This contribution presents a novel probabilistic approach for the generation of discretionary lane change proposals with a focus on highway driving situations. The developed model is based on the quantification of the utility of driving lanes. It generates a lane change proposal if the current driving lane is unsatisfactory in the sense that the desired velocity of the automated vehicle is undershot because of a slow preceding vehicle. A driving simulator study was conducted to create a dataset for the optimization of the model parameters. The optimization goal is to accurately match the timings of the lane change intentions of all participants. Finally, the applicability of the model is shown on real data from a test vehicle.

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.

This paper presents a method to simulate various automotive sensors based on functional properties. An extraction of the road users in the surroundings is determined by Deep Learning based object detectors to generate a sufficient perception of the vehicle environment. Simultaneously, it is important that the data processing runs in real time and with a high detection accuracy to ensure a safe maneuvering of the vehicle. For this reason the object detectors considered in this paper are the single shot multibox detector with MobileNet as base network for image processing and the Sparsely Embedded Convolutional Detection detector for lidar point clouds. The dataset used for the evaluation is KITTI. The detection accuracies for different object distances are investigated and compared. To create a functional simulation for the different sensor systems, the evaluation of the object detection for various sensor types is used to model the respective sensor behavior and its properties. The modeled object detectors are evaluated with respect to the detection accuracies of the real detectors. These sensor models can then be used in simulations for different sensor topologies.

Menschliche Fahrfehler stellen die Hauptursache für Unfälle im Straßenverkehr dar. Die automatische Verkehrsüberwachung bietet einen Beitrag, um die Vision des unfallfreien Straßenverkehrs zu erreichen. Eine solche Infrastruktur erhöht unmittelbar die Verkehrssicherheit insbesondere vor dem Hintergrund einer langwierigen Durchdringung des Fahrzeugbestandes durch neue Assistenzsysteme. Das Multi Funktionale Detektions System erkennt an Autobahnabfahrten sowie Park- und Rastanlagen potenzielle Falschfahrer bei der Auffahrt auf die falsche Richtungsfahrbahn. Typischerweise werden dabei ortsfeste Messplattformen eingesetzt, hierfür bieten Kamerasensoren besonders günstige Voraussetzungen. Der vorliegende Beitrag betrachtet die Fahrzeugdetektion in Kamerabildern für die Integration in ein Falschfahrerwarnsystem. Für diese Anwendung ist die Realisierung von möglichst schnellen Verarbeitungszeiten entscheidend. Gängige Methoden der kamerabasierten Objektdetektion führen zu diesem Zweck eine vollständige Abtastung des aufgezeichneten Bildes durch. Aufgrund der feststehenden Kameraposition und der bekannten statischen Verkehrselemente lässt sich jedoch der Suchraum deutlich einschränken. Der Ansatz basiert auf einer reduzierten Suchstrategie, welche die vorteilhaften Eigenschaften herkömmlicher Verfahren zur Objektdetektion erhält.

The development of algorithms for automated driving is a very challenging task. Recent progress in machine learning suggests that many algorithms will have a hybrid structure composed of deterministic or optimization and learning based elements. To train and validate such algorithms, realistic simulations are required. They need to be interaction based, incorporate intelligent surrounding traffic and the other traffic participants behavior has to be probabilistic. Current simulation environments for automotive systems often focus on vehicle dynamics. There are also microscopic traffic simulations that on the other hand don’t take vehicle dynamics into account. The few simulation software products that combine both elements still have at least one major problem. That is because lane change trajectories disregard human driving dynamics during such maneuvers. Consequently, machine learning algorithms developed and trained in simulations hardly generalize to non-synthetic data and therefore to real-world applications.

This letter presents automated driving using polygon clipping (A-D-PolyC), a novel framework for lane change behavior planning of automated vehicles on highways. It assumes that a mission planning layer generates lane change requests. In crowded traffic scenes, various variants for the lane change execution arise. The developed algorithm identifies the maneuver variants deterministically and in bounded runtime using polygon clipping in spatiotemporal domain. The variants are represented using a graph that captures the scene topology. Afterward, it efficiently samples optimal lateral and longitudinal trajectories with regard to vehicle dynamics. The satisfaction of hard constraints is checked. Finally, a scene prediction is conducted for a subset of the sampled trajectories. In comparison with most state of the art approaches, the framework accounts for traffic interaction and collects meaningful features. A novel measure based on generalized kinetic energies for the impact of a maneuver execution on the whole traffic scene is introduced. A-D-PolyC is a hierarchical, holistic, modular, and parallelizable concept that can be coupled with different scene prediction engines and lower-level local trajectory planners. The performance of A-D-PolyC is statistically evaluated using randomized simulation runs and compared to a state of the art approach.

Zusammenfassung Der vorliegende Beitrag untersucht die Erkennung benachbarter Fahrstreifen auf Grundlage von Kamerabildern. Hierbei wird sowohl die Anzahl befahrbarer Fahrstreifen als auch deren Verlauf innerhalb eines festgelegten Bereichs vor dem Fahrzeug bildbasiert geschätzt. Die Erkennung erfolgt durch Convolutional Neural Networks. Der Beitrag bewertet die Güte der Schätzung und Genauigkeit der Fahrstreifenerkennung. Zur Interpretation der Entscheidungen werden die durch das neuronale Netz gelernten Merkmale und Zwischenrepräsentationen visualisiert.

This contribution presents A-D-PolyC QP (Automated Driving using Polygon Clipping and Quadratic Programming), a framework for lane change behavior planning of automated vehicles on highways. It solves the tactical decision problem that arises through the presence of more than one lane change maneuver variant in traffic scenes. A-D-PolyC QP derives all variants deterministically using polygon clipping in spatiotemporal domain and is able to deal with lane changes of surrounding traffic participants. All variants are encoded in a directed graph. The nodes in the graph correspond to spatiotemporal free space polygons and allow the derivation of constraints for trajectory optimization in a Frenet-Serret coordinate frame. Novel, linear time-variant Time-To-Collision and Time Gap constraints based on geometric boundaries are introduced. Interaction-awareness is incorporated by forward simulation of the optimized trajectories. Finally each maneuver is assessed and features are calculated that allow the decision for one lane change maneuver by a higher-level function.

Automated driving is a key technology for the future of transportation. There are several motivations to develop automated vehicles. First and foremost, it promises to reduce the number of traffic accidents. Figure 1 shows the accidents recorded by the German police over the past years ([1]) ranging back to 1960.

In this contribution, we introduce an algorithm that allows the estimation of a transformation between two sensor systems. The primary system consists of non-imaging sensors like for example lidar and radar or even a fused combination of these. A camera always represents the secondary system. The algorithm aims at automating the process of calibration entirely during driving. This way, much flexibility and robustness are guaranteed. Changing the placement of the camera remains possible without the need to do a cumbersome manual calibration.