A Comprehensive Dataset for Victim Detection in SAR Operations Using Robot-Mounted UWB-Radar Sensors

Project Description The dataset was collected at the Antenna Laboratory of the University of West Attica (UniWA), in an office-classroom environment, as part of research on through-wall victim detection for Search and Rescue (SAR) operations, as a continuation of our previous work [1],[2]. Six human subjects were placed behind a 27 cm-thick wall at two distances (1 m and 2 m) and in four body orientations. Radar measurements were acquired using an ultra-wideband (UWB) pulsed radar sensor (SLMX4) mounted on a robotic platform, which moved across 11 predefined positions using odometry and LiDAR for accurate localization. The collected radar signals were later analyzed using both machine learning (ML) and deep learning (DL) techniques to detect human presence and estimate distance through the wall. A detailed description of the setup and methodology is provided in [2]. Dataset Description The dataset comprises 528 sessions of radar data corresponding to human presence and 11 sessions corresponding to human absence. Specifically, data were collected from six (6) human subjects ("victim actors") across eleven (11) predefined robot positions and four (4) different orientations: supine (face up), prone (face down), facing toward the wall, and facing away from the wall, as described in [3]. Each presence session lasts approximately 3 minutes, while each absence session lasts 30 minutes. In total, the dataset includes approximately 1,584 minutes of human presence and 330 minutes of human absence, recorded at a refresh rate of approximately 19 radar frames per second. All the data that has been collected corresponds to 2,179,281 rows of absence and presence, with presence having 1,806,240 and absence 373,041 making the dataset in total approximately 18 GB and almost 32 hours. Dataset Contents Each CSV file is assigned an individual identifier Ni, where i = 1...6 corresponds to the subject number and N0 corresponds to subject absence. Inside each CSV, there are 4 columns, starting with the unique identifier of each session (explanation in [3]); the next column has an array of arrays of complex data, with 180 different complex numbers, each for each range bin (with each bin having a distance from the previous of 0.0512 m); and finally, the last two columns contain the annotation with presence (0 for absence and 1 for presence) and the distance the victim is. Model Architecture The models used on the current dataset are described in detail in [3]. For victim detection, two methods were applied: breathing detection using a 1D CNN, and movement detection using an XGBoost model with the matrix of the standard deviation (STD) of the corresponding data. For victim localization, a simple mathematical approach was employed, based on finding the maximum STD value of the 1D matrix. [1] A.-P. Michalopoulos, E. N. Paliodimos, F. Papadopoulos, G. Nikolaou, C. Patrikakis, and S. A. Mytilinaios, “Victim Detection Using a Robot-Mounted UWB-Radar Platform,” in 2025 14th International Conference on Modern Circuits and Systems Technologies (MOCAST), June 2025, pp. 1–4. doi: 10.1109/MOCAST65744.2025.11083953. [2] A. P. Michalopoulos, E. Paliodimos, F. Papadopoulos, G. Nikolaou, C. Patrikakis, and S. Mitilineos, “Victim Detection Using a Robot-Mounted UWB-Radar Platform.” Zenodo, Mar. 15, 2025. Accessed: Aug. 15, 2025. [Online]. Available: https://zenodo.org/records/15032859 [3] Paper of Nature in processing stage