Deep learning approach for crowd monitoring using radar maps

Details

Supervisors

Louveaux, Jérôme ; De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

Estimating the number of people from radar data is a challenging task due to noise, clutter, target proximity, and the limited availability of reliable ground truth. Traditional signal-processing techniques such as Constant False Alarm Rate (CFAR) detection offer interpretable and computationally efficient solutions, but are not designed for direct counting and tend to degrade in dense scenarios. In parallel, machine learning approaches provide greater modelling flexibility, but raise questions regarding feature design, generalisation, and robustness. This thesis investigates and compares classical CFAR-based methods and machine learning–based regression models for estimating the number of people from range-Doppler radar data derived form an FMCW radar system. A supervised learning pipeline is developed, incorporating handcrafted feature representations, dimensionality reduction, and regression models. A strict evaluation protocol is followed, with all design choices made on training data and final performance assessed on a held-out test set. In parallel, a CA-CFAR-based detection pipeline is implemented and analysed, and two strategies are explored to convert its detection output into a count estimate. Experimental results show that while CFAR-based approaches provide a strong and interpretable reference, their performance is limited in complex or dense scenarios. Machine learning-based models achieve improved performance and robustness when appropriate feature representations are used, highlighting the benefits of data-driven approaches for this task.