LiDARacks

A lightweight Python toolkit for injecting physically-grounded anomalies into LiDAR point clouds, intended for robustness evaluation and stress-testing of perception models.

LiDARacks operates directly on spherical-coordinate point clouds and implements four corruption models, each with five configurable severity levels.

Corruption Models

Type	Key	Physical phenomenon	Effect on the point cloud
Background Noise	Background	Adverse weather (rain, fog, dust)	Inserts spurious returns at ranges shorter than the true surface
Zero-Range Background	ZRB	Sensor obstruction (debris, water on the lens)	Inserts points at rho = 0 on randomly selected azimuths
Electromagnetic Interference	EMI	Electronic noise coupling into the range channel	Adds a sinusoidal range error across the scan, clipped to the valid range
Occlusion	Occlusion	Ray diffusion or absorption	Removes all returns along randomly selected azimuths

Severity is selected with severity_level in the range 0–4, mapped per corruption type as follows:

Type	Severity parameter	0	1	2	3	4
Background	Points added ≈ N / value	45	40	35	30	20
ZRB	Zero-range points added	1	2	3	4	5
EMI	Range-error amplitude (m)	0.5	1.0	1.5	2.0	2.5
Occlusion	Azimuths removed	2	3	4	5	6

Input Format

Point clouds are expected as a numpy.ndarray of shape (N, 3) in spherical coordinates, with columns ordered [rho, theta, phi]:

Column	Meaning	Range	Resolution
rho	Radial distance (m)	0 – 5	continuous
theta	Azimuth (deg)	-180 – 180	360 / 61 ≈ 5.90°
phi	Elevation (deg)	75 – 100	5°

Requirements

• Python 3.8+
• NumPy (developed and tested with 1.26.3)
• Matplotlib (only required to run the visualizations in the notebook)

pip install numpy matplotlib

Usage

import numpy as np
import LiDARacks

sph_point_clouds = np.load("path/to/sph_point_cloud.npy")   # shape (N, 3)

corrupted = LiDARacks.lidaracks(
    sph_point_clouds,
    severity_level=3,        # 0 to 4
    type_of_noise="EMI",     # 'EMI', 'Occlusion', 'Background', 'ZRB'
)

API

lidaracks(sph_point_clouds, severity_level=0, type_of_noise='EMI') -> np.ndarray

Parameters

• sph_point_clouds (np.ndarray): input cloud of shape (N, 3) in [rho, theta, phi] order.
• severity_level (int): corruption severity, 0–4.
• type_of_noise (str): one of 'EMI', 'Occlusion', 'Background', 'ZRB'.

Returns

A new np.ndarray of shape (M, 3). The input array is not modified. Note that M differs from N for the additive (Background, ZRB) and subtractive (Occlusion) corruptions; only EMI preserves the point count.

Raises

• TypeError if sph_point_clouds is not a NumPy array.
• ValueError if the input shape is not (N, 3), if severity_level is outside 0–4, or if type_of_noise is unrecognized.

Reproducibility

The library function does not seed the random number generator. Set a global seed before calling it if deterministic output is required:

np.random.seed(42)
corrupted = LiDARacks.lidaracks(sph_point_clouds, 3, "Background")

Repository Contents

File	Description
LiDARacks.py	Library module exposing the lidaracks function
LiDARacks.ipynb	Notebook walking through each corruption model with before/after plots
sph_point_cloud.npy	Sample spherical point cloud (234 points) for testing and demos