Marcel Häselich

Moving vehicle detection and tracking in unstructured environments

The detection and tracking of moving vehicles is a necessity for collision-free navigation. In natural unstructured environments, motion-based detection is challenging due to low signal to noise ratio. This paper describes our approach for a 14 km/h fast autonomous outdoor robot that is equipped with a Velodyne HDL-64E S2 for environment perception. We extend existing work that has proven reliable in urban environments. To overcome the unavailability of road network information for background separation, we introduce a foreground model that incorporates geometric as well as temporal cues. Local shape estimates successfully guide vehicle localization. Extensive evaluation shows that the system works reliably and efficiently in various outdoor scenarios without any prior knowledge about the road network. Experiments with our own sensor as well as on publicly available data from the DARPA Urban Challenge revealed more than 96% correctly identified vehicles.

Probabilistic terrain classification in unstructured environments

Autonomous navigation in unstructured environments is a complex task and an active area of research in mobile robotics. Unlike urban areas with lanes, road signs, and maps, the environment around our robot is unknown and unstructured. Such an environment requires careful examination as it is random, continuous, and the number of perceptions and possible actions are infinite. We describe a terrain classification approach for our autonomous robot based on Markov Random Fields (MRFs ) on fused 3D laser and camera image data. Our primary data structure is a 2D grid whose cells carry information extracted from sensor readings. All cells within the grid are classified and their surface is analyzed in regard to negotiability for wheeled robots. Knowledge of our robots egomotion allows fusion of previous classification results with current sensor data in order to fill data gaps and regions outside the visibility of the sensors. We estimate egomotion by integrating information of an IMU, GPS measurements, and wheel odometry in an extended Kalman filter. In our experiments we achieve a recall ratio of about 90% for detecting streets and obstacles. We show that our approach is fast enough to be used on autonomous mobile robots in real time.

Calibration of multiple cameras to a 3D laser range finder

Robots, especially autonomous systems, need a precise perception of the environment for path planning and manipulation tasks. Different sensors provide various data for this task. Camera to 3D laser range finder calibration allows algorithms to work efficiently on depth measurements, color and texture features and benefit from the signals of both sensor types at once in real-time. In this paper we describe our extensions to an existing calibration approach from Unnikrishnan and Herbert. The calibration is analyzed from a robotic perspective with focus on improving required time, practicality and simplification of the calibration. Field of application ranges from robotics and agricultural engines to industrial applications. The feasibility of the approach is discussed and resulting data fusion is presented.

Heat mapping for improved victim detection

Disasters, such as earthquakes or tsunamis, result in destroyed buildings and other dangerous scenarios for human rescuers. Remotely controlled robots are often used to aid humans. Since those robots require steady communication, they are limited to a certain range and might get stuck in case the connection is interrupted. To augment remote controlled robots, autonomous robots are needed. These robots are able to navigate in devastated areas to detect victims while creating a map of the environment at the same time. Victim detection based on thermal sensors is the most widely used approach. In this paper we present a novel approach based on low-cost thermal sensors, using the global heat distribution. Therefore we developed a 2D heat map, which is created by the combination of thermal and laser information during continuous autonomous exploration. The map is build from the history of all sensor readings over time resulting in a heat distribution. The main contribution of this paper is the introduction of a 2D heat map accumulating thermal sensor readings over time for improved victim detection.

Confidence-based pedestrian tracking in unstructured environments using 3D laser distance measurements.

Detection and tracking of pedestrians is an essential task for autonomous outdoor robots. Modern 3D laser range finders provide a rich and detailed 360 degree picture of the environment. Unstructured environments pose a difficult scenario where a variety of objects with similar shape to a human like shrubs or small trees occur. Especially in combination with partial occlusions, sensor noise, and conclusions from traversing rough terrain.

Definition of semantic maps for outdoor robotic tasks

In recent years, a lot of work was conducted in order to give robots the ability to gather semantic information about their environment, store it, represent it for the user and to perform high-level tasks based on the semantic information. Most of the systems use internal representations of the gathered information that is not intuitively understandable by humans and that is inadequate for learning from commonly available sources. The combination of object/place classification and common-sense knowledge to semantic maps found its way into indoor semantic mapping approaches to improve human-robot interaction. The aim is to assign complex task settings to the robot allowing it to guide the search for the solution by itself. In this paper, we present a formal common definition of semantic maps. We discuss different criteria for designing and classifying semantic maps and their appropriate challenges. Furthermore, we present an outdoor semantic mapping approach incorporating common-sense knowledge into the classification process and a suitable map representation for high-level tasks.

Spline templates for fast path planning in unstructured environments

This paper addresses the problem of navigating a heavyweight robot autonomously across unstructured environments. The data of a 3D laser range finder are partitioned in real time into equidistant grid cells to locate obstacles and to classify the negotiability of the surface terrain. In addition to the distinction between obstacles or free space, ground measurements are examined to determine the local terrain roughness for each cell. We present a novel path planning algorithm operating on these classified grid cells. The algorithm is able to avoid positive and negative obstacles (cells without ground measurements), as well as regarding the roughness of the terrain. By applying a cost function, the robot is able to prefer routes across smooth terrain, like streets or trails, over routes across rough and muddy areas. We can determine the necessarily free terrain cells for each spline in the forefront because the splines and the dimensions of the robot are set. In this way we can access the resulting spline templates during runtime and connect them very fast for the path planning. The key feature of this approach is the low computation cost compared to existing approaches.

Model-based recognition of domino tiles using TGraphs

This paper presents a case study showing that domino tile recognition using a model-based approach delivers results comparable to heuristic or statistical approaches. The knowledge on our models is modeled in TGraphs which are typed, attributed, and ordered directed graphs. Four task-independent rules are defined to create a domain independent control strategy which manages the object recognition. To perform the matching of elements found in the image and elements given by the model, a large number of hypotheses may arise. We designed several belief functions in terms of Dempster-Shafer in order to rate hypotheses emerging from the assignment of image to model elements. The developed system achieves a recall of 89.4% and a precision of 94.4%. As a result we are able to show that model based object recognition is on a competitive basis with the benefit of knowing the belief in each model. This enables the possibility to apply our techniques to more complex domains again, as it was tried and canceled 10 years ago.

Markov random field terrain classification of large-scale 3D maps

Simultaneous localization and mapping, drivability classification of the terrain and path planning represent three major research areas in the field of autonomous outdoor robotics. Especially unstructured environments require a careful examination as they are unknown, continuous and the number of possible actions for the robot are infinite. We present an approach to create a semantic 3D map with drivability information for wheeled robots using a terrain classification algorithm. Our robot is equipped with a 3D laser range finder, a Velodyne HDL-64E, as primary sensor. For the registration of the point clouds, we use a featureless 3D correlative scan matching algorithm which is an adaption of the 2D algorithm presented by Olson. Every 3D laser scan is additionally classified with a Markov random field based terrain classification algorithm. Our data structure for the terrain classification approach is a 2D grid whose cells carry information extracted from the laser range finder data. All cells within the grid are classified and their surface is analyzed regarding its drivability for wheeled robots. The main contribution of this work is the novel combination of these two algorithms which yields classified 3D maps with obstacle and drivability information. Thereby, the newly created semantic map is perfectly tailored for generic path planning applications for all kinds of wheeled robots. We evaluate our algorithms on large datasets with more than 137 million annotated 3D points that were labeled by multiple human experts. All datasets are published online and are provided for the community.

Pedestrian detection in outdoor images using color and gradients

Pedestrian Detection in digital images is a task of huge importance for the development of autonomous systems and for the improvement of robots interacting with their environment. The challenges such a system has to overcome are the high inter-class variance of pedestrians and the demands of unstructured environments. Outdoor environments contain unknown regions, inhomogeneous illumination, and parts of the pedestrians can be occluded. In this work, a complete system for pedestrian detection is realized according to state-of-the-art techniques. As main features, we use the “Histograms of Oriented Gradients” in combination with the “Color Self-Similarity” feature as proposed by Walk et al. We describe and evaluate our complete detection approach and our new structure element is able to accelerate the Color Self-Similarity computations by a factor of four.