LiDAR data processing basics in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding LiDAR Point Clouds

What does a LiDAR point cloud primarily represent in 3D space?

A A sequence of GPS coordinates without elevation data

B A collection of 3D points representing surfaces detected by laser pulses

C A 2D image captured by a camera attached to the LiDAR sensor

D A set of temperature readings from the environment

💡 Hint

Think about what LiDAR sensors measure when they send laser pulses.

LiDAR sensors emit laser pulses and measure the time it takes for the pulses to return after hitting surfaces. This creates a 3D point cloud representing the shape and position of objects.

❓ Predict Output

intermediate

Output of Basic LiDAR Data Filtering

Given the following Python code that filters LiDAR points based on height, what is the output?

Computer Vision

import numpy as np
points = np.array([[1,2,3], [4,5,1], [7,8,9], [10,11,0]])
filtered = points[points[:,2] > 2]
print(filtered)

A

[[ 4  5  1]
 [10 11  0]]

B []

C

[[ 1  2  3]
 [ 7  8  9]]

D

[[ 1  2  3]
 [ 4  5  1]
 [ 7  8  9]]

💡 Hint

Look at the condition points[:,2] > 2 and which points satisfy it.

The code selects points where the third coordinate (height) is greater than 2. Only points [1,2,3] and [7,8,9] meet this condition.

❓ Model Choice

advanced

Choosing a Model for LiDAR Semantic Segmentation

You want to classify each point in a LiDAR point cloud into categories like ground, vegetation, and buildings. Which model type is best suited for this task?

A Convolutional Neural Network (CNN) applied on 2D images

B Generative Adversarial Network (GAN) for image generation

C Recurrent Neural Network (RNN) for sequential data

D PointNet or PointNet++ designed for point cloud data

💡 Hint

Consider models that directly handle unordered 3D point sets.

PointNet and PointNet++ are specialized neural networks designed to process raw point clouds directly for tasks like semantic segmentation.

❓ Hyperparameter

advanced

Effect of Voxel Size in LiDAR Data Processing

When converting a LiDAR point cloud into a voxel grid for processing, what is the effect of choosing a very small voxel size?

A Higher spatial resolution but increased memory and computation cost

B Lower spatial resolution and faster processing

C No effect on resolution or computation

D Automatically removes noise from the point cloud

💡 Hint

Think about how voxel size relates to detail and data volume.

Smaller voxels capture finer details but create more voxels, increasing memory use and computation time.

❓ Metrics

expert

Evaluating LiDAR Classification Accuracy

You have a LiDAR point cloud classification model. After testing, you get the following confusion matrix for three classes (Ground, Vegetation, Building):

Ground: TP=90, FP=10, FN=15
Vegetation: TP=80, FP=20, FN=10
Building: TP=70, FP=15, FN=20

What is the overall precision of the model?

A Approximately 0.85

B Approximately 0.79

C Approximately 0.90

D Approximately 0.70

💡 Hint

Precision = TP / (TP + FP) summed over all classes.

Total TP = 90+80+70=240, total FP = 10+20+15=45, overall precision = 240 / (240+45) = 240/285 ≈ 0.842 ≈ 0.85.