Point Cloud

TLDR: A point cloud is a collection of 3D data points in space. Each point has X, Y, Z coordinates. Point clouds power 3D perception in autonomous vehicles and robots.

A point cloud is a discrete set of data points in three-dimensional space. Each point holds a position defined by Cartesian coordinates (X, Y, Z). Points may also carry additional attributes: color (RGB), intensity, or timestamps. Point clouds represent the surfaces of physical objects and environments. They are the primary output of LiDAR sensors and 3D scanners.

How Point Clouds Are Created

1. LiDAR Scanning: A LiDAR sensor fires laser pulses. It records the time of return for each. This time-of-flight data yields precise 3D positions.
2. Photogrammetry: Multiple overlapping images are processed to reconstruct 3D geometry.
3. Stereo Vision: Two offset cameras mimic binocular vision to estimate depth.
4. Structured Light: A projected pattern deforms over the object surface. A camera captures the deformation and computes depth.

Point Clouds in Autonomous Vehicles

Self-driving cars capture dense point clouds at up to 10 times per second. The point cloud shows the 3D layout of the road, other vehicles, and pedestrians. AI models trained on labeled point clouds learn to detect and classify objects in 3D. Multiple successive point cloud frames track moving objects over time. This is critical for safe path planning and collision avoidance.

Point Clouds in Robotics and Mapping

1. SLAM: Robots build and update 3D maps of unknown environments in real time.
2. Object Grasping: Robotic arms use point clouds to estimate object pose before picking.
3. Digital Twins: Factories and cities are scanned into high-fidelity 3D models.
4. GIS and Terrain Mapping: Airborne LiDAR point clouds generate digital elevation models.

Point Cloud Processing for AI

Raw point clouds are unstructured. AI models must learn directly from irregular 3D data. PointNet and PointNet++ are pioneering neural architectures for point cloud learning. Voxel-based methods convert point clouds to a regular 3D grid for CNN processing. Labeling point clouds is labor-intensive — each point in a frame may need a class annotation. Bright Data’s datasets include 3D sensor data to accelerate training data pipelines for computer vision and robotics.