Challenges & Opportunities for Building Crowdsourced Mapping Services for Autonomous Driving
summary
NVIDIA aggregates sensor data from millions of DRIVE Hyperion vehicles into NVIDIA Map and converts compressed on-car mapstreams into HD maps using CUDA‑optimized pipelines on DJX pods. Radar Point clouds are quantized, tiled, spatially differenced and Huffman encoded to cut bandwidth. Maps average 2–4 commuter drives per lane for Seattle coverage, support thousands of regional updates with elastic compute, and produce fused radar/localization maps for highway, suburban and downtown scenarios that successfully localize vehicles in testing.
AI-written summary · about
chapters (10)
- 0:00Crowdsourced HD maps intro
- 2:56Traditional survey challenges
- 3:58Crowdsourced mapping benefits
- 5:02Nvidia perception subsystem
- 6:50Processing vehicle map streams
- 8:34Stream conversion to HD maps
- 9:37Key crowdsourced mapping challenges
- 12:53Global mapping scale needs
- 13:14Radar maps case study
- 16:45Conclusion & demo
related talks
19mHow we leverage AI to build great maps at massive scale
20mAccelerating AI pipelines w/ NVIDIA on AWS: An autonomous driving case
19mAI Trends in Intelligent Transport Systems
31mDeep learning for self-driving vehicle technology
42mAI and Self-Driving Cars
04mInterview
05mInterview
32mData&AI in TomTom: How are are transforming mobility