Doctorate Thesis Defense of Asma OMRI

Doctorate thesis defense on october 1^st 2025 at 09H30 AM ,in Amphitheater Ibn Khaldoun, SUP'COM 2.

Entitled : AI-Enhanced FMCW Radar Perception for VRU Safety in ADAS Applications

 
Presented by : Asma OMRI

Committee

Abstract

Enhancing road safety, particularly for Vulnerable Road Users (VRUs) like pedestrians and cyclists, is a key challenge for Advanced Driver Assistance Systems (ADAS). Frequency-Modulated Continuous Wave (FMCW) radar is ideal due to its robustness in adverse conditions. This thesis leverages FMCW radar to improve safety via object classification, VRU detection, ghost object filtering, and human-centric applications. A two-level clustering approach (temporal-velocity projection + spatial clustering) is proposed to transform sparse radar data into cluster object representations, enabling coarse and fine-grained classification (e.g., "Car," "Pedestrian"). Incorporating temporal track IDs significantly improves classification accuracy by resolving class ambiguities.

We enhance radar perception by using clustering with machine learning to filter ghost objects, improving scene interpretation, especially in urban areas. Recognizing that conventional CNNs struggle with sparse radar data, we propose a graph-based object classification method using Graph Neural Networks (GNNs). A novel two- stage Graph Convolutional Network (GCN) classifies objects, separating VRUs first, then performing fine VRU classification. This hierarchical GCN outperforms prior methods. Finally, the thesis investigates human-centric radar applications: pedestrian motion classification and contactless heart rate estimation. A lightweight neural network accurately classifies pedestrian motions (walking, crossing, hesitating, hand up) despite low radar resolution. We also use a Long Short-Term Memory (LSTM) network to estimate heart rate from raw radar