Earlier this month, I joined the perception team at EUFS(Edinburgh University Formula Student). I learned many interesting topics in autonomous vehicle perception and teamwork tools. Today, I would like to share what I’ve learned about point cloud technology.

This is a new series of for my blog, stay tuned 😎！

What is Point Cloud?

A point cloud is a collection of millions of 3D points that represent everything around you. Usually we use a sensor called LiDAR(Light Detection and Ranging) to create these point clouds.

How Point Cloud work:

  1. The system emits a laser pulse (usually a single-wavelength infrared light)
  2. measures the time it takes for the laser to reflect back. This calculates the distance and generates a 3D point cloud.
  3. Each point contains only the spatial coordinates (x, y, z) and the reflection intensity, without RGB color information. pointcloud

Point Cloud property:

  1. unorder
  2. sparsity
  3. unstructured
  4. geometric transformability

How can we process point clouds

1.PCL

We use a popular library called PCL that provides tools for:

  • Filtering: Removing noise and irrelevant points
  • Segmentation: Grouping points that belong to the same object
  • Clustering : Finding groups of point that might be cones
  • Feature extraction: Calculating geometric properties like shapes and sizes.

2. Grid-Based Approaches

  • Divide the space around the car into a 2D grid
  • Count how many points fell into each grid cell
  • Look for patern that match cone shapes
  • Extract potential cone locations

3. PointPillars

See details in this blog: Understanding PointPillars

Common Point Cloud tasks:

  1. 3d shape classification: categorizing entire point clouds into object classes(car, pedestrain)
  2. 3d object detection and tracking: Locating objects with 3D bounding boxes and following them across frames
  3. 3d point cloud segmentation: Assigning a class label to every single point.

Common Library used in Point Cloud:

The point cloud ecosystem offers several powerful libraries:

  1. PCL: C++ library for classical algorithms
  2. VCGLib: Specialized for mesh processing and geometry
  3. CGAL: Computational geometry algorithms library
  4. Open3d: Python library with excellent visualization

High quality point cloud datasets:

  1. KITTI
  2. waymo
  3. nuScenes

Other Methods for 3D object detection:

while LiDAR is powerful, various other approaches exists for 3D perception.

  1. monocular camera
    1. D4LCN
    2. Mono3D-PLiDAR
  2. binocular camera
    1. Stereo R-CNN
    2. PSMNet
  3. RGB-D
    1. F-PointNets
  4. LiDAR
    1. PV-RCNN
    2. PointPillar
    3. PointRCNN
  5. RGB+Li-dar
    1. EPNet
    2. CLOCs
    3. 3D-CVF Each approach has trade-offs in term of accuracy, speed, cost and robustness.

Series Perception

  1. Introduction to Point Cloud