Project Description

High Definition (HD) maps are important for perception as they complement senor data. They are needed
for redundancy, robustness, and most of all to know the location of objects and shape of the roadways
that are outside the range of the sensors or that are occluded. The challenge with HD maps is to keep
them up-to-date. The students will use videos recorded from vehicles to update the HD maps, in
particular, Argoverse high-definition maps [1].

Cameras are ubiquitous and inexpensive; they are an effective tool to record comprehensive data about
a traffic network. Our research group has over 6 years of GPS tagged video data recorded in and around
Pittsburgh and we continue to collect it with smartphones and transit bus cameras. This is a rich dataset
to develop and test computer vision algorithms. The most important part is for the algorithm to compare
current video with the HD map and detect relevant changes. The second part is to quantify the change
and enable an actual update of the map. The students should come up with their own algorithms, but
they can base them on previous work. E.g. [2] is neural net based change detector, it can be trained to
detect any class of object that has changed, while ignoring seasonal, weather, or other irrelevant changes.
The proposed tasks of this project are:

  1. Learn the notations and standards of HD maps in general and Argoverse HD maps in particular.
  2. Develop an efficient change detector.
  3. Develop a method to quantify the change and update the map.
  4. Test the system in realistic scenarios.

[2] Alcantarilla, P.F., Stent, S., Ros, G., Arroyo, R. and Gherardi, R., 2018. Street-view change detection
with deconvolutional networks. Autonomous Robots, 42(7), pp.1301-1322.