{"id":13461,"date":"2026-06-01T10:33:04","date_gmt":"2026-06-01T14:33:04","guid":{"rendered":"https:\/\/blogs.mathworks.com\/student-lounge\/?p=13461"},"modified":"2026-06-01T10:33:04","modified_gmt":"2026-06-01T14:33:04","slug":"using-lidar-for-3d-object-detection","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/student-lounge\/2026\/06\/01\/using-lidar-for-3d-object-detection\/","title":{"rendered":"Using Lidar for 3D Object Detection"},"content":{"rendered":"
\n
Today, we’re going through each step of creating a 3D object detector using Lidar data. We’ll start with an introduction to 3D point clouds and how to process them, then we’ll label point clouds for deep learning, train an 3D object detector, and wrap up by discussing a few options for deploying the detector to your projects or hardware.<\/div>\n
The complete dataset, code, and videos we’ll be discussing in this blog can be found at this GitHub Repository<\/a>.<\/div>\n
    \n
  1. Understanding and Processing Point Clouds<\/a><\/li>\n
  2. Labeling Point Clouds for Deep Learning<\/a><\/li>\n
  3. Training a 3D Object Detector with Deep Learning<\/a><\/li>\n
  4. Deploying a 3D Object Detector<\/a><\/li>\n<\/ol>\n

    1. Understanding and Processing Point Clouds<\/h2>\n
    The dataset used in this example can be found in this ‘parkingLot’ folder<\/a> and will need to be downloaded in order to follow along. A video explanation of this topic can be found here<\/a>.<\/div>\n

    Understanding Point Clouds<\/h3>\n
    Lidar sensors record spatial data and store it as a point cloud<\/span>, or a collection of points that represent the world around the sensor. In a 3D point cloud, these points are stored as x,y,z coordnates and can be visualized, so if your environment looks like this:<\/div>\n
    \"\"<\/div>\n
    your point cloud might look like this:<\/div>\n
    \"\"<\/div>\n
    This is a way of numerically representing the world around us in 3 dimensions, which is helpful for developing algorithms that rely on things like distance from a robot or size of an object.<\/div>\n

    Processing Point Clouds<\/h3>\n
    Processing point clouds before using them for deep learning can often lead to better results. The Lidar Viewer App<\/a> lets you import, view, and process point clouds interactively, and is demonstrated in this video<\/a>.<\/div>\n
    You can also process point clouds programmatically. Load in a point cloud from our dataset using the pcread<\/span> function:<\/div>\n
    \n
    \n
    singlePtCld = pcread(‘.\/parkingLot\/frame_0001.pcd’<\/span>);<\/span><\/div>\n<\/div>\n
    \n
    pcshow(singlePtCld)<\/span><\/div>\n
    \n
    \n
    <\/div>\n
    <\/div>\n
    <\/div>\n
    <\/div>\n
    <\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n

    Denoising<\/h3>\n
    Point clouds can have “noise”, or data points that are outliers to what you’d expect to see in your point cloud. Use the pcdenoise<\/span> function to remove or reduce these. The purple points in the figure below represent points that were removed.<\/div>\n
    \n
    \n
    ptCloudDenoise = pcdenoise(singlePtCld);<\/span><\/div>\n<\/div>\n
    \n
    figure<\/span><\/div>\n<\/div>\n
    \n
    pcshowpair(singlePtCld, ptCloudDenoise)<\/span><\/div>\n
    \n
    \n
    <\/div>\n
    <\/div>\n
    <\/div>\n
    <\/div>\n
    <\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n

    Downsampling<\/h3>\n
    Depending on your LiDAR and the environment you collected data in, you may have a lot of data points, which can lead to slower computaton. Use the pcdownsample<\/span> function to make your point cloud less dense through downsampling. Use pcshowpair<\/span> to view which data points were removed (purple) and which stayed (green).<\/div>\n
    \n
    \n
    ptCloudDownSampled = pcdownsample(ptCloudDenoise,‘random’<\/span>,0.60);<\/span><\/div>\n<\/div>\n
    \n
    pcshowpair(ptCloudDenoise, ptCloudDownSampled)<\/span><\/div>\n
    \n
    \n
    <\/div>\n
    <\/div>\n
    <\/div>\n
    <\/div>\n
    <\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n
    Check out this documentation page<\/a> to see further options for processing lidar data.<\/div>\n
    It can be helpful to test out different preprocessing methods on a single point cloud to see how it affects your data. Once you have a combination of steps and settings that you’re happy with, you can apply it to the whole dataset. This is easy to do with the Lidar Viewer app, which exports the processed dataset to a folder called parkingLot_Edit by default.<\/div>\n

    2. Labeling Point Clouds for Deep Learning<\/h2>\n
    Once the data is processed, you can label the objects in the point clouds. First, split the data into a train and test dataset, which helps for the deep learning step.<\/div>\n
    \n
    \n
    pcd_files = dir(“.\/parkingLot_Edit\/*.pcd”<\/span>);<\/span><\/div>\n<\/div>\n
    \n
    <\/div>\n<\/div>\n
    \n
    test_data = pcd_files(1:366);<\/span><\/div>\n<\/div>\n
    \n
    train_data = pcd_files(367:807); <\/span><\/div>\n<\/div>\n
    \n
    <\/div>\n<\/div>\n
    \n
    if <\/span>~exist(“.\/parkingLotTrain”<\/span>, ‘dir’<\/span>)<\/span><\/div>\n<\/div>\n
    \n
    mkdir(“.\/parkingLotTrain”<\/span>);<\/span><\/div>\n<\/div>\n
    \n
    end<\/span><\/span><\/div>\n<\/div>\n
    \n
    <\/div>\n<\/div>\n
    \n
    if <\/span>~exist(“.\/parkingLotTest”<\/span>, ‘dir’<\/span>)<\/span><\/div>\n<\/div>\n
    \n
    mkdir(“.\/parkingLotTest”<\/span>);<\/span><\/div>\n<\/div>\n
    \n
    end<\/span><\/span><\/div>\n<\/div>\n
    \n
    <\/div>\n<\/div>\n
    \n
    for <\/span>i=1:length(train_data)<\/span><\/div>\n<\/div>\n
    \n
    fileName = fullfile(train_data(i).folder, train_data(i).name);<\/span><\/div>\n<\/div>\n
    \n
    copyfile(fileName, “.\/parkingLotTrain”<\/span>);<\/span><\/div>\n<\/div>\n
    \n
    end<\/span><\/span><\/div>\n<\/div>\n
    \n
    <\/div>\n<\/div>\n
    \n
    for <\/span>i=1:length(test_data)<\/span><\/div>\n<\/div>\n
    \n
    fileName = fullfile(test_data(i).folder, test_data(i).name);<\/span><\/div>\n<\/div>\n
    \n
    copyfile(fileName, “.\/parkingLotTest”<\/span>);<\/span><\/div>\n<\/div>\n
    \n
    end<\/span><\/span><\/div>\n<\/div>\n<\/div>\n
    For this example, we label cars, cones, and barrels using the Lidar Labeler app<\/a>. You can start a new labeling session by running the ‘lidarLabeler’ command below<\/div>\n
    \n
    \n
    lidarLabeler<\/span><\/div>\n<\/div>\n<\/div>\n
    or by selecting the Lidar Labeler app from the ‘APPS’ tab.<\/div>\n
    Watch this video<\/a> to see how we label the dataset and save the labels to a variable called gTruth:<\/span><\/div>\n
    <\/div>\n