High-Level Perception for Intelligent Transportation

In the last few years, intelligent transportation has attracted a lot of interest, both from academia and industry. In this context, environmental perception from the vehicle point of view is key for the successful adoption of these new technologies. In this thesis, some advances in the field of perception for intelligent transportation will be discussed, both regarding automated vehicles and infrastructure. Considering infrastructure for intelligent transportation, a Deep Neural Network based train pantograph head pose estimator is described. Moving to automated vehicles, one important question to ask is \emph{How does the vehicle behave moving away from ideal conditions}? Considering Deep Neural Networks for vision tasks, this question will be addressed from two different points of view: on one side, analysing the training dataset and possible errors in the labelling process; on the other one, analysing detection performance in non-ideal cases (like distortion and compression). Finally, two examples of complete stacks for autonomous robotics are presented: one for autonomous racing and another for Unmanned Aerial Vehicle search and rescue operations. In both cases, environmental perception is a key stage for enabling all other autonomous features (localisation, navigation, etc.). The work presented in this thesis is a step towards improving Intelligent Transformation Systems, and the findings are important for the safety and reliability of future automated vehicles.