CarMap: Fast 3D Feature Map Updates for Automobiles
summary
CarMap presents a lean 3D feature-map format that uploads map updates in 0.6 seconds end-to-end by sending only map landmarks (about 4% of keyframe features) and reconstructing indices on the cloud. It uses position hints (GPS+KD-tree) and multi-frame, 3D-position matching instead of brittle visual words, plus semantic segmentation to filter transient objects. Evaluations show a 75× reduction in map size vs. alternatives, 45× lower localization error in dynamic scenes, and robust cross-lane localization.
AI-written summary · about
chapters (16)
- 0:00Start
- 1:00Depth Perception Sensors Stereo Camera
- 1:39Short Timescale Events
- 3:31Challenge: Size of 3D Maps SLAM
- 4:48CarMap Contributions Challenges
- 5:40Background: Image Features
- 6:08Background: 3D Frames
- 6:26Background: Keyframe Features
- 7:10Background: Image Feature Matching
- 8:11Map Elements: Minimal Representation?
- 8:37CarMap's Lean Map Representation
- 9:31Background: Feature Matching Data Structures
- 9:58Background: Histogram-based Matching
- 13:00Challenge: Environmental Dynamics Map from rush hour
- 13:30Semantic Segmentation for Feature Selection
- 15:47Evaluation: Multi-Lane Localization Setup
related talks
28mThe Role of High Definition 3D LiDAR in Self Driving Cars
17mSwarmMap: Scaling Up Real-time Collaborative Visual SLAM at the Edge
26mFrom Panoramic Vision to Lightweight 3D Worlds for Semantic Visual Tracking
16m"Putting a New Spin on
39mDistributed Mapping Technologies for Autonomous Vehicles
31mDeep learning for self-driving vehicle technology
21mThe race to replace human drivers
21mAutonomous driving and handling the handover