Inverse synthetic aperture radar imaging of automotive targets

Abstract

Advanced driver assistance systems (ADAS) aim to improve road safety, reduce fatalities, and enable autonomous driving. Modern vehicles rely on multiple sensors such as lidar, cameras, thermal detectors, and radar to detect and classify road users. Among these, millimeter-wave automotive radars offer robust range and velocity estimation, all-weather operation, and unobtrusive bumper integration. High-resolution two-dimensional radar images, and in particular inverse synthetic aperture radar (ISAR) images, can provide detailed information on target size, shape, and motion. However, existing ISAR studies of ground vehicles at automotive radar frequencies have produced only limited datasets for a few targets under controlled conditions, and these datasets are not publicly available. As a result, there is a lack of realistic, large-scale ISAR data suitable for modern machine learning (ML) algorithms for automotive applications. This thesis develops a framework for simulating high-fidelity ISAR images of automotive targets at millimeter-wave (mm-wave) frequencies. The simulation model incorporates vehicle kinematics, radar scattering phenomenology, range–Doppler clutter, and receiver noise for a 77 GHz automotive radar. Static and dynamic mm-wave clutter for automotive scenarios is modelled using measurements acquired on Indian roads, taking into account different surface types and roughness conditions. The resulting clutter statistics are used to parameterise phenomenological clutter models in the simulator. The framework is validated qualitatively and quantitatively against measurement data gathered from real automotive radars. The simulated ISAR images are then used as inputs to traditional ML classifiers and deep neural networks for the classification of automotive targets; the results show that ISAR radar images are excellent features for accurately classifying different road vehicles. The thesis further investigates the reliability and interpretability of these classification decisions. It is shown that misclassifications can occur even when noise and clutter are relatively low. To analyse such cases, a method based on counterfactual explanations is proposed, using generative adversarial networks (GANs) to perturb ISAR images of a query class until they are classified as a distractor class, while enforcing that the perturbations remain realistic and consistent with the original class distribution. The resulting counterfactual images provide physics-based insight into which target regions and micro-Doppler features drive the classifier’s decisions. Finally, two application-oriented studies using ISAR images are presented: (i) an automated parking test framework in which an externally mounted radar generates high-resolution images of a car parking in a designated slot and a polynomial trajectory fit is used to assess parking performance; and (ii) an around-the-corner radar (ACR) proof-of-concept at 77 GHz for non-line-of-sight collision avoidance, where sparsity-based dictionary learning is used to separate overlapping range–Doppler returns and reconstruct the signatures of multiple dynamic targets in an urban NLOS configuration