Improving decision making in autonomous vehicles with semantics-enhanced object detection

Autonomous vehicles (AV) which are thought of as the future of mobility are now becoming reality. The sensors such as RADAR, LiDAR, camera which form the sensing unit are the foundations of AV. The input from the sensing units are used for perception tasks such as localization, detection, predictions, which then is used for making decisions in AVs for navigation. The perception task, objection detection, is a key to the safety in AV.

However, due to numerous factors, the input from the sensors are not always of sufficient quality; for example, missing data frames, sparse data. The quality of data directly affects the perception task, particularly object detection and thus the quality of decision making which relies on outcome of perception. Improving the object detection result, accuracy, we can improve the decision making in AV. Semantic technology, namely ontology and knowledge graphs, which have the capability of maintaining relationships, can be used to enrich the data, provide interoperability, represent the problems like in the real-world due to its graph structure. The objective of this thesis is to use semantic technology to improve the decision making by improving the quality of object detection together with deep learning.

Keywords: Deep Learning, Autonomous Vehicles, Robot Operating Systems (ROS), Knowledge Graphs