Segmentation

Segment point cloud data using deep learning and geometric algorithms

Semantic segmentation clusters the points of a 3-D point cloud by using their similar characteristics, and associates each point with a class label such as car, building, ground, or vegetation.

You can segment a point cloud based on edges, neighboring point properties, and geometric shapes such as cuboid, plane, and cylinder. Lidar Toolbox™ includes functions and workflows for geometric segmentation of point clouds. For more information, see the Terrain Classification for Aerial Lidar Data example.

Lidar Toolbox also supports semantic segmentation using deep learning. You can use the included pretrained PointSeg, SqueezeSegV2, and PointNet++ convolutional neural networks (CNNs) or develop custom segmentation models. For a segmentation workflow using a PointNet++ network, see Aerial Lidar Semantic Segmentation Using PointNet++ Deep Learning.

Semantic segmentation in lidar point clouds.

Functions

expand all

Geometrical

`segmentGroundSMRF`	Segment ground from lidar data using a SMRF algorithm
`segmentLidarData`	Segment organized 3-D range data into clusters
`segmentGroundFromLidarData`	Segment ground points from organized lidar data
`segmentCurbPoints`	Segment curb points from point cloud
`pcsegdist`	Segment point cloud into clusters based on Euclidean distance

Deep Learning

Load Training Data

`combine`	Combine data from multiple datastores
`countEachLabel`	Count occurrence of pixel or box labels
`groundTruth`	Ground truth label data
`imageDatastore`	Datastore for image data
`pixelLabelDatastore`	Datastore for pixel label data

Augment and Preprocess Training Data

`transform`	Transform datastore
`sampleLidarData`	Sample 3-D bounding boxes and corresponding points from training data
`pcBboxOversample`	Randomly augment point cloud data using objects

Define Layers

pointCloudInputLayer Point cloud input layer

Design Networks

`pointCloudInputLayer`	Point cloud input layer
`squeezesegv2Layers`	Create SqueezeSegV2 segmentation network for organized lidar point cloud
`pointnetplusLayers`	Create PointNet++ segmentation network

Segment Point Cloud

`pcsemanticseg`	Point cloud semantic segmentation using deep learning
`semanticseg`	Semantic image segmentation using deep learning
`segmentAerialLidarVegetation`	Segment vegetation points from aerial lidar data
`segmentAerialLidarBuildings`	Segment building points from aerial lidar data

Visualize Results

`labeloverlay`	Overlay label matrix regions on 2-D image
`pcshow`	Plot 3-D point cloud

Evaluate Results

`evaluateSemanticSegmentation`	Evaluate semantic segmentation data set against ground truth
`segmentationConfusionMatrix`	Confusion matrix of multi-class pixel-level image segmentation

Topics

Deep Learning with Point Clouds
Learn point cloud processing using deep learning.
Semantic Segmentation in Point Clouds Using Deep Learning
Assign class labels to each point inside a point cloud using deep learning.
Getting Started with PointNet++
Define a PointNet++ network and use it to perform semantic segmentation.
Datastores for Deep Learning (Deep Learning Toolbox)
Learn how to use datastores in deep learning applications.
List of Deep Learning Layers (Deep Learning Toolbox)
Discover all the deep learning layers in MATLAB^®.

Featured Examples

Aerial Lidar Semantic Segmentation Using PointNet++ Deep Learning

Train a PointNet++ deep learning network to perform semantic segmentation on aerial lidar data.

Open Live Script

Code Generation For Aerial Lidar Semantic Segmentation Using PointNet++ Deep Learning

Generate CUDA® MEX code for a PointNet++ [1] network for lidar semantic segmentation. This example uses a pretrained PointNet++ network that can segment unorganized lidar point clouds belonging to eight classes (buildings, cars, trucks, poles, power lines, fences, ground, and vegetation). For more information on PointNet++ network, see Getting Started with PointNet++.

Open Live Script

Lidar Point Cloud Semantic Segmentation Using PointSeg Deep Learning Network

Train a PointSeg semantic segmentation network on 3-D organized lidar point cloud data.

Open Live Script

Lidar Point Cloud Semantic Segmentation Using SqueezeSegV2 Deep Learning Network

Train a SqueezeSegV2 semantic segmentation network on 3-D organized lidar point cloud data.

Open Live Script

Code Generation for Lidar Point Cloud Segmentation Network

Generate CUDA® MEX code for a deep learning network for lidar semantic segmentation. This example uses a pretrained SqueezeSegV2 [1] network that can segment organized lidar point clouds belonging to three classes (background, car, and truck). For information on the training procedure for the network, see Lidar Point Cloud Segmentation Using SqueezeSegV2 Deep Learning Network. The generated MEX code takes a point cloud as input and performs prediction on the point cloud by using the DAGNetwork object for the SqueezeSegV2 network.

Open Live Script

Data Augmentations for Lidar Object Detection Using Deep Learning

Perform typical data augmentation techniques for 3-D object detection workflows with lidar data.

Open Live Script

Terrain Classification for Aerial Lidar Data

Segment and classify terrain in aerial lidar data as ground, building, and vegetation. The example uses a LAZ file captured by an airborne lidar system as input. First, classify the point cloud data in the LAZ file into ground and non-ground points. Then, further classify non-ground points into building and vegetation points based on normals and curvature features. This figure provides an overview of the process.

Open Live Script

Extract Forest Metrics and Individual Tree Attributes from Aerial Lidar Data

Extract forest metrics and individual tree attributes from aerial lidar data.

Open Live Script