Abstract: Radar-based pedestrian detection is a key enabler of advanced driver-assistance systems (ADAS) and future autonomous driving functions, particularly under conditions where vision sensors are limited by poor lighting or adverse weather. In this study, we present a structured dataset collected at the ZalaZONE proving ground, a state-of-the-art automotive testing facility in Hungary, using a Continental ARS 408‑21 automotive radar operating at 77 GHz . The dataset contains static radar measurements of both a real human subject and a pedestrian dummy, recorded at multiple controlled distances (5 m and 10 m) and orientations (front-facing and side-facing). Each radar scan includes radar cross-section (RCS), object distance, relative velocity, and metadata retrieved directly from the vehicle’s CAN bus interface. The results demonstrate that RCS values strongly depend on target type, orientation, and distance. Real pedestrians produce significantly higher variance due to posture, clothing materials, and micro‑movements, whereas dummy mannequins exhibit stable, narrow RCS distributions. Interestingly, measurements at longer distances show less negative RCS values, likely influenced by multipath reflections and environmental dispersion effects. The dataset is fully labeled and formatted for direct use in supervised machine learning pipelines, supporting classification models such as logistic regression, support vector machines, and neural networks. It also pro-vides a foundation for future extensions with dynamic scenes and temporal sequence modeling, enabling the development of more robust and generalizable radar-based pedestrian detection algorithms. By combining controlled measurements with realistic environmental variability, this dataset contributes to the advancement of radar sensing technologies for safe and reliable autonomous driving. Quantitatively, the mean RCS difference between real pedestrians and dummy targets reached 3.8 dB at 5 m and 4.5 dB at 10 m, confirming the discriminative potential of the dataset for classification tasks.
Rózsás et al. (Wed,) studied this question.