Skip to main content
Back

Xiamen University's LEADER Cuts LiDAR Localization Error by 73.9% Without GPS

A new scene coordinate regression framework for LiDAR relocalization cuts position error by up to 73.9% while running in real time.

Pro

  • LEADER is a new LiDAR relocalization framework hitting 0.31m position error on NCLT, the first implicit method under 0.5m.
  • A cylindrical projection turns yaw rotations into translations, giving the sparse convolutional encoder built-in rotation invariance.
  • The Truncated Relative Reliability loss lets the network downweight unreliable points like ground, leaves, and moving objects.
  • Position error drops 24.1% on Oxford RobotCar and 73.9% on NCLT versus prior state-of-the-art.
  • Runs at 46-48ms per frame on an RTX 3090, real-time for both benchmarks; code available at github.com/JiansW/LEADER.
  • The TRR loss transfers to other architectures, cutting SGLoc's orientation error by 26% with no other changes.

Figuring out where a vehicle sits on a map from a single LiDAR sweep sounds solved until you try it in a rainy parking lot with a rotated sensor. A team from Xiamen University has released LEADER, a learning-based LiDAR relocalization framework that pushes accuracy well below the meter mark and, in some cases, into the tens of centimeters. The code is on GitHub.

The system pairs a Robust Projection-based Geometric Encoder that captures multi-scale geometric features with a new loss function that tells the network which points to trust and which to ignore. Together they deliver 24.1% and 73.9% relative reductions in position error over prior techniques on the two standard benchmarks.

When every point gets equal say

LiDAR relocalization estimates the 6-DoF pose (position and orientation) of a sensor from a single scan, which matters most where GPS breaks down: tunnels, parking garages, dense urban canyons. The dominant learning approach is Scene Coordinate Regression (SCR), where a network predicts the global 3D coordinate of every point in the scan and RANSAC then solves for the pose from those correspondences.

Two things wreck this in practice. Cars turn. Vehicles undergo yaw rotations during driving rather than maintaining a fixed heading, and most SCR networks were never built to be rotation-invariant, so predictions drift when the viewpoint changes. On top of that, some points deserve less trust than others. Ground pavement, tree leaves, and moving objects give ambiguous or misleading cues, yet a standard loss weighs their errors the same as those from stable buildings.