Lightweight Real-Time LiDAR Localization

Lightweight real-time localization in pre-mapped environments is a key capability for many robotic applications, especially in scenarios where GNSS is unavailable or only partially reliable.

Our approach addresses this challenge by identifying scene-specific landmarks that serve as stable anchors for localization within a given environment. During training, these landmarks are automatically discovered while a neural network is simultaneously trained to detect them. To enable efficient learning, LiDAR point clouds are represented as bird’s-eye-view density images.

Figure 1: Scene landmark-based localization in a 550 m × 400 m campus area. Left: BEV density image of a local point cloud with predicted landmark locations. Right: Scene landmarks, shown as blue squares. Lines indicate correctly predicted correspondences, and the red arrow denotes the estimated pose. Real-time localization is achieved using only a compact 20 MB representation consisting of the network and the landmark list.

At inference time, our method does not require access to a dense map. Instead, localization is performed using only the trained network and a list of learned landmarks, resulting in a highly compact representation of about 20 MB.

With this framework, we achieve real-time global LiDAR localization at up to 35 fps.

The full training code is available on GitHub:
https://github.com/davidskdds/BEV-SLD

A working demo that does not require dataset download can be found here:
https://github.com/davidskdds/BEV-SLD/tree/demo

Paper:
https://arxiv.org/abs/2603.17159