This paper presents an FPGA implementation of a machine performing exact Bayesian inference using stochastic bitstreams. We revisited stochastic computing, not to perform better computations with unreliable hardware, but to perform approximate computations with less hardware. The underlying trade-off is between precision and computation time. An automatic design of probabilistic machines that compute soft inferences with an arithmetic based on stochastic bitstreams is presented. The computation tree provided by a Bayesian inference software is used to define the stochastic circuit. Tests were performed and results presented concerning accuracy and resource usage of the stochastic computing implementation of Bayesian machines performing exact inference. An application example is given of a Bayesian sensorimotor system that performs obstacle avoidance for an autonomous robot, fully implemented on an FPGA. Some conclusions were drawn on the followed approach, providing insights for future implementations.

In this paper, we present a Bayesian framework for the active multimodal perception of 3-D structure and motion. The design of this framework finds its inspiration in the role of the dorsal perceptual pathway of the human brain. Its composing models build upon a common egocentric spatial configuration that is naturally fitting for the integration of readings from multiple sensors using a Bayesian approach. In the process, we will contribute with efficient and robust probabilistic solutions for cyclopean geometry-based stereovision and auditory perception based only on binaural cues, modeled using a consistent formalization that allows their hierarchical use as building blocks for the multimodal sensor fusion framework. We will explicitly or implicitly address the most important challenges of sensor fusion using this framework, for vision, audition, and vestibular sensing. Moreover, interaction and navigation require maximal awareness of spatial surroundings, which, in turn, is obtained through active attentional and behavioral exploration of the environment. The computational models described in this paper will support the construction of a simultaneously flexible and powerful robotic implementation of multimodal active perception to be used in real-world applications, such as human-machine interaction or mobile robot navigation.

This work presents a representation of 3D object shape using a probabilistic volumetric map derived from in-hand exploration. The exploratory procedure is based on contour following through the fingertip movements on the object surface. We first consider the simple case of having single hand exploration of a static object. The cumulative pose data provides a 3D point cloud that is quantized to the probabilistic volumetric map. For each voxel we have a probability distribution for the occupancy percentage. This is then extended to in-hand exploration of non-static objects. Since the object is moving during the in-hand exploration, and we also consider the use of the other hand for re-grasping, object pose has to be tracked. By keeping track of object motion we can register data to the initial pose to build a consistent object representation. An object centered representation is implemented using the computed object center of mass to define its frame of reference. Results are presented for in-hand exploration of both static and non-static objects that show that valid models can be obtained. The 3D object probabilistic representation can be used in several applications related with grasp generation tasks.

Robotic implementations of gaze control and image stabilization have been previously proposed, that rely on fusing inertial and visual sensing modalities. Human and biological system also combine the two sensing modalities for the same goal. In this work we build upon these previous results and, with the contribution of psychophysical studies, attempt a more bio-inspired approach to the robotic implementation. Since Bayesian models have been successfully used to explain psychophysical experimental findings, we propose a robotic implementation using Bayesian inference.

This article presents a technique for registration and segmentation of dense depth maps provided by a stereo vision system. The vision system uses inertial sensors to give a reference for camera pose. The maps are registered using a modified version of the ICP - iterative closet point algorithm to register dense depth maps obtained from a stereo vision system. The proposed technique explores the integration of inertial sensor data for dense map registration. Depth maps obtained by vision systems, are very point of view dependent, providing discrete layers of detected depth aligned with the camera. In this work we use inertial sensors to recover camera pose, and rectify the maps to a reference ground plane, enabling the segmentation of vertical and horizontal geometric features and map registration. We propose a real-time methodology segmentation of structures, object recognition, robot navigation or any other task that requires a three-dimensional representation of the physical environment. The aim of this work is a fast real-time system, which can be applied to autonomous robotic systems or to automated car driving systems, for modelling the road, identifying obstacles and roadside features in real-time.

This paper explores the combination of inertial sensor data with vision. Visual and inertial sensing are two sensory modalities that can be explored to give robust solutions on image segmentation and recovery of 3D structure from images, increasing the capabilities of autonomous robots and enlarging the application potential of vision systems. In biological systems, the information provided by the vestibular system is fused at a very early processing stage with vision, playing a key role on the execution of visual movements such as gaze holding and tracking, and the visual cues aid the spatial orientation and body equilibrium. In this paper, we set a framework for using inertial sensor data in vision systems, and describe some results obtained. The unit sphere projection camera model is used, providing a simple model for inertial data integration. Using the vertical reference provided by the inertial sensors, the image horizon line can be determined. Using just one vanishing point and the vertical, we can recover the camera's focal distance and provide an external bearing for the system's navigation frame of reference. Knowing the geometry of a stereo rig and its pose from the inertial sensors, the collineations of level planes can be recovered, providing enough restrictions to segment and reconstruct vertical features and leveled planar patches.

We propose a real-time system that extracts information from dense relative depth maps. This method enables the integration of depth cues on higher level processes including segmentation of structures, object recognition, robot navigation or any other task that requires a 3D representation of the physical environment. Inertial sensors coupled to a vision system can provide important inertial cues for the ego-motion and system pose. In this work we explore the integration of inertial sensor data in vision systems. Depth maps obtained by vision systems, are very point of view dependant, providing discrete layers of detected depth aligned with the camera. We use inertial sensors to recover the camera pose, and rectify the maps to a reference ground plane, enabling the segmentation of vertical and horizontal geometric features. The aim of this work is a fast real-time system, so that it can be applied to autonomous robotic systems or to automated car driving systems, for modelling the road, identifying obstacles and roadside features in real-time.

Active vision systems can be used in robotic systems for navigation. The active vision system provides data on the robot's environment. In mobile systems the position and attitude of the cameras relative to the world can be hard to determine. Inertial sensors coupled to the active vision system can provide valuable information to aid the image processing task. In human and other animals the vestibular system plays a similar role. In this article, we explain our recent steps in the integration of inertial data with an active vision system. The active vision system has a set of stereo cameras capable of vergence, with a common baseline, pan and tilt. A process of visual fixation has already been implemented, enabling symmetric vergence on any selected point. An inertial system prototype, based on low-cost sensors was built. It is used to keep track of the gravity vector, allowing the identification of the vertical in the images. By performing visual fixation of a ground plane point, and knowing the 3D vector normal to level ground we can determine the ground plane. The image can therefore be segmented, and the ground plane along which the robot can move identified. For on-the-fly visualisation of the segmented images and the detected points a VRML viewer is used.