In-Vehicle Change Detection for Self- Healing HD Maps Stephen - - PowerPoint PPT Presentation

▶
in vehicle change detection for self healing hd maps
SMART_READER_LITE
LIVE PREVIEW

In-Vehicle Change Detection for Self- Healing HD Maps Stephen - - PowerPoint PPT Presentation

Jan 11, 2023 235 likes •569 views

In-Vehicle Change Detection for Self- Healing HD Maps Stephen OHara, Ph.D. Principal Engineer, Edge Perception HERE Technologies HERE in numbers 100M 100M HERE Maps on board of ,000 + 8,000 200 200 vehicles and counting Countries

slide-1
SLIDE 1

Stephen O’Hara, Ph.D. Principal Engineer, Edge Perception HERE Technologies

In-Vehicle Change Detection for Self- Healing HD Maps

slide-2
SLIDE 2

HERE in numbers

200 200

Countries mapped 30

30+

Years of experience transforming location technology

8,000 ,000+

Employees in 56 countries focused on delivering the world’s best map and location technologies HERE Maps on board of

100M 100M

vehicles and counting

28 28

TB TB map data collected per day

4of

  • f5

In-car navigation systems in Europe and North America use HERE maps

+

collecting data for maps

400 400

HERE cars 3D data points per second per car

700, 700,000 000

555k+ 555k+ km

for Autonomous Driving HD Live Map covering

slide-3
SLIDE 3

Continued HD Map Leadership

#1 Global leader in map coverage & build out

555,555+km

slide-4
SLIDE 4

HERE HD Live Map

Cloud based Layered Tiled

A cloud service enabling scale and AI Integrated, scalable,

  • ff the shelf offering ready

for application consumption Allows for over the air updates in an efficient (data size) manner Leveraging crowdsourced data to ensure freshness Automated vehicles require far more detail, accuracy and reliability

Fresh Precise

slide-5
SLIDE 5

The Self-Healing Map

slide-6
SLIDE 6

The Self-Healing Map

slide-7
SLIDE 7

The Self-Healing Map

slide-8
SLIDE 8

Industrial Capture (HERE TRUE)

Gold standard for mapping, with high accuracy but low revisit rates.

Eyedrop

Mid-range capture device with broader reach than HERE True.

Cyclops

Leverage the masses using commodity hardware found in mobile phones.

Vehicle OEM Observations

Healing the map using observations provided by Vehicle OEM partners.

Other data providers

Healing the map using observations provided by 3rd party data providers.

HD Map Maintenance Strategy

slide-9
SLIDE 9

Perception for Build-Out

Processing HERE True Data

slide-10
SLIDE 10

Feature detection using deep learning

slide-11
SLIDE 11

Feature detection using deep learning

slide-12
SLIDE 12

Deep Learning – Scanline Detectors

1 2 3 4 4 3 2 1

slide-13
SLIDE 13

Material Selection – Which Images to Label?

Existing image New image

  • t-SNE dimensionality

reduction

  • Select clusters with few

existing labeled images

  • The testing accuracy improves

from 92.31 % to 98.80 %

slide-14
SLIDE 14

Edge Perception

Spotcam | Eyedrop | Cyclops

slide-15
SLIDE 15

Eyedrop Cyclops HERE Edge Computing Devices

slide-16
SLIDE 16

Common Compute Architecture

slide-17
SLIDE 17

Eyedrop

NVIDIA Xavier Custom PCB

slide-18
SLIDE 18

Cyclops

Phase I Phase II Phase III

slide-19
SLIDE 19

Edge Feature Detection Models

27k 27k

Dense s segmenta tati tion

350 350k

Images l labeled

4m 4m

Individual l labels 9 Cl Classes

slide-20
SLIDE 20

Sampled Image

Material Selection II

slide-21
SLIDE 21

Detection, Tracking, Localization

slide-22
SLIDE 22

Closing the Loop

Maplets | Change Detection | Aggregation | Quality Index

slide-23
SLIDE 23

From change detection in the cloud to change detection in the car

Observations Localize & Aggregate Observations Change Updated Map Tiles Localization & Change Detection Maplets Aggregation

slide-24
SLIDE 24

Maplet Structure

Content Container Compression Vendor-neutral

slide-25
SLIDE 25

OEM Backend In-Vehicle

Overview of In-Vehicle Processing for OEM Partners

HERE Maplets HD Live Map (optional) Change Detection (optional) Compression & Collection Policy (optional) Cameras Lidar Radar Ultrasonic … OEM Sensor Specific Interaction (Drivers)

slide-26
SLIDE 26

Overview of In-Vehicle Processing for OEM Partners OEM Backend

HERE Maplets HERE Maplets HERE Maplets

slide-27
SLIDE 27

Latitude: 48.0844, Longitude: 11.74003

April 11th: 1 Sign May 15th: 2 Signs

slide-28
SLIDE 28

Example Scenario: Artifact introduced in the map

Artificial Kink Lane Markings in the map

slide-29
SLIDE 29

Example Scenario: Artifact introduced in the map

slide-30
SLIDE 30

Change Detection of Lane Markings

White lanes lines failed to associate with the existing map

Indicates that there is a change

slide-31
SLIDE 31

HD Live Map Quality Index

Accuracy and reliability of data

[Classification] Will the object be classified correctly? [Accuracy] Will it be in the right position?

𝐵𝑐𝑡𝑝𝑚𝑣𝑢𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑝𝑜 𝐹𝑠𝑠𝑝𝑠

/0123454 =

(𝑦9 − 𝑦;)=+(𝑧9 − 𝑧;)=+(𝑨9 − 𝑨;)=

  • Where 𝑠
0 =

𝑦BC

D − 𝑦EC =

+ (𝑧BC

D − 𝑧EC)=+(𝑨BC D − 𝑨EC)=
  • 𝑆𝑓𝑚𝑏𝑢𝑗𝑤𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑝𝑜 𝐹𝑠𝑠𝑝𝑠 = 1

𝑂 K 𝑠

L 0MN

𝑆𝑓𝑑𝑏𝑚𝑚 = 𝑈𝑠𝑣𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑤𝑓𝑡 𝑈𝑠𝑣𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑤𝑓𝑡 + 𝐺𝑏𝑚𝑡𝑓 𝑂𝑓𝑕𝑏𝑢𝑗𝑤𝑓𝑡 Determine binary state of discrete map object attribute: Correctly Classified or Incorrectly Classified

[Existence] Will there be an object?

[A0,E0] [CA,] [CA,] [A0,E0] [CA,] [CA,] [A0,E0] [CA,] [CA,] [CA,][A0,E0] [A0,E0] [A0,E0] [CA,][A0,E0]

A published score for HDLM features representing the predicted quality of the HDLM

slide-32
SLIDE 32

Stronger with our partners