Page 135 - Dynamic Vision for Perception and Control of Motion

4.3 Situations as Precise Decision Scenarios 119

mental processing” algorithm. These low-frequency results should be made avail-
able to all other processes by providing special slots in the DOB and depositing the
values with proper time stamps. The situation assessment algorithm has to check
these values for decision-making regularly.
The specialist processes for visual perception should also have a look at them to
adjust parameters in their algorithms for improving results. In the long run, a direct
feedback component for learning may be derived. Perceiving weather conditions
through textures may be very computer-intensive; once the other basic perception
tasks for road and other vehicles run sufficiently reliable, additional computing
power becoming available may be devoted to this task, which again can run at a
very low rate. Building up a knowledge base for the inference from distributed tex-
tures in the images toward environmental conditions will require a large effort.
This includes transitions in behavior required for safe mission performance.

4.3.2 Objects/Subjects of Relevance
A first essential step is to direct attention (by gaze control and corresponding im-
age evaluation) to the proper environmental regions, depending on the mission
element being performed. This is, of course, different for simple roadrunning, for
preparing lane changes, or for performing a turnoff maneuver. Turning off to the
left on roads with oncoming (right-hand) traffic is especially demanding since their
lane has to be crossed.
Driving in urban environments with right-of-way for vehicles on crossroads
coming from the right also requires special attention (looking into the road). Enter-
ing traffic circles requires checking traffic in the circle, because these vehicles
have the right-of-way. Especially difficult are 4-way-stops in use in some coun-
tries; here the right-of-way depends on the time of reaching the stop–lines on all
four incoming roads.
Humans may be walking on roads through populated areas and in stop-and-go
traffic. On state, urban and minor roads, humans may ride bicycles, may be roller
skating, jogging, walking, or leisurely strolling. Children may be playing on the
road. Recognizing these situations with their semantic context is actually out of
range for machine vision. However, detecting and recognizing moving volumes
(partially) filled with massive bodies is in the making and will become available
soon for real-time application. Avoiding these areas with a relatively large safety
margin may be sufficient for driver assistance and even for autonomous driving.
Some nice results for assistance in recognizing humans crossing in front of the ve-
hicle (walking or biking) have been achieved in the framework of the project “In-
vent” [Franke et al. 2005].
With respect to animals on the road, there are no additional principal difficulties
for perception except the perhaps erratic motion behavior some of these animals
may show. Birds can both move on the ground and lift off for flying; in the transi-
tion period there are considerable changes in their appearance. Both their shapes
and the motion characteristics of their limbs and wings will change to a large ex-
tent.