Consideration of Safety Aspects in a Camera-Aided, Radar-Based Free Space Detection

Tim Brühl, Aravind Kumar Shanmuganathan, Lukas Ewecker, Robin Schwager, Tin Stribor Sohn, Ondrej Vaculín, Sören Hohmann

Abstract: This paper presents a framework for radar point cloud processing and enrichment with camera information to determine the available maneuver space for automated parking functions. This maneuver space is bounded by walls, fences as well as parking cars or objects which can occur randomly in the operational design domain. Firstly, a method for free space representation which is also compatible with the physical properties of radar point clouds, containing polar coordinate points, is presented. Radar sensors suffer from a relatively high noise level, and the received point cloud is usually sparse. As a mistakenly optimistic free space estimation can lead to safety hazards, the concept of virtual radar points is introduced in this work to improve the point cloud's density, and to constrain the free space in a safety-aware manner. Radar reflections are aggregated via clustering algorithms to improve the spatial stability of the objects. However, conventional clustering algorithms decide suboptimal if the discrepancy between the reflection distribution and the actual object shape is too high. We propose to improve the cluster accuracy by using additional camera data. This method is investigated in several experiments and can be further extended. Eventually, a parametric spline which is fitted around the points, defines the shape of the free space in a concise way. Future work should cover the question what to do with radar points which appear to be noise points, but are critical for maneuvering at the same time. One additional aspect is the investigation of ground plan information's potential for the clustering.

External IDs:dblp:conf/itsc/BruhlSESSVH24