My researches were focus on Traffic Sign Detection and Recognition and Pavement Distress Management. I made a broad review of state of art methods in Traffic Sign Detection and Recognition and wrote a proposal. Then, I switched to Pavement Distress Management, where problems were multiples. I was working on 3D profiles of roads obtained from a laser set up on a truck.
 |
| A 3D profile with a crack and a pothole |
 | |
| Detection of the drop-off on the road sides using 3D profiles. On the right, we can see a severe example of raveling. |
 | ||
| Detection of the white marks using intensity images. |
Raveling is a
pavement distress characterized by the loss of materials (stones) on
the road surface caused by the dislodging of aggregate particles and
loss of asphalt binder. Once small stones started to go away,
raveling grows rapidly until all the entire lift of asphalt pavement
disappears. Efficient detection of raveling is needed for road
distress management. Indeed, the lift of asphalt pavement is
essential for water drainage and a good adhesion of cars. We aim to
detect on a road surface what parts are raveled and for each part how
much is it raveled. Rutting is a major asphalt pavement distress that
affects pavement structure integrity and driving safety. It is a
permanent longitudinal depression that mainly forms because of
traffic loadings in the wheel paths of a road.
Raveling is often measured as the area or the volume originally
occupied by the missing stones. The difficulty when estimating the
rutting is to remove all the existing distresses in order to
approximate the actual 2D or 1D mean surface of the road. The
estimation of the mean surface of the road allows measuring the
longitudinal depression in the wheel paths. I propose, implemented and evaluated a new approach to
solve both the raveling detection and the rutting evaluation problem. It aims
to detect simultaneously the missing stones and the hypothetic undamaged
surface of the road. Thus, the detected missing stones can be used to
measure the raveling severity and the surface of the road the
rutting.
The idea was to
model a road profile as a mixture of models, where the models are the
same but shifted. The shift is a multiple of the average size of a
stone. As we know the maximum depth of the layer damaged by raveling
we can approximately set the number of models we need. Then the
Expectation-Maximization algorithm is used to fit the mixture of
model and the value of the mixing coefficients can be identified to
the number of missing stones. One of the difficulties is to find a
general regression model that can fit to the big variety of rutted
road surface. We first tried parametric models. Then, we developed a
non-parametric approach.