Page 46 - Autonomous Mobile Robots
P. 46
Visual Guidance for Autonomous Vehicles 29
When sensor outputs are read asynchronously, certain assumptions such as
being Linear Time Invariant (LTI) [38] can be made to propagate asynchron-
ized data to the upcoming sample time of the control system. Robl [38] showed
examplesofusingfirst-orderholdandthird-orderholdmethodstopredictsensor
values at desired times. When different resolution sensors are to be fused
at the data level (e.g., fusion of range images from ladar and stereo vision),
down-sampling of sensor data with higher spatial resolution by interpolation
is performed. For sensor fusion at the obstacle map level, spatial synchron-
ization is not necessary since a unique map representation is defined for all
sensors.
Example: Fusion of laser and stereo obstacle maps for false alarm suppression
Theoretically, pixel to pixel direct map fusion is possible if the calibra-
tion and synchronization of the geometrical constraints (e.g., rotation and
translation between laser and stereo system) remain unchanged after calib-
ration. Practically, however, this is not realistic, partially due to the fact that
sensor synchronization is not guaranteed at all times: CPU loading, terrain
differences, and network traffic for the map output all affect the synchroniza-
tion. Feature-based co-registration sensor fusion, alternatively, addresses this
issue by computing the best-fit pose of the obstacle map features relative to
multiple sensors which allows refinement of sensor-to-sensor registration. In
the following, we propose a localized correlation based approach for obstacle-
map-level sensor fusion. Assuming the laser map L ij and stereo map S ij is to be
merged to form F ij . A map element takes the value 0 for a traversable pixel, 1
for an obstacle, and anything between 0 and 1 for the certainty of the pixel to be
classified as an obstacle. We formulate the correlation-based sensor fusion as
L ij S ij = undefined
S ij
L ij = undefined
F ij =
(a 1 L ij + a 2 S i+m,j+n )/(a 1 + a 2 ) max(Corr(L ij S i+m,j+n )) m, n ∈
undefined S ij , L ij = undefined
(1.9)
where represents a search area and {a 1 , a 2 } are weighting factors. Corr(L, S)
is the correlation between L and S elements with window size w c :
w c /2 w c /2
Corr(L ij S i+m,j+n ) = L i+p,j+q S i+m+p,j+n+q (1.10)
p=−w c /2 q=−w c /2
© 2006 by Taylor & Francis Group, LLC
FRANKL: “dk6033_c001” — 2006/3/31 — 16:42 — page 29 — #29
