In-Vehicle Change Detection for Self- Healing HD Maps Stephen O’Hara, 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 mapped 30 Employees in 56 countries 28 28 focused on delivering the world’s best map and location technologies 30+ collected per day Years of experience TB map data TB transforming location technology 700,000 700, 000 3D data points per second per car In-car + 400 400 HERE cars 4 of of 5 navigation collecting HD Live Map covering systems in Europe 555k+ 555k+ km data for and North America maps use HERE maps for Autonomous Driving
Continued HD Map Leadership #1 Global 555,555+km leader in map coverage & build out
HERE HD Live Map Precise Cloud based Layered Fresh Tiled Automated vehicles A cloud service enabling Integrated, scalable, Allows for over the air Leveraging crowdsourced require far more detail, scale and AI off the shelf offering ready updates in an efficient data to ensure freshness accuracy and reliability for application consumption (data size) manner
The Self-Healing Map
The Self-Healing Map
The Self-Healing Map
HD Map Maintenance Strategy Vehicle OEM Other data Industrial Capture Cyclops Eyedrop Observations providers (HERE TRUE) Healing the map Healing the map Leverage the masses Gold standard for Mid-range capture using observations using observations using commodity mapping, with high device with broader provided by Vehicle provided by 3 rd party hardware found in accuracy but low reach than HERE OEM partners. data providers. mobile phones. revisit rates. True.
Perception for Build-Out Processing HERE True Data
Feature detection using deep learning
Feature detection using deep learning
Deep Learning – Scanline Detectors 4 3 2 1 1 2 3 4
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 %
Edge Perception Spotcam | Eyedrop | Cyclops
HERE Edge Computing Devices Eyedrop Cyclops
Common Compute Architecture
Eyedrop NVIDIA Xavier Custom PCB
Cyclops Phase I Phase II Phase III
Edge Feature Detection Models 350 350k 4m 4m labels 9 27k 27k Images l labeled Individual l Cl Classes Dense s segmenta tati tion
Material Selection II Sampled Image
Detection, Tracking, Localization
Closing the Loop Maplets | Change Detection | Aggregation | Quality Index
From change detection in the cloud to change detection in the car Localize & Aggregation Aggregate Observations Change Updated Map Tiles Observations Maplets Localization & Change Detection
Maplet Structure Content Container Compression Vendor-neutral
Overview of In-Vehicle Processing for OEM Partners Compression & Collection Policy (optional) Change Detection (optional) HD Live Map HERE Maplets (optional) OEM Sensor Specific Interaction (Drivers) Cameras Lidar Radar Ultrasonic … In-Vehicle OEM Backend
Overview of In-Vehicle Processing for OEM Partners OEM HERE Maplets HERE Maplets HERE Maplets Backend
Latitude: 48.0844, Longitude: 11.74003 April 11 th : 1 Sign May 15 th : 2 Signs
Example Scenario: Artifact introduced in the map Lane Markings Artificial in the map Kink
Example Scenario: Artifact introduced in the map
Change Detection of Lane Markings Indicates that there is a change White lanes lines failed to associate with the existing map
� � HD Live Map Quality Index [A 0, E 0 ] Accuracy and reliability of data [A 0, E 0 ] [C A, ] [C A, ] [C A, ][A 0, E 0 ] [Existence] Will there be an object? 𝑈𝑠𝑣𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑤𝑓𝑡 𝑆𝑓𝑑𝑏𝑚𝑚 = 𝑈𝑠𝑣𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑤𝑓𝑡 + 𝐺𝑏𝑚𝑡𝑓 𝑂𝑓𝑏𝑢𝑗𝑤𝑓𝑡 [A 0, E 0 ] [C A, ][A 0, E 0 ] [Accuracy] Will it be in the right position? A published score 𝐵𝑐𝑡𝑝𝑚𝑣𝑢𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑝𝑜 𝐹𝑠𝑠𝑝𝑠 /0123454 = (𝑦 9 − 𝑦 ; ) = +(𝑧 9 − 𝑧 ; ) = +(𝑨 9 − 𝑨 ; ) = for HDLM features 𝑆𝑓𝑚𝑏𝑢𝑗𝑤𝑓 𝑄𝑝𝑡𝑗𝑢𝑗𝑝𝑜 𝐹𝑠𝑠𝑝𝑠 = 1 L representing the = 𝑂 K 𝑠 Where 𝑠 D − 𝑧 E C ) = +(𝑨 B C D − 𝑨 E C ) = 0 = 𝑦 B C D − 𝑦 E C + (𝑧 B C 0 predicted quality of 0MN the HDLM [A 0, E 0 ] [Classification] Will the object be classified correctly? [C A, ] [C A, ] [A 0, E 0 ] Determine binary state of discrete map object attribute: [C A, ] [C A, ] Correctly Classified or Incorrectly Classified
Stronger with our partners
Recommend
More recommend
