Stefano Michieletto

Fast and Robust Multi-people Tracking from RGB-D Data for a Mobile Robot

This paper proposes a fast and robust multi-people tracking algorithm for mobile platforms equipped with a RGB-D sensor. Our approach features an efficient point cloud depth-based clustering, an HOG-like classification to robustly initialize a person tracking and a person classifier with online learning to manage the person ID matching even after a full occlusion. For people detection, we make the assumption that people move on a ground plane. Tests are presented on a challenging real-world indoor environment and results have been evaluated with the CLEAR MOT metrics. Our algorithm proved to correctly track 96% of people with very limited ID switches and few false positives, with an average frame rate of 25 fps. Moreover, its applicability to robot-people following tasks have been tested and discussed.

A Software Architecture for RGB-D People Tracking Based on ROS Framework for a Mobile Robot

This paper describes the software architecture of a distributed multi-people tracking algorithm for mobile platforms equipped with a RGB-D sensor. Our approach features an efficient point cloud depth-based clustering, an HOG-like classification to robustly initialize a person tracking and a person classifier with online learning to drive data association. We explain in details how ROS functionalities and tools play an important role in the possibility of the software to be real time, distributed and easy to configure and debug.

Learning how to approach industrial robot tasks from natural demonstrations

In the last years, Robot Learning from Demonstration (RLfD) [1] [2] has become a major topic in robotics research. The main reason for this is that programming a robot can be a very difficult and time spending task. The RLfD paradigm has been applied to a great variety of robots, but it is still difficult to make the robot learn a task properly. Often the teacher is not an expert in the field, and viceversa an expert could not know well enough the robot to be a teacher. With this paper, we aimed at closing this gap by proposing a novel motion re-targeting technique to make a manipulator learn from natural demonstrations. A RLfD framework based on Gaussian Mixture Models (GMM) and Gaussian Mixture Regressions (GMR) was set to test the accuracy of the system in terms of precision and repeatability. The robot used during the experiments is a Comau Smart5 SiX and a novel virtual model of this manipulator has also been developed to simulate an industrial scenario which allows valid experimentation while avoiding damages to the real robot.

A distributed perception infrastructure for robot assisted living

This paper presents an ambient intelligence system designed for assisted living. The system processes the audio and video data acquired from multiple sensors spread in the environment to automatically detect dangerous events and generate automatic warning messages. The paper presents the distributed perception infrastructure that has been implemented by means of an open-source software middleware called NMM. Different processing nodes have been developed which can cooperate to extract high level information about the environment. Examples of implemented nodes running algorithms for people detection or face recognition are presented. Experiments on novel algorithms for people fall detection and sound classification and localization are discussed. Eventually, we present successful experiments in two test bed scenarios.

Calibration of a dual-laser triangulation system for assembly line completeness inspection

In controlled industrial environments, laser triangulation is an effective technique for 3D reconstruction, which is increasingly being used for quality inspection and metrology. In this paper, we propose a method for calibrating a dual laser triangulation system - consisting of a camera, two line lasers, and a linear motion platform designed to perform completeness inspection tasks on an assembly line. Calibration of such a system involves the recovery of two sets of parameters - the plane parameters of the two line lasers, and the translational direction of the linear motion platform. First, we address these two calibration problems separately. While many solutions have been given for the former problem, the latter problem has been largely ignored. Next, we highlight an issue specific to the use of multiple lasers - that small errors in the calibration parameters can lead to misalignment between the reconstructed point clouds of the different lasers. We present two different procedures that can eliminate such misalignments by imposing constraints between the two sets of calibration parameters. Our calibration method requires only the use of planar checkerboard patterns, which can be produced easily and inexpensively.

Why teach robotics using ROS

This paper focuses on the key role played by the adoption of a framework in teaching robotics with a computer science approach in the master in Computer Engineering. The framework adopted is the Robot Operating System (ROS), which is becoming a standard de facto inside the robotics community. The educational activities proposed in this paper are based on a constructionist approach. The Mindstorms NXT robot kit is adopted to trigger the learning challenge. The ROS framework is exploited to drive the students programming methodology during the laboratory activities and to allow students to exercise with the major computer programming paradigms and the best programming practices. The major robotics topics students are involved with are: acquiring data from sensors, connecting sensors to the robot, and navigate the robot to reach the final goal. The positive effects given by this approach are highlighted in this paper by comparing the work recently produced by students with the work produced in the previous years in which ROS was not yet adopted and many different software tools and languages were used. The results of a questionnaire are reported showing that we achieved the didactical objectives we expected as instructors.

GMM-Based Single-Joint Angle Estimation Using EMG Signals

This paper aims to explore the possibility to use Electromyography (EMG) to train a Gaussian Mixture Model (GMM) in order to estimate the bending angle of a single human joint. In particular, EMG signals from eight leg muscles and the knee joint angle are acquired during a kick task from three different subjects. GMM is validated on new unseen data and the classification performances are compared with respect to the number of EMG channels and the number of collected trials used during the training phase. Achieved results show that our framework is able to obtain high performances even using few EMG channels and with a small training dataset (Normalized Mean Square Error: 0.96, 0.98, 0.98 for the three subjects, respectively), opening new and interesting perspectives for the hybrid control of humanoid robots and exoskeletons.

ROS-I Interface for COMAU Robots

The following paper presents the ROS-I interface developed to control Comau manipulators. Initially, the Comau controller allowed users to command a real robot thanks to motion primitives formulated through a Comau motion planning library. Now, either a ROS or a non ROS -compliant platform can move either a real or a virtual Comau robot using any motion planning library. Comau modules have been wrapped within ROS and a virtual model of a Comau robot has been created. The manufacturer controller has been innovatively used to drive both the real and the simulated automata.

Real-Time 3D Model Reconstruction with a Dual-Laser Triangulation System for Assembly Line Completeness Inspection

In this paper, we present an improved version of our Dual Laser Triangulation System, a low-cost color 3D model acquisition system built with commonly available machine vision products. The system produces a color point cloud model of scanned objects that can be used to perform completeness inspection tasks on assembly lines. In particular, we show that model acquisition and reconstruction can be achieved in real-time using such a low-cost solution. Our results demonstrate that 3D-based inspection can be achieved readily and economically in a real industrial production environment.

3DComplete: Efficient completeness inspection using a 2.5D color scanner

In this paper, we present a low-cost and highly configurable quality inspection system capable of capturing 2.5D color data, created using off-the-shelf machine vision components, open-source software libraries, and a combination of standard and novel algorithms for 2.5D data processing. The system uses laser triangulation to capture 3D depth, in parallel with a color camera and a line light projector to capture color texture, which are then combined into a color 2.5D model in real-time. Using many examples of completeness inspection tasks that are extremely difficult to solve with current 2D-based methods, we demonstrate how the 2.5D images and point clouds generated by our system can be used to solve these complex tasks effectively and efficiently. Our system is currently being integrated into a real production environment, showing that completeness inspection incorporating 3D technology can be readily achieved in a short time at low costs.