wuepix: Using Computer Vision to Count Pedestrians

#' @author Jeroen Staab
hog_list <- function(img.list, winStride = 4, padding = 8,
                     Mscale = 1.05, resize = 1, predictions = NULL) {
  #' @title Detect pedestrians using HOGDescriptor
  #' @description Detect objects using HOG+SVM (implemented in OpenCV) in all Files/Images of 'path'
  #' @details Python and OpenCV have to be installed. Tested on Linux only.
  #' @details Further ideas:
  #' [A] Add more 'hog.detectMultiScale' parameters: winStride=(4, 4), padding=(8, 8), scale=1.05)
  #' [B] Save predictions.png to a folder
  #' @usage hog(img.folder)
  #'
  #' @param img.folder Path to (preprocessed) image archive
  #' @param resize Numeric factor resizing image in integrated pre-processing
  #' step. E.g. 2 will double the image extent. People should be 100 pixels high.
  #' @param winStride Window stride. It must be a multiple of block stride.
  #' @param padding Not implemented yet!
  #' @param Mscale Numeric. Allows multi-scale detection. Coefficient of the detection
  #' window increase.
  #' @param predictions dir path to where to store prediction images. Must end with "/".
  #'
  #' @return Numeric vector with number of detected persons.
  #' @export hog_list

  # Check predictions folder
  if (!is.null(predictions) && !dir.exists(predictions)) {
    dir.create(predictions)
  }

  # Path to python script
  hog.bin <- paste0(system.file(package = "wuepix"), "/exec/hogdescriptor.py")

  # Write img.list
  img.list.tmp <- tempfile()
  cat(img.list, sep = ",", file = img.list.tmp)

  # Classification
  cmd <- paste(
    "python", hog.bin, "-i", img.list.tmp, "-x", resize,
    "-w", winStride, "-p", padding, "-s", Mscale
  )

  if (!is.null(predictions)) {
    cmd <- paste(cmd, "-o", predictions)
  }
  out <- system(cmd, intern = TRUE)

  rtn <- as.numeric(out)
  invisible(rtn)
}


hog_install <- function() {
  #' @title How to install HOG-Descriptor?
  #' @description hog_list() depends on a functional OpenCV installation.
  #' This is how I installed it on the LSFE workstation (Linux). OS-specific
  #' @description OpenCV: \code{sudo apt install python-opencv}
  #' @description Package Manager: \code{sudo apt install python-pip}
  #' @description HOG Dependency: \code{pip install imutils}
  #' @description CUDA GPU: \code{sudo apt-get install nvidia-cuda-dev nvidia-cuda-toolkit nvidia-nsight}
  #' @seealso http://docs.opencv.org/trunk/df/d65/tutorial_table_of_content_introduction.html
  #' @usage ?hog_install()
  #' @export hog_install
  ?hog_install()
}