Kristijan Lenac

Pymote: High Level Python Library for Event-Based Simulation and Evaluation of Distributed Algorithms

In recent years we have witnessed strong development and widespread use of powerful wirelessly connected platforms, thus the set of the related problems that need to be solved by distributed algorithms is growing rapidly. Some of them present large obstacles in harnessing the full potential of this new technology, so there is an imminent need for a fast and easy evaluation of new ideas and approaches. Simulation is a fundamental part of distributed algorithm design and evaluation process. In this paper, we present a library for event-based simulation and evaluation of distributed algorithms. This library provides a set of simple but powerful tools with a goal to ease virtual setup of a complex system such as a distributed network of communicating entities and to define, simulate, and analyze its behavior. In order to reduce a huge problem space inherent in such systems, our library is using a high level of abstraction. This is made possible by a strict and complete definition of the distributed computing environment. The library is implemented in Python whose simple and expressive syntax provides a possibility of minimal implementations and a mild learning curve. In addition to executing automated simulations or larger experiments, the library fully supports interactive mode along with a step-by-step execution, which can be a very powerful combination.

Fast Genetic Scan Matching Using Corresponding Point Measurements in Mobile Robotics

In this paper we address the problem of aligning two partially overlapping surfaces represented by points obtained in subsequent 2D scans for mobile robot pose estimation. The measured points representation contains incomplete measurements. We solve this problem by minimizing an alignment error via a genetic algorithm. Moreover, we propose an alignment metric based on a look-up table built during the first scan. Experimental results related to the convergence of the proposed algorithm are reported. We compare our approach with other scan matching algorithms proposed in the literature, and we show that our approach is faster and more accurate than the others.

SPATIAL MAP BUILDING USING FAST TEXTURE ANALYSIS OF ROTATING SONAR SENSOR DATA FOR MOBILE ROBOTS

This paper presents a novel, fast algorithm for accurate detection of the shape of targets around a mobile robot using a single rotating sonar element. The rotating sonar yields an image built up by the reflections of an ultrasonic beam directed at different scan angles. The image is then interpreted with an image-understanding approach based on texture analysis. Several important tasks are performed in this way, such as noise removal, echo correction and restoration. All these processes are obtained by estimating and restoring the degree of texture continuity. Texture analysis, in fact, allows us to look at the image on a large scale thus giving the possibility to infer the overall behavior of the reflection process. The algorithm has been integrated in a mobile robot. However, the algorithm is not suitable for working during the mobile robot movement, rather it can be used during the period when the robot stays in a fixed position.

Integrating Blink Click interaction into a head tracking system: implementation and usability issues

In the area of human–computer interaction, contemporary head tracking systems are often used as camera-based mouse emulators. While head movement detection provides the basis for related mouse shifting and positioning, standard click actions are usually emulated using stillness counter techniques such as Dwell Click (DC). However, these techniques can be a source of enlarged interaction burden, as users often have to struggle with time-consuming repetitive UI actions. This paper proposes a novel version of Blink Click (BC) action called B2C, based on double eye blink detection, as a valuable supplement for faster mouse click emulation. The integration of the proposed BC action into an existing head tracking system is presented, and implementation issues are thoroughly analyzed. Usability testing of the proposed B2C interaction model, along with the already embedded DC model, has been carried out, providing both quantitative and qualitative outcomes. The results show efficiency improvement as well as a higher level of users’ satisfaction when using the proposed version of BC, thus making it a strong candidate to become a standard feature within the computer-vision-based mouse emulation.

Robust and accurate genetic scan matching algorithm for robotic navigation

In this paper we propose a scan matching algorithm for robotic navigation based on the combination of ICP and genetic optimization. Since the genetic algorithm is robust but not very accurate, and ICP is accurate but not very robust, it is natural to use the two algorithms in a cascade fashion: first we run a genetic optimization to find an approximate but robust matching solution and then we run ICP to increase accuracy. The proposed genetic algorithm is very fast due to a lookup table formulation and very robust against large errors in both distance and angle during scan data acquisition. It is worth mentioning that large scan errors arise very commonly in mobile robotics due, for instance, to wheel slippage. We show experimentally that the proposed algorithm successfully copes with large localization errors.

Fast Genetic Scan Matching in Mobile Robotics

In this chapter, we address the problem of aligning two partially overlapping two-dimensional maps represented by data sets acquired using range sensors. The measured data may be incomplete and noisy. To solve this problem, we used a genetic algorithm for minimizing an alignment error. A lookup-based fitness function was devised. The considered range devices are laser and focalized ultrasonic scanners. Scan matching is often considered for mobile robot displacement and/or pose estimation tasks. We experimentally show that the algorithm is robust against noise and incomplete measurements and that it can be used for both the mentioned tasks. Moreover, the proposed algorithm is suitable for both local and global robot pose estimation. Experimental results related to the convergence, accuracy and speed of the proposed algorithm with different coding approaches are reported. We compare our approach with other scan matching algorithms proposed in the literature, and we show that our approach is faster and more accurate than the others.

An effective and efficient hybrid scan matching algorithm for mobile object applications

In this paper we analyze hybrid scan matching algorithms and we test their performances in typical mobile applications. Since the genetic algorithm is robust but not very accurate, and ICP is accurate but not very robust, it is natural to use the two algorithms in a cascade fashion: first we run a genetic optimization to find an approximate but robust matching solution and then we run the Iterative Closest Point (ICP) algorithm to increase the accuracy. The proposed genetic algorithm is very fast due to a look-up table formulation and very robust against large errors in both distance and angle during scan data acquisition. It is worth mentioning that large scan errors arise very commonly in mobile object applications due, for instance, to wheel slippage or when closing loops. We show experimentally that the proposed algorithm successfully copes with large localization errors.

Virtual Reality Applications with Oculus Rift and 3D Sensors

In this paper we describe our experiences with skeletal tracking using Unreal Engine 4 with Oculus Rift and Xbox 360 Kinect while building a tool for rehabilitation of patients with impaired motor skills. We give an overview of the implemented solution, describe the problems encountered and how they were solved.

Short-term people re-identification using image classification

This work addresses the problem of short-term people re-identification, an interesting field with applications in different domains like banks, airports, border crossings etc. For the purpose of this research, an algorithm is presented based on image classification. It is divided in three parts: building a dataset, creating a model, and real-time people re-identification. For every person in dataset, bag of features is used for image classification and features are extracted using MinEigen (Minimum Eigenvalue Algorithm) and SURF (Speeded Up Robust Features). Features are further trained using SVM (Support Vector Machine) and then exported as a model which is used for real time prediction and person recognition.

People tracking using synthetically generated depth maps

Modern 3D sensors using time-of-flight and structured light technology are increasingly being used in a number of interesting applications. One such application is people tracking where the available 3D information enables more robust human detection and tracking methods than those based on classic video cameras. A large data set is typically required both for testing and tuning of the solution. We describe a methodology for efficient development of 3D applications using synthetically generated depth maps to complement the available data set obtained with real cameras. To demonstrate the effectiveness of the approach, a case study in the area of people counting with 3D sensors is presented.