Deep Learning based Turn Light Detection at Road Intersections for Autonomous Driving

Autonomous driving promises to revolutionise transportation by making it more efficient, cheaper, safer and climate friendlier. Bringing autonomous vehicles onto roads requires effectively sensing the surrounding environment and integrating derived information into a safe and efficient driving policy. This involves, amongst others, capturing the intent of other traffic participants as signalled by turn lights. Contributing to the development of autonomous systems as part of the publicly funded MEC-View project at Bosch, this thesis aims at detecting turn signals with an emphasis on oncoming vehicles at road intersections. Specifically, using images taken from cameras attached to side mirrors of an autonomous vehicle, a Convolutional LSTM Neural Network detects and classifies turn signals of approaching vehicles during both day and night time. As such, this work differs in important aspects from existing literature. Results obtained from experiments and detailed analyses of test cases indicate that the devised model performs competitively despite data scarcity and strong label imbalance, with a weighted F1 score of 80.8%. Hence, this thesis lays promising ground work in the domain of autonomous driving and identifies potential future improvements.Detta examensarbete syftar till att utveckla autonoma system som en del av det offentligt finansierade MEC-View-projektet hos Bosch och upptäcker blinkersignaler med tonvikt på mötande fordon vid vägkorsningar. Specifikt, med bilder tagna från kameror som är fästa vid sidospeglar på ett autonomt fordon, upptäcker och klassificerar ett LSTM-faltningsnätverk blinkersignaler för ankommande fordon under både dag och natt. Resultat som erhållits från experiment och detaljerade analyser av testfall indikerar att den planerade modellen presterar konkurrenskraftigt trots dataknapphet och stark etikettobalans, med en vägd F1-score på 80,8%