R/fun_ChangeDetection.r

In georoen/wuepix: Using Computer Vision to Count Pedestrians

#' @author Jeroen Staab
# getImage
getImage <- function(filename, extend = NULL, plot = FALSE) {
  ## load JPEGs and preprocess them
  #' filename Path to first image
  #' extend Optional extend to crop image, numeric vector.
  #' plot Optional plot loaded image, boolean.
  #' @import jpeg
  #' @export getImage
  file <- jpeg::readJPEG(filename)
  if (!is.null(extend)) {
    file <- file[extend[1]:extend[2], extend[3]:extend[4], ]
  }
  if (plot) {
    JPEG_plot(file, filename)
  }
  return(file)
}

CD_single <- function(file.now, file.old, Min = 0.2, Max = 1, predictions = NULL,
                      extend = NULL, plot = FALSE, method = "diff") {
  #' @title Change Detection
  #' @title  Detect changes between two images using image differencing
  #'
  #' @param now Path to first image
  #' @param old Path to second image
  #' @param Min Threshold for positive classification
  #' @param Max Threshold for positive classification
  #' @param predictions dir path to where to store prediction images
  #' @param extend DEPECATED!
  #' Used to crop images. Has been moved to a seperate preprocess step.
  #' @param method Select change detection method.
  #' "ratio" Image Rationing. "diff" Image Differencing, absolute changes in
  #' both directions. "diff+" Image Differencing, positive changes only.
  #'
  #' @return Classification result. Here work is in progess...
  #' @export CD_single


  # Load Images
  now <- getImage(file.now, extend, plot)
  old <- getImage(file.old, extend, plot)

  # Intensety Normalization. As in RADKE 2005 eq.1
  normalizeIntensety <- function(now, old) {
    for (k in dim(old)[3]) {
      old[k] <- sd(now[k]) / sd(old[k]) * (old[k] - mean(old[k])) + mean(now[k])
    }
    old
  }
  old <- normalizeIntensety(now, old)

  # Define change detection methods
  method_diff_pos <- function(now, old, Min, Max) {
    # Image difference. Positives Only
    dif <- now
    dif[] <- now[] - old[]
    # Keep positve changes only (now > old). Darker Parts will be classified in next iteration.
    hum <- dif
    hum[which(dif < 0)] <- 0
    # Classify significant changes with tresholds.
    classified <- hum[, , 1] < Max & hum[, , 1] > Min & hum[, , 2] < Max & hum[, , 2] > Min & hum[, , 3] < Max & hum[, , 3] > Min
    hum[classified] <- 1
    hum[!classified] <- 0
    hum[, , 1]
  }

  method_diff <- function(now, old, Min, Max) {
    # Image Difference. Absolute changes (both)
    dif <- now
    dif[] <- now[] - old[]
    # Absolute Values cuz change.dif Direction doesn't matter. Due to different ligth exposures.
    hum <- abs(dif)
    # Classify significant changes with tresholds.
    classified <- hum[, , 1] < Max & hum[, , 1] > Min & hum[, , 2] < Max & hum[, , 2] > Min & hum[, , 3] < Max & hum[, , 3] > Min
    hum[classified] <- 1
    hum[!classified] <- 0
    hum[, , 1]
  }

  method_ratio <- function(now, old, Min, Max) {
    # Image Ratio. Absolute changes (both)
    dif <- now
    dif[] <- log((now[] + 0.0001) / (old[] + 0.0001)) # As in BERBERGOLU 2008 p.48
    # dif[] <- atan(now[]/old[]) - pi/4  # As in ILSEVER 2012 p.11
    # Absolute Values cuz change.dif Direction doesn't matter. Due to different ligth exposures.
    hum <- abs(dif)
    # Classify significant changes with tresholds.
    classified <- hum[, , 1] < Max & hum[, , 1] > Min & hum[, , 2] < Max & hum[, , 2] > Min & hum[, , 3] < Max & hum[, , 3] > Min
    hum[classified] <- 1
    hum[!classified] <- 0
    hum[, , 1]
  }

  method_t_test <- function(now, old, Min, Max) {
    # As in RADKE 2005 p. 299
    hum <- t.test(now[], old[])$p.value
    # hum <- hum > Min & hum < Max
    as.numeric(hum)
  }

  # Select change detion method
  method <- paste0("method_", method)
  if (!method %in% ls(environment())) {
    stop(cat("Undefined method selected. Implemented: ", lsf.str(environment())))
  }
  method <- get(method)

  hum <- method(now, old, Min, Max)

  if (!is.null(predictions)) {
    dir.create(predictions)
    jpeg::writeJPEG(hum, file.path(predictions, basename(file.now)))
  }

  return(sum(as.vector(hum)))
}


CD_list <- function(img.list, ...) {
  #' @title Change Detection
  #' @description Detect changes using image differencing for a list of images.
  #' Includes parallel processing.
  #'
  #' @param img.list file path to images.
  #' @param ... Arguments passed to CD_single().
  #'
  #' @return Classification result. Here work is in progess...
  #'
  #' @import doParallel
  #' @import foreach
  #' @export CD_list
  `%dopar%` <- foreach::`%dopar%`

  cores <- parallel::detectCores() - 1
  cl <- parallel::makeCluster(cores)
  doParallel::registerDoParallel(cl)
  old.stop <- length(img.list)
  new.stop <- old.stop - 1

  act.Data <- foreach::foreach(
    now = img.list[1:new.stop],
    old = img.list[2:old.stop],
    .combine = c
  ) %dopar%
    wuepix::CD_single(now, old, ...)

  parallel::stopCluster(cl)

  act.Data <- c(NA, act.Data)
  act.Data
}