-1

Recently I've been doing some study on the topic Road Recognition. There is a problem that if you have illuminant variant images, the recognition results will not be satisfactory if you just use the normal segmentation approach. Here, shadow removal will improve the results a lot, but for the image, part of which is taken under strong lighting condition, the part of the road will be hard to recognize, like in this picture in the lower-left corner:

enter image description here

Could someone please suggest some approaches of solving this problem?

eckes
  • 10,103
  • 1
  • 59
  • 71
GoC
  • 31
  • 8
  • This image is pathologically hard! Not only is there no left hand side to the road you're trying to detect but the lower left corner isn't just bright, it's overexposed - there's no texture or anything visible in that region that would help to establish that it's contiguous with the road. And that shadow is the strongest straight line in the image. You could try to use a sliding window and subtract mean intensity for each region. That may help somewhat. Good luck! – Dave Durbin May 12 '15 at 00:43

1 Answers1

0

when an image is too bright (or too dark) you could try contrast enhancement.

edit looking at your example image it looks like the features are completely washed out and contrast enhancement will not help detect the white line marking the edge of the road. another alternative might be to find this line using a line detector, and then to simply extend it to the bottom of the image. We just assume that the line continues the same path.

If you are doing a road recognition I'm guessing you have (or will) implement a method for lane tracking even over dotted lines in the road. The same logic you used for the dotted line case could be applied to this image. The bright area would be seen as a dotted line, but your algorithm should still know its the same lane

andrew
  • 2,451
  • 1
  • 15
  • 22
  • If you post an example pic or even just links to pictures, people might be able to offer more detailed recommendations – andrew May 11 '15 at 19:15
  • yeah, I want to attach a picture but it says I need 10 reputation to do that :( – GoC May 11 '15 at 19:16
  • You can post a picture on a website, like imgur, tinybucket, photobucket, etc and just add the link to your post – andrew May 11 '15 at 19:21
  • it looks like contrast enhancement will not work on this part of the image. It's completely white so we can't extract information about it. But see my edit – andrew May 12 '15 at 18:04
  • thank you again for your reply! The images that I have are of different types, there are structured roads like highway and unstructured roads such as dirt road in the country, and they are not always straight. What I'm thinking is to apply an region segmentation based approach. – GoC May 12 '15 at 18:12
  • Thats a good approach too I dont know what language you are using but matlab has plocbroc command it can easily be copied in python http://stackoverflow.com/questions/5073767/how-can-i-efficiently-process-a-numpy-array-in-blocks-similar-to-matlabs-blkpro and Im sure opencv has some code examples somewhere. But just as Dave commented, the sensor was simply over saturated, and in this particular case you can't extract any information from the bright portion. it's simply not there to extract – andrew May 12 '15 at 18:20