Automotive LiDAR

General Motors R&DAriel Lipson

Overview

¶ How LiDARs work (Automotive)

¶ State of play – current devices, costs

¶ Alternative technologies / approaches and future outlook

¶ Advantages of LiDAR-enabled perception

¶ Technical challenges for LiDAR – interference, form factor etc.

High Performance LiDAR in Automotive

Google Toyota Audi Volkswagen

Integrated data from Velodyne 64HD lidar and 5 SICK lidars using an

Oxford Technical Solutions GPS package. Data was taken around 2007,

during preparation to the DARPA Urban Challenge Competition

Robo Taxis - 2016

4

LiDAR Theory – Time of Flight (TOF)

Laser

𝑃𝑟 =𝑃𝑡 𝐷

2 𝜌

4𝑅2𝜂𝑠𝑦𝑠𝜂𝑎𝑡𝑚𝑐𝑜𝑠𝛼

Detector

𝑅𝑎𝑛𝑔𝑒 =𝑇𝑖𝑚𝑒×𝑠𝑝𝑒𝑒𝑑 𝑜𝑓 𝑙𝑖𝑔ℎ𝑡

2

10𝑛𝑠×3∙108𝑚/𝑠

2= 15𝑚

LiDAR

Laser

Detector

Rotating LiDAR

• Resolution: separable measurement points

• We would like 3 points on a motorcyclist at 100 m

•0.6 𝑚

100𝑚= 6𝑚𝑟𝑎𝑑 → 2𝑚𝑟𝑎𝑑 = 0.1 𝑑𝑒𝑔

Coherent Detection - FMCW

• FMCW modulation is better suited for LiDAR using lower peak

power semiconductor lasers

ADANY et al.: CHIRPED LIDAR USING SIMPLIFIED HOMODYNE

DETECTION

System Considerations

1500 nm905 nm

Illumination

• Type: laser / LED

• Power: mW KW

• Wavelength

• Eye safety

• Field of view

• Pulses / CW

Detection

• Material: Si, InGaAs, Ge

• Type: PIN, APD, SPAD

• Field of view

• Pulses / CW

• Detection scheme:

• Direct TOF

• Coherent, FMCW

Sun Reduction

• Reduce field of view

• Spectral filters

• AC coupling

• Longer wavelength

System (automotive)

• Large field of view

• High resolution

• Small form factor

• Low cost

• Robust

• Simple integration

When Price is Less Relevant

• 360 Field of view

• Long focal length

• High resolution

• Low ambient light interference

• Large collection lenses

• 64 high power lasers

• Long range

• High resolution encoders

From Spinning to Solid State

360 FOVMechanically

rotating

120 FOV

Velodyne

Ibeo Scala

90 FOV

Solid StateQuanergy

ASC / Continental

Conti, Aerostar

• No moving parts

• Compact in size

• Lower cost

• Automotive grade

• <120 Field of view

• Robustness issues

• Not designed for MP

• No flexibility

• Hard to integrate on a vehicle

Trends

Price

$10000 $100

Size (cm)

10s 1s

Performance

Longer range

Denser

Mechanical Scanner

Pros

360 FOV, hi res

Low light interference

Small number of

expensive components

Cons

Mechanically rotating

Inherently large

Vertical resolution

DetectorSilicon, simple

Examples of Lidar systems

Mechanical Scanner Flash

Pros

360 FOV, hi res

Low sun interference

Small number of

expensive components

No time distortion

Low sun interference

No moving parts

Cons

Mechanically rotating

Inherently large

Vertical resolution

High peak power laser

complex detector

Range?

DetectorSilicon, simple Non silicon / hi res -

complex

Examples of Lidar systems

Mechanical Scanner Flash Phase Array

Pros

360 FOV, hi res

Low sun interference

Small number of

expensive components

No time distortion

Low sun interference

No moving parts

small

No moving parts

Random access

performance

Cons

Mechanically rotating

Inherently large

Vertical resolution

High peak power laser

complex detector

Range?

Complex technology

Side lobes

2D very complex

Detector

Silicon, simple Non silicon / hi res -

complex

Needs a 2D array to

remove side lobes

ambiguity

Martijn J.R. Heck; published by De Gruyter

Nature 493, 195–199 (10 January 2013)

IEEE Spectrum –

MIT Photonic

Microsystems

Group

Examples of Lidar systems

Low Cost solid state LiDAR Development

15

• What is the #1 problem with low cost LiDAR development?

• Everyone is used to the HD data from rotating Systems

LiDAR-Enabled Perception

The ability to become aware of something

through the senses

LiDAR Application – Forward Looking

10 FOV

100-200m

0.1 resolution STO

P

180 FOV

100-200m = 6 sec at 50 kph

0.1 resolution

LiDAR-Enabled Perception

Occupancy Grid

representationClassification Tracking

Himmelsbach et al. ,2009 IEEE/RSJ International C

CogInfoCom 2013

LiDAR-Enabled Localization

Building occupancy grid Localization based on stored data

Velodyne LIDAR

Automated Ground Truthing

Use Lidar to train a neural network. No manual annotation needed

Result example

Technical Challenges

¶ Long Range

¶ High Resolution, vertical as well.

¶ Performance in

bad weather conditions

¶ Dark colour / reflective objects

¶ Interference between vehicles

¶ Small form factor

¶ Low cost

The Robo-Taxi Era The consumer Era

No signal from road

>70m

20m

Fusion

LiDAR

Radar

Camera

Thank You