This video presents the Autonomous City Explorer (ACE) project. Its goal was to create a robot capable of navigating unknown urban environments without the use of GPS data or prior map knowledge. The robot had to find its way solely by interacting with pedestrians and building a topological representation of its surroundings. This video outlines the necessary ingredients for successful low-level navigation on sidewalks, information retrieval from pedestrians as well as the construction of a semantic representation of an urban environment. A system architecture for outdoor localization, traversability assessment, path planning, behavior selection and topological abstraction in urban environments is presented.

In the autonomous city explorer (ACE) project a mobile robot is developed, which is capable of finding its way to a given destination in an unknown urban environment. An exemplary mission is to find the way from our institute to the Marienplatz, a public place in the center of Munich, without any prior knowledge or GPS information. Inspired by the behavior of humans in unknown environments, ACE must find its way by asking pedestrians. The route is about 1.5 kilometers far and includes heavily traveled roads and crowded public places. In order to navigate safely in an unknown urban environment, some challenges arise for the vision system. Robust human detection, tracking and the estimation of human body poses is essential for natural interaction with pedestrians. Furthermore, the robot needs to be able to detect sidewalk and crossroads. A visual odometry system is used to support the conventional navigation. Outdoor experiments were conducted twice successfully. After about 5 hours and interacting with 25 and 38 persons respectively, ACE arrived the Marienplatz. This paper describes both, an architecture of the vision system used for ACE and the algorithms used to deal with the described challenges.

Knowledge about the environment is essential for humanoid and mobile robots to move and act safely. The most intuitive way to perceive information about the environment is through the vision system. However, the accuracy provided by stereo vision is insufficient for many tasks. A more accurate representation is created by a laser range-finder, which delivers no color information. This paper describes a novel approach to merge data obtained by stereo vision and laser range-finders by using a genetic ICP algorithm, which is able to register noisy point clouds with different resolutions and a small overlap. Furthermore, it describes an easy-to-use and robust method to calibrate the extrinsic parameters of two or more laser range-finders.

Inspired by the expectation-based perception of humans, a surprise-driven active vision system is proposed. This vision system not only considers spatial saliency of objects in the environment, but also investigates temporal novelty in the neighborhood. Surprise is defined as the difference of the saliency probability distributions of two consecutive input images, which is measured using Kullback-Leibler divergence. The high-speed gaze shift capability of the camera platform and the parallel computation with the aid of GPUs enable a real-time tracking of the surprising event.

Estimating the human body pose is of great interest for many tasks, such as human robot interaction, people tracking and surveillance. During the recent years, several approaches have been presented, which still have weaknesses regarding occlusions or complex scenes. In this paper, we present a novel algorithm for human body pose estimation using any three-dimensional representation of the environment, like stereo vision. The presented algorithm is able to leave out body parts and is therefore able to deal with occluded body parts. In a first step, possible humans need to be detected, e.g. by using a skin color filter. A disparity map containing depth information is computed using a stereo matching algorithm. It leads to a three-dimensional representation of the scene. Starting with the detected skin parts, our algorithm segments this point cloud into smaller clusters. The possible matches are then verified, and the body pose is estimated using a kinematic human model with 28 degrees of freedom. As our algorithm is capable of dealing with arbitrary three-dimensional representations, it can easily be adapted to use a three-dimensional laser range finder instead of a stereo camera system.