Abstract — This paper presents a new technique for the update of a probabilistic spatial occupancy grid map using a forward sensor model. Unlike currently popular inverse sensor models, forward sensor models can be found experimentally and can represent sensor characteristics better. The formulation is applicable to both 2D and 3D range sensors and does not have some of the theoretical and practical problems associated with the current approaches which use forward models. As an illustration of this procedure, a new prototype 3D forward sensor model is derived using a beam represented as a spherical sector. Furthermore, this model is used for fusion of point-clouds obtained from different 3D sensors, in particular, time-of-flight sensors (Swiss-ranger, laser range finders), and stereo vision cameras. Several techniques are described for an efficient data-structure representation and implementation. The range beams from different sensors are fused in a common local Cartesian occupancy map. Experimental results of this fusion are presented and evaluated using Hough-transform performed on the grid. I.

This paper presents a method for precise surface reconstruction with time-of-flight (TOF) cameras. A novel calibration approach which simplifies the calibration task and doubles the camera’s precision is developed and compared to current calibration methods. Remaining errors are tackled by applying filter and error distributing methods. Thus, a reference object is circumferentially reconstructed with an overall mean precision of approximately 3mm in translation and 3° in rotation. The resulting model serves as quantification of achievable reconstruction precision with TOF cameras. This is a major criteria for the potential analysis of this sensor technology, that is firstly demonstrated within this work.

This paper presents a novel method for object detection and classification in 3D laser range data that is acquired by an autonomous mobile robot. Unrestricted objects are learned using classification and regression trees (CARTs) and using an Ada Boost learning procedure. Off-screen rendered depth and reflectance images serve as an input for learning. The performance of the classification is improved by combining both sensor modalities, which are independent from external light. This enables highly accurate, fast and reliable 3D object localization with point matching. Competitive learning is used for evaluating the accuracy of the object localization.

This paper addresses the question of generating large-scale 3d models from a set of 2d laser scans. We investigate two configurations: a rotating laser scanner and a setup of two orthogonally mounted laser range finders. Both systems are used to combine data from several viewpoints into a map using global registration. A feature-based and a raw-data based method for pose estimation and global registration are presented. These approaches are applied to data from indoor and outdoor scenes. For data acquisition the Biba-robot and a Smart-car were used. With additional information from a panoramic camera a textured 3d model of a part of EPFL campus was created. The precision of the resulting 3d model is evaluated by comparing it to an orthophoto of the environment. 1

This paper presents a novel scheme for fast color invariant ball detection in the RoboCup context. Edge filtered camera images serve as an input for an Ada Boost learning procedure that constructs a cascade of classification and regression trees (CARTs). Our system is capable to detect different soccer balls in the RoboCup and other environments. The resulting approach for object classification is real-time capable and reliable.

Mobile robots, autonomous exploration, behaviour-based, indoor environments. Besides safe motion control the gain of environmental knowledge is a key for a reliable home or office service robot. When being set into a completely unknown environment the robot has to be able to derive a certain abstract internal representation of this world without any user interaction. This knowledge enables the robot to know how to get from its actual place in one room to a target position in another room as a prerequisite for transportation tasks for example. In this context, the combination of a behavior-based motion control system and an abstract topological map based on geometric representations of rooms seems promising. As the concept of motion and exploration behaviors facilitates to compete with noisy sensor information and geometrically imprecise maps, it has been used to develop completely autonomous exploration strategies for deriving topological representations of common indoor environments. The only prescribed world knowledge is the fact that these environments are composed of rectangular entities (rooms) which are connected by openings (doors). The developed system has successfully been tested in simulation and reality. 1

In this paper, we present a new visual attention system for robotic applications capable of processing data from different sensor modes simultaneously. The consideration of several sensor modalities is an obvious approach to regard a variety of object properties. Nevertheless, conventional attention systems only consider the processing of camera images. We present a bimodal system that processes two sensor modes simultaneously and is easily extensible to additional modes. In contrast to other systems, the input data to our system is provided by a bimodal 3D laser scanner, mounted on top of an autonomous mobile robot. In a single 3D scan pass, the scanner yields range as well as reflectance data. Both data modes are illumination independent, yielding a robust approach that enables all day operation. Data from both laser modes are fed into our attention system built on principles of one of the standard models of visual attention by Koch & Ullman. The system computes conspicuities of both modes in parallel and fuses them into one saliency map. The focus of attention is directed to the most salient points in this map sequentially. We present results on recorded scans of indoor and outdoor scenes showing the respective advantages of the sensor modalities enabling the mode-specific detection of different object properties. Furthermore, we show as an application of the attention system the recognition of objects for building semantic 3D maps of the robot’s environment. Key words: visual attention, saliency detection, bimodal sensor fusion, 3D laser scanner

Robotic navigation requires that the robotic platform have an idea of its location and orientation within the environment. This localization is known as pose estimation, and has been a much researched topic. There are currently two main categories of pose estimation techniques: pose from hardware, and pose from video (PfV). Hardware pose estimation utilizes specialized hardware such as Global Positioning Systems (GPS) and Inertial Navigation Systems (INS) to estimate the position and orientation of the platform at the specified times. PfV systems use video cameras to estimate the pose of the system by calculating the inter-frame motion of the camera from features present in the images. These pose estimation systems are readily integrated, and can be used to augment and/or supplant each other according to the needs of the application. Both pose from video and hardware pose estimation have their uses, but each also has its degenerate cases in which they fail to provide reliable data. Hardware solutions can provide extremely accurate data, but are usually quite pricey and can be restrictive in their environments of operation. Pose from video solutions can be implemented with low-cost off-the-shelf components, but the accuracy of the PfV results can be degraded by noisy imagery, ambiguity in the feature matching process, and moving objects. This paper attempts to evaluate the cost/benefit comparison between pose from video and hardware pose estimation experimentally, and to provide a guide as to which systems should be used under certain scenarios.

A basic issue of mobile robotics is generating environment maps automatically. Outdoor terrain is challenging since the ground is uneven and the surrounding is structured irregularly. In earlier work, we have introduced 6D SLAM (Simultaneous Localization and Mapping) as a method to taking all six DOF of robot poses (x, y, z translation; roll, pitch, yaw angles) into account. This paper adds to 6D SLAM a method for extracting drivable surfaces in the 3D maps while they are being generated. Experiments have

Abstract. 3D map building is a complex robotics task which needs mathematical robust models. From a 3D point cloud, we can use the normal vectors to these points to do feature extraction. In this paper, we will present a robust method for normal estimation and unconstrained 3D-mesh generation from a not-uniformly distributed point cloud. 1