Webseite Institut für Kraftfahrzeuge (ika)
Topic and Goals of the Thesis
For the development and safety validation of automated vehicles, huge trajectory datasets of road users are necessary. One solution to create those datasets is the use of drone imagery of traffic on e.g. highways or urban intersections.
For this purpose, all road users must be detected with high accuracy in the images, for which typically deep neural networks are used. For a neural network to achieve a high recognition accuracy, a lot of training data is typically required. However, the creation of these annotations is very time-consuming and costly. A promising alternative is the use of semi-supervised learning. Here, an existing network and heuristics are applied to large amounts of unannotated video data in order to automatically detect the shortcomings of the network and generate new training data.
· Literature research on approaches on semi-supervised learning
· Design and implementation of a solution suitable for drone imagery
· Evaluation of the solution on a large amount of existing videos
· Good English and/or German language skills
· Reliability, commitment and enjoyment of working independently
· Basic machine learning and/or computer vision experience
· Advanced programming experience in Python and/or Matlab
Um sich für diesen Job zu bewerben, sende deine Unterlagen per E-Mail an firstname.lastname@example.org