We present a novel vehicle detection and tracking system that works solely on 3D LiDAR information. Our approach segments vehicles using a dual-view representation of the 3D LiDAR point cloud on two independently trained convolutional neural networks, one for each view. A bounding box growing algorithm is applied to the fused output of the networks to properly enclose the segmented vehicles. Bounding boxes are grown using a probabilistic method that takes into account also occluded areas. The final vehicle bounding boxes act as observations for a multi-hypothesis tracking system which allows to estimate the position and velocity of the observed vehicles. We thoroughly evaluate our system on the KITTI benchmarks both for detection and tracking separately and show that our dual-branch classifier consistently outperforms previous single-branch approaches, improving or directly competing to other state of the art LiDAR-based methods.

Today's complex robotic designs comprise in some cases a large number of degrees of freedom, enabling for multi-objective task resolution (e.g., humanoid robots or aerial manipulators). This paper tackles the local stability problem of a hierarchical closed-loop inverse kinematics algorithm for such highly redundant robots. We present a method to guarantee this system stability by performing an online tuning of the closed-loop control gains. We define a semi-definite programming problem (SDP) with these gains as decision variables and a discrete-time Lyapunov stability condition as a linear matrix inequality, constraining the SDP optimization problem and guaranteeing the local stability of the prioritized tasks. To the best of authors' knowledge, this work represents the first mathematical development of an SDP formulation that introduces these stability conditions for a multi-objective closed-loop inverse kinematic problem for highly redundant robots. The validity of the proposed approach is demonstrated through simulation case studies, including didactic examples and a Matlab toolbox for the benefit of the community.

This paper proposes the use of piecewise Cnsmooth curve for mobile-base motion planning and control, coined Timed-Elastic Smooth Curve (TESC) planner. Based on a Timed-Elastic Band, the problem is defined so that the trajectory lies on a spline in SE(2) with non-vanishing n-th derivatives at every point. Formulated as a multi-objective nonlinear optimization problem, it allows imposing soft constraints such as collision-avoidance, velocity, acceleration and jerk limits, and more. The planning process is realtime-capable allowing the robot to navigate in dynamic complex scenarios. The proposed method is compared against the state-of-the-art in various scenarios. Results show that trajectories generated by the TESC planner have smaller average acceleration and are more efficient in terms of total curvature and pseudo-kinetic energy while being produced with more consistency than state-of-the-art planners do.

In this paper we introduce the Boosted Random Ferns (BRFs) to rapidly build discriminative classifiers for learning and detecting object categories. At the core of our approach we use standard random ferns, but we introduce four main innovations that let us bring ferns from an instance to a category level, and still retain efficiency. First, we define binary features on the histogram of oriented gradients-domain (as opposed to intensity-), allowing for a better representation of intra-class variability. Second, both the positions where ferns are evaluated within the sliding window, and the location of the binary features for each fern are not chosen completely at random, but instead we use a boosting strategy to pick the most discriminative combination of them. This is further enhanced by our third contribution, that is to adapt the boosting strategy to enable sharing of binary features among different ferns, yielding high recognition rates at a low computational cost. And finally, we show that training can be performed online, for sequentially arriving images. Overall, the resulting classifier can be very efficiently trained, densely evaluated for all image locations in about 0.1 seconds, and provides detection rates similar to competing approaches that require expensive and significantly slower processing times. We demonstrate the effectiveness of our approach by thorough experimentation in publicly available datasets in which we compare against state-of-the-art, and for tasks of both 2D detection and 3D multi-view estimation.

Localization is the key perceptual process closing the loop of autonomous navigation, allowing self-driving vehicles to operate in a deliberate way. To ensure robust localization, autonomous vehicles have to implement redundant estimation processes, ideally independent in terms of the underlying physics behind sensing principles. This paper presents a stereo radar odometry system, which can be used as such a redundant system, complementary to other odometry estimation processes, providing robustness for long-term operability. The presented work is novel with respect to previously published methods in that it contains: (i) a detailed formulation of the Doppler error and its associated uncertainty; (ii) an observability analysis that gives the minimal conditions to infer a 2D twist from radar readings; and (iii) a numerical analysis for optimal vehicle sensor placement. Experimental results are also detailed that validate the theoretical insights.

We propose a novel method for robotic exploration that evaluates paths that minimize both the joint path and map entropy per meter traveled. The method uses Pose SLAM to update the path estimate, and grows an RRT* tree to generate the set of candidate paths. This action selection mechanism contrasts with previous appoaches in which the action set was built heuristically from a sparse set of candidate actions. The technique favorably compares agains the classical frontier-based exploration and other Active Pose SLAM methods in simulations in a common publicly available dataset.

This paper develops a simple and low-cost method for 3D, high-rate vehicle state estimation, specially designed for free-flying Micro Aerial Vehicles (MAVs). We fuse observations from inertial measurement units and the recently appeared low-cost optical flow smart cameras. These smart cameras integrate a sonar altimeter, a triaxial gyrometer and an optical flow sensor, and directly provide metric ego-motion information in the form of body velocities and altitude. Compared to state-of-the-art visual-inertial odometry methods, we are able to drastically reduce the computational load in the main processor unit, and obtain an accurate estimation of the vehicle state at a high update rate of 100Hz. We thus extend the current use of these smart cameras from hovering purposes to odometry estimation. In order to propose a simple algorithmic solution, we investigate the performances of two Kalman filters, in the extended and error-state flavors, alongside a large number of algorithm variations, using simulations and real experiments with precise ground-truth. We observe that the marginal performance gain attained with these algorithm improvements does not pay for the effort of implementing them. We conclude that a classical EKF in its simplest form is sufficient for providing motion estimates that coherently exploit the available measurements.

We introduce LETHA (Learning on Easy data, Test on Hard), a new learning paradigm consisting of building strong priors from high quality training data, and combining them with discriminative machine learning to deal with low-quality test data. Our main contribution is an implementation of that concept for pose estimation. We first automatically build a 3D model of the object of interest from high-definition images, and devise from it a pose-indexed feature extraction scheme. We then train a single classifier to process these feature vectors. Given a low quality test image, we visit many hypothetical poses, extract features consistently and evaluate the response of the classifier. Since this process uses locations recorded during learning, it does not require matching points anymore. We use a boosting procedure to train this classifier common to all poses, which is able to deal with missing features, due in this context to self-occlusion. Our results demonstrate that the method combines the strengths of global image representations, discriminative even for very tiny images, and the robustness to occlusions of approaches based on local feature point descriptors.

An information-driven autonomous robotic exploration method on a continuous representation of unknown environments is proposed in this paper. The approach conveniently handles sparse sensor measurements to build a continuous model of the environment that exploits structural dependencies without the need to resort to a fixed resolution grid map. A gradient field of occupancy probability distribution is regressed from sensor data as a Gaussian process providing frontier boundaries for further exploration. The resulting continuous global frontier surface completely describes unexplored regions and, inherently, provides an automatic stop criterion for a desired sensitivity. The performance of the proposed approach is evaluated through simulation results in the well-known Freiburg and Cave maps.

Industrial environments are rarely static and often their configuration is continuously changing due to the material transfer flow. This is a major challenge for infrastructure free localization systems. In this paper we address this challenge by introducing a localization approach that uses a dual-timescale approach. The proposed approach - Dual-Timescale Normal Distributions Transform Monte Carlo Localization (DT-NDT-MCL) - is a particle filter based localization method, which simultaneously keeps track of the pose using an apriori known static map and a short-term map. The short-term map is continuously updated and uses Normal Distributions Transform Occupancy maps to maintain the current state of the environment. A key novelty of this approach is that it does not have to select an entire timescale map but rather use the best timescale locally. The approach has real-time performance and is evaluated using three datasets with increasing levels of dynamics. We compare our approach against previously proposed NDT-MCL and commonly used SLAM algorithms and show that DT-NDT-MCL outperforms competing algorithms with regards to accuracy in all three test cases.

Abstract — We propose a method for the computation of entropy decrease in C-space. These estimates are then used to evaluate candidate exploratory trajectories in the context of autonomous mobile robot mapping. The method evaluates both map and path entropy reduction and uses such estimates to compute trajectories that maximize coverage whilst min-imizing localization uncertainty, hence reducing map error. Very efficient kernel convolution mechanisms are used to evaluate entropy reduction at each sensor ray, and for each possible robot position and orientation, taking frontiers and obstacles into account. In contrast to most other exploration methods that evaluate entropy reduction at a small number of discrete robot configurations, we do it densely for the entire C-space. The computation of such dense entropy reduction maps opens the window to new exploratory strategies. In this paper we present two such strategies. In the first one we drive exploration through a gradient descent on the entropy decrease field. The second strategy chooses maximal entropy reduction configurations as candidate exploration goals, and plans paths to them using RRT*. Both methods use PoseSLAM as their estimation backbone, and are tested and compared with classical frontier-based exploration in simulations using common publicly available datasets. I.

Presentado al 24th BMVC celebrado en Bristol (UK) del 9 al 13 de septiembre 2013.-- The copyright of this document resides with its authors.

This paper develops a new method for uncalibrated image-based visual servoing. In contrast to traditional image-based visual servo, the proposed solution does not require a known value of camera focal length for the computation of the image Jacobian. Instead, it is estimated at run time from the observation of the tracked target. The technique is shown to outperform classical visual servoing schemes in situations with noisy calibration parameters and for unexpected changes in the camera zoom. The method's performance is demonstrated both in simulation experiments and in a ROS implementation of a quadrotor servoing task. The developed solution is tightly integrated with ROS and is made available as part of the IRI ROS stack.

We propose HRA*, a new randomized path planner for complex 3D environments. The method is a modified A* algorithm that uses a hybrid node expansion technique that combines a random exploration of the action space meeting vehicle kinematic constraints with a cost to goal metric that considers only kinematically feasible paths to the goal. The method includes also a series of heuristics to accelerate the search time. These include a cost penalty near obstacles, and a filter to prevent revisiting configurations. The performance of the method is compared against A*, RRT and RRT* in a series of challenging 3D outdoor datasets. HRA* is shown to outperform all of them in computation time, and delivering shorter paths than A* and RRT.

We propose a novel approach for the estimation of the pose and focal length of a camera from a set of 3D-to-2D point correspondences. Our method compares favorably to competing approaches in that it is both more accurate than existing closed form solutions, as well as faster and also more accurate than iterative ones. Our approach is inspired on the EPnP algorithm, a recent O(n) solution for the calibrated case. Yet we show that considering the focal length as an additional unknown renders the linearization and relinearization techniques of the original approach no longer valid, especially with large amounts of noise. We present new methodologies to circumvent this limitation termed exhaustive linearization and exhaustive relinearization which perform a systematic exploration of the solution space in closed form. The method is evaluated on both real and synthetic data, and our results show that besides producing precise focal length estimation, the retrieved camera pose is almost as accurate as the one computed using the EPnP, which assumes a calibrated camera.

The maps that are built by standard feature-based simultaneous localization and mapping (SLAM) methods cannot be directly used to compute paths for navigation, unless enriched with obstacle or traversability information, with the consequent increase in complexity. Here, we propose a method that directly uses the Pose SLAM graph of constraints to determine the path between two robot configurations with lowest accumulated pose uncertainty, i.e., the most reliable path to the goal. The method shows improved navigation results when compared with standard path-planning strategies over both datasets and real-world experiments.

We present an active exploration strategy that complements Pose SLAM [1] and optimal navigation in Pose SLAM [2]. The method evaluates the utility of exploratory and place revisiting sequences and chooses the one that minimizes overall map and path entropies. The technique considers trajectories of similar path length taking marginal pose uncertainties into account. An advantage of the proposed strategy with respect to competing approaches is that to evaluate information gain over the map, only a very coarse prior map estimate needs to be computed. Its coarseness is independent and does not jeopardize the Pose SLAM estimate. Moreover, a replanning scheme is devised to detect significant localization improvement during path execution. The approach is tested in simulations in a common publicly available dataset comparing favorably against frontier based exploration.

We propose an efficient method for object localization and 3D pose estimation. A two-step approach is used. In the first step, a pose estimator is evaluated in the input images in order to estimate potential object locations and poses. These candidates are then validated, in the second step, by the corresponding pose-specific classifier. The result is a detection approach that avoids the inherent and expensive cost of testing the complete set of specific classifiers over the entire image. A further speedup is achieved by feature sharing. Features are computed only once and are then used for evaluating the pose estimator and all specific classifiers. The proposed method has been validated on two public datasets for the problem of detecting of cars under several views. The results show that the proposed approach yields high detection rates while keeping efficiency. © 2011. The copyright of this document resides with its authors.

The probabilistic belief networks that result from standard feature-based simultaneous localization and map building cannot be directly used to plan trajectories. The reason is that they produce a sparse graph of landmark estimates and their probabilistic relations, which is of little value to find collision free paths for navigation. In contrast, we argue in this paper that Pose SLAM graphs can be directly used as belief roadmaps. We present a method that devises optimal navigation strategies by searching for the path in the pose graph with lowest accumulated robot pose uncertainty, independently of the map reference frame. The method shows improved navigation results when compared to shortest paths both over synthetic data and real datasets.

We present a new approach for building an efficient and robust classifier for the two class problem, that localizes objects that may appear in the image under different orientations. In contrast to other works that address this problem using multiple classifiers, each one specialized for a specific orientation, we propose a simple two-step approach with an estimation stage and a classification stage. The estimator yields an initial set of potential object poses that are then validated by the classifier. This methodology allows reducing the time complexity of the algorithm while classification results remain high. The classifier we use in both stages is based on a boosted combination of Random Ferns over local histograms of oriented gradients (HOGs), which we compute during a preprocessing step. Both the use of supervised learning and working on the gradient space makes our approach robust while being efficient at run-time. We show these properties by thorough testing on standard databases and on a new database made of motorbikes under planar rotations, and with challenging conditions such as cluttered backgrounds, changing illumination conditions and partial occlusions.

We propose a new algorithm for detecting multiple object categories that exploits the fact that different categories may share common features but with different geometric distributions. This yields an efficient detector which, in contrast to existing approaches, considerably reduces the computation cost at runtime, where the feature computation step is traditionally the most expensive. More specifically, at the learning stage we compute common features by applying the same Random Ferns over the Histograms of Oriented Gradients on the training images. We then apply a boosting step to build discriminative weak classifiers, and learn the specific geometric distribution of the Random Ferns for each class. At runtime, only a few Random Ferns have to be densely computed over each input image, and their geometric distribution allows performing the detection. The proposed method has been validated in public datasets achieving competitive detection results, which are comparable with state-of-the-art methods that use specific features per class.

Time-of-Flight (ToF) cameras deliver 3D images at 25 fps, offering great potential for developing fast object modeling algorithms. Surprisingly, this potential has not been extensively exploited up to now. A reason for this is that, since the acquired depth images are noisy, most of the available registration algorithms are hardly applicable. A further difficulty is that the transformations between views are in general not accurately known, a circumstance that multi-view object modeling algorithms do not handle properly under noisy conditions. In this work, we take into account both uncertainty sources (in images and camera poses) to generate spatially consistent 3D object models fusing multiple views with a probabilistic approach. We propose a method to compute the covariance of the registration process, and apply an iterative state estimation method to build object models under noisy conditions.

Pose SLAM is the variant of simultaneous localization and map building (SLAM) is the variant of SLAM, in which only the robot trajectory is estimated and where landmarks are only used to produce relative constraints between robot poses. To reduce the computational cost of the information filter form of Pose SLAM and, at the same time, to delay inconsistency as much as possible, we introduce an approach that takes into account only highly informative loop-closure links and nonredundant poses. This approach includes constant time procedures to compute the distance between poses, the expected information gain for each potential link, and the exact marginal covariances while moving in open loop, as well as a procedure to recover the state after a loop closure that, in practical situations, scales linearly in terms of both time and memory. Using these procedures, the robot operates most of the time in open loop, and the cost of the loop closure is amortized over long trajectories. This way, the computational bottleneck shifts to data association, which is the search over the set of previously visited poses to determine good candidates for sensor registration. To speed up data association, we introduce a method to search for neighboring poses whose complexity ranges from logarithmic in the usual case to linear in degenerate situations. The method is based on organizing the pose information in a balanced tree whose internal levels are defined using interval arithmetic. The proposed Pose-SLAM approach is validated through simulations, real mapping sessions, and experiments using standard SLAM data sets.

A method for evaluating, at video rate, the quality of actions for a single camera while mapping unknown indoor environments is presented. The strategy maximizes mutual information between measurements and states to help the camera avoid making ill-conditioned measurements that are appropriate to lack of depth in monocular vision systems. Our system prompts a user with the appropriate motion commands during 6-DOF visual simultaneous localization and mapping with a handheld camera. Additionally, the system has been ported to a mobile robotic platform, thus closing the control-estimation loop. To show the viability of the approach, simulations and experiments are presented for the unconstrained motion of a handheld camera and for the motion of a mobile robot with nonholonomic constraints. When combined with a path planner, the technique safely drives to a marked goal while, at the same time, producing an optimal estimated map.

In this work we present a robust detection method in outdoor scenes under cast shadows using color based invariant gradients in combination with HoG local features. The method achieves good detection rates in urban scene classification and person detection outperforming traditional methods based on intensity gradient detectors which are sensible to illumination variations but not to cast shadows. The method uses color based invariant gradients that emphasize material changes and extract relevant and invariant features for detection while neglecting shadow contours. This method allows to train and detect objects and scenes independently of scene illumination, cast and self shadows. Moreover, it allows to do training in one shot, that is, when the robot visits the scene for the first time.

This paper introduces an approach that reduces the size of the state and maximizes the sparsity of the information matrix in exactly sparse delayed-state SLAM. We propose constant time procedures to measure the distance between a given pair of poses, the mutual information gain for a given candidate link, and the joint marginals required for both measures. Using these measures, we can readily identify non redundant poses and highly informative links and use only those to augment and to update the state, respectively. The result is a delayed-state SLAM system that reduces both the use of memory and the execution time and that delays filter inconsistency by reducing the number of linearization introduced when adding new loop closure links. We evaluate the advantage of the proposed approach using simulations and data sets collected with real robots.

Outdoor camera networks are becoming ubiquitous in critical urban areas of large cities around the world. Although current applications of camera networks are mostly limited to video surveillance, recent research projects are exploiting advances on outdoor robotics technology to develop systems that put together networks of cameras and mobile robots in people assisting tasks. Such systems require the creation of robot navigation systems in urban areas with a precise calibration of the distributed camera network. Despite camera calibration has been an extensively studied topic, the calibration (intrinsic and extrinsic) of large outdoor camera networks with no overlapping view fields, and likely to suffer frequent recalibration, poses novel challenges in the development of practical methods for user-assisted calibration that minimize intervention times and maximize precision. In this paper we propose the utilization of Laser Range Finder (LRF) data covering the area of the camera network to support the calibration process and develop a semi-automated methodology allowing quick and precise calibration of large camera networks. The proposed methods have been tested in a real urban environment and have been applied to create direct mappings (homographies) between image coordinates and world points in the ground plane (walking areas) to support person and robot detection and localization algorithms.

We present an approach to the problem of 3D map building in urban settings for service robots, using three-dimensional laser range scans as the main data input. Our system is based on the probabilistic alignment of 3D point clouds employing a delayed-state information-form SLAM algorithm, for which we can add observations of relative robot displacements efficiently. These observations come from the alignment of dense range data point clouds computed with a variant of the iterative closest point algorithm. The datasets were acquired with our custom built 3D range scanner integrated into a mobile robot platform. Our mapping results are compared to a GIS-based CAD model of the experimental site. The results show that our approach to 3D mapping performs with sufficient accuracy to derive traversability maps that allow our service robots navigate and accomplish their assigned tasks on a urban pedestrian area.

Wire-based tracking devices are an affordable alternative to costly tracking devices. They consist of a fixed base and a platform, attached to the moving object, connected by six wires whose tension is maintained along the tracked trajectory. One important shortcoming of these devices is that they are forced to operate in reduced workspaces so as to avoid singular configurations. Singularities can be eliminated by adding more wires, but this causes more wire interferences, and a higher force exerted on the moving object by the measuring device itself. This paper shows how, by introducing a rotating base, the number of wires can be reduced to three, and singularities can be avoided by using an active sensing strategy. This also permits reducing wire interference problems and the pulling force exerted by the device.

In this paper we present an observability analysis for a mobile robot performing SLAM with a single monocular camera. The aim is to get a better understanding of the well known intuitive behavior of these systems, such as the need for triangulation to features from different positions in order to get accurate relative pose estimates. The characterisation of the unobservable directions is made using the nullspace basis of the stripped observability matrix. This allows us to identify which vehicle motions are required to maximise the number of observable states in the system, which in turn affects accuracy in the estimation process. The analysis is performed by modelling the system in the continuous time domain as piecewise constant. Simulation results using an extended information filter are shown to verify the results of the observability analysis.

This paper shows results on outdoor vision-based loop closing for simultaneous localization and mapping. Our experiments show that for loops of over 50 m, the pose estimates maintained with a delayed-state extended information filter are consistent enough to guarantee assertion of vision- based pose constraints for loop closure, provided no necessary information links are added to the estimator. The technique computes relative pose constraints via a robust least squares minimization of 3D point correspondences, which are in turn obtained from the matching of SIFT features over candidate image pairs. We propose a loop closure test that checks both for closeness of means and for highly informative updates at the same time.

In this paper we consider a single hand-held camera performing SLAM at video rate with generic 6DOF motion. The aim is to optimise both the localisation of the sensor and building of the feature map by computing the most appropriate control actions or movements. The actions belong to a discrete set (e.g. go forward, go left, go up, turn right, etc), and are chosen so as to maximise the mutual information gain between posterior states and measurements. Maximising the mutual information helps the camera avoid making ill-conditioned measurements appropriate to bearing-only SLAM. Moreover, orientation changes are determined by maximising the trace of the Fisher information matrix. In this way, we allow the camera to continue looking at those landmarks with large uncertainty, but from better-posed directions. Various position and gaze control strategies are first tested in a simulated environment, and then validated in a video-rate implementation. Given that our system is capable of producing motion commands for a real-time 6DOF visual SLAM, it could be used with any type of mobile platform, without the need of other sensors

In this article we propose an algorithm to reduce the effects caused by linearization in the typical EKF approach to SLAM. The technique consists in computing the vehicle prior using an Unscented Transformation. The UT allows a better nonlinear mean and variance estimation than the EKF. There is no need however in using the UT for the entire vehicle-map state, given the linearity in the map part of the model. By applying the UT only to the vehicle states we get more accurate covariance estimates. The a posteriori estimation is made using a fully observable EKF step, thus preserving the same computational complexity as the EKF with sequential innovation. Experiments over a standard SLAM data set show the behavior of the algorithm.

This paper presents an analysis of the fully correlated approach to the simultaneous localization and map building (SLAM) problem from a control systems theory point of view, both for linear and nonlinear vehicle models. We show how partial observability hinders full reconstructibility of the state space, making the final map estimate dependent on the initial observations. Nevertheless, marginal filter stability guarantees convergence of the state error covariance to a positive semidefinite covariance matrix. By characterizing the form of the total Fisher information, we are able to determine the unobservable state space directions. Moreover, we give a closed-form expression that links the amount of reconstruction error to the number of landmarks used. The analysis allows the formulation of measurement models that make SLAM observable.

In this article, we show that partial observability hinders full reconstructibility of the state space in SLAM, making the final map estimate dependent on the initial observations, and not guaranteeing convergence to a positive semi-definite covariance matrix. By characterizing the form of the total Fisher information we are able to determine the unobservable state space directions. To overcome this problem, we formulate new fully observable measurement models that make SLAM stable.

In this article, we show marginal stability in SLAM, guaranteeing convergence to a non-zero mean state error estimate bounded by a constant value. Moreover, marginal stability guarantees also convergence of the Riccati equation of the one-step ahead state error covariance to at least one psd steady state solution. In the search for real-time implementations of SLAM, covariance inflation methods produce a suboptimal filter that eventually may lead to the computation of an unbounded state error covariance. We provide tight constraints in the amount of decorrelation possible, to guarantee convergence of the state error covariance, and at the same time, a linear-time implementation of SLAM.

Current techniques to concurrent map building and localization (CML) have been devised for static environments, and lack robustness in more realistic situations. In this communication we provide new ideas that extend the typical stochastic estimation approach to CML, to take into account the dynamics of the environment. The basic idea consists on using the history of data association mismatches for the computation of the likelihood of future data association. The incorporation of a novel temporal landmark quality test, together with the spatial compatibility tests already available, help alleviate the difficulty of data association. We propose a pair of temporal landmark quality functions to aid in those situations in which landmark observations might not be consistent in time; and show how by incorporating these functions, the overall estimation-theoretic approach to CML is improved. Special attention is paid in that the removal of landmarks from the map does not violate the basic convergence properties of the localization and map building algorithms already described in the literature. Namely, asymptotic convergence and full correlation.

A system that builds and maintains a dynamic map for a mobile robot is presented. A learning rule associated to each observed landmark is used to compute its robustness. The position of the robot during map construction is estimated by combining sensor readings, motion commands, and the current map state by means of an Extended Kalman Filter. The combination of landmark strength validation and Kalman filtering for map updating and robot position estimation allows for robust learning of moderately dynamic indoor environments.

Describes a method to localize faces in color images based on the fusion of the information gathered from a stereo vision system and the analysis of color images. Our method generates a depth map of the scene and tries to fit a head model taking into account the shape of the model and skin color information. The method is tailored for its use in factory automation applications where the detection and localization of humans is necessary for the completion or interruption of a particular task, such as robot manipulator safety or the interaction of service robots with humans.

Different data acquisition methods are tailored at extracting particular characteristics from a scene and by combining their results a more robust scene description can be created. A method to fuse perceptual groupings extracted from color-based segmentation and depth information from stereo using supervised classification is presented. The merging of data from these two acquisition modules allows for a spatially coherent blend of smooth regions and detail in an image. Depth cues are used to limit the area of interest in the scene and to improve perceptual grouping solving subsegmentation and oversegmentation of the original images. The complexity of the algorithm does not exceed that of the individual acquisition modules. The resulting scene description can then be fed to an object recognition modules for scene interpretation.