This paper describes human classifiers that are 'viewpoint specific', meaning specific to subjects being observed by a particular camera in a particular scene. The advantages of the approach are (a) improved human detection in the presence of perspective foreshortening from an elevated camera, (b) ability to handle partial occlusion of subjects e.g. partial occlusion by furniture in an indoor scene, and (c) ability to detect subjects when partially truncated at the top, bottom or sides of the image. Elevated camera views will typically generate truncated views for subjects at the image edges but our viewpoint specific method handles such cases and thereby extends overall detection coverage. The approach is - (a) define a tiling on the ground plane of the 3D scene, (b) generate training images per tile using virtual humans, (c) train a classifier per tile (d) run the classifiers on the real scene. The approach would be prohibitive if each new deployment required real training images, but it is feasible because training is done with a virtual humans inserted into a scene model. The classifier is a linear SVM and HOGs. Experimental results provide a comparative analysis with existing algorithms to demonstrate the advantages described above.

We present a generic algorithm for self calibration of robotic systems that utilizes two key innovations. First, it uses information theoretic measures to automatically identify and store novel measurement sequences. This keeps the computation tractable by discarding redundant information and allows the system to build a sparse but complete calibration dataset from data collected at different times. Second, as the full observability of the calibration parameters may not be guaranteed for an arbitrary measurement sequence, the algorithm detects and locks unobservable directions in parameter space using a truncated QR decomposition of the Gauss-Newton system. The result is an algorithm that listens to an incoming sensor stream, builds a minimal set of data for estimating the calibration parameters, and updates parameters as they become observable, leaving the others locked at their initial guess. Through an extensive set of simulated and real-world experiments, we demonstrate that our method outperforms state-of-the-art algorithms in terms of stability, accuracy, and computational efficiency.

This paper presents a novel approach to tracking dynamic objects in 3D range data. Its key contribution lies in the generative object detection algorithm which allows the tracker to robustly extract objects of varying sizes and shapes from the observations. In contrast to tracking methods using discriminative detectors, we are thus able to generalize over a wide range of object classes matching our assumptions. Whilst the generative model underlying our framework inherently scales with the complexity and the noise characteristics of the environment, all parameters involved in the detection process obey a clean probabilistic interpretation. Nevertheless, our unsupervised object detection and tracking algorithm achieves real-time performance, even in highly dynamic scenarios covering a significant amount of moving objects. Through an application to populated urban settings, we are able to show that the tracking performance of the presented approach yields results which are comparable to state-of-the-art discriminative methods.

In this paper, we address the problem of curb detection for a pedestrian robot navigating in urban environments. We devise an unsupervised method that is mostly view-independent, makes no assumptions about the environment, restricts the set of hand-tuned parameters, and builds on sound probabilistic reasoning from the input data to the outcome of the algorithm. In our approach, we construct a piecewise planar model of the environment and determine curbs at plane segment boundaries. Initially, we sense the environment with a nodding laser range-finder and project the 3D measurements into an efficient Digital Elevation Map (DEM). Each cell of the DEM maintains an error model that is propagated throughout the entire algorithm. Plane segments are further estimated with a mixture of linear regression models on the DEM. Here, we propose an original formulation of the standard Expectation-Maximization (EM) algorithm for mixture models. Specifically, in the E-step, responsibilities are computed with a Conditional Random Field (CRF) that introduces dependencies between the covariates of the mixture model. A graph-based segmentation of the DEM provides an estimate of the number of planes and initial parameters for the EM. We show promising results of the algorithm on simulated and real-world data.

This paper presents a novel self-supervised online learning method to discover driving behaviors from data acquired with an inertial measurement unit (IMU) and a camera. Both sensors where mounted in a car that was driven by a human through a typical city environment with intersections, pedestrian crossings and traffic lights. The presented system extracts motion segments from the IMU data and relates them to visual cues obtained from camera data. It employs a Bayesian on-line estimation method to discover the motion segments based on change-point detection and uses a Dirichlet Compound Multinomial (DCM) model to represent the visual features extracted from the camera images. By incorporating these visual cues into the on-line estimation process, labels are computed that are equal for similar motion segments. As a result, typical traffic situations such as braking maneuvers in front of a red light can be identified automatically. Furthermore, appropriate actions in form of observed motion changes are associated to the discovered traffic situations. The approach is evaluated on a real data set acquired in the center of Zurich.

Most large-scale public environments provide direction signs to facilitate the orientation for humans and to find their way to a goal location in the environment. Thus, for a robot operating in the same environment, it would be beneficial to interpret such signs correctly for a safe and efficient navigation. In this work, we propose a novel approach to infer the meaning of direction signs and to use that for navigation, i.e., to find a mapping of a detected sign to a motion direction. Our method uses a hierarchical extension of the Implicit Shape Model framework called HISM that does not require any hand-labeled training data to detect the signs. On the lower level of this two-stage hierarchy, ISM is applied to image descriptors as in the standard approach. On the higher level, ISM operates on subparts of signs called tokens, using weights learned from data. The interpretation of the signs is inferred by associating navigation data to direction instructions. We conducted experiments from image data acquired in an airport terminal, aiming towards the implementation of a robotic guide, with promising results.

The capability to overcome terrain irregularities or obstacles, named terrainability, is mostly dependant on the suspension mechanism of the rover and its control. For a given wheeled robot, the terrainability can be improved by using a sophisticated control, and is somewhat related to minimizing wheel slip. The proposed control method, named torque control, improves the rover terrainability by taking into account the whole mechanical structure. The rover model is based on the Newton-Euler equations and knowing the complete state of the mechanical structures allows us to compute the force distribution in the structure, and especially between the wheels and the ground. Thus, a set of torques maximizing the traction can be used to drive the rover. The torque control algorithm is presented in this paper, as well as tests showing its impact and improvement in terms of terrainability. Using the CRAB rover platform, we show that the torque control not only increases the climbing performance but also limits odometric errors and reduces the overall power consumption.

We present a Bluetooth scatternet protocol (SNP) that provides the user with a serial link to all connected members in a transparent wireless Bluetooth (BT) network. By using only local decision making we can reduce the overhead of our scatternet protocol dramatically. We show how our SNP software layer simplifies a variety of tasks like the synchronization of central pattern generator controllers for actuators, collecting sensory data and building modular robot structures. The whole BT software stack including our new scatternet layer is implemented on a single Bluetooth and memory chip. To verify and characterize the SNP we provide data from experiments using real hardware instead of software simulation. This gives a realistic overview of the scatternet performance showing higher order effects that are difficult to be simulated correctly and guarantees the correct function of the SNP in real world applications.