Search

In this paper, we investigate data structures for grid mapping in autonomous driving. While the traversed environment can be virtually infinitely large, one is mainly interested in the immediate surroundings of the vehicle. Thus, a world-fixed, square grid map centered roughly at the vehicle is used. For an efficient data handling, this grid map is subdivided into multiple submaps, which can be added and removed in order to move the currently considered window with the vehicle. We evaluate different choices for the underlying data structures (array, hashmap, quadtree) as well as different submap sizes on both a theoretical and practical level: We provide a complexity analysis and a discussion on possible overheads. Furthermore, we present empirical results regarding memory consumption and runtime of common operations (mapping, scan matching, map retrieval) on real-world datasets of different scenarios (sub-urb, rural area, highway). Finally, we give recommendations depending on the actual requirements of the used algorithm.

Algorithms for controlling fully autonomous systems must meet especially high requirements with respect to safety and robustness. A particularly challenging example are autonomous deep space missions, which we investigated in several projects. In this context, we showed that a safe and robust autonomous system can be realized through nonlinear model predictive control approaches using optimization techniques in combination with multi-sensor fusion based on an extended representation of uncertainty. The focus of this paper is on demonstrating the versatility of that concept by transferring the corresponding algorithms to the also very challenging application of autonomous driving. In particular, we propose a system concept for a self-driving car based on our methodology. Furthermore, we present results of a real world research vehicle that autonomously explores a parking lot, dynamically takes obstacles into account, and finally performs a parking maneuver.

Relative localization in GNSS-denied environments is an essential task in advanced driver assistance systems or autonomous driving. In this paper, we present a loosely coupled visual-multi-sensor odometry algorithm for relative localization. Inertial Measurement Unit (IMU), vehicle speed, and steering angle measurements are fused in an Unscented Kalman Filter (UKF) to estimate the vehicle’s pose and velocity as well as the IMU biases. Relative pose estimates between two camera frames are used to update the UKF and further increase precision in the estimated localization. Our system is able to localize a vehicle in real-time from arbitrary states such as an already moving car which is a challenging scenario. We evaluate our visual-multi-sensor algorithm on real-world datasets recorded in inner-city and rural areas and compare it two state-of-the-art Visual-Inertial Odometry (VIO) algorithms. We report a lower relative odometry error in particular at the start of motion estimation with lower computational cost.

Object detection in real-world applications such as autonomous driving scenarios is a challenging issue since objects often occlude each other. 3D object detection has achieved high accuracy and efficiency, but detecting small object instances and occluded objects are the most challenging issues to deploy detectors in crowded scenes. Our main focus in this paper is deep multi-modal based object detector in an automated driving system with early fusion on 3D object detection utilizing both Light Detection and Ranging (LiDAR) and image data. We aim at obtaining highly accurate 3D localization and recognition of objects in the road scene and try to improve the performance. In this regard, our basic architecture follows an established two-stage architecture, Aggregate View Object Detection-Feature Pyramid Network (AVOD-FPN), one of the best among sensor fusion-based methods. AVOD-FPN has yielded promising results especially for detecting small instances. Moreover, another main challenging issue in autonomous driving is detecting the occluded objects. So we try to address this difficulty by integrating attention network into the multi-modal 3D object detector. Experiments are shown to produce state-of-the-art results on the KITTI 3D sensor fusion-based object detection benchmark.