How dongjae0107/elite works
Project Type & Purpose
• ELite is an open-source, Python-based LiDAR continual mapping framework that incorporates *ephemerality* (local & global) into a multi-stage pipeline. It is neither a web application nor a microservice suite, but a domain-specific processing toolchain for aligning successive LiDAR “sessions,” removing dynamic elements, and incrementally updating a persistent map.
File Structure & Architectural Insights
• Top-level scripts & config:
• run_elite.py – single entry point
• config/*.yaml – per-session parameters
• Core modules (all under the Python package root):
• alignment/ – session alignment & scan matching
• dynamic_removal/ – ephemerality-based map cleaner
• map_update/ – incremental map builder
• utils/ – logging, session abstraction, I/O
• External dependencies & build:
• requirements.txt & setup.py – pip installation
• scripts/download_parkinglot.sh – data bootstrap
Main Components & Responsibilities
a. Input Layer
• Session Loader (utils/session.py & session_map.py)
• Configuration Parser (YAML → Python dict)
• Raw LiDAR scans (.pcd) and pose files
b. Alignment Module
• GlobalRegistration (alignment/global_registration.py) – coarse transform
• MapZipper (alignment/map_zipper.py) – merges sessions in time
• ScanMatcher sub-components (base_scan_matcher, open3d_scan_matcher, pygicp_scan_matcher) – fine ICP adjustments
c. Ephemerality Calculators
• Local & Global ephemerality scoring baked into alignment & dynamic_removal
d. Dynamic Removal Module
• MapRemover (dynamic_removal/map_remover.py) – filters out transient points
e. Update Module
• MapUpdater (map_update/map_updater.py) – fuses cleaned data into the persistent map
f. Output & Persistence
• Session outputs directory (prev_outputs_dir)
• Logs & intermediate files
g. External Libraries
• Open3D (point cloud operations)
• PyGICP or fast_gicp (CUDA-accelerated ICP)
Architectural Patterns & Principles
• Pipeline/Staged Processing: clear separation of alignment, filtering, and fusion steps
• Strategy Pattern: interchangeable ScanMatcher implementations (Open3D vs. PyGICP)
• Configuration-driven: YAML files determine branch logic (first vs. subsequent session)
• Modular & extensible: each submodule communicates via simple pointcloud + metadata objects