Antonio A. F. Oliveira

Tracing patterns and attention: humanoid robot cognition

Humanoid robots promise to lead us toward more effective and informative interactions between humans and robotic devices. The authors introduce mechanisms for attention control and pattern categorization as the basis for cognition in a humanoid robot.

Dual-t-snakes model for medical imaging segmentation

The Dual-T-Snakes model plus dynamic programming (DP) techniques is a powerful methodology for boundary extraction and segmentation of 2D images. However, the original Dual-T-Snakes is not efficient for noisy images due to nonconvexity problems. In this paper we improve the model through multigrid and region growing methods to get more robustness against local minima. Besides, we demonstrate the advantage of using pass-band filtering methods and a fuzzy segmentation technique plus Dual-T-Snakes. We test these methods for artificial and cell images.

Optimal Covering of Plane Domains by Circles Via Hyperbolic Smoothing

We consider the problem of optimally covering plane domains by a given number of circles. The mathematical modeling of this problem leads to a min–max–min formulation which, in addition to its intrinsic multi-level nature, has the significant characteristic of being non-differentiable. In order to overcome these difficulties, we have developed a smoothing strategy using a special class C∞ smoothing function. The final solution is obtained by solving a sequence of differentiable subproblems which gradually approach the original problem. The use of this technique, called Hyperbolic Smoothing, allows the main difficulties presented by the original problem to be overcome. A simplified algorithm containing only the essential of the method is presented. For the purpose of illustrating both the actual working and the potentialities of the method, a set of computational results is presented.

A boundary extraction method based on Dual-T-Snakes and dynamic programming

The original proposal of active contour models, also called snakes, for image segmentation, suffers from a strong sensitivity to its initial position and can not deal with topological changes. The sensitivity to initialization can be addressed by dynamic programming (DP) techniques which have the advantage of guaranteeing the global minimum and of being more stable numerically than the variational approaches. Their disadvantages are the storage requirements and computational complexity. In this paper we address these limitations of DP by reducing the region of interest (search space) through the use of the Dual-T-Snake approach. The solution of this method consists of two curves enclosing each object boundary which allows the definition of a more efficient search space for a DP technique. The resulting method (Dual-T-Snake plus DP) inherits the capability of changing the topology and avoiding local minima from the Dual-T-Snake and the global optimal properties of the dynamic programming. It can be also extended for 3D.

Learning policies for attentional control

We propose two behaviourally active policies for attentional control. These policies must act based on a multi-modal sensory feedback. Two approaches are used to derive the policies: 1) a simple straightforward strategy, and 2) using Q-learning to learn a policy based on the perceptual state of the system. As a practical result of both algorithms, a robotic agent is capable to select a region of interest and perform shifts of attention focusing on the selected region. Then, a multi-feature extraction can take place allowing the system to identify or recognize a pattern representing that region of interest. Also, the policies have the desired property that all objects in the environment are visited at least once, although some of them can be visited more. In this way a robotic agent can relate sensed information to actions, abstracting and providing a feedback (categorization and mapping) for environmental stimuli.

Neural mechanisms for learning of attention control and pattern categorization as basis for robot cognition

We present mechanisms for attention control and pattern categorization as the basis for robot cognition. For attention, we gather information from attentional feature maps extracted from sensory data constructing salience maps to decide where to foveate. For identification, multi-feature maps are used as input to an associative memory, allowing the system to classify a pattern representing a region of interest. As a practical result, our robotic platforms are able to select regions of interest and perform shifts of attention focusing on the selected regions, and to construct and maintain attentional maps of the environment in an efficient manner.

A Collision Detection and Response Scheme for Simplified Physically Based Animation

In this paper we describe a system for physical animation of rigid and deformable objects. These are represented as groups of particles linked by linear constraints, while a Verlet integrator is used for motion computation. Unlike traditional approaches, we accomplish physical simulation without explicitly computing orientation matrices, torques or inertia tensors. The main contribution of our work is related to the way collisions are handled by the system, which employs different approaches for deformable and rigid bodies. In particular, we show how collision detection using the GJK algorithm [9] and bounding sphere hierarchies can be combined with the projection based collision response technique described by Jakobsen [14].

Multi-modal stereognosis

In this work, vision and touch (artificial) senses are integrated in a cooperative active system. Multi-modal sensory information acquired on-line is used by a robotic agent to perform real-time tasks involving categorization of objects. The visual-touch system proposed is able to foveate (verge) the eyes onto an object, to move the arms to touch an object, and to choose another object by shifting its focus of attention. Also, the system can detect changes occurred on a previously visited region. We propose a hybrid computational system in which an associative memory remembers the visual and tactile signature of objects for recognition and reinforcement learning tasks (Q-learning) are formulated to learn active sensing policies. Several vision architectures have been proposed recently. Kosslyn’s architecture [6], based on results in experimental neuro-psychology, seems to be technically possible with special hardware. Despite practical issues it is intuitively attractive to attach computational mechanisms to that architecture. Ferrell [3] uses registered, multi-modal, topographically organized maps of the sensory-motor space to orient a robot (COG) towards environmental stimuli. In our approach we also use topographically organized (visual and arm) maps for feature extraction. But, instead of a pixel (patch) representation we have regions of interest (ROIs) segmented in those maps, as suggested in Kosslyn [6]. At each time, only one region is selected to be processed by the high level processes. This selective aspect provides a substantial reduction in the amount of computational efforts necessary for feature extraction.

A semi-automatic surface reconstruction framework based on T-Surfaces and isosurface extraction methods

In this paper we present a new approach which integrates the T-surfaces model and isosurface generation methods in a general framework for surface reconstruction in 3D medical images. T-surfaces is a deformable model based on a triangulation of the image domain, a discrete surface model and an image threshold. Two types of isosurface generation methods are considered: the continuation ones and the marching ones. The former is useful during the reparameterization of T-surfaces while the later is suitable to initialize the model closer the boundary. Specifically, in a first stage, the T-surfaces grid and the threshold are used to define a coarser image resolution. This field is thresholded to get a 0-1 function which is processed by a marching method to generate polygonal surfaces whose interior may contain the desired objects. If a polygonal surface involves more than one object, then the resolution is increased in that region and the marching applied again. Next, we apply T-surfaces to improved the result. If the obtained topology remains incorrect, we enable the user to modify the topology by an interactive method based on the T-surfaces framework. Finally, we demonstrate the utility of diffusion methods for our approach.

A boundary extraction approach based on multi-resolution methods and the T-Snakes framework

We present a new approach which integrates the T-Snakes model and a multi-resolution method in a unified framework for segmentation and boundary extraction. In a first stage, a local scale property of the objects is used to define a triangulation of the image domain and a sampling (coarsest resolution) of the image field. The low resolution image is thresholded to get a 0-1 field which is processed by a simple continuation method to generate polygonal curves whose interior contain the desired objects. If the polygonal curve involves more than one object, then the resolution is increased in that region and the method will be applied again. This stage gives a rough approximation of the desired boundaries which will be improved by the T-Snakes to get the final result. We demonstrate the method for 2D medical imaging in the experimental results and indicate how it can be extended to 3D in future work.