(A) study on efficient auto-labeling leveraging 4d radar dataset properties

Abstract

In autonomous driving systems, object detection technology identifies the location and type of obstacles around the vehicle. Since autonomous driving systems conduct path plans and behavior decisions based on object detection, reliable object detection technology across various conditions is vital for safety. Recently, 4D Radar sensors have gained interest in object detection studies for their robustness against poor weather conditions such as rain and snow. However, training networks for object detection requires large datasets and manual labeling of vast amounts of sensor data. Moreover, since 4D Radar sensor data is not intuitive for humans, annotating accurate 3D labels is difficult and expensive labor. To address these challenges, this paper addresses an auto-labeling method utilizing pre-trained deep learning networks to effectively train 4D Radar object detection networks.