Main Content

Detection and Tracking

Object detection, shape fitting, and tracking in lidar point cloud data

Object detection is a technique that identifies and locates objects in a scene. This enables you to detect 3-D objects in a point cloud. Lidar Toolbox™ includes functionality that enables you to detect objects using geometric shape fitting or deep learning with convolutional neural networks.

  • Geometric shape fitting — Detect the 3-D geometry of the objects in the point cloud by using ground segmentation and plane-fitting algorithms. You can detect the location, dimensions and direction of each object. You can use the detected objects for downstream workflows such as tracking, path planning and labeling.

  • Deep learning — A deep learning approach to object detection uses convolutional neural networks to perform object detection. Lidar Toolbox includes object detection workflows that use neural networks such as PointPillars and Complex-YOLO v4. You can train a custom object detection model, or use the available pretrained networks and further tune it for your application. The toolbox also supports CUDA® MEX code generation for PointPillars and SqueezeSegV2 networks.

Object tracking is a technique that estimates and tracks the movement of objects across multiple scans of a scene. Object tracking consists of assigning a unique ID to detected objects and tracking their movement across point cloud frames. Lidar Toolbox includes detection and tracking workflows for vehicles, road lanes, and curbs. Most of these workflows use the joint probabilistic data association (JPDA) tracker.

Deep learning-based object detection in lidar point clouds.

Functions

expand all

pcfitcuboidFit cuboid over point cloud
pcfitplaneFit plane to 3-D point cloud
pcnormalsEstimate normals for point cloud
planeModelObject for storing parametric plane model
cuboidModelParametric cuboid model

Load Training Data

groundTruthGround truth label data
combineCombine data from multiple datastores
fileDatastoreDatastore with custom file reader
boxLabelDatastoreDatastore for bounding box label data

Augment and Preprocess Training Data

randomAffine3dCreate randomized 3-D affine transformation
bboxwarpApply geometric transformation to bounding boxes
pctransformTransform 3-D point cloud

Object Detection

lidarObjectDetectorTrainingDataCreate training data for lidar object detection
pointPillarsObjectDetectorPointPillars object detector
trainPointPillarsObjectDetectorTrain PointPillars object detector
detectDetect objects using PointPillars object detector
detectLOAMFeaturesDetect LOAM feature points from 3-D lidar data

Visualize Results

showShapeDisplay shapes on image, video, or point cloud
pcshowPlot 3-D point cloud

Evaluate Results

evaluateDetectionAOSEvaluate average orientation similarity metric for object detection
bboxOverlapRatioCompute bounding box overlap ratio

Topics