Road detection and machine learning

asked 2017-12-09 03:44:37 -0600

procton gravatar image

My goal is to detect roads from high resolution aerial images. I would like to adopt some kind of cascade classification (Haar, LBP) but I am stuck on some key decisions:

  1. Overall, is cascade approach fine for this kind of detection ? I found tons of examples related to object detection, but in my case objects are roads that cannot be represented with rectangles containing detected object;
  2. I have 3 channel images with 16 bit each channel (16*3=48 bit total). Usual viewers cannot display such kind of images (I tried also Gimp without success). So how positive samples should be created manually ? Moreover cropping could led to resolution loss. Is this a problem ?
  3. Which is the optimal contents for positive images ? Just road texture or larger images that contain roads ? Positive samples must have same size ?
  4. I have ground truth data made of a set of points that represent roads as lines and junctions between crossing lines. Maybe this data can replace image cropping for training ?


edit retag flag offensive close merge delete


"Overall, is cascade approach fine for this kind of detection ?" -- no, not at all. it only works with rigid things with a well defined texture, and a single pose.

berak gravatar imageberak ( 2017-12-09 03:52:08 -0600 )edit

OK, this answer solves almost all questions. So which should be the correct approach ? Maybe random forest using ground truth data for training ?

procton gravatar imageprocton ( 2017-12-09 04:10:07 -0600 )edit

idk. it also seems more a "segmentation" task, not so much an "object detection" one, no ?

do some research

btw, openstreetmap has vectorized data for almost the whole planet.

berak gravatar imageberak ( 2017-12-09 04:23:37 -0600 )edit