Graph-based SLAM Research Articles

PADLoC: LiDAR-Based Deep Loop Closure Detection and Registration Using Panoptic Attention

A key component of graph-based SLAM systems is the ability to detect loop closures in a trajectory to reduce the drift accumulated over time from the odometry. Most LiDAR-based methods achieve this goal by using only the geometric information, disregarding the semantics of the scene. In this work, we introduce PADLoC for joint loop closure detection and registration in LiDAR-based SLAM frameworks. We propose a novel transformer-based head for point cloud matching and registration, and to leverage panoptic information during training time. In particular, we propose a novel loss function that reframes the matching problem as a classification task for the semantic labels and as a graph connectivity assignment for the instance labels. During inference, PADLoC does not require panoptic annotations, making it more versatile than other methods. Additionally, we show that using two shared matching and registration heads with their source and target inputs swapped increases the overall performance by enforcing forward-backward consistency. We perform extensive evaluations of PADLoC on multiple real-world datasets demonstrating that it achieves state-of-the-art results.

The Deep Convolutional Neural Network Role in the Autonomous Navigation of Mobile Robots (SROBO)

The ability to navigate unstructured environments is an essential task for intelligent systems. Autonomous navigation by ground vehicles requires developing an internal representation of space, trained by recognizable landmarks, robust visual processing, computer vision and image processing. A mobile robot needs a platform enabling it to operate in an environment autonomously, recognize the objects, and avoid obstacles in its path. In this study, an open-source ground robot called SROBO was designed to accurately identify its position and navigate certain areas using a deep convolutional neural network and transfer learning. The framework uses an RGB-D MYNTEYE camera, a 2D laser scanner and inertial measurement units (IMU) operating through an embedded system capable of deep learning. The real-time decision-making process and experiments were conducted while the onboard signal processing and image capturing system enabled continuous information analysis. State-of-the-art Real-Time Graph-Based SLAM (RTAB-Map) was adopted to create a map of indoor environments while benefiting from deep convolutional neural network (Deep-CNN) capability. Enforcing Deep-CNN improved the performance quality of the RTAB-Map SLAM. The proposed setting equipped the robot with more insight into its surroundings. The robustness of the SROBO increased by 35% using the proposed system compared to the conventional RTAB-Map SLAM.

Real-Time 3D Object Detection and SLAM Fusion in a Low-Cost LiDAR Test Vehicle Setup

Recently released research about deep learning applications related to perception for autonomous driving focuses heavily on the usage of LiDAR point cloud data as input for the neural networks, highlighting the importance of LiDAR technology in the field of Autonomous Driving (AD). In this sense, a great percentage of the vehicle platforms used to create the datasets released for the development of these neural networks, as well as some AD commercial solutions available on the market, heavily invest in an array of sensors, including a large number of sensors as well as several sensor modalities. However, these costs create a barrier to entry for low-cost solutions for the performance of critical perception tasks such as Object Detection and SLAM. This paper explores current vehicle platforms and proposes a low-cost, LiDAR-based test vehicle platform capable of running critical perception tasks (Object Detection and SLAM) in real time. Additionally, we propose the creation of a deep learning-based inference model for Object Detection deployed in a resource-constrained device, as well as a graph-based SLAM implementation, providing important considerations, explored while taking into account the real-time processing requirement and presenting relevant results demonstrating the usability of the developed work in the context of the proposed low-cost platform.

What if there was no revisit? Large-scale graph-based SLAM with traffic sign detection in an HD map using LiDAR inertial odometry

What if there was no revisit? Large-scale graph-based SLAM with traffic sign detection in an HD map using LiDAR inertial odometry

Low Data Overlap Rate Graph-Based SLAM with Distributed Submap Strategy

Low Data Overlap Rate Graph-Based SLAM with Distributed Submap Strategy

Moving object removal and surface mesh mapping for path planning on 3D terrain

ABSTRACT This paper proposes a method of moving object removal and map building for path planning on 3D terrain. Our method introduces a map representation named surface mesh maps, which is built from 3D LIDAR points using graph-based SLAM with moving object removal and polygon mesh reconstruction. We need moving object removal since moving objects, such as pedestrians, are included in raw maps built by SLAM in dynamic environments. For path planning, it is desirable that only stationary (static) objects remain in the maps. Occupancy voxel filtering can be used to remove moving objects. Stationary objects have high occupancy probability, and moving objects have low occupancy probability. However, due to shallow incidence of the LIDAR beams in 3D space, it is difficult to determine the correct probability of occupancy. Hence, we improve moving object removal using a new inverse measurement model with incident angles and a two-pass scheme to properly calculate the occupancy probability. In addition, we propose a path planning method using a graph search and the graph structures of the surface mesh maps. Path costs are given to the nodes and the arcs of the graph since path planning on 3D terrain requires an appropriate cost calculation. We introduce node costs and arc costs based on slopes, roughness, height differences, and travel distances. Path planning is performed by a graph search using the costs and graph pruning using collision detection and mesh boundary detection. The main contribution of our method is to combine SLAM with moving object removal and a graph search for path planning on large-scale 3D terrain. We conducted experiments by traveling over 5 km in outdoor dynamic environments. The results showed that the proposed method is capable of moving object removal and surface mesh mapping for path planning on 3D terrain.

3D mapping of outdoor environments by scan matching and motion averaging

Building 3D maps of surroundings is a fundamental and important task for autonomous robots operating in large previously unknown semi-structured or unstructured outdoor environments, where external methods for localization like GPS are not always available. Most of existing 3D mapping approaches are based on the local scan matching algorithm, which are difficult to achieve accurate mapping with good efficiency. In this paper, we propose an effective graph-based SLAM approach for 3D mapping of outdoor environments without odometry measurement. It first utilizes scan matching algorithm to sequentially estimate initial robot poses and progressively construct a pose graph, where each vertex denotes the robot pose at a point in time and each edge indicates the constraint between two robot poses. As scan matching algorithm cannot be error-free, the estimation of robot poses suffers from the error accumulation problem. To address this issue, the hypothesis of loop closure can be generated by an adaptive search circle and then verified by the global scan matching algorithm. Once the loop closure is verified, more edges should be added into the pose graph, which is subsequently optimized by the weighted motion averaging algorithm so as to obtain accurate robot poses for 3D mapping. Experimental results carried out on multiple outdoor environment data sets show that the proposed approach can build 3D maps with good performance.

EyeSAM: graph-based localization and mapping of retinal vasculature during intraocular microsurgery.

Robot-assisted intraocular microsurgery can improve performance by aiding the surgeon in operating on delicate micron-scale anatomical structures of the eye. In order to account for the eyeball motion that is typical in intraocular surgery, there is a need for fast and accurate algorithms that map the retinal vasculature and localize the retina with respect to the microscope. This work extends our previous work by a graph-based SLAM formulation using a sparse incremental smoothing and mapping (iSAM) algorithm. The resulting technique, "EyeSAM," performs SLAM for intraoperative vitreoretinal surgical use while avoiding spurious duplication of structures as with the previous simpler technique. The technique also yields reduction in average pixel error in the camera motion estimation. This work provides techniques to improve intraoperative tracking of retinal vasculature by handling loop closures and achieving increased robustness to quick shaky motions and drift due to uncertainties in the motion estimation.

Trajectory Estimation and Crowdsourced Radio Map Establishment From Foot-Mounted IMUs, Wi-Fi Fingerprints, and GPS Positions

Trajectory estimation is a problem derived from a common indoor positioning scenario: the user is perceiving the environment while in motion. This paper focuses on an instance: the dead reckoning data of the user are provided by foot-mounted inertial positioning and the observations in the environment are position estimated from global positioning system (GPS) and Wi-Fi received signal strengths. An approach is proposed to merge the three types of data from multiple users using the graph-based SLAM framework. In constructing the graph, Wi-Fi-based edges are established adaptively and GPS-based edges are established as prior information for the corresponding positions from the dead reckoning generated trajectories. The users' trajectories can be aligned and calibrated using the proposed graph construction strategies. Afterwards, a crowsourced Wi-Fi-based radio map (RM) can be established for other users who need the fingerprinting-based Wi-Fi indoor positioning service. The accuracy of the users' trajectories and the crowdsourced RM are tested to validate the proposed approach.

Efficient Velodyne SLAM with point and plane features

This paper develops and tests a plane based simultaneous localization and mapping algorithm capable of processing the uneven sampling density of Velodyne-style scanning LiDAR sensors in real-time. The algorithm uses an efficient plane detector to rapidly provide stable features, both for localization and as landmarks in a graph-based SLAM. When planes cannot be detected or when they provide insufficient support for localization, a novel constraint tracking algorithm selects a minimal set of supplemental point features to be provided to the localization solver. Several difficult indoor and outdoor datasets, totaling 6981 scans, each with $$\sim $$ 70,000 points, are used to analyze the performance of the algorithm without the aid of any additional sensors. The results are compared to two competing state-of-the-art algorithms, GICP and LOAM, showing up to an order of magnitude faster runtime and superior accuracy on all datasets, with loop closure errors of 0.14–0.95 m, compared to 0.44–66.11 m.

A Compact and Portable Implementation of Graph-based SLAM

This paper presents our compact and portable implementation of graph-based SLAM. Nowadays g2o, a popular graph-based SLAM implementation with state-of-the-art performance, is publicly available. However, g2o has a problem about portability because it depends on a number of external libraries. On contrary, our implementation depends solely on Eigen algebra library. By using Eigen, we were able to implement the SLAM algorithm as a cross-platform, header-only library with small amount of code. We evaluated the performance of our implementation using two public datasets. Our implementation was only two times slower than g2o and faster than iSAM, and we consider it reasonable for such a compact implementation.

On the Representation of Planes for Efficient Graph-based SLAM with High-level Features

On the Representation of Planes for Efficient Graph-based SLAM with High-level Features

Dual quaternion-based graphical SLAM

This paper presents a new parameterization approach for the graph-based SLAM problem and reveals the differences of two popular over-parameterized ways in the optimization procedure. In the SALM problem, constraints or relative transformations between any two poses are generally separated into translations plus 3D rotations, which are then described in a homogeneous transformation matrix (HTM) to simplify computational operations. This however introduces added complexities in frequent conversions between the HTM and state variables, due to their different representations. This new approach, unit dual quaternion (UDQ), describes a spatial transformation as a screw with only 8 elements. We show that state variables can be directly represented by UDQs, and how their relative transformations can be written with the UDQ product, without the trivial computations of HTM. Then, we explore the performances of the unit quaternion and the axis–angle representations in the graph-based SLAM problem, which have been successfully applied to over parameterize perturbations under the assumption of small errors. Based on public synthetic and real-world datasets in 2D and 3D environments, experimental results show that the proposed approach reduces greatly the computational complexity while obtaining the same optimization accuracies as the HTM-based algorithm, and the axis–angle representation is superior to be the quaternion in the case of poor initial estimations.

RGB-D SLAM Combining Visual Odometry and Extended Information Filter.

In this paper, we present a novel RGB-D SLAM system based on visual odometry and an extended information filter, which does not require any other sensors or odometry. In contrast to the graph optimization approaches, this is more suitable for online applications. A visual dead reckoning algorithm based on visual residuals is devised, which is used to estimate motion control input. In addition, we use a novel descriptor called binary robust appearance and normals descriptor (BRAND) to extract features from the RGB-D frame and use them as landmarks. Furthermore, considering both the 3D positions and the BRAND descriptors of the landmarks, our observation model avoids explicit data association between the observations and the map by marginalizing the observation likelihood over all possible associations. Experimental validation is provided, which compares the proposed RGB-D SLAM algorithm with just RGB-D visual odometry and a graph-based RGB-D SLAM algorithm using the publicly-available RGB-D dataset. The results of the experiments demonstrate that our system is quicker than the graph-based RGB-D SLAM algorithm.

Robust linear pose graph-based SLAM

This paper addresses a robust and efficient solution to eliminate false loop-closures in a pose-graph linear SLAM problem. Linear SLAM was recently demonstrated based on submap joining techniques in which a nonlinear coordinate transformation was performed separately out of the optimization loop, resulting in a convex optimization problem. This however introduces added complexities in dealing with false loop-closures, which mostly stems from two factors: (a) the limited local observations in map-joining stages and (b) the non block-diagonal nature of the information matrix of each submap. To address these problems, we propose a Robust Linear SLAM by (a) developing a delayed optimization for outlier candidates and (b) utilizing a Schur complement to efficiently eliminate corrupted information block. Based on this new strategy, we prove that the spread of outlier information does not compromise the optimization performance of inliers and can be fully filtered out from the corrupted information matrix. Experimental results based on public synthetic and real-world datasets in 2D and 3D environments show that this robust approach can cope with the incorrect loop-closures robustly and effectively.

SLAM 기술의 과거와 현재

This paper surveys past and state-of-the-art SLAM technologies. The standard methods for solving the SLAM problem are the Kalman filter, particle filter, graph, and bundle adjustment-based methods. Kalman filters such as EKF (Extended Kalman Filter) and UKF (Unscented Kalman Filter) have provided successful results for estimating the state of nonlinear systems and integrating various sensor information. However, traditional EKF-based methods suffer from the increase of computation burden as the number of features increases. To cope with this problem, particle filter-based SLAM approaches such as FastSLAM have been widely used. While particle filter-based methods can deal with a large number of features, the computation time still increases as the map grows. Graph-based SLAM methods have recently received considerable attention, and they can provide successful real-time SLAM results in large urban environments.

Improving Point Cloud Accuracy Obtained from a Moving Platform for Consistent Pile Attack Pose Estimation

We present a perception system for enabling automated loading with waist-articulated wheel loaders. To enable autonomous loading of piled materials, using either above-ground wheel loaders or underground load-haul-dump vehicles, 3D data of the pile shape is needed. However, using common 3D scanners, the scan data is distorted while the wheel loader is moving towards the pile. Existing methods that make use of 3D scan data (for autonomous loading as well as tasks such as mapping, localisation, and object detection) typically assume that each 3D scan is accurate. For autonomous robots moving over rough terrain, it is often the case that the vehicle moves a substantial amount during the acquisition of one 3D scan, in which case the scan data will be distorted. We present a study of auto-loading methods, and how to locate piles in real-world scenarios with nontrivial ground geometry. We have compared how consistently each method performs for live scans acquired in motion, and also how the methods perform with different view points and scan configurations. The system described in this paper uses a novel method for improving the quality of distorted 3D scans made from a vehicle moving over uneven terrain. The proposed method for improving scan quality is capable of increasing the accuracy of point clouds without assuming any specific features of the environment (such as planar walls), without resorting to a "stop-scan-go" approach, and without relying on specialised and expensive hardware. Each new 3D scan is registered to the preceding using the normal-distributions transform (NDT). After each registration, a mini-loop closure is performed with a local, per-scan, graph-based SLAM method. To verify the impact of the quality improvement, we present data that shows how auto-loading methods benefit from the corrected scans. The presented methods are validated on data from an autonomous wheel loader, as well as with simulated data. The proposed scan-correction method increases the accuracy of both the vehicle trajectory and the point cloud. We also show that it increases the reliability of pile-shape measures used to plan an efficient attack pose when performing autonomous loading.

Lifelong Map Learning for Graph-based SLAM in Static Environments

In this paper, we address the problem of lifelong map learning in static environments with mobile robots using the graph-based formulation of the simultaneous localization and mapping problem. The pose graph, which stores the poses of the robot and spatial constraints between them, is the central data structure in graph-based SLAM. The size of the pose graph has a direct influence on the runtime and the memory complexity of the SLAM system and typically grows over time. A robot that performs lifelong mapping in a bounded environment has to limit the memory and computational complexity of its mapping system. We present a novel approach to prune the pose graph so that it only grows when the robot acquires relevant new information about the environment in terms of expected information gain. As a result, our approach scales with the size of the environment and not with the length of the trajectory, which is an important prerequisite for lifelong map learning. The experiments presented in this paper illustrate the properties of our method using real robots.