Evaluation of Deep Neural Networks for Radar Based Point Cloud Classification

Alexander Sandelius () and Emil Ronkainen ()

Abstract

Thanks to the ever-increasing computational power, machine learning has become
a staple in computer science. Many computer vision methods have become so
reliable that their implementation in the surveillance industry is now the de facto standard for many object detection and classification tasks. The use of machine learning for surveillance is not confined to images and videos, however, it is also applicable to RADAR. Axis Communications produces several RADAR-based solutions, some use RADAR only, whereas others provide RADAR and video fusion
technology. RADAR generates unordered sets of points, or point clouds. Point
cloud classification is a fairly unexplored area of machine learning, in particular in the context of RADAR generated point clouds. In this thesis, we address the classification of moving clusters of RADAR point cloud data, which requires the classifier to consider several instances of an input cluster, where spatial as well as temporal information is present. We examine how two main classifier architectures perform on this data type and compare them to the existing classifier at Axis. Empirically, they show promising performance on the available data. However, the results also indicate that a more robust study might be required before employing the classifiers.