Neural Networks in Autonomous Vehicles – Artificial Experts or Simple Pattern Matching?

Authors

Matthias Pollach, Dr. Daniel Clarke, Johannes Mauthe,
Mentor Graphics, a Siemens Business, Munich

Summary

Highly assisted and autonomous vehicles rely on different sensor systems to observe the environment, with machine learning providing the semantic context of the detected objects. Over the last decade, deep learning has risen to prominence with the ability to map the complex, non-linear relationships between large volumes of labelled observations and effectively classify these observations. However, given the bounds of efficiency and practicality in automotive applications, there remain a number of challenges for the effective deployment of object classification systems, not the least of which is the computational complexity. Within this paper, we propose a solution for undertaking object classification in automotive environments which is expected to be
computationally efficient and effective. Our proposition uses a cascading hierarchy of traditional (i.e. shallow) classifiers, focusing on the use of expert domain knowledge both to build the classifiers and to motivate the correct selection of classifiers. This paper outlines the motivation and scope of the problem, then introduces the proposition and outlines some preliminary results.

Members Login