The increasing angular resolution of modern automotive radar sensors enables a more detailed and more accurate perception of the environment. This has implications for the measured target detections of extended objects such as vehicles, which cause complex backscatter signatures. Simulation-based approaches strive to enable efficient testing concepts, but require high-fidelity sensor models. To assess the impact of different sensors on the virtual replication of sensor data, it is important to examine the variations between measurements from various sensors. In this letter, radar detections of a passenger vehicle using three radars with different characteristics are presented. Similarities between the spatial distributions of the detections are revealed and differences in the number of target detections and the accuracy of their localization within the bounding box of the vehicle are identified. Furthermore, the spatial fluctuations of point clouds between succeeding measurement cycles are investigated. The results suggest that existing data-driven modeling approaches can be applied to different sensors as well, but that particular attention must be paid to the distinct spatial spread of the detections and to the fluctuations of point clouds.

This paper lines out the trajectory of automotive radar from its first research steps in 1970 to enable first automated cruise control (ACC) systems until today as the backbone of modern driver assistance systems and explains how it became the essential key sensor technology to enable autonomous driving in the future in a mass market environment of privately owned passenger cars. Socially it contributes to Vision Zero and has led to a drastic reduction of injuries and fatalities along its journey. Over the years, radar has also become an economic factor in the range of several billion € and has great potential to develop in the future. The paper also shows that consequent and massive public funded research activities were one key success factor and that cross-border research over different market and application areas stimulated the big success of Radar. Today, the journey of Radar made it clear, that consequent HW-SW Co-design combined with sacrificing traditional paradigm on Radar signal processing paved the ground for this tremendous success story of automotive Radar.

This paper introduces BAAS, a new Extended Object Tracking (EOT) and fusion-based label annotation framework for radar detections in autonomous driving. Our framework utilizes Bayesian-based tracking, smoothing and eventually fusion methods to provide veritable and precise object trajectories along with shape estimation to provide annotation labels on the detection level under various supervision levels. Simultaneously, the framework provides evaluation of tracking performance and label annotation. If manually labeled data is available, each processing module can be analyzed independently or combined with other modules to enable closed-loop continuous improvements. The framework performance is evaluated in a challenging urban real-world scenario in terms of tracking performance and the label annotation errors. We demonstrate the functionality of the proposed approach for varying dynamic objects and class types

Ensuring that virtual sensor models are capable of generating realistic data is a major challenge. It is of particular importance to investigate the properties of real-world sensor data, in order to incorporate them into a consistent simulation tool chain. Radar sensors are influenced by a variety of physical effects which lead to strong fluctuations on the detection interface. Probability distributions can be used to evaluate aggregated measurement data, but lack information about the dynamic behavior of point clouds within measurements. This paper introduces an additional validation criterion, which describes the characteristics of radar target detections from single measurement cycles. Therefore, the spatial fluctuations of radar point clouds detected from a target vehicle are quantified and correlations with scenario-parameters are analyzed. A holistic validation approach is proposed to evaluate the fidelity of such synthetic radar detections.

A key aspect of the imaging capability of radar systems is the angular resolution, which is determined by the aperture size of the antenna array. Therefore technologies such as MIMO and especially radar networks consisting of multiple independent MIMO radar sensors seek to maximize the virtual aperture size. Depending on the range and velocity resolution of the MIMO radar network, multistatic aspects must be accounted for. So far, those multistatic effects were seen as errors, which must be compensated for in order to restore the classical DoA properties of the virtual aperture, described by the narrowband beam pattern. This paper shows that new virtual aperture designs with larger antenna spacings are possible while still preserving the angular ambiguity range of smaller antenna spacings, as long as the multistatic effects of distributed radar networks, namely radar networks whose virtual aperture is large in comparison to the range resolution, are correctly accounted for. The larger antenna element spacing enables larger aperture sizes leading to higher angular resolution. This paper illustrates that the well-known, Fourier Tranformation-based signal processing is unable to exploit this potential of distributed radar networks, and an computationally efficient approximated matched filter is proposed. This article presents a signal model for distributed radar networks, suitable signal processing, and a comparison to the well-known Fourier Transformation-based signal processing for compact radar networks. Both the signal model and the proposed signal processing are verified by measurements with a radar sensor network composed of 2 MIMO radar sensors operating in the automotive frequency range of $76 \,\mathrm{G}\mathrm{Hz}\,\mathrm{to}\, 81 \,\mathrm{G}\mathrm{Hz}$ providing 64 virtual channels with a range resolution of $0.03 \,\mathrm{m}$ . The virtual aperture size of the radar network is ${\sim }0.5 \,\mathrm{m}$ with virtual antenna spacing of twice the wavelength, but the proposed signal processing still allows unambiguous DoA estimation within the full $180 \,\mathrm{^{\circ }}$ range.

Automotive self-localization is an essential task for any automated driving function. This means that the vehicle has to reliably know its position and orientation with an accuracy of a few centimeters and degrees, respectively. This paper presents a radar-based approach to self-localization, which exploits fully polarimetric scattering information for robust landmark detection. The proposed method requires no input from sensors other than radar during localization for a given map. By association of landmark observations with map landmarks, the vehicle's position is inferred. Abstract point- and line-shaped landmarks allow for compact map sizes and, in combination with the factor graph formulation used, for an efficient implementation. Evaluation of extensive real-world experiments in diverse environments shows a promising overall localization performance of $0.12 \text{m}$ RMS absolute trajectory and $0.43 {}^\circ$ RMS heading error by leveraging the polarimetric information. A comparison of the performance of different levels of polarimetric information proves the advantage in challenging scenarios.

MIMO radar networks consisting of multiple independent radar sensors offer the possibility to create large virtual apertures and therefore provide high angular resolution for automotive radar systems. In order to increase the angular resolution, the network must be able to process all data phase coherently. Establishing phase coherency, without distributing the transmitted RF signal to all sensors, poses a significant challenge in the automotive frequency range of $\text{76 GHz} \,\text{to}\, \text{81 GHz}$. This paper presents a signal model for uncoupled and low frequency coupled radar networks. The requirements for phase coherent processing for uncoupled radar sensors are systematically derived from the signal model. The proposed signal processing methods, which establish coherency, are sub-aperture based. Both the signal model and the proposed signal processing methods are verified by measurements with radar sensor networks composed of 2 and 3 radar sensors, providing 768 and 1728 virtual channels respectively. Measurements verify that phase noise is insignificant in the process of establishing coherency in uncoupled and low frequency coupled radar networks.

Uncoupled radar networks offer the advantage of enabling large virtual apertures without the need for a costly RF link connecting individual radar sensors. To fully harness the potential of the large virtual aperture of such a network, it is imperative to address hardware impairments, including gain and phase imbalances among individual virtual channels. The distinctive signal structure inherent in uncoupled MIMO radar networks introduces novel challenges in the estimation of phase imbalances. This paper describes a sub-aperturebased calibration method designed to address these challenges effectively, facilitating the calibration of virtual apertures within uncoupled MIMO radar networks. The presented method is validated by measurements with an uncoupled MIMO radar network consisting of 2 sensors using CS-FMCW and operating at 76GHz, realizing a virtual aperture with a total of 768 virtual channels.

High angular resolution provides improved environmental perception and increases the detection quality of extended targets. It is therefore a key requirement towards future radar systems for autonomous driving. The angular resolution of a radar system fundamentally depends on its antenna array aperture size. It is technically difficult and economically challenging to realize a large aperture radar system as a single sensor. Radar networks, consisting of multiple individual radar sensors, mitigate the challenges caused by creating a large aperture radar system. This paper presents a radar network consisting of two individual MIMO radar sensors equipped with L-shaped physical antenna arrays. L-shaped arrays for the individual sensors are chosen to achieve a rectangular equally spaced radar network virtual aperture. Furthermore, the paper discusses the performance of the resulting virtual aperture in the context of DoA estimation. Measurements of a bicycle, conducted with a coherently coupled radar network consisting of 768 virtual channels, demonstrate the performance of a high angular resolution radar system.

Perceiving the road environment as robust and complete as possible is a fundamental requirement on the way to fully automated driving. Crucial components of perception include detection and classification of road users as well as estimating their extensions and orientation. Using data from a newly available automotive polarimetric radar, this work presents a neural network model using pre-CFAR data as input to detect road users and additionally predict their oriented bounding box. The model is trained to detect static and dynamic road users in diverse urban and rural scenarios on a large data set. A major improvement in detection performance is shown when comparing against CFAR based detection approaches. Additionally, the benefit of polarimetric information is evaluated by optimizing the model on two representations of polarimetric information and comparing it to a model on data without polarimetric information. Results show further promising performance increases when polarimetric information is available.

The radar scattering characteristics of extended objects are an important parameter for perception and tracking algorithms in automated driving tasks. Therefore, high-fidelity sensor models are required to simulate and evaluate typical driving scenarios in virtual testing applications. While the general analysis of typical scattering centers of passenger cars is well studied, there are only a few publicly available reports that analyze specific features of the scattering characteristics of different vehicle types. Hence, this work presents detection distributions derived from systematic measurements for six different vehicle types, conducted with a commercial automotive radar on a proving ground. In particular, the contribution of underbody reflections to the respective radar signatures is analyzed, which are caused by multipath propagation via the road surface. The measurements reveal distinctive differences between the scattering characteristics of different vehicles, which are attributed to the respective underbody geometry.

Virtual validation methods strive to reduce the testing efforts of automated driving functions significantly. However, assuring the fidelity of deployed sensor models based on objective criteria remains an unsolved challenge, especially for radar target detection point clouds, since they are subject to major stochastic fluctuations.This work focuses on the scenario-based derivation of requirements for synthetic radar target detections, which are deduced from sensor data recorded in real-world test drives. Based on these reference data, deviations between the radar point clouds from different recordings are quantified using metrics for both point clouds from single measurement cycles and relative probability distributions for accumulated point clouds. Finally, the suitability of the calculated reference values as adequate metrics of deviations between recorded and simulated sensor data is evaluated.

Radar sensors play an important role in automated driving technologies. However, the rising number of sensors deployed to enable autonomous driving functions leads to enormous validation efforts. While simulations are a possible approach to accelerate the validation process, the development effort for realistic sensor models increases significantly. Data-driven sensor models offer the possibility to replicate sensor data accurately and efficiently. Using real measurement data, the sensor output can be simulated without the detailed parametric modeling of the wave propagation and sensor effects. In this paper, the radar signatures of a passenger vehicle under a constant aspect angle are analyzed in real measurements. Then, a data-driven approach for stochastically modeling the radar target detections is presented. The model is trained with real sensor data to achieve a high degree of realism. A qualitative comparison between the simulated and measured detections reveals promising results.

With increasing levels of vehicle automation, the requirements for the sensors that are used for environment perception are rising at least as steadily. In automotive radars, the capability to perceive polarimetric information inspired some recent work to exploit this aspect in order to improve perception of static and dynamic surroundings. Such utilization requires a well-calibrated radar system. This paper demonstrates how to achieve the necessary stability in the scattering information of an automotive multiple input multiple output (MIMO) millimeter wave radar. The parameters of a fully polarimetric calibration are derived based on pole and dihedral target measurements in an automated test facility. Furthermore, the impact on the polarimetric quantities estimation due to mounting the radar behind a radome is analyzed.

For higher levels of driving automation, a vehicle has to perceive the static environment in addition to other road users to ensure safe and reliable operation. This paper presents a radar-based method utilizing a lower data level than typical peak detection point clouds (e.g. Constant False Alarm Rate (CFAR) detection) to address some challenges in static environment perception. Using dense, image-like radar data overcomes CFAR's disadvantage of masking on extended objects as well as the suppression of low reflection amplitude targets. The improved performance of the proposed technique is demonstrated in a direct comparison to a CFAR approach based on real-world experiments.

Automotive radar perception is an integral part of automated driving systems. Radar sensors benefit from their excellent robustness against adverse weather conditions such as snow, fog, or heavy rain. Despite the fact that machine-learning-based object detection is traditionally a camera-based domain, vast progress has been made for lidar sensors, and radar is also catching up. Recently, several new techniques for using machine learning algorithms towards the correct detection and classification of moving road users in automotive radar data have been introduced. However, most of them have not been compared to other methods or require next generation radar sensors which are far more advanced than current conventional automotive sensors. This article makes a thorough comparison of existing and novel radar object detection algorithms with some of the most successful candidates from the image and lidar domain. All experiments are conducted using a conventional automotive radar system. In addition to introducing all architectures, special attention is paid to the necessary point cloud preprocessing for all methods. By assessing all methods on a large and open real world data set, this evaluation provides the first representative algorithm comparison in this domain and outlines future research directions.

Radars are one of the sensor modalities employed in systems for automated driving, due to their robustness, instantaneous radial velocity estimation capability of targets and moderate costs. With the availability of polarimetric radars for automotive applications, the scattering matrices for detections on road users can be measured additionally. However, with limited sampling time and calibration errors resulting in sidelobes, the scattering matrices of objects located close to each other can mix and a wrong scattering mechanism is inferred for the targets, which is especially observable in the angular dimension. This paper presents an approach to decrease this mixing by using gradients along the dimension of interest to shift amplitudes and evaluates it on real world polarimetric radar data.

Grid maps are widely established for the representation of static objects in robotics and automotive applications. Though, incorporating velocity information is still widely examined because of the increased complexity of dynamic grids concerning both velocity measurement models for radar sensors and the representation of velocity in a grid framework. In this paper, both issues are addressed: sensor models and an efficient grid framework, which are required to ensure efficient and robust environment perception with radar. To that, we introduce new inverse radar sensor models covering radar sensor artifacts such as measurement ambiguities to integrate automotive radar sensors for improved velocity estimation. Furthermore, we introduce UNIFY, a multiple belief Bayesian grid map framework for static occupancy and velocity estimation with independent layers. The proposed UNIFY framework utilizes a grid-cell-based layer to provide occupancy information and a particle-based velocity layer for motion state estimation in an autonomous vehicle's environment. Each UNIFY layer allows individual execution as well as simultaneous execution of both layers for optimal adaption to varying environments in autonomous driving applications. UNIFY was tested and evaluated in terms of plausibility and efficiency on a large real-world radar data-set in challenging traffic scenarios covering different densities in urban and rural sceneries.

This paper presents a new adaptive multi-hypothesis clustering method for extended objects on radar data. The proposed method provides several clustering hypotheses per object for a given measurement set efficiently by ordering the data set similar to the HDBSCAN and extracting clusters from the ordered data set with the help of prior knowledge obtained from Extended Object Tracking (EOT) and fusion. The performance of the proposed method is tested on a manually labeled real-world data set. The dependency on accurate prior knowledge is reduced compared to previously introduced adaptive clustering methods.

Human gesture classification using radar sensors is becoming indispensable in the era of autonomous driving. Challenging automotive-related gestures need different techniques for accurate and quick classification. In this paper, classifications based on temporal radar data using CNNs (Convolutional Neural Networks), RNNs (Recurrent Neural Networks), and BRNNs (Bidirectional Recurrent Neural Networks) are going to be introduced and compared. RNNs have the advantage of getting the classification decision even before the arrival of the total data stream. Twelve challenging scenarios are presented, together with the signal processing chain. Classification is performed on successive range-velocity diagrams, which spares the step of time-frequency processing already used in other applications. Four different models are introduced which result in high classification accuracies. This shows a high potential of employing radar for gesture classification using temporal information.