Object counting (Pléiades, QB, Ikonos)

Specification

• Count compact homogeneous objects on optical high resolution images
  • Examples of objects: buildings, isolated trees, vehicles, etc.
• Inputs for the algorithm
  • The image to process
  • A set of ROIs selected by an operator on examples of objects to count
• Outputs of the algorithm
  • The number of objects found
  • A list of polygons which fit the detected objects
• Algorithm description
  • A binary image resulting from a supervised classification is generated from the input image and the training ROIs
    • A simplified version of this can be a spectral angle with respect to the mean of all ROIs (or the min SAM to the mean of each ROI) and an automatic threshold (otsu)
  • The input image is clustered using a MeanShift filter. The boundaries are also generated from the MeanShift.
    • The classified image and the mean shift results are fused in order to generate a set of disconnected regions belonging to the detected class
    • The edges of the regions are converted to polygons
    • The polygons can be refined in order to better fit the objects

Use case

• • Open an optical HR image
  • Select examples of objects to count
  • Trigger the counting
  • Display on a text field the resulting number of objects found
  • Overlay polygons on top of the detected objects
    • Put a text label on the polygon with the object id

Outputs

• • A raster image with a unique label per object (unsigned short)
  • A vector data file with one polygon per object

States of the application and transitions