Jan Hartmann

A comparison of feature descriptors for visual SLAM

Feature detection and feature description plays an important part in Visual Simultaneous Localization and Mapping (VSLAM). Visual features are commonly used to efficiently estimate the motion of the camera (visual odometry) and link the current image to previously visited parts of the environment (place recognition, loop closure). Gradient histogram-based feature descriptors, like SIFT and SURF, are frequently used for this task. Recently introduced binary descriptors, as BRIEF or BRISK, claim to offer similar capabilities at lower computational cost. In this paper, we will compare the most popular feature descriptors in a typical graph-based VSLAM algorithm using two publicly available datasets to determine the impact of the choice for feature descriptor in terms of accuracy and speed in a realistic scenario.

Sonar-based FastSLAM in an underwater environment using walls as features

A lot of research has been done in the area of Simultaneous Localization and Mapping (SLAM). It is particularly challenging in underwater environments where many of the self-localization methods used at land no longer work. This paper presents a method to use a rotating scanning sonar as the sole sensor to perform SLAM in harbors or natural waterways. We present a feature extraction process which is capable of extracting walls of arbitrary shape. These extracted wall features are then used to perform SLAM using the well-known FastSLAM algorithm. We show that SLAM is possible given this type of sensor and using our feature extraction process. The algorithm was validated on an open water test site and will be shown to provide localization accuracy generally within the error of the GPS ground truth.

Graph-based visual SLAM and visual odometry using an RGB-D camera

In this paper we present a real-time graph-based visual SLAM approach. The presented visual SLAM algorithm can be separated into three parts: feature extraction, data association, and SLAM back-end. We use FAST for feature detection and the Binary Robust Independent Elementary Features (BRIEF) as feature descriptor, which together provide a fast and stable feature extraction. The data association is solved using Locality Sensitive Hashing (LSH), which uses local hash tables and profits from binary feature descriptors. As SLAM back-end we use the general graph optimization framework g2o, which is designed to provide solutions to several SLAM variants. We further provide a novel approach to visual odometry by combining recent sensor measurements into a small pose graph and optimizing it using g2o. For finding potential neighbour nodes and loop closures we introduce the Global Feature Repository (GFR). GFR searches for loop closures and potential neighbours independent of their position in the graph. Finally, we show the accuracy and real-time ability of our algorithm by comparing it to a recently published benchmark dataset. We further provide some large-scale datasets using state-of-the-art laser localization as ground-truth.

A unified visual graph-based approach to navigation for wheeled mobile robots

The emergence of affordable 3D cameras in recent years has led to an increased interest in camera-based navigation solutions. Yet, while there have been significant efforts in the field of visual simultaneous localization and mapping (VSLAM), a complete navigation package that could rival popular laser-based solutions is not available. In this paper, we will therefore introduce visual solutions to SLAM, localization, and path planning in a unified graph-based framework with the main target of wheeled robots in industrial applications. Novel solutions will be introduced in the fields of place recognition and loop closing as well as localization. Our algorithms will be built for the Robot Operating System (ROS) and fully replace the popular gmapping and AMCL algorithms.

Self-adaptation for mobile robot algorithms using organic computing principles

Many mobile robot algorithms require tedious tuning of parameters and are, then, often suitable to only a limited number of situations. Yet, as mobile robots are to be employed in various fields from industrial settings to our private homes, changes in the environment will occur frequently. Organic computing principles such as self-organization, self-adaptation, or self-healing can provide solutions to react to new situations, e.g. provide fault tolerance. We therefore propose a biologically inspired self-adaptation scheme to enable complex algorithms to adapt to different environments. The proposed scheme is implemented using the Organic Robot Control Architecture (ORCA) and Learning Classifier Tables (LCT). Preliminary experiments are performed using a graph-based Visual Simultaneous Localization and Mapping (SLAM) algorithm and a publicly available benchmark set, showing improvements in terms of runtime and accuracy.

HANSE—A Low-Cost Autonomous Underwater Vehicle

HANSE is a low-cost Autonomous Underwater Vehicle (AUV) capable of solving many common underwater challenges. In this paper we will present HANSE’s modular and expandable hardware and software design, the underwater simulator MARS, as well as robust and efficient sonar-based localization and vision-based object detection algorithms, with which we have successfully participated in the Student Autonomous Underwater Vehicle Challenge in Europe (SAUC-E) 2011.

Slippery and sandy ground detection for hexapod robots based on organic computing principles and somatosensory feedback

Insect-like walking of six-legged robots on unstructured and rough terrain is considered a challenging task. Furthermore, the properties of the walking ground are considered an important issue and a challenge to insure stable adaptive walking. This paper will shed light on the applied decentralized controller approach for detecting slippery and sandy ground and also presents the proposed strategies to overcome these challenges. The novelty of our approach is the evaluation of the local current consumption and angular position of each legs joint as somatosensory feedback. Backward walking is proposed as a reflex reaction once a slippery ground is detected and an adaptive walking as soon as the robot detects sandy ground. Our approach is based on an organic computing architecture and was tested on a low-cost version of the OSCAR walking robot.

An approach to self-learning multicore reconfiguration management applied on Robotic Vision

Robotic Vision combined with real-time control imposes challenging requirements on embedded computing nodes in robots, exhibiting strong variations in computational load due to dynamically changing activity profiles. Reconfigurable Multiprocessor System-on-Chip offers a solution by efficiently handling the robots resources, but reconfiguration management seems challenging. The goal of this paper is to present first ideas on self-learning reconfiguration management for Reco nfigurable multicore computing nodes with dynamic reconfiguration of soft-core CPUs and HW accelerators, to support dynamically changing activity profiles in Robotic Vision scenarios.