Automotive radar sensors have become irreplaceable not only when it comes to autonomous driving, but also in terms of driver assistance functionalities already available, e.g. for automatic cruise control on highways. Due to multipath propagation, wave interference and ambiguities, extracting target states from raw radar measurement data remains a challenge. To make fast evaluation of tracking and detection methods possible, a radar sensor simulation model has been developed in previous projects. This thesis applies a Track Before Detect (TBD) multi-target tracking algorithm to real and synthetic sensor data to evaluate the aforementioned radar simulation model.

In recent years, Bayes filter methods in the labeled random finite set formulation have become increasingly powerful in the multi-target tracking domain. One of the latest outcomes in this area is the Generalized Labeled Multi-Bernoulli (GLMB) filter which allows for stable cardinality and target state estimation as well as target identification in a unified framework. In contrast to the initial context of the GLMB filter, this thesis makes use of it in the TBD framework and thus runs the tracking algorithm directly on raw data. Therefore, the applied method avoids information loss due to thresholding and other preprocessing steps. Besides evaluating the filter performance in various scenarios on synthetic and real sensor data, different measurement models corresponding to Swerling 0 and 1 targets are implemented and compared. Hence, in addition to the assessment of the simulation model in different scenarios, multiple filter designs are proposed which allow to contrast real and synthetic data outcomes from differing perspectives.

Consequently, the major contributions of this work are the evaluation of the developed radar sensor simulation model under different circumstances and the development of a TBD GLMB filter under the separable likelihood assumption that can be applied to real world scenarios and data. Beyond schemes already discussed in literature, an adaptive birth model and a track merging strategy consistent with the Gibbs sampling version of the GLMB filter are presented. Moreover, to the best of the author’s knowledge, the GLMB filter is applied to real world radar measurement data in a TBD framework for the first time.

This thesis provides a working TBD GLMB filter design and implementation for the automotive radar context. Its applicability to real sensor data is demonstrated in various scenarios. Moreover, the synthetic sensor data is shown to generate meaningful data through its usage as input to the same tracking algorithm. In addition, further improvement strategies are suggested for the radar simulation model based on a Fourier tracing approach as introduced by Holder et al. (2019). The main results include the discussion of the size of the regions in the measurement space influenced by individual targets, power mitigation effects in multipath propagation processes and the range dependency of the received power measurements.