165 / 2019-12-15 14:08:00

Low Level Data Analysis for Automotive Radar Perception System in Real World

IEEE SAM 2020: Special Sessions > Automotive Radar Sensing

Autonomous driving poses unique challenges for vehicle environment perception due to complicated driving scenarios. Precise knowledge of dynamic and feature information about surrounding objects is one of the key tasks. Radar, as one of the major environment sensing system, has been developing to be more and more high resolution and image-like. This provides richer information on object shape, semantic, and kinematics to satisfy autonomous driving purpose. In this paper, to fully utilize the feature information, we present a model-matched tracking system using Radar-only information whilst providing core part for automatic Radar data annotation of complete work pipeline. As will be shown, the tracking and the auto-annotation are mutually beneficial in Radar perception systems. We start with raw Radar point clouds, then followed by clutter removal, clustering, labeling (from associating corresponding Lidar semantics), and tracking. We put special effort on dealing with an intriguing conflict between theory and practice called likelihood credibility issue, which introduced by sensor limitation in real world applications. To alleviate the impact, we present a Fibonacci Fuzzy logic to handle the corresponding track existence probability.