From Signals to Point Clouds: Enhancing Radar Target Classification with Neural Networks

Advances in AI are transforming radar-based ATR for security and surveillance. Traditional radar relies on signal processing but struggles in high-dimensional settings where AI excels. This thesis integrates neural networks with 1D, 2D, and 3D radar data to enhance classification and detection in safety-critical environments. The 1D radar section focuses on ultra-wideband echo classification to detect weapons hidden under clothing. Neural networks process late time response data to distinguish between benign and threatening items, enabling non-invasive, real-time monitoring. In the 2D section, the study addresses the “black box” nature of neural networks by incorporating explainable AI techniques. Methods like local interpretable model-agnostic explanations and generative adversarial networks improve transparency and support synthetic data generation, mitigating data shortages. The 3D chapter introduces a transformer-based ATR system for three-dimensional interferometric inverse synthetic aperture radar point cloud data. This transformer architecture captures intricate object details and delivers high classification accuracy even with limited real-world data, thanks to a curated synthetic dataset that enhances model generalization. Overall, this work advances ATR technology by combining AI techniques across multiple radar dimensions, improving detection accuracy and bolstering the deployment of AI-driven algorithms in radar classification applications.