Perception for driverless vehicles

1
Perception for driverless vehicles

• Rodrigo Benenson

2
(No Transcript)
3
Perception is the process of transforming the
measures into an internal representation
4
Perception depends on the application

• What to measure ?
• What to transform ?
• Which representation ?
• Has to be useful for
• Route planning
• Trajectory planning
• Control

5
(No Transcript)
6
We require

• Localization
• to follow the trajectory
• Mapping
• to consider already sawn static objects
• Moving objects tracking
• to consider moving objects

7
We require to do
SLAMMOT
(Simultaneous localization, mapping and moving
objects tracking)
in city sized urban scenes
we follow the works of Bob Wang (2004)
8
SLAMMOT is a hard problem
SLAMMOT ? SLAM ? Localization ? Bayesian
filtering ? Bayesian fusion
It is a most probable explanation problem
Managing uncertainty is a core issue
Computational tractability is a core issue
9
Which is the simplest system that could work ?
10
(No Transcript)
11
Divide and conquer
12

• SLAMMOT ??
• Small area SLAM
• Moving objects detection
• Moving objects tracking
• Large area SLAM

13
We make some base assumptions

• The world is flat
• The world is semi-structured (a city)
• We can construct a model of the vehicles dynamics
  and of the sensor uncertainties
• Considerable computational power is available
• Online and offline computation are performed

14
Work in progress
(using pre-recorded laser scanner data)
15
Work in progress
16
SLAMMOT Planning Control Vehicle moving
alone
17
SLAMMOT PMP Control Vehicle moving alone
18
However
19
Only using visible obstaclesmay
createoverconfident decisions
20
(No Transcript)
21
(No Transcript)
22
(No Transcript)
23
Believing that the world is aggressive
createsconservative trajectories
24
We can use SLAMMOTresults to createmoving
objects maps
25
(No Transcript)
26
We can use moving objects mapsto create more
natural trajectories
A
B
27
We can use moving objects mapsto create more
natural trajectories
A
B
28
Solving the SLAMMOT problemopens the door for
aninfrastructure free deploymentof driverless
vehicles
29
(No Transcript)
30
But
31
Even in theory, SLAMMOTis not enough for real
world
32
We need to detect the road

• Free space ? allowed space Can we do better
  than white lines? Texture approaches seems
  promising

33
Can we do it cheaper ?

• We need
• Detection of occupied space
• Detection of free space
• Accurate and robust estimate the displacement
  between two consecutive moments
• Capacity to recognize a visited place
• Can we do SLAMMOT with vision ?

34
What about road planning ?

• It needs a graph description of the city roads
• Can we deduce streets graphs from static and
  moving objects maps ?

35
(No Transcript)
36
Collaborative perception is the next big thing

• Enhances the range of perception
• Mitigates occlusion problems
• Allows more robust and precise estimations
• Allows better prediction of communicating
  entities

37
Collaborative perception is the next big problem

• In the SLAMMOT framework, how to put in relation
  the different references frames ?
• What is more important to transmit ?
• How to coordinate communicating entities ?
• Wireless networking and security issues

38
Work plan

• Next weeks
• Finish the implementation for driverless tests
• Middle 2006
• Collaborative perception for planning
• End 2006
• Validation of the approach,
• detect and solve problems

39
Questions ?

• Rodrigo Benenson
• http//www-rocq.inria.fr/benenson

40
(No Transcript)