Simultaneous Localization and Mappingfor Pedestrians
using only Foot-Mounted Inertial Sensors
Patrick Robertson, Michael Angermann, Bernhard KrachGerman Aerospace Center (DLR)B. Krach is now with EADS Germany
Raw NavShoe Odometry Results
NavShoe INS produced reasonable resultsstand alone, but still unbounded error growth
NavShoe INS had larger heading slips;unbounded error begins to rise earlier
Algorithm: Extended Kalman Filter with Zero Velcocity Updates (Foxlin)
State of the Art: Use Maps
Inertial sensors used indoors achieve accurate positioning when used in conjunction with maps
Krach, Robertson: WPNC 08, PLANS 08+Widyawan, Klepal, Beauregard: WPNC 08Woodman, Harle: UbiComp 2008
But what if the mapis unknown?
So, could we derive a map from this?
Naïve approach:
“Transfer the raw odometry trace to a piece of wire and bend it bit by bit so that similar areas overlap”
SLAM in Robotics
Simultaneous Localization and Mapping - identified by robotics community in mid ‘80s!Premise:
Localization using odometry and sensing of known landmarks is easy!
Mapping of landmarks given known location and orientation (pose) is easy!
Simultaneous Localization and Mapping is hard!
What about SLAM for Humans?Human pedestrians are not robots but share some similarities with them
Visual sensors (eyes)'Odometry' (in humans: sensed by proprioception)Path and planning and execution
In humans, we usually have little or no direct 'access' to most of these senses and functionsOur central assumption:
The pedestrian is able to actively control motion without violating physical constraints (i.e. walls, etc)
A Person Processes Numerous Visual Inputs
P
U
Zu E
Int
Vis
Map
Visual information„what the person sees“
Intention„what the person wants to do“
Error states of the odometry
U: Actual step taken (pose change vector)
PoseP
U
Zu E
Int
Vis
P
U
Zu E
Int
Vis
Time k-1 Time k Time k+1
MeasuredStep Zu
“Environment” = Map … constant over time
Bayesian Formulation: DBN
Intuitive Explanation of the Sequential Monte Carlo Algorithm
FootSLAM lets particles, or hypotheses, explore the state space of odometry errors, like evolution of driftIn this way, every particle is trying a slightly “differently bent piece of wire”Particles are weighted by their “compatibility” with
their individual mapoptional sensor readings, such as GPS, magnetometer
We can show that this is optimal in the Bayesian sense!
Experiments and Results
Measurement data taken from a pedestrian wearing a foot mounted IMU
Two scenarios:Indoor onlyOutdoor – indoor - outdoor sequence
Video
See
http://www.kn-s.dlr.de/indoornav
Resulting Maps
Resulting Maps
Relative Position Accuracy - Indoors No Scale Adaptation was Performed
Outdoor-Indoor-Outdoor
Concluding NotesFootSLAM effectively bounds the otherwise unbounded error growth without the need for pre-existing maps!
FootSLAM (like all forms of SLAM) is inherently invariant to rotation, translation and scale
In mixed scenarios, the resulting maps are globally and precisely anchored using GPS
Our future work:Map building with multiple users; “crowdsourcing” collaborative mapping
Movies: http://www.kn-s.dlr.de/indoornav/
Thank you!
Movies: http://www.kn-s.dlr.de/indoornav/