Detection and Estimation of Dispersive Target Signals

The first hypothesis proposes to utilize the dispersiveness of indoor multipath environments to enhance the detection of occupants. This study specifically investigates using UWB radars to detect car occupants based on their respiratory motion. By employing a point-target assumption and pinhole channel model, we show that the respiratory motion and multipath channel response can be factorized. Two detection algorithms are derived from this factorized model: an estimator-correlator (EC) and a detector based on variational inference (VI). Both of these detectors are shown to outperform the baseline fast Fourier transform (FFT) approach, which assumes a line-of-sight-only model. Furthermore, a data-driven approach is also applied to this use case based on the widely-used ResNet architecture. A measurement campaign was conducted to obtain a dataset for training the network. The dataset comprises radar channel impulse responses (CIRs) of participants engaged in a range of motion activities within the vehicle. The data-driven detection algorithm outperforms the model-based approaches. However, since the dataset contains only data from adults, the ResNet’s generalization capabilities should be validated on cases not contained in the dataset, such as infants or small children.
The second hypothesis proposes to enhance the detection and estimation performance of sparsity-based line spectral estimation (LSE) for structured line spectra using group-sparsity. Structured line spectra, i.e. spectra comprised of clusters of related spectral lines, arise in several applications. E.g., the time-dispersed backscattered signal from extended objects is approximated by a structured line spectrum. However, the majority of LSE algorithms assume the spectral lines to be unstructured, i.e. they assume all spectral lines are uncorrelated. We derive two sparse Bayesian learning (SBL)-based LSE algorithms which utilize group-sparsity to enhance the estimation performance for structured line spectra. The fast variational block-SBL algorithm is based on the assumption that the group-size is known and the dictionary fixed. The more sophisticated group-sparse super-resolution algorithm incorporates the estimation of dictionary parameters (on a continuum) and the group sizes into the estimation procedure. Through simulations, we show that our group-sparse super-resolution algorithm increases the detection probability of the spectral lines related to extended objects in low-signal-to-noise ratio (SNR) scenarios. As a consequence, the parameters of the extended object, e.g. its center of mass and extent, can be estimated more accurately. Furthermore, structured line spectra are also observed in other applications. We apply the fast variational block-SBL algorithm to multiple measurement vector (MMV) direction of arrival (DOA) estimation using data from an antenna array, and the group-sparse super-resolution algorithm to multi-pitch estimation of audio data. The algorithms presented in this thesis outperform the current state-of-the-art methods in both cases.
In summary, this thesis contributes to the detection and estimation of dispersive target signals. It does so both practically, by developing detection algorithms utilizing the dispersiveness of dense multipath channels to enhance the detection performance in the car occupancy detection use case, and theoretically, by developing LSE algorithms tuned to the detection and estimation of structured line spectra.