy, grad_x, grad_y, ns, delta, 2.0 * dt, typeofcell);
}
X = getsampling(Cx, ns);
for (i = 0; i < Cy->m; i++)
m_set_val(Cy_temp, i, 0, m_get_val(Cy, i, 0) - 40.0);
Y = getsampling(Cy_temp, ns);
Ix1 = linear_interp2(grad_x, X, Y);
Iy1 = linear_interp2(grad_x, X, Y);
Xs = getfdriv(Cx, ns);
Ys = getfdriv(Cy, ns);
Nx = v_get(Ys->dim);
for (i = 0; i < Ys->dim; i++)
v_set_val(Nx, i, v_get_val(Ys, i) / sqrt(v_get_val(Xs, i)*v_get_val(Xs, i) + v_get_val(Ys, i)*v_get_val(Ys, i)));
Ny = v_get(Xs->dim);
for (i = 0; i < Xs->dim; i++)
v_set_val(Ny, i, -1.0 * v_get_val(Xs, i) / sqrt(v_get_val(Xs, i)*v_get_val(Xs, i) + v_get_val(Ys, i)*v_get_val(Ys, i)));
W = v_get(Nx->dim);
for (i = 0; i < Nx->dim; i++)
v_set_val(W, i, m_get_val(Ix1, 0, i) * v_get_val(Nx, i) + m_get_val(Iy1, 0, i) * v_get_val(Ny, i));
V[n] = mean(W) / std_dev(W);
// Find the cell centers by computing the means of X and Y values for all snaxels of the spline contour
QAX_CENTERS[k_count] = mean(X);
QAY_CENTERS[k_count] = mean(Y) + TOP;
k_count++;
// Free memory
v_free(W);
v_free(Ny);
v_free(Nx);
v_free(Ys);
v_free(Xs);
m_free(Iy1);
m_free(Ix1);
v_free(Y);
v_free(X);
m_free(Cy_temp);
m_free(Cy);
m_free(Cx);
}
// Free memory
v_free(y_row);
v_free(x_row);
m_free(y);
m_free(x);
}
}
// Free memory
free(gicov_mem);
free(strel);
free(V);
free(ccol);
free(crow);
free(GICOV_spots);
free(t);
free(G);
free(x_result);
free(y_result);
m_free(A);
m_free(celly);
m_free(cellx);
m_free(img_dilated);
m_free(gicov);
m_free(grad_y);
m_free(grad_x);
// Report the total number of cells detected
printf("Cells detected: %d\n\n", k_count);
// Report the breakdown of the detection runtime
printf("Detection runtime\n");
printf("-----------------\n");
printf("GICOV computation: %.5f seconds\n", ((float) (GICOV_end_time - GICOV_start_time)) / (1000*1000));
printf(" GICOV dilation: %.5f seconds\n", ((float) (dilate_end_time - dilate_start_time)) / (1000*1000));
printf(" Total: %.5f seconds\n", ((float) (get_time() - program_start_time)) / (1000*1000));
// Now that the cells have been detected in the first frame,
// track the ellipses through subsequent frames
if (num_frames > 1) printf("\nTracking cells across %d frames\n", num_frames);
else printf("\nTracking cells across 1 frame\n");
long long tracking_start_time = get_time();
int num_snaxels = 20;
ellipsetrack(cell_file, QAX_CENTERS, QAY_CENTERS, k_count, radius, num_snaxels, num_frames);
printf(" Total: %.5f seconds\n", ((float) (get_time() - tracking_start_time)) / (float) (1000*1000*num_frames));
// Report total program execution time
printf("\nTotal application run time: %.5f seconds\n", ((float) (get_time() - program_start_time)) / (1000*1000));
return 0;
}
void ellipseevolve(MAT *f, double *xc0, double *yc0, double *r0, double *t, int Np, double Er, double Ey) {
/*
% ELLIPSEEVOLVE evolves a parametric snake according
% to some energy constraints.
%
% INPUTS:
% f............potential surface
% xc0,yc0......initial center position
% r0,t.........initial radii & angle vectors (with Np elements each)
% Np...........number of snaxel points per snake
% Er...........expected radius
% Ey...........expected y position
%
% OUTPUTS
% xc0,yc0.......final center position
% r0...........final radii
%
% Matlab code written by: DREW GILLIAM (based on work by GANG DONG /
% NILANJAN RAY)
% Ported to C by: MICHAEL BOYER
*/
// Constants
double deltax = 0.2;
double deltay = 0.2;
double deltar = 0.2;
double converge = 0.1;
double lambdaedge = 1;
double lambdasize = 0.2;
double lambdapath = 0.05;
int iterations = 1000; // maximum number of iterations
int i, j;
// Initialize variables
double xc = *xc0;
double yc = *yc0;
double *r = (double *) malloc(sizeof(double) * Np);
for (i = 0; i < Np; i++) r[i] = r0[i];
// Compute the x- and y-gradients of the MGVF matrix
MAT *fx = gradient_x(f);
MAT *fy = gradient_y(f);
// Normalize the gradients
int fh = f->m, fw = f->n;
for (i = 0; i < fh; i++) {
for (j = 0; j < fw; j++) {
double temp_x = m_get_val(fx, i, j);
double temp_y = m_get_val(fy, i, j);
double fmag = sqrt((temp_x * temp_x) + (temp_y * temp_y));
m_set_val(fx, i, j, temp_x / fmag);
m_set_val(fy, i, j, temp_y / fmag);
}
}
double *r_old = (double *) malloc(sizeof(double) * Np);
VEC *x = v_get(Np);
VEC *y = v_get(Np);
// Evolve the snake
int iter = 0;
double snakediff = 1.0;
while (iter < iterations && snakediff > converge) {
// Save the values from the previous iteration
double xc_old = xc, yc_old = yc;
for (i = 0; i < Np; i++) {
r_old[i] = r[i];
}
// Compute the locations of the snaxels
for (i = 0; i < Np; i++) {
v_set_val(x, i, xc + r[i] * cos(t[i]));
v_set_val(y, i, yc + r[i] * sin(t[i]));
}
// See if any of the points in the snake are off the edge of the image
double min_x = v_get_val(x, 0), max_x = v_get_val(x, 0);
double min_y = v_get_val(y, 0), max_y = v_get_val(y, 0);
for (i = 1; i < Np; i++) {
double x_i = v_get_val(x, i);
if (x_i < min_x) min_x = x_i;
else if (x_i > max_x) max_x = x_i;
double y_i = v_get_val(y, i);
if (y_i < min_y) min_y = y_i;
else if (y_i > max_y) max_y = y_i;
}
if (min_x < 0.0 || max_x > (double) fw - 1.0 || min_y < 0 || max_y > (double) fh - 1.0) break;
// Compute the length of the snake
double L = 0.0;
for (i = 0; i < Np - 1; i++) {
double diff_x = v_get_val(x, i + 1) - v_get_val(x, i);
double diff_y = v_get_val(y, i + 1) - v_get_val(y, i);
L += sqrt((diff_x * diff_x) + (diff_y * diff_y));
}
double diff_x = v_get_val(x, 0) - v_get_val(x, Np - 1);
double diff_y = v_get_val(y, 0) - v_get_val(y, Np - 1);
L += sqrt((diff_x * diff_x) + (diff_y * diff_y));
// Compute the potential surface at each snaxel
MAT *vf = linear_interp2(f, x, y);
MAT *vfx = linear_interp2(fx, x, y);
MAT *vfy = linear_interp2(fy, x, y);
// Compute the average potential surface around the snake
double vfmean = sum_m(vf ) / L;
double vfxmean = sum_m(vfx) / L;
double vfymean = sum_m(vfy) / L;
// Compute the radial potential surface
int m = vf->m, n = vf->n;
MAT *vfr = m_get(m, n);
for (i = 0; i < n; i++) {
double vf_val = m_get_val(vf, 0, i);
double vfx_val = m_get_val(vfx, 0, i);
double vfy_val = m_get_val(vfy, 0, i);
double x_val = v_get_val(x, i);
double y_val = v_get_val(y, i);
double new_val = (vf_val + vfx_val * (x_val - xc) + vfy_val * (y_val - yc) - vfmean) / L;
m_set_val(vfr, 0, i, new_val);
}
// Update the snake center and snaxels
xc = xc + (deltax * lambdaedge * vfxmean);
yc = (yc + (deltay * lambdaedge * vfymean) + (deltay * lambdapath * Ey)) / (1.0 + deltay * lambdapath);
double r_diff = 0.0;
for (i = 0; i < Np; i++) {
r[i] = (r[i] + (deltar * lambdaedge * m_get_val(vfr, 0, i)) + (deltar * lambdasize * Er)) /
(1.0 + deltar * lambdasize);
r_diff += fabs(r[i] - r_old[i]);
}
// Test for convergence
snakediff = fabs(xc - xc_old) + fabs(yc - yc_old) + r_diff;
// Free temporary matrices
m_free(vf);
m_free(vfx);
m_free(vfy);
m_free(vfr);
iter++;
}
// Set the return values
*xc0 = xc;
*yc0 = yc;
for (i = 0; i < Np; i++)
r0[i] = r[i];
// Free memory
free(r);
free(r_old);
v_free( x);
v_free( y);
m_free(fx);
m_free(fy);
}
void ellipsetrack(avi_t *video, double *xc0, double *yc0, int Nc, int R, int Np, int Nf) {
/*
% ELLIPSETRACK tracks cells in the movie specified by 'video', at
% locations 'xc0'/'yc0' with radii R using an ellipse with Np discrete
% points, starting at frame number one and stopping at frame number 'Nf'.
%
% INPUTS:
% video.......pointer to avi video object
% xc0,yc0.....initial center location (Nc entries)
% Nc..........number of cells
% R...........initial radius
% Np..........nbr of snaxels points per snake
% Nf..........nbr of frames in which to track
%
% Matlab code written by: DREW GILLIAM (based on code by GANG DONG /
% NILANJAN RAY)
% Ported to C by: MICHAEL BOYER
*/
int i, j;
// Compute angle parameter
double *t = (double *) malloc(sizeof(double) * Np);
double increment = (2.0 * PI) / (double) Np;
for (i = 0; i < Np; i++) {
t[i] = increment * (double) i ;
}
// Allocate space for a snake for each cell in each frame
double **xc = alloc_2d_double(Nc, Nf + 1);
double **yc = alloc_2d_double(Nc, Nf + 1);
double ***r = alloc_3d_double(Nc, Np, Nf + 1);
double ***x = alloc_3d_double(Nc, Np, Nf + 1);
double ***y = alloc_3d_double(Nc, Np, Nf + 1);
// Save the first snake for each cell
for (i = 0; i < Nc; i++) {
xc[i][0] = xc0[i];
yc[i][0] = yc0[i];
for (j = 0; j < Np; j++) {
r[i][j][0] = (double) R;
}
}
// Generate ellipse points for each cell
for (i = 0; i < Nc; i++) {
for (j = 0; j < Np; j++) {
x[i][j][0] = xc[i][0] + (r[i][j][0] * cos(t[j]));
y[i][j][0] = yc[i][0] + (r[i][j][0] * sin(t[j]));
}
}
// Keep track of the total time spent on computing
// the MGVF matrix and evolving the snakes
long long MGVF_time = 0;
long long snake_time = 0;
// Process each frame
int frame_num, cell_num;
for (frame_num = 1; frame_num <= Nf; frame_num++) {
printf("\rProcessing frame %d / %d", frame_num, Nf);
fflush(stdout);
// Get the current video frame and its dimensions
MAT *I = get_frame(video, frame_num, 0, 1);
int Ih = I->m;
int Iw = I->n;
// Set the current positions equal to the previous positions
for (i = 0; i < Nc; i++) {
xc[i][frame_num] = xc[i][frame_num - 1];
yc[i][frame_num] = yc[i][frame_num - 1];
for (j = 0; j < Np; j++) {
r[i][j][frame_num] = r[i][j][frame_num - 1];
}
}
// Split the work among multiple threads, if OPEN is defined
#ifdef OPEN
#pragma omp parallel for num_threads(omp_num_threads) private(i, j)
#endif
// Track each cell
for (cell_num = 0; cell_num < Nc; cell_num++) {
// Make copies of the current cell's location
double xci = xc[cell_num][frame_num];
double yci = yc[cell_num][frame_num];
double *ri = (double *) malloc(sizeof(double) * Np);
for (j = 0; j < Np; j++) {
ri[j] = r[cell_num][j][frame_num];
}
// Add up the last ten y-values for this cell
// (or fewer if there are not yet ten previous frames)
double ycavg = 0.0;
for (i = (frame_num > 10 ? frame_num - 10 : 0); i < frame_num; i++) {
ycavg += yc[cell_num][i];
}
// Compute the average of the last ten y-values
// (this represents the expected y-location of the cell)
ycavg = ycavg / (double) (frame_num > 10 ? 10 : frame_num);
// Determine the range of the subimage surrounding the current position
int u1 = max(xci - 4.0 * R + 0.5, 0 );
int u2 = min(xci + 4.0 * R + 0.5, Iw - 1);
int v1 = max(yci - 2.0 * R + 1.5, 0 );
int v2 = min(yci + 2.0 * R + 1.5, Ih - 1);
// Extract the subimage
MAT *Isub = m_get(v2 - v1 + 1, u2 - u1 + 1);
for (i = v1; i <= v2; i++) {
for (j = u1; j <= u2; j++) {
m_set_val(Isub, i - v1, j - u1, m_get_val(I, i, j));
}
}
// Compute the subimage gradient magnitude
MAT *Ix = gradient_x(Isub);
MAT *Iy = gradient_y(Isub);
MAT *IE = m_get(Isub->m, Isub->n);
for (i = 0; i < Isub->m; i++) {
for (j = 0; j < Isub->n; j++) {
double temp_x = m_get_val(Ix, i, j);
double temp_y = m_get_val(Iy, i, j);
m_set_val(IE, i, j, sqrt((temp_x * temp_x) + (temp_y * temp_y)));
}
}
// Compute the motion gradient vector flow (MGVF) edgemaps
long long MGVF_start_time = get_time();
MAT *IMGVF = MGVF(IE, 1, 1);
MGVF_time += get_time() - MGVF_start_time;
// Determine the position of the cell in the subimage
xci = xci - (double) u1;
yci = yci - (double) (v1 - 1);
ycavg = ycavg - (double) (v1 - 1);
// Evolve the snake
long long snake_start_time = get_time();
ellipseevolve(IMGVF, &xci, &yci, ri, t, Np, (double) R, ycavg);
snake_time += get_time() - snake_start_time;
// Compute the cell's new position in the full image
xci = xci + u1;
yci = yci + (v1 - 1);
// Store the new location of the cell and the snake
xc[cell_num][frame_num] = xci;
yc[cell_num][frame_num] = yci;
for (j = 0; j < Np; j++) {
r[cell_num][j][frame_num] = ri[j];
x[cell_num][j][frame_num] = xc[cell_num][frame_num] + (ri[j] * cos(t[j]));
y[cell_num][j][frame_num] = yc[cell_num][frame_num] + (ri[j] * sin(t[j]));
}
// Output the updated center of each cell
//printf("%d,%f,%f\n", cell_num, xci[cell_num], yci[cell_num]);
// Free temporary memory
m_free(IMGVF);
free(ri);
}
// Output a new line to visually distinguish the output from different frames
//printf("\n");
}
// Free temporary memory
free(t);
free_2d_double(xc);
free_2d_double(yc);
free_3d_double(r);
free_3d_double(x);
free_3d_double(y);
// Report average processing time per frame
printf("\n\nTracking runtime (average per frame):\n");
printf("------------------------------------\n");
printf("MGVF computation: %.5f seconds\n", ((float) (MGVF_time)) / (float) (1000*1000*Nf));
printf(" Snake evolution: %.5f seconds\n", ((float) (snake_time)) / (float) (1000*1000*Nf));
}
template <typename PointInT, typename PointOutT> void
pcl::IntegralImageNormalEstimation<PointInT, PointOutT>::computePointNormalMirror (
const int pos_x, const int pos_y, const unsigned point_index, PointOutT &normal)
{
float bad_point = std::numeric_limits<float>::quiet_NaN ();
const int width = input_->width;
const int height = input_->height;
// ==============================================================
if (normal_estimation_method_ == COVARIANCE_MATRIX)
{
if (!init_covariance_matrix_)
initCovarianceMatrixMethod ();
const int start_x = pos_x - rect_width_2_;
const int start_y = pos_y - rect_height_2_;
const int end_x = start_x + rect_width_;
const int end_y = start_y + rect_height_;
unsigned count = 0;
sumArea<unsigned>(start_x, start_y, end_x, end_y, width, height, boost::bind(&IntegralImage2D<float, 3>::getFiniteElementsCountSE, &integral_image_XYZ_, _1, _2, _3, _4), count);
// no valid points within the rectangular reagion?
if (count == 0)
{
normal.normal_x = normal.normal_y = normal.normal_z = normal.curvature = bad_point;
return;
}
EIGEN_ALIGN16 Eigen::Matrix3f covariance_matrix;
Eigen::Vector3f center;
typename IntegralImage2D<float, 3>::SecondOrderType so_elements;
typename IntegralImage2D<float, 3>::ElementType tmp_center;
typename IntegralImage2D<float, 3>::SecondOrderType tmp_so_elements;
center[0] = 0;
center[1] = 0;
center[2] = 0;
tmp_center[0] = 0;
tmp_center[1] = 0;
tmp_center[2] = 0;
so_elements[0] = 0;
so_elements[1] = 0;
so_elements[2] = 0;
so_elements[3] = 0;
so_elements[4] = 0;
so_elements[5] = 0;
sumArea<typename IntegralImage2D<float, 3>::ElementType>(start_x, start_y, end_x, end_y, width, height, boost::bind(&IntegralImage2D<float, 3>::getFirstOrderSumSE, &integral_image_XYZ_, _1, _2, _3, _4), tmp_center);
sumArea<typename IntegralImage2D<float, 3>::SecondOrderType>(start_x, start_y, end_x, end_y, width, height, boost::bind(&IntegralImage2D<float, 3>::getSecondOrderSumSE, &integral_image_XYZ_, _1, _2, _3, _4), so_elements);
center[0] = float (tmp_center[0]);
center[1] = float (tmp_center[1]);
center[2] = float (tmp_center[2]);
covariance_matrix.coeffRef (0) = static_cast<float> (so_elements [0]);
covariance_matrix.coeffRef (1) = covariance_matrix.coeffRef (3) = static_cast<float> (so_elements [1]);
covariance_matrix.coeffRef (2) = covariance_matrix.coeffRef (6) = static_cast<float> (so_elements [2]);
covariance_matrix.coeffRef (4) = static_cast<float> (so_elements [3]);
covariance_matrix.coeffRef (5) = covariance_matrix.coeffRef (7) = static_cast<float> (so_elements [4]);
covariance_matrix.coeffRef (8) = static_cast<float> (so_elements [5]);
covariance_matrix -= (center * center.transpose ()) / static_cast<float> (count);
float eigen_value;
Eigen::Vector3f eigen_vector;
pcl::eigen33 (covariance_matrix, eigen_value, eigen_vector);
flipNormalTowardsViewpoint (input_->points[point_index], vpx_, vpy_, vpz_, eigen_vector[0], eigen_vector[1], eigen_vector[2]);
normal.getNormalVector3fMap () = eigen_vector;
// Compute the curvature surface change
if (eigen_value > 0.0)
normal.curvature = fabsf (eigen_value / (covariance_matrix.coeff (0) + covariance_matrix.coeff (4) + covariance_matrix.coeff (8)));
else
normal.curvature = 0;
return;
}
// =======================================================
else if (normal_estimation_method_ == AVERAGE_3D_GRADIENT)
{
if (!init_average_3d_gradient_)
initAverage3DGradientMethod ();
const int start_x = pos_x - rect_width_2_;
const int start_y = pos_y - rect_height_2_;
const int end_x = start_x + rect_width_;
const int end_y = start_y + rect_height_;
unsigned count_x = 0;
unsigned count_y = 0;
sumArea<unsigned>(start_x, start_y, end_x, end_y, width, height, boost::bind(&IntegralImage2D<float, 3>::getFiniteElementsCountSE, &integral_image_DX_, _1, _2, _3, _4), count_x);
sumArea<unsigned>(start_x, start_y, end_x, end_y, width, height, boost::bind(&IntegralImage2D<float, 3>::getFiniteElementsCountSE, &integral_image_DY_, _1, _2, _3, _4), count_y);
if (count_x == 0 || count_y == 0)
{
normal.normal_x = normal.normal_y = normal.normal_z = normal.curvature = bad_point;
return;
}
Eigen::Vector3d gradient_x (0, 0, 0);
Eigen::Vector3d gradient_y (0, 0, 0);
sumArea<typename IntegralImage2D<float, 3>::ElementType>(start_x, start_y, end_x, end_y, width, height, boost::bind(&IntegralImage2D<float, 3>::getFirstOrderSumSE, &integral_image_DX_, _1, _2, _3, _4), gradient_x);
sumArea<typename IntegralImage2D<float, 3>::ElementType>(start_x, start_y, end_x, end_y, width, height, boost::bind(&IntegralImage2D<float, 3>::getFirstOrderSumSE, &integral_image_DY_, _1, _2, _3, _4), gradient_y);
Eigen::Vector3d normal_vector = gradient_y.cross (gradient_x);
double normal_length = normal_vector.squaredNorm ();
if (normal_length == 0.0f)
{
normal.getNormalVector3fMap ().setConstant (bad_point);
normal.curvature = bad_point;
return;
}
normal_vector /= sqrt (normal_length);
float nx = static_cast<float> (normal_vector [0]);
float ny = static_cast<float> (normal_vector [1]);
float nz = static_cast<float> (normal_vector [2]);
flipNormalTowardsViewpoint (input_->points[point_index], vpx_, vpy_, vpz_, nx, ny, nz);
normal.normal_x = nx;
normal.normal_y = ny;
normal.normal_z = nz;
normal.curvature = bad_point;
return;
}
// ======================================================
else if (normal_estimation_method_ == AVERAGE_DEPTH_CHANGE)
{
if (!init_depth_change_)
initAverageDepthChangeMethod ();
int point_index_L_x = pos_x - rect_width_4_ - 1;
int point_index_L_y = pos_y;
int point_index_R_x = pos_x + rect_width_4_ + 1;
int point_index_R_y = pos_y;
int point_index_U_x = pos_x - 1;
int point_index_U_y = pos_y - rect_height_4_;
int point_index_D_x = pos_x + 1;
int point_index_D_y = pos_y + rect_height_4_;
if (point_index_L_x < 0)
point_index_L_x = -point_index_L_x;
if (point_index_U_x < 0)
point_index_U_x = -point_index_U_x;
if (point_index_U_y < 0)
point_index_U_y = -point_index_U_y;
if (point_index_R_x >= width)
point_index_R_x = width-(point_index_R_x-(width-1));
if (point_index_D_x >= width)
point_index_D_x = width-(point_index_D_x-(width-1));
if (point_index_D_y >= height)
point_index_D_y = height-(point_index_D_y-(height-1));
const int start_x_L = pos_x - rect_width_2_;
const int start_y_L = pos_y - rect_height_4_;
const int end_x_L = start_x_L + rect_width_2_;
const int end_y_L = start_y_L + rect_height_2_;
const int start_x_R = pos_x + 1;
const int start_y_R = pos_y - rect_height_4_;
const int end_x_R = start_x_R + rect_width_2_;
const int end_y_R = start_y_R + rect_height_2_;
const int start_x_U = pos_x - rect_width_4_;
const int start_y_U = pos_y - rect_height_2_;
const int end_x_U = start_x_U + rect_width_2_;
const int end_y_U = start_y_U + rect_height_2_;
const int start_x_D = pos_x - rect_width_4_;
const int start_y_D = pos_y + 1;
const int end_x_D = start_x_D + rect_width_2_;
const int end_y_D = start_y_D + rect_height_2_;
unsigned count_L_z = 0;
unsigned count_R_z = 0;
unsigned count_U_z = 0;
unsigned count_D_z = 0;
sumArea<unsigned>(start_x_L, start_y_L, end_x_L, end_y_L, width, height, boost::bind(&IntegralImage2D<float, 1>::getFiniteElementsCountSE, &integral_image_depth_, _1, _2, _3, _4), count_L_z);
sumArea<unsigned>(start_x_R, start_y_R, end_x_R, end_y_R, width, height, boost::bind(&IntegralImage2D<float, 1>::getFiniteElementsCountSE, &integral_image_depth_, _1, _2, _3, _4), count_R_z);
sumArea<unsigned>(start_x_U, start_y_U, end_x_U, end_y_U, width, height, boost::bind(&IntegralImage2D<float, 1>::getFiniteElementsCountSE, &integral_image_depth_, _1, _2, _3, _4), count_U_z);
sumArea<unsigned>(start_x_D, start_y_D, end_x_D, end_y_D, width, height, boost::bind(&IntegralImage2D<float, 1>::getFiniteElementsCountSE, &integral_image_depth_, _1, _2, _3, _4), count_D_z);
if (count_L_z == 0 || count_R_z == 0 || count_U_z == 0 || count_D_z == 0)
{
normal.normal_x = normal.normal_y = normal.normal_z = normal.curvature = bad_point;
return;
}
float mean_L_z = 0;
float mean_R_z = 0;
float mean_U_z = 0;
float mean_D_z = 0;
sumArea<float>(start_x_L, start_y_L, end_x_L, end_y_L, width, height, boost::bind(&IntegralImage2D<float, 1>::getFirstOrderSumSE, &integral_image_depth_, _1, _2, _3, _4), mean_L_z);
sumArea<float>(start_x_R, start_y_R, end_x_R, end_y_R, width, height, boost::bind(&IntegralImage2D<float, 1>::getFirstOrderSumSE, &integral_image_depth_, _1, _2, _3, _4), mean_R_z);
sumArea<float>(start_x_U, start_y_U, end_x_U, end_y_U, width, height, boost::bind(&IntegralImage2D<float, 1>::getFirstOrderSumSE, &integral_image_depth_, _1, _2, _3, _4), mean_U_z);
sumArea<float>(start_x_D, start_y_D, end_x_D, end_y_D, width, height, boost::bind(&IntegralImage2D<float, 1>::getFirstOrderSumSE, &integral_image_depth_, _1, _2, _3, _4), mean_D_z);
mean_L_z /= float (count_L_z);
mean_R_z /= float (count_R_z);
mean_U_z /= float (count_U_z);
mean_D_z /= float (count_D_z);
PointInT pointL = input_->points[point_index_L_y*width + point_index_L_x];
PointInT pointR = input_->points[point_index_R_y*width + point_index_R_x];
PointInT pointU = input_->points[point_index_U_y*width + point_index_U_x];
PointInT pointD = input_->points[point_index_D_y*width + point_index_D_x];
const float mean_x_z = mean_R_z - mean_L_z;
const float mean_y_z = mean_D_z - mean_U_z;
const float mean_x_x = pointR.x - pointL.x;
const float mean_x_y = pointR.y - pointL.y;
const float mean_y_x = pointD.x - pointU.x;
const float mean_y_y = pointD.y - pointU.y;
float normal_x = mean_x_y * mean_y_z - mean_x_z * mean_y_y;
float normal_y = mean_x_z * mean_y_x - mean_x_x * mean_y_z;
float normal_z = mean_x_x * mean_y_y - mean_x_y * mean_y_x;
const float normal_length = (normal_x * normal_x + normal_y * normal_y + normal_z * normal_z);
if (normal_length == 0.0f)
{
normal.getNormalVector3fMap ().setConstant (bad_point);
normal.curvature = bad_point;
return;
}
flipNormalTowardsViewpoint (input_->points[point_index], vpx_, vpy_, vpz_, normal_x, normal_y, normal_z);
const float scale = 1.0f / sqrtf (normal_length);
normal.normal_x = normal_x * scale;
normal.normal_y = normal_y * scale;
normal.normal_z = normal_z * scale;
normal.curvature = bad_point;
return;
}
// ========================================================
else if (normal_estimation_method_ == SIMPLE_3D_GRADIENT)
{
PCL_THROW_EXCEPTION (PCLException, "BORDER_POLICY_MIRROR not supported for normal estimation method SIMPLE_3D_GRADIENT");
}
normal.getNormalVector3fMap ().setConstant (bad_point);
normal.curvature = bad_point;
return;
}
void ellipsetrack(avi_t *video, double *xc0, double *yc0, int Nc, int R, int Np, int Nf) {
/*
% ELLIPSETRACK tracks cells in the movie specified by 'video', at
% locations 'xc0'/'yc0' with radii R using an ellipse with Np discrete
% points, starting at frame number one and stopping at frame number 'Nf'.
%
% INPUTS:
% video.......pointer to avi video object
% xc0,yc0.....initial center location (Nc entries)
% Nc..........number of cells
% R...........initial radius
% Np..........number of snaxels points per snake
% Nf..........number of frames in which to track
%
% Matlab code written by: DREW GILLIAM (based on code by GANG DONG /
% NILANJAN RAY)
% Ported to C by: MICHAEL BOYER
*/
// Compute angle parameter
double *t = (double *) malloc(sizeof(double) * Np);
double increment = (2.0 * PI) / (double) Np;
int i, j;
for (i = 0; i < Np; i++) {
t[i] = increment * (double) i ;
}
// Allocate space for a snake for each cell in each frame
double **xc = alloc_2d_double(Nc, Nf + 1);
double **yc = alloc_2d_double(Nc, Nf + 1);
double ***r = alloc_3d_double(Nc, Np, Nf + 1);
double ***x = alloc_3d_double(Nc, Np, Nf + 1);
double ***y = alloc_3d_double(Nc, Np, Nf + 1);
// Save the first snake for each cell
for (i = 0; i < Nc; i++) {
xc[i][0] = xc0[i];
yc[i][0] = yc0[i];
for (j = 0; j < Np; j++) {
r[i][j][0] = (double) R;
}
}
// Generate ellipse points for each cell
for (i = 0; i < Nc; i++) {
for (j = 0; j < Np; j++) {
x[i][j][0] = xc[i][0] + (r[i][j][0] * cos(t[j]));
y[i][j][0] = yc[i][0] + (r[i][j][0] * sin(t[j]));
}
}
// Allocate arrays so we can break up the per-cell for loop below
double *xci = (double *) malloc(sizeof(double) * Nc);
double *yci = (double *) malloc(sizeof(double) * Nc);
double **ri = alloc_2d_double(Nc, Np);
double *ycavg = (double *) malloc(sizeof(double) * Nc);
int *u1 = (int *) malloc(sizeof(int) * Nc);
int *u2 = (int *) malloc(sizeof(int) * Nc);
int *v1 = (int *) malloc(sizeof(int) * Nc);
int *v2 = (int *) malloc(sizeof(int) * Nc);
MAT **Isub = (MAT **) malloc(sizeof(MAT *) * Nc);
MAT **Ix = (MAT **) malloc(sizeof(MAT *) * Nc);
MAT **Iy = (MAT **) malloc(sizeof(MAT *) * Nc);
MAT **IE = (MAT **) malloc(sizeof(MAT *) * Nc);
// Keep track of the total time spent on computing
// the MGVF matrix and evolving the snakes
long long MGVF_time = 0;
long long snake_time = 0;
// Process each frame sequentially
int frame_num;
for (frame_num = 1; frame_num <= Nf; frame_num++) {
printf("\rProcessing frame %d / %d", frame_num, Nf);
fflush(stdout);
// Get the current video frame and its dimensions
MAT *I = get_frame(video, frame_num, 0, 1);
int Ih = I->m;
int Iw = I->n;
// Initialize the current positions to be equal to the previous positions
for (i = 0; i < Nc; i++) {
xc[i][frame_num] = xc[i][frame_num - 1];
yc[i][frame_num] = yc[i][frame_num - 1];
for (j = 0; j < Np; j++) {
r[i][j][frame_num] = r[i][j][frame_num - 1];
}
}
// Sequentially extract the subimage near each cell
int cell_num;
for (cell_num = 0; cell_num < Nc; cell_num++) {
// Make copies of the current cell's location
xci[cell_num] = xc[cell_num][frame_num];
yci[cell_num] = yc[cell_num][frame_num];
for (j = 0; j < Np; j++) {
ri[cell_num][j] = r[cell_num][j][frame_num];
}
// Add up the last ten y values for this cell
// (or fewer if there are not yet ten previous frames)
ycavg[cell_num] = 0.0;
for (i = (frame_num > 10 ? frame_num - 10 : 0); i < frame_num; i++) {
ycavg[cell_num] += yc[cell_num][i];
}
// Compute the average of the last ten values
// (this represents the expected location of the cell)
ycavg[cell_num] = ycavg[cell_num] / (double) (frame_num > 10 ? 10 : frame_num);
// Determine the range of the subimage surrounding the current position
u1[cell_num] = max(xci[cell_num] - 4.0 * R + 0.5, 0 );
u2[cell_num] = min(xci[cell_num] + 4.0 * R + 0.5, Iw - 1);
v1[cell_num] = max(yci[cell_num] - 2.0 * R + 1.5, 0 );
v2[cell_num] = min(yci[cell_num] + 2.0 * R + 1.5, Ih - 1);
// Extract the subimage
Isub[cell_num] = m_get(v2[cell_num] - v1[cell_num] + 1, u2[cell_num] - u1[cell_num] + 1);
for (i = v1[cell_num]; i <= v2[cell_num]; i++) {
for (j = u1[cell_num]; j <= u2[cell_num]; j++) {
m_set_val(Isub[cell_num], i - v1[cell_num], j - u1[cell_num], m_get_val(I, i, j));
}
}
// Compute the subimage gradient magnitude
Ix[cell_num] = gradient_x(Isub[cell_num]);
Iy[cell_num] = gradient_y(Isub[cell_num]);
IE[cell_num] = m_get(Isub[cell_num]->m, Isub[cell_num]->n);
for (i = 0; i < Isub[cell_num]->m; i++) {
for (j = 0; j < Isub[cell_num]->n; j++) {
double temp_x = m_get_val(Ix[cell_num], i, j);
double temp_y = m_get_val(Iy[cell_num], i, j);
m_set_val(IE[cell_num], i, j, sqrt((temp_x * temp_x) + (temp_y * temp_y)));
}
}
}
// Compute the motion gradient vector flow (MGVF) edgemaps for all cells concurrently
long long MGVF_start_time = get_time();
MAT **IMGVF = MGVF(IE, 1, 1, Nc);
MGVF_time += get_time() - MGVF_start_time;
// Sequentially determine the new location of each cell
for (cell_num = 0; cell_num < Nc; cell_num++) {
// Determine the position of the cell in the subimage
xci[cell_num] = xci[cell_num] - (double) u1[cell_num];
yci[cell_num] = yci[cell_num] - (double) (v1[cell_num] - 1);
ycavg[cell_num] = ycavg[cell_num] - (double) (v1[cell_num] - 1);
// Evolve the snake
long long snake_start_time = get_time();
ellipseevolve(IMGVF[cell_num], &(xci[cell_num]), &(yci[cell_num]), ri[cell_num], t, Np, (double) R, ycavg[cell_num]);
snake_time += get_time() - snake_start_time;
// Compute the cell's new position in the full image
xci[cell_num] = xci[cell_num] + u1[cell_num];
yci[cell_num] = yci[cell_num] + (v1[cell_num] - 1);
// Store the new location of the cell and the snake
xc[cell_num][frame_num] = xci[cell_num];
yc[cell_num][frame_num] = yci[cell_num];
for (j = 0; j < Np; j++) {
r[cell_num][j][frame_num] = 0;
r[cell_num][j][frame_num] = ri[cell_num][j];
x[cell_num][j][frame_num] = xc[cell_num][frame_num] + (ri[cell_num][j] * cos(t[j]));
y[cell_num][j][frame_num] = yc[cell_num][frame_num] + (ri[cell_num][j] * sin(t[j]));
}
// Output the updated center of each cell
// printf("\n%d,%f,%f", cell_num, xci[cell_num], yci[cell_num]);
// Free temporary memory
m_free(Isub[cell_num]);
m_free(Ix[cell_num]);
m_free(Iy[cell_num]);
m_free(IE[cell_num]);
m_free(IMGVF[cell_num]);
}
#ifdef OUTPUT
if (frame_num == Nf)
{
FILE * pFile;
pFile = fopen ("result.txt","w+");
for (cell_num = 0; cell_num < Nc; cell_num++)
fprintf(pFile,"\n%d,%f,%f", cell_num, xci[cell_num], yci[cell_num]);
fclose (pFile);
}
#endif
free(IMGVF);
// Output a new line to visually distinguish the output from different frames
//printf("\n");
}
// Free temporary memory
free_2d_double(xc);
free_2d_double(yc);
free_3d_double(r);
free_3d_double(x);
free_3d_double(y);
free(t);
free(xci);
free(yci);
free_2d_double(ri);
free(ycavg);
free(u1);
free(u2);
free(v1);
free(v2);
free(Isub);
free(Ix);
free(Iy);
free(IE);
// Report average processing time per frame
printf("\n\nTracking runtime (average per frame):\n");
printf("------------------------------------\n");
printf("MGVF computation: %.5f seconds\n", ((float) (MGVF_time)) / (float) (1000*1000*Nf));
printf(" Snake evolution: %.5f seconds\n", ((float) (snake_time)) / (float) (1000*1000*Nf));
printf("CAUTION: cpu_offset: %d time: %lf mseconds\n", cpu_offset, ((float) (MGVF_time)) / (float) (1000*1000*Nf)*1000);
}
ImageRAII canny( IplImage * image, std::pair< int, int > thresh, double sigma )
{
const char * WINDOW_NAME = "Basic Canny Edge Detector";
ImageRAII grayscale( cvCreateImage( cvGetSize( image ), image->depth, 1 ) );
ImageRAII destination( cvCreateImage( cvGetSize( image ), image->depth, grayscale.image->nChannels ) );
ImageRAII gaussian( cvCreateImage( cvGetSize( image ), image->depth, grayscale.image->nChannels ) );
ImageRAII gradient_x( cvCreateImage( cvGetSize( image ), image->depth, grayscale.image->nChannels ) );
ImageRAII gradient_y( cvCreateImage( cvGetSize( image ), image->depth, grayscale.image->nChannels ) );
ImageRAII gradient( cvCreateImage( cvGetSize( image ), image->depth, grayscale.image->nChannels ) );
ImageRAII orientation( cvCreateImage( cvGetSize( image ), image->depth, grayscale.image->nChannels ) );
// convert image to grayscale
cvCvtColor( image, grayscale.image, CV_BGR2GRAY );
// gaussian smoothing
cvSmooth( grayscale.image, gaussian.image, CV_GAUSSIAN, GAUSSIAN_X, GAUSSIAN_Y, sigma );
// find edge strength
cvSobel( gaussian.image, gradient_x.image, 1, 0, 3 );
cvSobel( gaussian.image, gradient_y.image, 0, 1, 3 );
// find edge orientation
CvSize image_size = cvGetSize( gaussian.image );
for( int i = 0; i < image_size.width; i++ )
{
for( int j = 0; j < image_size.height; j++ )
{
double x = cvGet2D( gradient_x.image, j, i ).val[0];
double y = cvGet2D( gradient_y.image, j, i ).val[0];
float angle;
if( x == 0 )
{
if( y == 0 )
angle = 0;
else
angle = 90;
}
else
angle = cvFastArctan( y, x );
CvScalar g;
CvScalar a;
g.val[0] = cvSqrt( pow( x, 2 ) + pow( y, 2 ) );
a.val[0] = find_angle( angle );
cvSet2D( destination.image, j, i, g );
cvSet2D( orientation.image, j, i, a );
}
}
ImageRAII suppressed_image = nonMaxSup( destination.image, orientation.image );
ImageRAII hysteresis_image = hysteresis( suppressed_image.image, orientation.image, thresh );
cvNamedWindow( WINDOW_NAME );
cvShowImage( WINDOW_NAME, destination.image );
cvMoveWindow( WINDOW_NAME, image_size.width, 0 );
return hysteresis_image;
}
