ROBOVIS 2020 Abstracts

Area 1 - Computer Vision

Short Papers

Paper Nr:	14
Title:	Evaluating Person Re-identification Performance on GAN-enhanced Datasets
Authors:	Daniel Hofer and Wolfgang Ertel
Abstract:	Person re-identification remains a hard task for AI systems because high intra-class variance across different cameras, angles and lighting conditions make it difficult to create a reliable re-identification system. Since only small datasets for person re-id tasks are available, in recent years Generative Adversarial Networks (GANs) have become popular to improve intra-class variance to train more robust re-identification frameworks. In this work we evaluate an Inception-ResNet-v2 using triplet loss, introduced by (Weinberger and Saul, 2009), which works very well for face re-identification and use it for full-body person re-identification. The network is trained without GAN generated images to get a baseline accuracy of the network. In further experiments, the network is trained by adding constantly rising amounts of synthetic images produced by two image generators using different generating approaches.
Download

Area 2 - Intelligent Systems

Short Papers

Paper Nr:	24
Title:	Fuzzy Logic-based Adaptive Cruise Control for Autonomous Model Car
Authors:	Khaled Alomari, Ricardo C. Mendoza, Stephan Sundermann, Daniel Goehring and Raúl Rojas
Abstract:	One of the most critical challenges for the driver during highway driving is to adjust the vehicle speed continuously to maintain safe distance in respect to the heading vehicles or highway traffic. Neglecting a safe distance can cause deadly collisions, especially at high velocities. Thus, car speed must adapt smoothly and efficiently in relation to the velocity of the vehicle in front and the headway distance. Adaptive Cruise Control (ACC) is an Advanced Driver Assistant System that is used to control both velocity and distance at the same time. The system needs either a PID controller per state or a MIMO system. In this paper, we propose an ACC using a Fuzzy Logic approach for an autonomous model car called “AutoMiny.” AutoMiny was developed at the Dahlem Center for Machine Learning and Robotics at Freie Universität Berlin. It navigates by correcting its orientation error given by a global localization system and a pre-built grid map. The proposed controller can handle two states with differently designed profiles, and we will compare the performance of our approach with that of a PID controller.
Download

Area 3 - Robotics

Full Papers

Paper Nr:	8
Title:	A Visual Intelligence Scheme for Hard Drive Disassembly in Automated Recycling Routines
Authors:	Erenus Yildiz, Tobias Brinker, Erwan Renaudo, Jakob J. Hollenstein, Simon Haller-Seeber, Justus Piater and Florentin Wörgötter
Abstract:	As the state-of-the-art deep learning models are taking the leap to generalize and leverage automation, they are becoming useful in real-world tasks such as disassembly of devices by robotic manipulation. We address the problem of analyzing the visual scenes on industrial-grade tasks, for example, automated robotic recycling of a computer hard drive with small components and little space for manipulation. We implement a supervised learning architecture combining deep neural networks and standard pointcloud processing for detecting and recognizing hard drives parts, screws, and gaps. We evaluate the architecture on a custom hard drive dataset and reach an accuracy higher than 75% in every component used in our pipeline. Additionally, we show that the pipeline can generalize on damaged hard drives. Our approach combining several specialized modules can provide a robust description of a device usable for manipulation by a robotic system. To our knowledge, we are the pioneers to offer a complete scheme to address the entire disassembly process of the chosen device. To facilitate the pursuit of this issue of global concern, we provide a taxonomy for the target device to be used in automated disassembly scenarios and publish our collected dataset and code.
Download

Paper Nr:	15
Title:	Deployment of Multi-agent Pathfinding on a Swarm of Physical Robots Centralized Control via Reflex-based Behavior
Authors:	Ján Chudý, Nestor Popov and Pavel Surynek
Abstract:	Multi-agent pathfinding is a problem of finding paths for multiple agents from their initial configuration to their goal configuration that results in a plan execution without collisions. In this paper, we deploy MAPF solutions on a swarm of small mobile robots. During the plan execution, we mitigate the problem of desynchronization that comes with the plan execution on physical hardware using the reflex-based behavior of the robots. Such deployment can help researchers and educators to demonstrate and test their findings in the physical world. The robot has a line-following capability that can be used for simulation of discrete MAPF solutions. The control curves are displayed in real-time on a display on which the robots move during their path execution. A prototype of the deployment was built and tested experimentally.
Download

Paper Nr:	23
Title:	GALNet: An End-to-End Deep Neural Network for Ground Localization of Autonomous Cars
Authors:	Ricardo C. Mendoza, Bingyi Cao, Daniel Goehring and Raúl Rojas
Abstract:	Odometry based on Inertial, Dynamic and Kinematic data (IDK-Odometry) for autonomous cars has been widely used to compute the prior estimation of Bayesian localization systems which fuse other sensors such as camera, RADAR or LIDAR. IDK-Odometry also gives the vehicle information by way of emergency when other methods are not available. In this work, we propose the use of deep neural networks to estimate the relative pose of the car given two timestamps of inertial-dynamic-kinematic data. We show that a neural network can find a solution to the optimization problem employing an approximation of the Vehicle Slip Angle (VSA). We compared our results to an IDK-Odometry system based on an Unscented Kalman Filter and Ackermann-wheel odometry. To train and test the network, we used a dataset which consists of ten driven trajectories with our autonomous car. Moreover, we successfully improved the results of the network employing collected data with a model autonomous car in order to increase the trajectories with high VSA.
Download

Paper Nr:	25
Title:	Intelligent Algorithms for Non-parametric Robot Calibration
Authors:	Marija Turković, Marko Švaco and Bojan Jerbić
Abstract:	In this paper, a novel method for non-parametric robot calibration which uses intelligent algorithms is proposed. The non-parametric calibration should prove very useful, because it does not need to identify the geometric parameters of the robot as is the case in parametric calibration. Instead, only the position measurements need to be provided. This could potentially lead to a cheaper and faster calibration process which could simplify its application on different and unique robot geometries. The biggest issue of using neural networks is that they require a lot of data, while for the process of robot calibration a very limited number of measurements is usually collected. In this experiment, the improvement of the hyperparameters of the neural network was attempted by using the genetic algorithms. Simulations also showed that the parametric optimization converges faster and that feed-forward back-propagating neural networks could not correctly simulate the behaviour of complex robots, or problems which used small datasets. However, for simple robot geometries and massive datasets, the neural network successfully simulated the behaviour of the robot. Although the number of measurements needed was well beyond the scope for real world applications, a few possible improvements were suggested for future research.
Download

Short Papers

Paper Nr:	3
Title:	DCNN-based Screw Classification in Automated Disassembly Processes
Authors:	Erenus Yildiz and Florentin Wörgötter
Abstract:	E-waste recycling is thriving yet there are many challenges waiting to be addressed until high-degree, device-independent automation is possible. One of these challenges is to have automated procedures for screw classification. Here we specifically address the problem of classification of the screw heads and implement a universal, generalizable, and extendable screw classifier which can be deployed in automated disassembly routines. We selected the best performing state-of-the-art classifiers and compared their performance to that of our architecture, which combines a Hough transform with the top-performing state-of-the-art deep convolutional neural network proven by our experiments. We show that our classifier outperforms currently existing methods by achieving 97% accuracy while maintaining a high speed of computation. Data set and code of this study are made public.
Download

Paper Nr:	10
Title:	Robotic Finger Design Workflow for Adaptable Industrial Assembly Tasks
Authors:	Adam Wolniakowski, Anders P. Lindvig, Nicolai Iversen, Norbert Krüger and Aljaž Kramberger
Abstract:	In this work, we introduce a web-based system connected to a simulation framework that can be used to facilitate the design of industrial fingers. We provide an overview of the state of the art and of the currently used manual gripper finger design methods prevailing in the industry. With a concrete use case we demonstrate the advantages in terms of quality and saved time for designing the fingers when utilizing our presented framework compared to a common manual method of designing the gripper fingers.
Download

Paper Nr:	16
Title:	Improving Learning in a Mobile Robot using Adversarial Training
Authors:	Todd W. Flyr and Simon Parsons
Abstract:	This paper reports research training a mobile robot to carry out a simple task. Specifically, we report on experiments in learning to strike a ball to hit a target on the ground. We trained a neural network to control a robot to carry out this task with data from a small number of trials with a physical robot. We compare the results of using this neural network with that of using a neural-network trained with this dataset plus the output of a generative adversarial network (GAN) trained on the same data. We find that the neural network that uses the GAN-generated data provides better performance. This relationship holds as we present the robot with generalized versions of this task. We also find that we can produce acceptable results with an exceptionally small initial dataset. We propose that this is a possible way to solve the “big data” problem, where training a neural network to learn physical tasks requires a large corpus of labeled trial data that can be difficult to obtain.
Download

Paper Nr:	17
Title:	Airborne Visual Tracking of UAVs with a Pan-Tilt-Zoom Camera
Authors:	Athanasios Tsoukalas, Nikolaos Evangeliou, Nikolaos Giakoumidis and Anthony Tzes
Abstract:	The visual detection and tracking of UAVs using a Pan-Tilt-Zoom (PTZ) camera attached to another aerial platform is the scope of this article. The long-term tracker is performing image background subtraction using visual homography and is used to initialize a short term tracker that works in parallel to enhance the tracking for various motions. The moving UAV is detected using optical flow concepts and its bounding box encapsulates its detected features. A Kalman predictor provides a robust smooth tracking of the bounding box in the temporary absence of a detected UAV. The camera pans and tilts so as the tracked UAV is centered within its Field-of-View and zooms in order to expand the UAV’s view. Experimental results are offered using an evader-tracker UAV-group to validate the presented tracking algorithm.
Download

Paper Nr:	18
Title:	Building a Camera-based Smart Sensing System for Digitalized On-demand Aircraft Cabin Readiness Verification
Authors:	Luis Unzueta, Sandra Garcia, Jorge Garcia, Valentin Corbin, Nerea Aranjuelo, Unai Elordi, Oihana Otaegui and Maxime Danielli
Abstract:	Currently, aircraft cabin operations such as the verification of taxi, take-off, and landing (TTL) cabin readiness are done manually. This results in an increased workload for the crew, operational inefficiencies, and a non-negligible risk of human errors in handling safety-related procedures. For TTL, specific cabin readiness requirements apply to the passenger, to the position of seat components and cabin luggage. The usage of cameras and vision-based object-recognition algorithms may offer a promising solution for specific functionalities such as cabin luggage detection. However, building a suitable camera-based smart sensing system for this purpose brings many challenges as it needs to be low weight, with competitive cost and robust recognition capabilities on individual seat level, complying with stringent constraints related to airworthiness certification. This position paper analyzes and discusses the main technological factors that system designers should consider for building such an intelligent system. These include the sensor setup, system training, the selection of appropriate camera sensors and lenses, AI-processors, and software tools for optimal image acquisition and image content analysis with Deep Neural Network (DNN)-based recognition methods. Preliminary tests with pre-trained generalist DNN-based object detection models are also analyzed to assist with the training and deployment of the recognition methods.
Download

Paper Nr:	19
Title:	The Answer Is Blowing in the Wind: Directed Air Flow for Socially-acceptable Human-Robot Interaction
Authors:	Vincent Zhang, Natalie Friedman, David Goedicke, Dmitriy Rivkin, Michael Jenkin, Xue Liu and Gregory Dudek
Abstract:	A key problem for a robot moving within a social environment is the need to capture the attention of other people using the space. In most use cases, this capture of attention needs to be accomplished in a socially acceptable manner without loud noises or physical contact. Although there are many communication mechanisms that might be used to signal the need for a person’s attention, one particular modality that has received little interest from the robotics community is the use of controlled air as a haptic signal. Recent work has demonstrated that controlled air can provide a useful signal in the social robot domain, but what is the best mechanism to provide this signal? Here, we evaluate a number of different mechanisms that can provide this attention-seeking communication. We demonstrate that many different simple haptic air delivery systems can be effective and show that air on and air off haptic events have very similar time courses using these delivery systems.
Download

Paper Nr:	20
Title:	Towards Optical Flow Ego-motion Compensation for Moving Object Segmentation
Authors:	Renáta N. Elek, Artúr I. Károly, Tamás Haidegger and Péter Galambos
Abstract:	Optical flow is an established tool for motion detection in the visual scene. While optical flow algorithms usually provide accurate results, they can not make a difference between image-space displacements originated from moving objects in the space and the ego-motion of the moving viewpoint. In the case of optical flow-based moving object segmentation, camera ego-motion compensation is essential. Hereby, we show the preliminary results of a moving viewpoint optical flow ego-motion filtering method, using two dimensional optical flow, image depth information and the camera holder robot arm’s state of motion. We tested its accuracy through physical experiments, where the camera was fixed on a robot arm, and a test object was attached onto an other robot arm. The test object and the camera were moved relative to each other along given trajectories in different scenarios. We validated our method for optical flow background filtering, which showed 94.88% mean accuracy in the different test cases. Furthermore, we tested the proposed algorithm for moving object state of motion estimation, which showed high accuracy in the case of translational and rotational movements without depth variation, but lower accuracy, when the relative motion produced change in depth, or the camera and the moving object move in the same directions. The proposed method with future work including outlier filtering and optimisation could become useful in various robot navigation applications and optical flow-based computer vision problems.
Download

Paper Nr:	26
Title:	Detection of Objects and Trajectories in Real-time using Deep Learning by a Controlled Robot
Authors:	Adil Sarsenov, Aigerim Yessenbayeva, Almas Shintemirov and Adnan Yazici
Abstract:	Nowadays, there are many different approaches to detect objects as well as to determine the trajectory of an object. Each of these approaches has its advantages and disadvantages in terms of real-time use for various applications. In this study, we propose an approach to detect objects in real-time using the YOLOv3 deep learning algorithm and plot the trajectory of an object using 2D LIDAR and depth cameras on a robot. The laser rangefinder allows us to find distances to objects from a certain angle, but does not provide accurate object detection of the object class. In order to detect the object in real-time and discover the class to which the object belongs, we formed YOLOv3 deep learning model using transfer learning on several classes from data sets of publicly accessible images. We also measured the distance to an object using a depth camera with LIDAR together to determine and estimate the trajectory of objects. In addition, these detected trajectories are smoothed by polynomial regression. Our experiments in a laboratory environment show that YOLOv3 with 2D LIDAR and depth camera on a controlled robot can be used fairly accurately and efficiently in real-time situations for the detection of objects and trajectories necessary for various applications.
Download