Holger Mielenz

Toward automated driving in cities using close-to-market sensors: An overview of the V-Charge Project

Future requirements for drastic reduction of CO2 production and energy consumption will lead to significant changes in the way we see mobility in the years to come. However, the automotive industry has identified significant barriers to the adoption of electric vehicles, including reduced driving range and greatly increased refueling times. Automated cars have the potential to reduce the environmental impact of driving, and increase the safety of motor vehicle travel. The current state-of-the-art in vehicle automation requires a suite of expensive sensors. While the cost of these sensors is decreasing, integrating them into electric cars will increase the price and represent another barrier to adoption. The V-Charge Project, funded by the European Commission, seeks to address these problems simultaneously by developing an electric automated car, outfitted with close-to-market sensors, which is able to automate valet parking and recharging for integration into a future transportation system. The final goal is the demonstration of a fully operational system including automated navigation and parking. This paper presents an overview of the V-Charge system, from the platform setup to the mapping, perception, and planning sub-systems.

Understanding and Applying the Concept of “Driver Availability” in Automated Driving

Several levels of automated driving functions require the human as a fallback driver in case system performance limits are exceeded. Human factors research in this area is especially concerned with human performance in these take-over situations and the influence of the driver state. Based on work of the publicly funded project Ko-HAF the paper introduces a comprehensive model of the transition process from automated driving to manual driving and specifies relevant time stamps and time windows. The concept of Driver Availability is regarded as a quantitative measure that relates the estimated time required to safely take-over manual control to the available time budget. A conceptual framework outlines potential influencing factors on driver availability as well as ways to apply the measure in a real-time application.

Precise vehicle localization in dense urban environments

In this contribution we introduce a framework for precise vehicle localization in dense urban environments which are characterized by high rates of dynamic and semi-static objects. The proposed localization method is specifically designed to handle inconsistencies between map material and sensor measurements. This is achieved by means of a robust map matching procedure based on the Fourier-Mellin transformation (FMT) for global vehicle pose estimation. Accurate and reliable relative localization is obtained from a LiDAR odometry. Consistency checks based on the cumulative sum (CUSUM) test are instrumented for rejection of inconsistent map matching results from the fusion procedure. Our key contributions are: i) Introduction and adaptation of a spectral map matching procedure based on the FMT for urban automated driving, ii) Presentation of a framework for efficient and robust localization in dense urban environments based on a novel LiDAR odometry, map matching, wheel odometry and GPS, iii) Proposal of a procedure for localization integrity monitoring which leads to significantly increased pose estimation accuracy. Evaluation results show the superior performance of the proposed approach compared to another state-of-the-art localization algorithm for a challenging urban dataset. All maps were recorded two years in advance of the evaluation test run. Furthermore, different LiDAR-based sensor setups were used for mapping and localization.

Localization accuracy estimation with application to perception design

Landmark-based localization in dynamic environments poses high demands on the perception system of a mobile robot. The pose estimate generally has to fulfill specific accuracy requirements which might be necessitated by dependent systems, such as behavior planning. Thus, in this contribution we focus on the model-based derivation of perception requirements, i.e. detectable landmark types and minimum detection rates, to enable global localization with a specified upper bound on uncertainty. To this end, we utilize stochastic geometry to accurately capture and explicitly consider characteristics of the dynamic environment (e.g. occlusions), and the perception system (e.g. missed detections). From this point our contributions are twofold: i) We propose an analytical model of upper bounds on localization uncertainty. For continuous pose tracking, the Kalman filter equations for intermittent observations are considered and ii) perception requirements, i.e. minimum detection rates, based on specified upper bounds on pose estimation uncertainty are derived. Monte Carlo simulations are used to demonstrate the performance of the proposed methods.

Model-Based Derivation of Perception Accuracy Requirements for Vehicle Localization in Urban Environments

In this contribution, we address the model-based derivation of perception requirements based on upper bounds on vehicle localization uncertainty for urban driver assistance (UDA) and urban automated driving (UAD). We show that a probabilistic model for the estimation of map-relative localization accuracy can be obtained and utilized for proper parametrization of a perception system. Therefore, the paper at hand entails two main contributions: i) Proposal of a probabilistic model for localization accuracy in closed form under the assumption of a generic measurement model with Gaussian noise and a stochastic landmark distribution, ii) Presentation of a framework for model-based derivation of perception requirements which permit desired localization performance. To exemplify the application of our method, sensor parameters for a stereo vision system (e.g. stereo base-width) are determined and verified via comprehensive simulation experiments. This is conducted in the context of an urban automated lane keeping system under explicit consideration of non-existent or occluded lane markings and curb stones.

GridTiles: A method for modelling a spatial unconstrained environment

The objective of this paper is the modelling of an unbounded environment of a human-driven car that may contain multilevel structures such as bridges or parking decks. Such a model might be used by a driver assistant system (DAS) where one drives through an urban environment, requests for an assistance and the DAS should immediately be able to give the user the required support. E.g. it can guide through a narrow passage or a turn. For such an assistance an environment model is needed that runs in real-time. But to keep the system at low cost the required memory should be as small as possible. Hence the algorithm should be optimized with respect to computational power and memory consumption. A new approach is proposed that models the environment by incrementally adding small tiles at places where obstacles created measurements. Each tile contains an occupancy grid and some neighbourhood relations. By modelling the world this way, a compact representation of the environment is created that is aligned at the user-given trajectory and the allocation of obstacles. By using a grid based algorithm efficient and hence fast techniques can be used to work on the world representation. Also an extension is introduced to restrict the required memory to a given limit and concurrently map the local obstacles but avoids any transformation of historic data. A comparision with state of the art algorithms was made and the capability of the proposed algorithm is demonstrated with some experimental results.

A data-driven approach for pedestrian intention estimation

In the context of future urban automated driving many important problems remain unsolved. A critical one is the analysis and prediction of pedestrian movements around urban roads. Especially the analysis of non-critical situations has not received much attention in the past. This paper focuses on analyzing and predicting movements of pedestrians approaching crosswalks, a very crucial pedestrian-vehicle interaction in urban scenarios. In our previous work, we analyzed the performance of a data-driven Support Vector Machine-based architecture, and the relevance of specific features to infer pedestrian crossing intentions. In this paper, we will use our previous results as baseline to compare against an architecture based on neural networks for time-series classification. In particular we analyze the effectiveness of dense and Long-Short-Term-Memory networks. Furthermore, we will be looking into enhancing our feature vectors by adding LiDAR based images to the classification process. Additionally the evaluation provides an estimate for the temporal prediction horizon. The approaches presented are validated with real world trajectories recorded in Germany. Our results show an average accuracy improvement of 10-20% with respect to our previous Support Vector Machine-based approach.

Perzeption für robuste Fahrzeuglokalisierung

Eine hinreichend genaue und zuverlassige Fahrzeuglokalisierung ist ein fundamentaler Bestandteil vieler moderner Fahrerassistenz‐ und hochautomatisierter Fahrzeugsysteme. Die Schatzung einer globalen Fahrzeugpose ist haufig erforderlich um zusatzliche Informationen, wie z.B. Fahrbahnverlaufe, aus einer digitalen Karte zu entnehmen und als Grundlage fur eine fundierte Verhaltensplanung und ‐entscheidung verwenden zu konnen. Minimalanforderungen an die Genauigkeit der Fahrzeuglokalisierung resultieren aus Art und Anwendungskontext der verwendeten Umfeldinformationen.

Effiziente Umfeldmodellierung für Fahrerassistenzsysteme im Niedergeschwindigkeitsbereich

Mit der steigenden Automatisierung von Fahrerassistenzsystemen wachsen auch die Anforderungen an die eingesetzten Verfahren zur Umfeldmodellierung. Soll sich ein Fahrzeug weitgehend selbststandig in einem vorab unbekannten und undefinierten Umfeld bewegen, so ist eine genaue Kenntnis der Aufenthaltsorte von statischen sowie dynamischen Objekten von hoher Relevanz, um Assistenzfunktionen, wie Kollisionsvermeidung oder Pfadplanung, durchfuhren zu konnen. Die Problematik der Umfeldmodellierung ist in der Robotik seit vielen Jahren ein bekanntes Problem und gut untersucht [7]. Demnach ist es bereits eine hochkomplexe Aufgabe, wenn ein autonomer Roboter in einem gut strukturierten Gebaude interagiert, die Zeit hat Messungen in ein Umfeldmodell zu integrieren und in der Lage ist, seine Bewegung einzustellen, um Messungen verwerten [5].

Multi level modeling and loop closing with GridTiles

The objective of this paper is to use GridTiles to model a complex multi level scenario and additionally use the resulting model to help solving the loop closing problem. The GridTile algorithm is a memory and computational efficent method to map an unbounded outdoor environment of a human driven car. GridTiles are a set of small occupancy grids with a neighbourhood relation in the four directions. This paper discusses four of the major problems in modelling: the local and the global mapping, multi level modelling and loop closing. For each of these problems possible solutions will be shown. Local and global maps can be directly modelled with the standard GridTile algorithm. Because the neighbourhood of the tiles is only known locally, multi level scenarios can be modelled implicitly. Additionally a loop closing solution is proposed to correct the built model afterwards in a memory efficient way. Instead of the classic approach that has to store all measurements over an unpredictable duration, the GridTiles are used as a data storage.