First thing I did was to Binarize the image. I tried several different variations, but ChanVeseBinarize worked the best. I picked gray as my foreground color and black as the background. The streets are mostly in the shadows, and I used gray on a hunch. Later, I tried DominantColors on the image, and the 2nd and 3rd colors were two grays, so I think there is a good reason that gray works well.
Then, run EdgeDetect on the Binarized image. Making the range too big or too small affected the line fitting and the angles of the final lines. (This image is omitted because I can only post 2 links.)
Finally, run ImageLines and composite the image.
I tried several other preprocessing steps, but with my limited skills, the above worked the best. I would think there is some combination of manipulating color and contrast at first, then binarizing the image, and finally dilating/eroding the resulting black and white image in the right way to help accent the streets. However, I couldn't beat the above combination.