dataset & benchmark · 49 actions · 4 LiDARs · RGB-D-IR

Four LiDARs watch one action.

Every LiDAR sees the world differently. LaserAct records the same human actions simultaneously with four scanning principles and a depth camera — the first controlled benchmark for cross-sensor action recognition on laser range data.

velodyne · spinning · 360°

livox · rosette · non-repetitive

cepton · micro-motion

blickfeld · MEMS raster

2,341sequences

49action classes

9subjects

4 + 1LiDARs + RGB-D-IR

3.1 hsynchronized recording

2D posesincluded

The same action, six sensors

A static rig streams a spinning Velodyne HDL-32E and three solid-state LiDARs — Livox MID-100, Cepton Vista-P60, Blickfeld Cube 1 — simultaneously with an Orbbec Astra Pro (RGB, registered depth, IR). All streams share one clock and full extrinsic calibration; only the sensor changes between the views below.

One action seen by RGB, depth, and four LiDAR point clouds — One take, all modalities. Point clouds colored by distance (near → far along the range gradient).

S01 · Applauding — with estimated 2D pose overlay

S10 · Waving — every sequence ships such a preview

The 49 classes span daily activities, traffic gestures, safety-critical events (falling, stumbling, lying) and person-person interactions. Because the LiDAR modality carries no facial imagery, the benchmark directly quantifies how much action understanding survives in a privacy-preserving sensing stack.

Samples per action class, stacked by subject — Sequences per class, colored by subject.

Explore the sensors

One sequence, five point-of-view panels — all four LiDARs and the depth camera as a point cloud, synchronized on the common clock. Drag any panel to orbit (cameras are linked), pick an action, scrub the timeline. Points are colored by distance.

Protocols

Three official evaluation protocols; splits ship with the data.

cross-subject

Split by person — multi-session subjects never straddle the boundary. 5 persons train, 4 test.

cross-take

The chronologically last take of every session × action is held out.

cross-sensor

Train on one sensor's point clouds, test on another — including depth-camera point clouds. The scan pattern is the only variable.

Baselines

Deliberately simple, fully reproducible references: a PointNet-style per-frame encoder with temporal pooling, a set-abstraction + temporal-transformer model (P4Transformer-style), a Pose-GRU on the shipped 2D skeletons, late LiDAR+depth fusion, and a density-equalization ablation separating point count from scan pattern.

baseline grid is training — the full per-sensor table and cross-sensor matrix land here with the paper.

Download

The dataset is packaged per modality on Hugging Face in two tiers:

Open tier (CC BY-NC 4.0): huggingface.co/datasets/raphaelmemmesheimer/laseract — the four LiDAR streams, depth, 2D poses, labels, splits, calibration, and face-free previews (~250 GB extracted).

Gated tier (research-only access): huggingface.co/datasets/raphaelmemmesheimer/laseract-rgb — RGB video and IR frames, granted on request under the accompanying access terms.

A dependency-light loader (numpy-only, optional PyTorch wrapper), a standalone HTML sequence viewer, and a quickstart example ship in the companion toolkit alongside the dataset; every archive is covered by published SHA256 checksums.

BibTeX

@inproceedings{laseract,
  title     = {LaserAct: A Multi-LiDAR and RGB-D Dataset
               for Human Action Recognition},
  author    = {Anonymous},
  booktitle = {under review},
  year      = {2026}
}