Millimeter-wave (mmWave) radar sensors have gained significant attention in applications involving target detection and monitoring. Owing to its pro-privacy capability against vision sensors, mmWaves are widely used for human centered applications where accurate target detection is crucial for downstream tasks such as tracking and activity recognition. mmWave radars generate point clouds representing object reflections. These point clouds are sparse and often contain outliers or ghost points due to multipath effects, making robust target detection and localization challenging. Existing clustering methods used for target detection rely on point cloud density assumptions and prior scene knowledge, making them unreliable on sparse mmWave point clouds. This thesis investigates learning-based frameworks to extract target localization features from sparse and noisy point clouds generated by off-the-shelf mmWave radar sensors. To address the limitations of conventional clustering methods, compact neural network models are proposed in this thesis which are capable of learning human target states (position estimates) from sparse mmWave point clouds without requiring prior knowledge of the scene for adjusting their parameters as required in clustering techniques. A novel grid transformation technique is introduced to convert point cloud detections into 2-dimensional grids, enabling convolutional neural networks to effectively capture spatial information for improved localization accuracy. In addition, this thesis addresses the challenge of collecting large-scale labeled data required for supervised learning by presenting a data surrogate technique that leverages the statistical distribution of single-target point clouds to synthesize realistic multi-target scenarios. Based on that, the study presents a learning-based hierarchical framework for multi-target detection driven entirely by synthetically generated mmWave radar point clouds which is able to generalize well to real-world point clouds. Extensive experiments are conducted to optimize neural network architectures, hyperparameters, and loss functions for improved localization accuracy. Validation on real-world over-the-air datasets shows that the proposed learning frameworks are robust and can generalize across different environments. Furthermore, the learning approach is extended to a target tracking framework where a data-driven based linear regression model is demonstrated to show improved tracking performance against conventional Kalman and Particle filters, hence reducing the need for prior knowledge of targets' motion and sensor's measurement models. This thesis also demonstrates that large-scale deep learning models designed for point clouds such as PointNet and DGCNN struggles with sparse mmWave point clouds due to their high model complexity. Therefore, this work highlights the potential of compact and efficient learning models tailored for mmWave sparse point clouds. The findings in this thesis contribute to advancing research and setting benchmarks for learning-based target localization in mmWave radar point clouds.
| Date of Award | 1 Jun 2026 |
|---|
| Original language | English |
|---|
| Sponsors | Marie Skłodowska-Curie Actions |
|---|
| Supervisor | Qasim Ahmed (Main Supervisor) & Pavlos Lazaridis (Co-Supervisor) |
|---|