Robust 4D Millimeter-Wave Radar Perception in Adversarial Environments

Abstract

This thesis investigates the robustness of 4D mmWave radar perception for autonomous driving, emphasizing real-time, point-cloud-based object detection in adverse and enclosed environments. Unlike conventional radar studies that rely on range--Doppler or heatmap representations, this work leverages the native 4D radar point cloud as the sole sensing modality. This design enhances compatibility with modern 3D perception architectures, reduces computational overhead, and enables seamless integration within existing autonomous driving stacks.

The study begins with a comprehensive analysis of perception sensing modalities---camera, lidar, and radar---to contextualize their relative strengths, limitations, and degradation mechanisms under visibility-challenged conditions. A system-level characterization of 4D radar measurements is presented, highlighting their unique spatio--temporal properties, the preprocessing pipeline, and the effects of dust, multipath interference, and metallic reflections in operational environments.

Two complementary perception pipelines are developed. The first, a model-driven approach, integrates adaptive noise filtering, unsupervised clustering, and rule-based 3D classification. It demonstrates strong real-time performance in harsh indoor environments but reveals a limitation: the inability to detect fully static pedestrians, inherent to Doppler-reliant sensing. The second, a learning-based framework, adapts lidar-style 3D detectors through a radar pillar feature encoder, enabling effective pretraining on public datasets and fine-tuning on custom indoor scenarios. The fine-tuned model achieves a substantial gain in pedestrian detection accuracy, confirming the advantage of data-driven radar perception.

Together, these results establish a unified and robust framework for standalone 4D mmWave radar perception, illustrating both its feasibility and its remaining challenges toward deployment in safety-critical autonomous and industrial applications.