Efficient Processing of Radar Data Through Transformers for Robust Object Detection and Tracking

Details

Supervisors

Macq, Benoît ; Craeye, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enAutonomous driving perception refers to the ability of autonomous driving systems to effectively perceive the surrounding environment, enabling vehicles to understand and respond to various driving scenarios. Among the sensors used in this field, radar has attracted significant attention owing to its robustness to adverse weather conditions and its low cost. However, radar data is sparse and unstructured, making it challenging to train deep learning models for object detection. This master’s thesis addresses the following research question: How can radar data be efficiently processed for robust object detection and tracking? To this end, RadViT, a Vision Transformer-based architecture, was designed around three guiding principles: (i) enabling fair comparisons between multiple radar input representations, (ii) exploiting the multi-antenna structure of radar data, and (iii) remaining independent of the specific radar hardware, so as to facilitate generalisation across different sensor configurations. In a second part, the K-MD2 FMCW radar sensor was experimentally characterised through measurements conducted in an anechoic chamber, providing a deeper understanding of the physical phenomena and signal artefacts underlying radar data. A crucial step in ensuring true operational robustness for radar object detection and enabling the quantification of the domain shift between the radar used for training and the K-MD2.