Graph-Based SLAM and Open Source Tools. Giorgio Grisetti

Transcription

1 Graph-Based SLAM and Open Source Tools Giorgio Grisetti

2 SLAM SLAM= Simultaneous Localization and Mapping Estimate: the map of the environment the trajectory of a moving device using a sequence of sensor measurements.

3 SLAM SLAM= Simultaneous Localization and Mapping Estimate: these quantities the map of the environment are correlated the trajectory of a moving device using a sequence of sensor measurements.

4 Why SLAM is so Important? Most applications require to localize a device in a map. A map cannot always be provided. Do SLAM! industrial applications Google street view service robotics autonomous cars

5 History 1960 Bundle Adjustment (~10 images) 1970 Recursive Partitioning (~1000 images) time and size of the environment 1990 (SLAM is born) 1993 Scan-Matching, Iconic maps 1997 Graph-SLAM 2000 Modern Sparse Matrix Techniques for BA 2002 FastSLAM 2003 ESDF, Treemap, TJTF 2005 SAM 2006 Appearance-Based Localization 2006 Efficient Graph-Based SLAM Towards the unification of SfM and SLAM

6 Graph-based SLAM in a Nutshell Node: robot position and sensor measurement. Edge: spatial transformation between nodes depends on the matching of scans

7 Graph-based SLAM in a Nutshell Node: robot position and sensor measurement. Edge: spatial transformation between nodes depends on the matching of scans

8 Graph-based SLAM in a Nutshell The graph abstracts away the measurements The most likely is trajectory obtained by optimization.

9 Graph-based SLAM in a Nutshell The graph abstracts away the measurements The most likely is trajectory obtained by optimization. like this

10 Graph-based SLAM in a Nutshell and thus the map.

11 An Example using Lasers

12 ... or Vision

13 Front-end and Back-end Front-end: extracts constraints from the sensor data Back-end: optimizes the pose-graph to reduce the error caused by the constraints raw data graph construction (front-end) node positions edges graph optimization (back-and) Insight: intermediate solutions are needed to make good data associations

14 How Does the Graph Look Like? It has n nodes x=x 1:n Each node x i is a 2D or 3D transformation representing the pose of the robot at time t i. There is a constraint e ij between the node x i and the node x j if either the robot observed the same part of the environment from both x i and x j and, via this common observation it constructs a virtual measurement about the position of x j seen from. Or the positions are subsequent in time and there is an odometry measurement between the two.

15 How Does the Graph Look Like? It has n nodes x=x 1:n Each node x i is a 2D or 3D transformation representing the pose of the robot at time t i. There is a constraint e ij between the node x i and the node x j if either the robot observed the same part of the environment from both x i and x j and, via this common observation it constructs a virtual measurement about the position of x j seen from. Or the positions are subsequent in time and there is an odometry measurement between the two. Measurement from x i x i x j Measurement from x J

16 How Does the Graph Look Like? It has n nodes x=x 1:n Each node x i is a 2D or 3D transformation representing the pose of the robot at time t i. x i x j There is a constraint e ij between the node x i and the node x j if either the robot observed the same part of the environment from both x i and x j and, via this common observation it constructs a virtual measurement about the position of x j seen from. Or the positions are subsequent in time and there is an odometry measurement between the two. The edge represents the position of x j seen from x i, based on the observations

17 How Does the Graph Look Like? It has n nodes x=x 1:n Each node x i is a 2D or 3D transformation representing the pose of the robot at time t i. There is a constraint e ij between the node x i and the node x j if either the robot observed the same part of the environment from both x i and x j and, via this common observation it constructs a virtual measurement about the position of x j seen from. Or the positions are subsequent in time and there is an odometry measurement between the two. x i X i+1 The edge represents the odometry measurement

18 Pose Graph The input for the optimization procedure is a graph annotated as follows: observation of from edge Goal: nodes error Find the assignment of poses to the nodes of the graph which minimizes the negative log likelihood of the observations:

19 OpenSLAM.org A platform to share SLAM code Research Oriented Not only full solutions, also subsystems can be uploaded Enables for comparative experiments As a SLAM user Provide a set of mapping systems to people interested in developing high level applications Reference implementations Promoting the use of SLAM technologies in industrial contexts As a SLAM developer Benchmarking/comparison Improving and extending the existing solutions

20 Open Source Tools For 3D Full Solutions 6 DoF SLAM system [Nuechter et al.] (laser-only) PTAM [Klein et al.] (camera-only) VSLAM stack in ROS (stereo + mono camera) Bundler [Snavely et al.] (unordered set of uncalibrated images) Back Ends g2o [Kuemmerle et al.] SAM [Kaess et al.] Hogman [Grisetti et al.] (only pose graphs) SBA [] Front ends PCL [Rusu et al.] (a veeeery general library on point clouds) FAB-Map [Cummings et al.] (Loop Closing on images) on

21 Some Applications

22 Conclusions SLAM is an active field The graph-based formulation allows for an efficient and intuitive formulation of the problem There are open source solutions to approach SLAM and all its subproblems Some of them are available on