In KarlEdwards/fixed-camera-surveillance: Fixed Camera Surveillance

knitr::opts_chunk$set( cache = TRUE )

knitr::opts_chunk$set( collapse = TRUE, comment = "#>" )
options( tibble.print_min = 4L, tibble.print_max = 4L )
library( dplyr )

# --- Configuration
img_path <- '/Users/Karl/Dropbox/Projects/Video-Captioning/pipelines/box2/'
image_filenames <- dir( path=img_path )

Introduction

The Fixed Camera Surveillance (FCS) Package contains functions to cluster images in various ways, then display similar images together and evaluate which feature extraction and clustering methods work best.

Common tasks for Fixed Camera Surveillance include:

• Background Subtraction

• Foreground Segmentation

• Object Identification

---

Background Subtraction

The general approach is:

1. Extract features from the images, for example, by using HoGs or hashes.

2. Cluster the images, based on their features

3. Display images for each cluster in its own collage.

Outline

A. Feature extraction

1. Hashing

   a. average

   b. phash

   c. dhash

   d. ( invariant hash compares two images )

2. HoGs

   a. cells

   b. bins

B. Clustering 1. K-means

1. Linear SVM

C. Display

---

Examples

Feature Extraction

Hashes

images_found <- as.logical( length( image_filenames ))
if( !images_found ) cat( sprintf( 'No image files found\n'))
if( images_found ){
  hashes <- hash_batch( img_path )
  all_hashes <- purrr::map( hashes, 2 )
}

if( images_found ){
  hashes$names
  all_hashes[['values']][['files']][1:4]
  all_hashes[['values']][['hashes']][1:4]
}

HoGs

if( images_found ){
  hogs <- hog_batch( IMAGE_PATH = img_path )
  all_hogs <- purrr::map( hogs, 2 )
}

if( images_found ){
  cat( sprintf( 'Results:\n' ))
  print( hogs$names )
  print( all_hogs[['values']][['files']][1:4] )
  print( all_hogs[['values']][['hogs']][1:4] )
}

Clustering Examples

#   3 cells
#   | 6 bins
#   | |
hogs3c6b <- OpenImageR::HOG_apply( img_path, cells = 3, orientations = 6 )

get_clusters <- function( data, filenames, algorithm = 'adaptive', N = 2 ){
  clusters <- exemplar_image_names <- NULL

  if( algorithm == 'adaptive' ){
    # adaptive clustering
    clusters <- ADPclust::adpclust( data )
    seq( clusters$nclust ) %>%
      purrr::map_int( ~clusters$centers[ .x ] ) -> indexes
  }

  if( algorithm == 'kmeans' ){
    # K-means clustering
    clusters <- kmeans( data, centers = N )
    seq( N ) %>%
      purrr::map_int( ~match( .x, clusters$cluster )) -> indexes
  }

  exemplar_image_names <- if( any( grepl( 'indexes', ls() )) ) purrr::map_chr( indexes, ~filenames[ .x ] )

  try(
    if( length( clusters ) > 0 ){
      # --- Get the name of the variable used to identify clusters
      cluster_var <- grep( 'cluster*', names( clusters ), value=TRUE )

      # --- Now use the cluster variable to get cluster IDs
      clusters <- clusters[[ cluster_var ]]
    }
  )
  list( IDs=clusters, exemplars=exemplar_image_names )
}

clustered <- get_clusters(
    data = hogs3c6b[[ 2 ]]
  , filenames = image_filenames
  , algorithm = 'kmeans'
  , N = 9
)
clustered$IDs
clustered$exemplars

KarlEdwards/fixed-camera-surveillance documentation built on May 17, 2019, 11:06 a.m.