Road network extraction

 

State of the art

General hypothesis :

• the radiometry within the road is locally homogeneous
• there is a significant contrast between the road and the background
• its width is varying slowly
• its curvature is regular (i.e. a road may be considered locally rectilinear)

 
The main unsolved difficulties :

• the noise in the image
• interruptions w.r.t. the road network are caused by trees, shadows, cars etc.
• false alarms may be generated by the presence of buildings and the similarity of the radiometry of an urban region or some crop fields

 
 
Proposed solutions :

• dynamic programming
• active contours
• information theory : twenty questions game
• neural networks
• geographical data

 
Conclusion :

• supervised techniques
• pixel approach : important sensitivity to noise and to local minima

 
Idea : an object approach using the stochastic geometry framework