SPM-ICP: Incremental Registration using the ICP Algorithm and Sparse Point Maps


Sancta Simplicitas -- On the efficiency and achievable results of SLAM using ICP-Based Incremental Registration

Abstract -- This paper presents an efficient combination of algorithms for SLAM in dynamic environments. The overall approach is based on range image registration using the ICP algorithm. Different extensions to this algorithm are used to incrementally construct point models of the robot's workspace. A simple heuristic allows for determining which points in a newly acquired range image are already contained in the point model and for adding only those points that provide new information. Furthermore, the means for dealing with environment dynamics are presented which allow for continuously conducting SLAM and updating the point model according to changes in a dynamic environment. The achievable results of the overall approach are compared to Rao-Blackwellized Particle Filters as a state-of-the-art solution to the SLAM problem and evaluated using a recently published benchmark by Burgard et al. (2009).






Additional Material

Manuscript (pdf): Submitted to ICRA'10 and accepted for publication..
Videos+Logfiles: see below






Results

The incremental registration procedure using the Iterative Closest Point (ICP) algorithm together with the concept of Sparse Point Maps (SPMs) and the removal of less probable measurements according to a probabilistic grid maps allows for constructing globally consistent maps of small and medium-sized indoor environments.


(Red: constructed map was globally not consistent)
(1: Taking the best out of 10-20 runs from GMapping, rev. 40)
(2: As provided by Burgard et al.: GMapping + Pre-processing)






Benchmarking: the ACES data set

Data set recorded by Patrick Beeson in the ACES building at the University of Texas in Austin.
(Slight inconsistency in upper left → minor errors in local pose shift, global estimate is corrected!)

   
Constructed grid map Video visualization of incremental registration


Map, detail view of inconsistency and individual deviations from ground truth relations.

Trans. error distribution Rot. error distribution Measured Runtimes

Original Logfile: aces.clf
Logfile by ICP+SPM: aces_icp.clf
Logfile by gmapping: aces_gmapping.clf
Relations and Evaluator: http://kaspar.informatik.uni-freiburg.de/~slamEvaluation






Benchmarking: the INTEL data set

Data set recorded by Dirk Hähnel at the Intel Research Lab in Seattle.

   
Constructed grid map Video visualization of incremental registration

Trans. error distribution Rot. error distribution Measured Runtimes

Original Logfile: intel.clf
Logfile by ICP+SPM: intel_icp.clf
Logfile by gmapping: intel_gmapping.clf
Relations and Evaluator: http://kaspar.informatik.uni-freiburg.de/~slamEvaluation






Benchmarking: the CSAIL data set

Data set recorded by Cyrill Stachniss at the MIT CSAIL building.

   
Constructed grid map Video visualization of incremental registration

Trans. error distribution Rot. error distribution Measured Runtimes

Original Logfile: mitcsail.clf
Logfile by ICP+SPM: mitcsail_icp.clf
Logfile by gmapping: mitcsail_gmapping.clf
Relations and Evaluator: http://kaspar.informatik.uni-freiburg.de/~slamEvaluation






Benchmarking: the FR079 data set

Data set recorded by Cyrill Stachniss in Building 079 at the University of Freiburg.
This data set contains several errors and jumps in the odometry estimates. By using these erroneous pose shifts when sampling from the motion model, gmapping is not able to produce a globally consitent map (tested over 10 runs using 50 and 100 particles). ICP+SPM corrects these estimates and constructs a globally consistent map.
In domestic and office-like environments where the robot never traverses longer distances through previously unmodeled terrain but repeatedly encounters already modeled terrain, smaller registration errors do not considerably accumulate. Hence, the constructed maps are locally and globally consitent, and the robot is always well localized.

   
Constructed grid map Video visualization of incremental registration


Correction of large jumps in odometry estimates.
Original Logfile: fr079.clf
Logfile by ICP+SPM: fr079_icp.clf
Logfile by gmapping: fr079_gmapping.clf
Relations and Evaluator: http://kaspar.informatik.uni-freiburg.de/~slamEvaluation






References