LiDAR Based 3D Perception Guide

Spinning lasers paint exact points day and night. LiDAR perception turns those returns into boxes that cameras cannot match at range.

Laser beams sweep rings that thin with range while intensity marks surfaces

Why Does This Exist?

Cameras fail at night and guess depth; radar is coarse. LiDAR exists as the measured middle: active lasers return range plus intensity regardless of sun, feeding 3D object detection with metric truth. It costs more and hates rain, which is why fusion persists.

Detectors built on it are VoxelNet, PointPillars, CenterPoint, and SECOND. Points themselves are point clouds.

Think of It Like This

Bats flying with laser rulers

Bats click and hear walls in the dark. Give each bat a laser ruler that also reports wall paint brightness, and it maps caves at noon or midnight alike.

Clicks are laser pulses and echoes are returns. The analogy stops at weather: sound crosses fog while $905$ nm light scatters, so rain blinds lidar first.

How It Actually Works

A $64$ -beam unit spins at $10$ Hz with vertical spacing about $0.4$ degrees, recording $(x, y, z)$ plus intensity and elongation per return. Density falls as $1/r^2$ : a car gives about $800$ points at $10$ m but $50$ at $40$ m since $(10/40)^2 = 1/16$ . Detectors voxelize, pillar-encode, or point-process these returns, then regress boxes with height from direct $z$ instead of guessing.

Worked numbers

Angular step $0.4$ degrees spans $0.07$ m at $10$ m ( $10 \times 0.007$ rad) but $0.28$ m at $40$ m. A $1.8$ m tall car covers about $25$ beams near and $6$ far. Intensity $0.9$ on plates versus $0.2$ on tires separates parts before any network runs.

Code

1def point_count(near: int, r0: float, r1: float) -> float:2    return near * (r0 / r1) ** 2  # inverse square thinning3
4
5print(point_count(800, 10.0, 40.0))  # -> 50.0

Watch Out For

Trusting lidar in heavy spray

Rain and exhaust scatter returns into ghost points that detectors box as obstacles. The symptom is phantom braking in storms. Filter by intensity and multi-return flags, and lean on radar when spray peaks.

The Quick Version

LiDAR returns measured range plus intensity day or night.
Sixty-four beams at $0.4$ degrees set the angular grid.
Density falls with range squared: $800$ near becomes $50$ far.
Height comes measured, not guessed like cameras.
Rain scatter creates ghost points to filter.
Voxel, pillar, and point heads all start from these returns.