Paulo Fernando Ferreira Rosa

Identification system for smart homes using footstep sounds

This work proposes an identification system for smart buildings using footstep information. Basically, the system has three modules: pre-processing, feature extraction and selection, and classification of the information. Gait frequency, spectral envelope, cepstral and mel-cepstral analysis and loudness compose the set of parameters used to identify subjects in the smart building. In order to select features, the Fishers criterion was used. For the classification task, K-Means and K-NN algorithms were performed along with different methodologies for test and validation. Moreover, the best sets were gathered, increasing the accuracy of classification. Several experiments were performed with 15 subjects wearing 6 different footwear. The classification rates were in the range from 33.5% to 97.5%.

A surveillance task for a UAV in a natural disaster scenario

This paper presents a design for an unmanned aerial vehicle (UAV) with a modular embedded architecture. This research is structured in three stages: (a) embedded systems, (b) communications link and (c) inertial navigation system. Systems integration was satisfactorily implemented and several remotely controlled flight tests were conducted. Finally, real time experiments were conducted during an environmental disaster in mountainous region, making aerial reconnaissance of areas devastated by mudslides. In this test the system has shown to be adequate and robust.

Environment exploration with multiple vehicles and FastSLAM technique

In this work we investigate the exploration of an environment with multiple vehicles using a strategy based on occupancy grids and a technique of simultaneous localization and mapping (SLAM). The exploration strategy uses concepts of costs and utility from frontier cells. Besides, the used SLAM method is based on a FastSLAM algorithm with landmarks extracted from visual sensors and a common map of characteristics. Both activities are coordinated by a central agent. The results obtained by simulation show that when two vehicles can communicate with a central agent building a common map of characteristics, the exploration task becomes more efficient than that performed with only one vehicle. This occurs because the exploration time is reduced while the accuracy of vehicle position and orientation is maintained or even improved with the same number of particles.

Exploration with FastSLAM technique

In this work we investigate the exploration of an environment with multiple vehicles using a strategy based on occupancy grids and a technique of simultaneous localization and mapping (SLAM). The exploration strategy uses concepts of costs and utility from frontier cells. Besides, the used SLAM method is based on a FastSLAM algorithm with landmarks extracted from visual sensors and a common map of characteristics. Both activities are coordinated by a central agent. The results obtained by simulation show that when two vehicles can communicate with a central agent building a common map of characteristics, the exploration task becomes more efficient than that performed with only one vehicle. This occurs because the exploration time is reduced while the accuracy of vehicle position and orientation is maintained or even improved with the same number of particles. In this paper, we present new results in the data association of FastSLAM algorithm.

Integration of People Detection and Simultaneous Localization and Mapping Systems for an Autonomous Robotic Platform

This paper presents the implementation of a people detection system for a robotic platform able to perform Simultaneous Localization and Mapping (SLAM), allowing the exploration and navigation of the robot considering people detection interaction. The robotic platform consists of a Pioneer 3DX robot equipped with an RGB-D camera, a Sick Lms200 sensor laser and a computer using the robot operating system ROS. The idea is to integrate the people detection system to the simultaneous localization and mapping (SLAM) system of the robot using ROS. Furthermore, this paper presents an evaluation of two different approaches for the people detection system. The first one uses a manual feature extraction technique, and the other one is based on deep learning methods. The manual feature extraction method in the first approach is based on HOG (Histogram of Oriented Gradients) detectors. The accuracy of the techniques was evaluated using two different libraries. The PCL library (Point Cloud Library) implemented in C ++ and the VLFeat MatLab library with two HOG variants, the original one, and the DPM (Deformable Part Model) variant. The second approaches are based on a Deep Convolutional Neural Network (CNN), and it was implemented using the MatLab MatConvNet library. Tests were made objecting the evaluation of losses and false positives in the peoples detection process in both approaches. It allowed us to evaluate the people detection system during the navigation and exploration of the robot, considering the real time interaction of people recognition in a semi-structured environment.

SLAM solution based on particle filter with outliers filtering in dynamic environments

Our work presents a proposal for SLAM (Simultaneous Localization and Mapping) problem in dynamic environments. A method for robot self-localization is described. During operations in dynamic environments, only static landmarks are used. For this purpose, an outliers filter was designed. It separates static and mobile landmarks that are included in a set of neglectable marks. A modified particle filter is proposed which estimates the pose of a robot by a new way. Furthermore, it resamples a new set of particles at each estimative of the pose of the robot. Set of particles is based on current observations made by robot using RGBD camera. ROS (Robot Operating System) platform is used in order to integrate all hardware and our application.

Fixed-wing state level HIL via factor graph incremental smoothing

Many robotics problems such as simultaneous localization and mapping (SLAM) and bundle adjustment (BA) can be formulated as a nonlinear least squares graph optimization. In these strategies, each node in the graph represents a robot position or a sensor measurement obtained at that position. Edges represent constraints between these nodes. The first part of the problem known as front-end is generating the graph. The second part, known as back-end, aims to find a configuration that best explains the constraints usually modeled as error functions. Clearly this is an expensive operation since every movement or measurement adds nodes to the graph that may become optimization hard for limited real-time applications. In order to overcome this problem, an innovative method called incremental smoothing, optimizes only a small subset of the graph nodes, reducing the computational load considerably. Besides, to explore this novel algorithm we propose a state-level hardware-in-the loop (HIL) for a small fixed-wing UAV. It consists of an aerial flight simulator extended with our sensor fusion plugin and the UAV aerodynamic model, an autopilot and a ground control station. MAVLink communications protocol interfaces these modules. Furthermore, we validate our approach with three mission scenarios exploiting different maneuvers ranging from a simple square flight to a much more challenging lawn-mower search pattern. Experiments have shown that the proposed method performed satisfactorily.

Data-driven analysis of kinaesthetic and tactile information for shape classification

Humans sense of touch consists in a complexity of sensors and nervous system. The information inferred by this system enables the daily dexterous manipulation tasks. In biological systems, there is no conscious prioritization of sensors while performing tactile exploration and the selection of exploratory movements is driven by learning instincts and data gathered by previous movements. The development of artificial systems tries to mimic such systems with engineered sensors and strategies for movement selection. This paper presents a data-driven analysis to the problem of sensor selection in the contour following for shape discrimination task. This task consists of a 4-DOF robotic finger exploring a set of 7 synthetic shapes. The data collected from the motors, inertial measurement unit, and magnetometer was analyzed applying principal component analysis and a multilayer perceptron neural network. Results show the variation of classification rate depending on the fingertip material and sensor considered. It is worth to observe that the magnetometer was the most robust in both cases.

Cooperating robots for mapping tasks with a multilayer perceptron

This paper proposes a Multilayer Perceptron application in a FastSLAM solution for landmarks update aimed at improve computational performance. The algorithm extracts landmarks from visual sensors, and uses a common features map for two robots. Both activities (i. e. location for the robots and exploration task) are coordinate by central agent. Landmarks update a than are done with a neural network using an Extended Kalman Filter, as usual in several research works in literature. Experiments have shown that generated errors obtained are equivalent in both methods. However, processing time is 10-12 times lower when using our proposed method. This contributes to attend real time requirements during autonomous robot operation.

3D Object Handling Support System in a CAVE Setup

To mimic reality through a computational system is not a trivial task. It requires that one manages resources, displays coherent imagery, as well as allowing real-time interaction. This work describes the implementation of a system which allows the display of content in a CAVE system. The system provides support to non conventional devices, such as Data Gloves, Position Tracker as well as devices such as a joystick and 3D Mouse. To display the functionality of the system we developed a prototype in which a user can handle a virtual object using his/her own hand. We discuss the system implementation based on the Instant Reality architecture and the X3D standard. We describe the development of a set of plugins for a Data Glove and Position Tracker. Through such plugins the user actions are detected and processed, which allows that such a user handles virtual content with his own hands.