Hi everyone. My name is Matías, from Uruguay.
I am working to help improve procedures for a government office dedicated to road maintenance. This office has a truck equipped with a camera that can take pictures of the road and georeference each one. However, people are not using this tool because they don’t have any means to analyze the material.
I am trying to find a trained algorithm to detect and classify different types of pavement failures, but I haven’t found anything. I need an algorithm that can classify 20 different types of failures and measure them (area and length).
Do you know anything similar to what I am searching for? I was thinking of an algorithm like YOLO, first training it to detect all types of failures, and then training it again to classify the different types (for the 20 classifications).
Any help would be useful!
Thanks a lot!
YOLO does object location (bounding boxes) and classification simultaneously at run time. Some versions train more efficiently if the two are separated, learning classification first and then transfer learning refining that model to also perform location. YOLO is particularly useful when fast runtime throughput is important, which may not be the case for this application. It doesn’t seem that road condition needs to be analyzed on the fly near real time. Training YOLO models from scratch also requires a large collection of labeled training data, with lots of augmentation to account for different orientations, weather conditions (wet vs dry) and lighting etc. Do you know of a source for that kind of data?
Interweb search using deep learning pavement failure classification suggests others have traveled this road (see what I did there?) before. Maybe take a read through some of those papers, see what worked or not, and look particularly for discussions about training data. Let us know what you find?