Automatisiertes Fahren 2020

With infinitely many situations that an autonomous vehicle could face in real life, it is beneficial to combine different development methodologies with the V-model for the development of autonomous driving functions in a more robust and efficient way. To meet the complex challenges of developing safe and robust functions for autonomous driving, IAV proposes an approach in which different stages of V-model are combined with agile software development methodology and DevOps philosophy. This leads to a process model, which is more flexible as compared to conventional V-model and has ability to take into the account the information obtained at later stages of project into the development process efficiently.

In this contribution, a methodology for simulating multiple vehicles under test with driving dynamics models is presented. The virtual test drives that are conducted in this setup are used for the evaluation of cooperative adaptive cruise control systems. This pre-series function for collaborative platooning relies on environment sensors as well as car-2-car communication and permits very low headways. The multi-ego simulation setup proves to be suitable for the development tasks and the presented results indicate that the consideration of vehicle dynamics is crucial for the development of such control algorithms.

The complexity of technical systems increased immensely over the past decades. Where those technical systems were often mainly mechanical ones in the past, hardware and especially software have become the main driving forces today.

In the 1970s, defects of a system which resulted in injuries, severe injuries or even deaths could be mainly attributed to technical component failures. The “low” complexity of systems at that time didn’t reach a level, where losses due to component interactions where of huge concern. Today, the safety industries (automotive, avionics, medical, etc.) have become quite experienced in dealing with issues related to component failures. With the use of qualitative and quantitative safety analysis like FMEA, FTA, HAZOP, FMEDA etc., losses due to technical component failures have been heavily reduced. But, where today we reduced the risks of component failures, the risks of losses due to complex system designs, and therefore complex component interactions heavily increased. This paper will show, how the method of system-theoretic process analysis can help to cover these issues. How the gaps in our safety efforts, regarding the belief of the system about its environment, can be closed, rather than just focusing on failure of components and their performance.

Emergency braking systems have been introduced successfully in the commercial vehicle industry. They significantly improved the safety on motorways, where large quantities of trucks are present. Moreover the sensors have matured together with the algorithms that are required to filter and interpret sensor data to understand the traffic situation. In consequence of this evolution new use cases come into focus. Collision mitigation systems (CMS) or advanced emergency braking systems (AEBS) for public transportation are such Advanced Driver-Assistance Systems (ADAS) as they were excluded from regulation for AEBS (United Nations, 2013, p. 4) due to the fact, that city buses operate in different environment and automatic braking may cause serious injuries to standing passengers. However, it makes sense to let also vehicles which carry a lot of persons benefit from progress in safety systems. In this paper we summarize the challenges of designing such a CMS / AEBS for public transportation starting from some accident statistics as design target and finding a “safe” system design which incorporates the impact to the passengers in case of false interventions. We propose a new design of a braking profile which to our investigations improves controllability by standing passengers. This new design of the braking request allows a higher deceleration and jerk request and thus has the potential to further close the gap between targeted scenarios and technical feasibility with respect to risk of passenger harm. However, we conclude and point out that solving the remaining fatal accidents with pedestrians that are reported by some sources will be a tough challenge and not solvable by onboard technology such as ADAS or automated driving alone.

Autonomous Driving is creating a revolution within the metro bus industry, offering significant opportunities to reduce costs, improve operational efficiencies, and modernize public transportation.

The AUTOSAR community has grown to over 280 partner companies since the first AUTOSAR Classic Platform specification was released more than twelve years ago. Based on the well-established high-quality standards, mature processes and strong communication channels inside this organization, AUTOSAR has been developing a completely new approach to cope with the challenging market trends in the automotive industry such as internet access in cars, highly automated driving and vehicle to vehicle communication. The result of these activities is an intelligent and flexible software infrastructure which therefore is named AUTOSAR Adaptive Platform. The software platform runs on high-end computing hardware and supports parallel processing on many core systems and GPUs. Since AUTOSAR has its roots in the automotive field, the partnership’s prioritization of safety and security features is as self-evident as the compatibility to systems based on the AUTOSAR Classic Platform. A suitable software framework for safe and secure applications is now available.

Automated driving generates a huge amount of data that is not owned by anyone. Numerous companies from almost all industries can monetize the automated data. However, different regulations must be observed for the different types of data: depending on whether they are personal or non-personal, the GDPR applies or not. Within the scope of application of the GDPR, driver consent will usually be required. Data can be commercialized by increasing revenue, reducing costs and increasing security. Legally difficult to assess is the access to the data for third parties, which is necessary for the sale of the data. Solutions can be neutral or proprietary platform models, but the GDPR-compliant implementation remains a legal challenge because it is questionable if driver consent is really given voluntarily. The permanent data exchange between manufacturer and car has a cost-reducing effect and advertisers can save money by tailored and well-placed adverts. However, the monetization of automotive data brings new competitors onto the market, especially IT companies. In order not to be left behind technologically, manufacturers should form strategic (crossindustry) alliances with start-ups. Automotive data could also accelerate and simplify criminal and civil proceedings by facilitating the production of evidence, which would also allow the data to be used commercially. But to make automated driving possible worldwide, legislation must break new ground and internationalization of law is necessary for cross-border traffic.

The auto industry is on the cusp of its greatest transformation since the advent of the internal combustion engine. Electrification, car sharing, connectivity and autonomy may forever change the way will build and use vehicles, and if we use this opportunity well, we will finally be able to build a sustainable road transportation system. Today’s system results in major environmental problems, congestion and more than 1.35 million road fatalities every single year.

In the contribution, a model predictive trajectory tracking approach is presented. Due to the utilization of an accurate prediction model, which considers not only the vehicle dynamics but also the limited actuator dynamics, the approach can be used even in emergency collision avoidance systems. The approach explicitly predicts a trajectory set for defined control inputs. Out of the set, the trajectory which is closest to the reference is selected. Two different objective functions are defined, each of them selecting the optimum input variable for trajectory tracking. On the one hand, the selection is based on the predicted position trajectories and, on the other hand, on the speed and yaw rate of the trajectory set. The evaluation is carried out in the simulation with a vehicle model for which different error sources, like sensor errors, sensor noise, and static friction are modeled using data from a real vehicle. This development method allows a direct and fast transferability into the real test vehicle.

Gated imaging has become a promising technology for self-driving cars under adverse weather conditions because this technology is able to suppress backscatter efficiently. Moreover, gated images do not only provide intensity images but can also generate perfectly aligned depth maps. Recently, a benchmark for depth estimation in adverse weather conditions has been recorded in a fogchamber with limited length. To evaluate and benchmark gated depth estimation in these conditions, we propose a novel short-range gating scheme that is adapted to the fogchamber range. We show that gated depth estimation performs significantly more stable in adverse weather conditions compared to other stateof-the-art 3D sensing methods such as monocular depth estimation, stereo vision, and LiDAR depth completion.

Driver assistance systems and automated driving functions use environmental sensor systems to obtain input information for the realisation of various customer functions. This requires sensor integration into the overall vehicle, dealing with competing requirements. These include the consideration of the sensors specific requirements, seamless integration onto the vehicle and robustness of the sensor function against minor accidents, repairs and misuse. The focus of interest is the performance of the sensor system regarding various integration conditions. Appropriate tests are realized during development process. Manual test procedures partly executed under extreme environmental conditions are state of the art.

An automated test system is developed, based on test procedures of environmental sensors used for parking assistance that works with ultrasonic sensors. Due to a high degree of automation, it is possible to increase the test frequency and reduce the costs per test execution at the same time. Furthermore, lower tolerances and a higher reproducibility improve the quality of the measurements. In addition, the user benefits from an automated processing of the measured values. The test system is applicable in the entire parameter interval of the environmental conditions of the known manufacturer and working circuit specifications.

Based on test series the technical and economic goals using the automated test system are evaluated. This article presents a detailed description and discussion of achieved results regarding a comparison of manual and automatized tests. Therefore, the comparison focusses on test frequency and measurement results.

The described concept of the test automation for ultrasonic sensor systems, known as EDscene, has a modular design. This allows an adaption to test systems of various sensor systems with minimal effort. Furthermore, the development of this test system offers the potential to carry out multiple test sequences in parallel. Currently, tests are executed sequentially in the front and in the back of the vehicle with a need of an initial setup for each procedure. In the future, a transformation towards a circulating system could allow the users to test environmental sensors with a single initial setup. However, the measurements of the article are based on a portable version of EDscene. A stationary, laboratory setup could additionally increase both efficiency and accuracy.

In recent years, neural networks have shown astonishing results as data generators, especially in settings of generative adversarial networks (GANs). One special application is the field of style transfer, where the goal is to transform a data sample with a certain style (e.g. from sunny to rainy weather conditions) while keeping its content unchanged. This application is particularly interesting in the field of highly autonomous driving, where a major task is to generate appropriate data sets to verify and validate various kinds of algorithms. In real-world measurements, the data is usually recorded in the form of unpaired training examples, meaning that contentequivalent samples from both domains do not exist. In this case, the required GAN arrangement is complex and limits the highest processible data resolution. In this paper, we solve this issue by processing the data on the level of overlapping subsamples and present several high-quality results on image data as well as radar point clouds.

Radar sensors are a key component of automated vehicles. The requirements for radar perception modules are growing more demanding. At the same time, the radar sensors themselves are becoming increasingly sophisticated. Both developments lead to the progression of very complex algorithms. In the field of machine learning, increased task difficulty is often managed by using various types of deep neural networks (DNN). Deeper and more complex network structures allow for achieving results that had, until recently, been considered unattainable. In order to make use of this new set of machine learning algorithms, particular attention must be paid to the quality of the input data. This article gives an overview of some of the most promising ideas that will define the near to mid-term future in the field of DNNs in automotive radar perception. Contrary to image- or lidar-based approaches, the main challenge towards using DNNs on automotive radar data is information sparsity at a perception level. This currently prevents riding on the wave of recent successes in the object detection area. Another difficulty is the need for large amounts of labeled data. For information sparsity, important solutions such as high resolution processing or the utilization of low-level data layers and polarimetric radars are discussed. Furthermore, the annotation problem is derived using an example and a practical solution for the realization of an auto-labeling system is described.

Due to the growing number of ADAS in modern vehicles and the advent of autonomous driving, the testing effort will massively increase. The only way to overcome this will be a significant parallelization of the testing effort. Due to a lack of manpower, available proving ground space, an ever-decreasing number of test vehicles and simple cost reasons, this can only be achieved with massive transition towards virtualization.

Besides various legal and societal constraints, the technical requirements of a virtual test environment need to be set up, understood and solved. It needs to comprise all properties of a real test environment. It is therefore essential to embed a virtual development process in the overall development process instead of as a separate niche application.

Even though all of the above might be in place, it is not enough to run large-scale tests. In the case of large numbers of deterministic scenarios, the virtual development landscape needs to be moved to HPC and the cloud. This is even more crucial when it comes to stochastic / random / non-deterministic testing.

Using centrally managed and structured data, ZF Friedrichshafen, PDTec and IPG Automotive have developed a data management solution which addresses the challenges related to the cross-domain use of simulation for the development and validation of automated driving functions.

The presentation will give an overview of what such a virtual development landscape might look like and how it can be embedded in the overall development process at OEMs and Tier 1s.

Scenario-based virtual testing is seen as a key element to bring the overall safety validation effort for automated driving functions to an economically feasible level. In this work, a generic and modular architecture for simulation of automotive perception sensors is introduced, as part of the overall virtual testing pipeline. It is based on the functional decomposition of real world perception sensors. All interfaces between the individual modules of the model architecture are oriented on internationally recognized standards and therefore facilitate a high degree of interchangeability. In addition, a wrapper framework handles all outer communication and enables a profile-based parameterization of the model, where every profile reflects a specific set of parameters tailored to the specifications and use case of the end user.

This paper evaluates different deep learning based depth estimation algorithms. We propose improvements for a state-of-the-art unguided depth completion method where the number of necessary parameters can be more than halved at unvarying accuracy. Based on the results of the depth estimation evaluation, we consider the performance of semantic segmentation methods. We investigate if the completion improves the accuracy of point cloud based segmentation. The results are compared to the segmentation accuracy using only measured sensor data. Moreover, we give a comparison to the segmentation based solely on predicted depths of a monocular camera. Here, we depict the differences in accuracy when a costly lidar sensor is economized. The results are further validated on a self-provided dataset recorded with the institute’s own test vehicle.

The development of autonomous driving and driver assistance systems that involve the lateral guidance of the car has the crucial challenge of trajectory calculation. Under the many possible solutions to this problem, the use of a model predictive controller (MPC) is one of the most popular. This article presents a method of calculating a trajectory using an iterative moving time horizon MPC combined with a forwardbackward velocity estimation to guess the initial solution. The discretization in the time domain provides the possibility to easily extend the method for the use with multiple vehicles while the linearization provides very high performance combined with low model complexity. The drawback of the iterative solution which usually would lead to a loss of safety due to possible errors by unfeasible problem formulation is compensated with the efficient guessing of an initial solution. To evaluate the performance of the method it is applied to the problem of global trajectory calculation for autonomous racecars. The presented approach proves to be superior to other methods that only use the geometry of the road while still being able to calculate a global trajectory for a whole circuit. The comparison with other methods like a minimum curvature optimization and the driver of the vehicle simulation program IPG Carmaker shows clear advantages both in terms of lap time and computation time while only using a very simple problem formulation. Also the results show a strong relationship between the driver, the trajectory and the driving performance which is especially relevant to emergency evasive maneuvers where the best use of the potential of the car is crucial for the driver’s safety.

At Advanced Driver Assist Systems (ADAS), and especially in the case of automated vehicles, the reliable functioning of sensors is absolutely prerequisite to safe mobility. But, for various reasons, sensor abilities to sense and interpret the surroundings can get lost about the lifetime of cars. Possible causes are ageing, fault, disassembly respectively assembly of sensors without adjustment, subsequently bonded wrap film, collisions and repairs. This paper presents the results of a study to the impact of body repair on radar sensors. Tested system was a Lane Change Assistants. The detection of vehicles is frequently carried out via radar sensors. To testing the functionality of ADAS and automated vehicles, test runs under real conditions were carried out. The results show, that body repairs can have a significant impact on sensor performance. Therefore, adequate repair methods and whose professional realization extremely important. In order to implement a professional repair clearly instructing in OEM repair manuals are extremely important. This information must be easily found for body shops, insurers and experts. In addition, repair and body shops are required to invest into employee training and qualification, adequate (manufacturer-specific and cost-intensive) adjustment devices (calibration targets) and diagnosis tools to execute a professional repair. Especially multibrand repair and body shops are affected by this challenge.