EdgeFirst LiDAR Publisher v2.1.0

EdgeFirst LiDAR Publisher v2.1.0

Installation

Download Binaries

Platform	Binary
Linux x86_64	edgefirst-lidarpub-linux-x86_64
Linux ARM64	edgefirst-lidarpub-linux-aarch64

# Download and install (ARM64 example for Maivin/Raivin)
wget https://github.com/EdgeFirstAI/lidarpub/releases/download/v2.1.0/edgefirst-lidarpub-linux-aarch64
chmod +x edgefirst-lidarpub-linux-aarch64
sudo mv edgefirst-lidarpub-linux-aarch64 /usr/local/bin/edgefirst-lidarpub

What's Changed

Added

• IMU-guided ground plane filter (src/ground.rs): Region-wise PCA on a polar
  grid (12 azimuth × 6 range sectors) detects and removes floor points before
  clustering, preventing nearby objects from being merged via shared ground points.
  Uses IMU accelerometer data (Robosense E1R built-in) to determine gravity vector
  and compensate for sensor tilt. EMA smoothing for stable detection across frames.
• Voxel connected-component clustering: Fast BFS over occupied voxels as an
  alternative to point-level DBSCAN. ~5× faster on aarch64 (~13ms vs ~70ms for
  E1R ~25k points) with coarser but adequate segmentation.
• Bridge threshold for both DBSCAN and voxel clustering: Points/voxels with
  fewer than bridge_pts neighbors are labeled as part of a cluster but do not
  propagate expansion, preventing thin structures (railings, wires, ropes) from
  merging separate objects.
• Pipeline instrumentation: Per-stage timing with std::time::Instant
  accumulators (logged every 100 frames) and tracing::info_span!() wrappers
  for Tracy profiler visualization.
• --clustering algorithm selection: "" (disabled), "dbscan", or "voxel"
• --clustering-bridge CLI argument for bridge threshold control
• --ground-filter flag to enable IMU-guided ground removal
• --ground-thickness for ground slab thickness (millimeters)
• --sensor-height for fixed sensor height override (skips PCA auto-detection)
• Real-data PCD test files (testdata/e1r_frame0.pcd, testdata/os1_frame0.pcd)
  for ground filter and clustering integration tests
• 7 ground filter unit tests (synthetic flat floor, tilted sensor, no IMU,
  known height, wall rejection, reuse, real E1R data)
• 16 clustering tests covering DBSCAN, voxel, real-data, and cross-algorithm
  agreement

Changed

• --clustering changed from boolean flag to string algorithm selector
  ("", "dbscan", "voxel") with PossibleValuesParser validation
• DBSCAN distance comparison changed from < to <= to match standard
  DBSCAN semantics (affects both scalar and NEON SIMD paths)
• Cluster thread no longer catches panics silently; unrecoverable errors
  terminate the process with a clear error message
• Updated ARCHITECTURE.md with ground filter, clustering algorithms, and
  pipeline instrumentation documentation
• Updated README.md with user-level documentation for ground filter, clustering
  options, bridge threshold, and troubleshooting
• Updated TESTING.md with per-module test tables, manual test scenarios, and
  on-target deployment instructions
• Updated lidarpub.default configuration template with all new parameters

Performance

• Ground filter: ~5-9ms on E1R ~25k points (aarch64 Cortex-A)
• Voxel clustering: ~13ms on E1R ~25k points (aarch64 Cortex-A)
• DBSCAN clustering: ~70ms on E1R ~25k points (aarch64 Cortex-A)
• Full pipeline (ground + voxel): ~29ms total, well within 100ms (10Hz) budget

---

Full Documentation: https://github.com/EdgeFirstAI/lidarpub#readme