4D Radar Object Localization and Classification
Year: 2026
Type: conference-paper
Abstract: Autonomous mobile robots require functional safety perception systems, such as 2D laser scanners, which perform reliably indoors, but degrade outdoors, where snow, rain, and airborne debris frequently trigger false stops. Sensor fusion with optical modalities (e.g., RGB images) is prone to performance drops in adverse weather and lighting conditions. To overcome these limitations, we propose a radar-only classification and localization approach that transforms 4D radar ADC data into Cartesian images with object height, intensity, and Doppler velocity channels. Unlike many high-resolution radar setups, this work intentionally uses a low angular resolution sensor to reflect embedded and cost constraints of practical AMR systems. With YOLOv8, we successfully detect and classify people and bicycles in unstructured outdoor environments based on these images. We achieve an$A P_{30}$of 0.53 for people on the SEW multimodal AMR dataset and demonstrate that multi-object detection is feasible with low angular resolution radar alone, even under challenging environmental conditions.
Institutions: SEW Eurodrive (Germany), Karlsruhe Institute of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7222449337
