Machine Learning for Autonomous Vehicle Behavior Adaptation in Complex Traffic Situations

Abstract

The work presented in this paper takes the latter route towards the development of data-driven learning-based approaches. More specifically, this work proposes machine-learning methods to develop learned models and algorithms for coping with the often encountered object behavior uncertainty. The topic is challenging and involves modeling and using techniques from machine learning that focus on handling limited and often incomplete label-to-data type data sets, as in this case only partial observation is received from an object on the road and the data likelihood is subject to the error in perception and other influences. Perception limitations such as occlusions and detection errors might occur frequently in real road settings. Consequently, this results in missing and less reliable information. In addition to this, other road user behavior might become difficult to track for the aforementioned reasons and also due to the predictions leading to a chain effect that directly costs the road users their robustness against errors [1].