Haselberger, Johann
Refine
Document Type
Language
- English (4)
Publication reviewed
- begutachtet (4)
Keywords
- Automated Driving (1)
- Ground Truth MB (1)
- KI (1)
- MLPaSSAD (1)
- advanced driver assistance systems (1)
- automated driving (1)
- automated lateral control (1)
- deep learning (1)
- driving experience (1)
- head-up display (1)
During the development of state-of-the-art driver assistance systems and highly autonomous driving functions, there is a demand for reliable research vehicle platforms that can be used in a variety of applications. Especially for data-driven machine learning approaches, a large amount of measurement data obtained from multimodal sensors is needed. This paper presents a Robot Operating System (ROS) based prototype vehicle that is built on a Porsche Cayenne, which provides a dedicated test environment for autonomous research. To bridge the gap between pure research and actual production vehicles, the platform features near-series placement of sensors and the use of the built-in camera and actuators. Open-source packages and a containerized software architecture make the system reusable and easy to extend in terms of hardware and algorithms. Furthermore, we describe our approach for data recording and long-term persistence.
Unsupervised Domain Adaptation demonstrates great potential to mitigate domain shifts by transferring models from labeled source domains to unlabeled target domains. While Unsupervised Domain Adaptation has been applied to a wide variety of complex vision tasks, only few works focus on lane detection for autonomous driving. This can be attributed to the lack of publicly available datasets. To facilitate research in these directions, we propose CARLANE, a 3-way sim-to-real domain adaptation benchmark for 2D lane detection. CARLANE encompasses the single-target datasets MoLane and TuLane and the multi-target dataset MuLane. These datasets are built from three different domains, which cover diverse scenes and contain a total of 163K unique images, 118K of which are annotated. In addition we evaluate and report systematic baselines, including our own method, which builds upon Prototypical Cross-domain Self-supervised Learning. We find that false positive and false negative rates of the evaluated domain adaptation methods are high compared to those of fully supervised baselines. This affirms the need for benchmarks such as CARLANE to further strengthen research in Unsupervised Domain Adaptation for lane detection. CARLANE, all evaluated models and the corresponding implementations are publicly available at https://carlanebenchmark.github.io.
The focus of this publication is on the development of lane-precise “Ground Truth” (GT) maps for the objective quality evaluation of automated driving functions. Therefore, attention is paid to the proper measurement of road geometry. The road geometry forms the basic layer of the HD maps.
A new map format Curved Regular Objects (CRO) is developed, which is based on the idea of OpenCRG®. For the evaluation of current Advanced Driver Assistance Systems (ADAS) an accurate High Definition (HD) maps as GT are necessary. This makes it possible to locate the high precision vehicle position and motion with centimeter accuracy. The aim is to achieve maximum accuracy of the absolute 3D positions when measuring lanes. This method for the generation of highly accurate GT maps promises an absolute accuracy of < ± 0.05 m. Various research activities benefit from the exact street reference at the Kempten University of Applied Sciences (UAS Kempten) in the Adrive Living Lab. First of all, the publication deals with the current Lane Keeping Assistant Systems (LKAS). The accuracy of the vehicle’s localization on the GT map and an objective evaluation of the LKAS is shown. In addition, the CRO data is used as a virtual sensor for the steering assistant in real time. Another application is the Visual Range Finder (VRF), which requires less computation power by using the CRO data. In addition, a current LKAS camera sensor performance is evaluated using CRO maps.
Deep learning for lateral vehicle control – an end-to-end trained multi-fusion steering model
(2019)
Deep Learning based behavior reflex methods found their way into modern vehicles. To model the human driving behavior it is not sufficient to rely solely on individual, noncontiguous camera frames without taking vehicle signals or road specific features into account. In this work four temporal fusion methods are evaluated based on three different Deep Learning models. The proposed spatio-temporal Mixed Fusion model extends the present end-to-end models and consist of multiple levels of fusions. The raw image data from a single front facing camera is mixed with recorded vehicle data and a map based predicted road bank angle gradient vector. The model accesses multiple time axes: temporal features of multiple image frames are extracted through a combination of Convolution and LSTM layers while it can also make assumptions about the future road condition with the use of upcoming Ground Truth road bank angle changes. Experiments are performed on a recorded data set of real world drivings. Results show, that this approach leads to an accurate imitation of the human driver with an inference capability of more than 60 FPS.