// Initialize the filter using some model
void pf_init_model(pf_t *pf, pf_init_model_fn_t init_fn, void *init_data)
{
int i;
pf_sample_set_t *set;
pf_sample_t *sample;
set = pf->sets + pf->current_set;
// Create the kd tree for adaptive sampling
pf_kdtree_clear(set->kdtree);
set->sample_count = pf->max_samples;
// Compute the new sample poses
for (i = 0; i < set->sample_count; i++)
{
sample = set->samples + i;
sample->weight = 1.0 / pf->max_samples;
sample->pose = (*init_fn) (init_data);
// Add sample to histogram
pf_kdtree_insert(set->kdtree, sample->pose, sample->weight);
}
// Re-compute cluster statistics
pf_cluster_stats(pf, set);
return;
}
// Initialize the filter using a guassian
void pf_init(pf_t *pf, pf_vector_t mean, pf_matrix_t cov)
{
int i;
pf_sample_set_t *set;
pf_sample_t *sample;
pf_pdf_gaussian_t *pdf;
set = pf->sets + pf->current_set;
// Create the kd tree for adaptive sampling
pf_kdtree_clear(set->kdtree);
set->sample_count = pf->max_samples;
pdf = pf_pdf_gaussian_alloc(mean, cov);
// Compute the new sample poses
for (i = 0; i < set->sample_count; i++)
{
sample = set->samples + i;
sample->weight = 1.0 / pf->max_samples;
sample->pose = pf_pdf_gaussian_sample(pdf);
// Add sample to histogram
pf_kdtree_insert(set->kdtree, sample->pose, sample->weight);
}
pf_pdf_gaussian_free(pdf);
// Re-compute cluster statistics
pf_cluster_stats(pf, set);
return;
}
void pf_update_current_cluster_stats(pf_t *pf){
pf_sample_set_t *c_set = pf->sets + pf->current_set;
int i;
pf_sample_t *sample;
//poses have changed - clear the kd tree and reinsert the nodes
pf_kdtree_clear(c_set->kdtree);
for (i = 0; i < c_set->sample_count; i++){
sample = c_set->samples + i;
pf_kdtree_insert(c_set->kdtree, sample->pose, sample->weight);
}
pf_cluster_stats(pf, c_set);
}
// Initialize the filter using some model
void pf_init_model(pf_t *pf, pf_init_model_fn_t init_fn, void *init_data)
{
int i;
pf_sample_set_t *set;
pf_sample_t *sample;
set = pf->sets + pf->current_set;
// Create the kd tree for adaptive sampling
pf_kdtree_clear(set->kdtree);
set->sample_count = pf->max_samples;
// Compute the new sample poses
for (i = 0; i < set->sample_count; i++)
{
sample = set->samples + i;
sample->weight = 1.0 / pf->max_samples;
sample->pose = (*init_fn) (init_data);
// Add sample to histogram
pf_kdtree_insert(set->kdtree, sample->pose, sample->weight);
}
pf->w_slow = pf->w_fast = 0.0;
// Re-compute cluster statistics
pf_cluster_stats(pf, set);
//set converged to 0
pf_init_converged(pf);
//allow converged to be 1 if we are actually converged from the start
pf_update_converged(pf);
return;
}
// Resample the distribution
void pf_update_resample(pf_t *pf)
{
int i;
double total;
//double *randlist;
pf_sample_set_t *set_a, *set_b;
pf_sample_t *sample_a, *sample_b;
//pf_pdf_discrete_t *pdf;
double r,c,U;
int m;
double count_inv;
set_a = pf->sets + pf->current_set;
set_b = pf->sets + (pf->current_set + 1) % 2;
// Create the discrete distribution to sample from
/*
total = 0;
randlist = calloc(set_a->sample_count, sizeof(double));
for (i = 0; i < set_a->sample_count; i++)
{
total += set_a->samples[i].weight;
randlist[i] = set_a->samples[i].weight;
}
*/
//printf("resample total %f\n", total);
// Initialize the random number generator
//pdf = pf_pdf_discrete_alloc(set_a->sample_count, randlist);
// Create the kd tree for adaptive sampling
pf_kdtree_clear(set_b->kdtree);
// Draw samples from set a to create set b.
total = 0;
set_b->sample_count = 0;
// Low-variance resampler, taken from Probabilistic Robotics, p110
count_inv = 1.0/set_a->sample_count;
#if defined (WIN32)
// TODO: this isn't quite the same behaviour: drand48 returns uniformly-distributed values
r = ((double) rand() / (double) RAND_MAX) * count_inv;
#else
r = drand48() * count_inv;
#endif
c = set_a->samples[0].weight;
i = 0;
m = 0;
while(set_b->sample_count < pf->max_samples)
{
U = r + m * count_inv;
while(U>c)
{
i++;
// Handle wrap-around by resetting counters and picking a new random
// number
if(i >= set_a->sample_count)
{
#if defined (WIN32)
// TODO: this isn't quite the same behaviour: drand48 returns uniformly-distributed values
r = ((double) rand() / (double) RAND_MAX) * count_inv;
#else
r = drand48() * count_inv;
#endif
c = set_a->samples[0].weight;
i = 0;
m = 0;
U = r + m * count_inv;
continue;
}
c += set_a->samples[i].weight;
}
//i = pf_pdf_discrete_sample(pdf);
sample_a = set_a->samples + i;
//printf("%d %f\n", i, sample_a->weight);
assert(sample_a->weight > 0);
// Add sample to list
sample_b = set_b->samples + set_b->sample_count++;
sample_b->pose = sample_a->pose;
sample_b->weight = 1.0;
total += sample_b->weight;
// Add sample to histogram
pf_kdtree_insert(set_b->kdtree, sample_b->pose, sample_b->weight);
//fprintf(stderr, "resample %d %d %d\n", set_b->sample_count, set_b->kdtree->leaf_count,
//pf_resample_limit(pf, set_b->kdtree->leaf_count));
// See if we have enough samples yet
if (set_b->sample_count > pf_resample_limit(pf, set_b->kdtree->leaf_count))
break;
m++;
}
//fprintf(stderr, "\n\n");
//pf_pdf_discrete_free(pdf);
//free(randlist);
// Normalize weights
for (i = 0; i < set_b->sample_count; i++)
{
sample_b = set_b->samples + i;
sample_b->weight /= total;
}
// Re-compute cluster statistics
pf_cluster_stats(pf, set_b);
// Use the newly created sample set
pf->current_set = (pf->current_set + 1) % 2;
return;
}
// Resample the distribution
void pf_update_resample(pf_t *pf)
{
int i;
double total;
pf_sample_set_t *set_a, *set_b;
pf_sample_t *sample_a, *sample_b;
//double r,c,U;
//int m;
//double count_inv;
double* c;
double w_diff;
set_a = pf->sets + pf->current_set;
set_b = pf->sets + (pf->current_set + 1) % 2;
// Build up cumulative probability table for resampling.
// TODO: Replace this with a more efficient procedure
// (e.g., http://www.network-theory.co.uk/docs/gslref/GeneralDiscreteDistributions.html)
c = (double*)malloc(sizeof(double)*(set_a->sample_count+1));
c[0] = 0.0;
for(i=0;i<set_a->sample_count;i++)
c[i+1] = c[i]+set_a->samples[i].weight;
// Create the kd tree for adaptive sampling
pf_kdtree_clear(set_b->kdtree);
// Draw samples from set a to create set b.
total = 0;
set_b->sample_count = 0;
w_diff = 1.0 - pf->w_fast / pf->w_slow;
if(w_diff < 0.0)
w_diff = 0.0;
//printf("w_diff: %9.6f\n", w_diff);
// Can't (easily) combine low-variance sampler with KLD adaptive
// sampling, so we'll take the more traditional route.
/*
// Low-variance resampler, taken from Probabilistic Robotics, p110
count_inv = 1.0/set_a->sample_count;
r = drand48() * count_inv;
c = set_a->samples[0].weight;
i = 0;
m = 0;
*/
while(set_b->sample_count < pf->max_samples)
{
sample_b = set_b->samples + set_b->sample_count++;
if(drand48() < w_diff)
sample_b->pose = (pf->random_pose_fn)(pf->random_pose_data);
else
{
// Can't (easily) combine low-variance sampler with KLD adaptive
// sampling, so we'll take the more traditional route.
/*
// Low-variance resampler, taken from Probabilistic Robotics, p110
U = r + m * count_inv;
while(U>c)
{
i++;
// Handle wrap-around by resetting counters and picking a new random
// number
if(i >= set_a->sample_count)
{
r = drand48() * count_inv;
c = set_a->samples[0].weight;
i = 0;
m = 0;
U = r + m * count_inv;
continue;
}
c += set_a->samples[i].weight;
}
m++;
*/
// Naive discrete event sampler
double r;
r = drand48();
for(i=0;i<set_a->sample_count;i++)
{
if((c[i] <= r) && (r < c[i+1]))
break;
}
assert(i<set_a->sample_count);
sample_a = set_a->samples + i;
assert(sample_a->weight > 0);
//景恆ZACH add 1205 START
if(drand48()<0.3){
pf_sample_t *sample_tmp = sample_a;
sample_tmp->pose.v[0] = sample_a->pose.v[0] + 0.01*(2*drand48()-1);
sample_tmp->pose.v[1] = sample_a->pose.v[1] + 0.01*(2*drand48()-1);
sample_tmp->pose.v[2] = sample_a->pose.v[2] + 0.01*(2*drand48()-1);
// Add sample to list
sample_b->pose = sample_tmp->pose;
}
else sample_b->pose = sample_a->pose;
//景恆ZACH add 1205 END
// Add sample to list
//sample_b->pose = sample_a->pose;
}
sample_b->weight = 1.0;
total += sample_b->weight;
// Add sample to histogram
pf_kdtree_insert(set_b->kdtree, sample_b->pose, sample_b->weight);
// See if we have enough samples yet
if (set_b->sample_count > pf_resample_limit(pf, set_b->kdtree->leaf_count))
break;
}
// Reset averages, to avoid spiraling off into complete randomness.
if(w_diff > 0.0)
pf->w_slow = pf->w_fast = 0.0;
//fprintf(stderr, "\n\n");
// Normalize weights
for (i = 0; i < set_b->sample_count; i++)
{
sample_b = set_b->samples + i;
sample_b->weight /= total;
}
// Re-compute cluster statistics
pf_cluster_stats(pf, set_b);
// Use the newly created sample set
pf->current_set = (pf->current_set + 1) % 2;
pf_update_converged(pf);
free(c);
return;
}