Terrain-based Bayesian Model

This module is a Bayesian model that uses terrain features and past human behavior data to predict the probability distribution of the likely places to find a missing person.


Internal Parameters:

  • Candidate distributions means and standard deviations
  • Burn (e.g. 1000)
  • Number of iterations (e.g. 100000)
  • Slope discretization theshold (e.g. s = 20 degrees)

Input Parameters:

  • Search region (defined by GPS coordinates)
  • Terrain features:
    • Topography (lake, hill, plane, etc.)
    • Vegetation density (Sparse, medium, dense)
    • Elevation (Uphill, flat, downhill)
    • Trail (On trail, off trail)
  • Last point seen (GPS coordinates converting to hex coordinates)
  • Intended destination (GPS coordinates converting to hex coordinates)
  • Prior Beliefs (transitional probabilities ←- Get from ParamMod)
  • Desired duration (time since missing person last seen in minute)

Output Parameters:

  • Probability distribution maps at each minute (Array of matrices)

Functions to be implemented

  • Use CUDA to speed up matrix multiplications.
    • Play with CUDA Net Sample Code
    • CUDA Test App
  • Interface to enter search region, last point seen, and (optionally) intended desitnation
  • Tool to download/encode topography info (create look up table)
  • Tool to label topography info (paint tool)
  • Tool to download/encode vegetation density (create look up table)
  • Tool to label trail info (paint tool)
  • Allow user to specify last point seen and intended destination by clicking a map.
  • Convert all GPS coordinates to hex coordinates.
  • Extend model to support trail nodes
  • Extend model to support intended destination

Validations to be implemented

  • Validate model correctness
    • Use n-fold cross validation with geocacher GPS track log data
    • Measure probability density for area within radius r of geocacher position
  • Validate accuracy of representation
    • Use past real cases and have both expert and model generate probability distribution maps.
    • Ask expert to compare maps and reason the differences.
    • Identify significant differences (exceeding 50% for a certain region)
  • Validate model usefulness
    • Use Charles Twardy's web site to evaluate models (prior and posterior).
    • Measure probability density for area within radius r of location where missing person was found.
    • Measure for multiple r values for sensitivity analysis.
    • Average densities across all cases to get general sense of how well the model is performing.

Current To Do List

  • Play with CUDA Sample Code
  • Create CUDA Test App
  • Implement CUDA in my app
wisar/tbmod.txt · Last modified: 2014/08/11 13:45 by tmburdge
Back to top
CC Attribution-Share Alike 4.0 International
chimeric.de = chi`s home Valid CSS Driven by DokuWiki do yourself a favour and use a real browser - get firefox!! Recent changes RSS feed Valid XHTML 1.0