Page 120 - Dynamic Vision for Perception and Control of Motion

3 Subjects and Subject Classes
104

Since perturbations are abundant in road traffic, toward the end of the lane
change maneuver after taking the new lane as reference for perception, a feedback
component is superimposed leading to automatic centering in the new lane. This
also takes care of curvature onsets during the lane change maneuver.

Driving on curved roads: The assumption is made that longitudinal speed V is
controlled in dependence of road curvature in order not to exceed lateral accelera-
tion limits. Speed acts as a parameter in selecting lateral behaviors as discussed
above.
The heading angle of the vehicle body with respect to inertial space is desig-
nated by ȥ abs and with respect to the local road tangent by ȥ rel. Between ȥ abs and
ȥ rel is the heading angle of the road Ȥ. The temporal change in road heading at
speed V is (see Equation 3.10)
dȤ / dt C ˜ 0h V . (3.46)
The visually recognizable curvature C 0h of the road at the actual location of ve-
hicle cg can be introduced as an additional term in the dynamic model (see Chapter
6). In the block diagram Figure 3.24 (center top), this has been used to decouple
local roadrunning from the absolute geodetic direction. Local heading ȥ rel times
speed V yields the lateral speed v on the road.
With representations like these, the linguistic symbol “lane keeping” is activated
by organizing the feedback control output computed by Equation 3.22 with a
proper matrix K to be used for the steering rate Ȝ-dot. Note that the visually deter-
mined quantities “road curvature C 0h”, lateral position in the lane y, relative head-
ing angle ȥ rel as well as the conventionally measured value of vehicle speed V are
used in the closed-loop action-perception cycle taking a dynamic model for the
motion process into account. It has been shown in linear control theory that com-
plete state feedback yields optimal control laws with respect to a chosen payoff
function.
This feedback control constitutes the behavioral capability “roadrunning” made
up of the perceptual capability road (lane) recognition with relative egostate (in-
cluding reconstruction of the slip angle ȕ not directly measurable) and the locomo-
tion capability lane keeping by state feedback. Since visual evaluation of the situa-
tion and control computation as well as implementation take their time (a few
tenths of a second), this time delay between measurement taking and control output
has to be taken into account when determining the control output. The spatio-
temporal models of the process allow doing this with well-known methods from
control engineering. Tuning all the parameters such that the abstract symbolic ca-
pabilities for roadrunning coincide with real-world behavior of subjects is the
equivalent of “symbol grounding”, often deplored in AI as missing.

3.4.6 Phases of Smooth Evolution and Sudden Changes

Similar to what has been discussed for “lane keeping” (by feedback control) and
for “lane change” (by feed-forward control), corresponding control laws and their
abstract representation in the system have to be developed for all behavioral capa-
bilities like turningoff, etc. This is not only true for locomotion but also for gaze