Welcome to Jetson Orin NANO Developer Kit repository
12.LiDAR 3D point cloud detection
TAO PointPillarNet LiDAR 3D Point Cloud Detection
TAO PointPillarNet LiDAR 3D Point Cloud Detection1. Case Goal and Processing Flow2. Pre-Run Checks3. Starting Point Cloud Detection4. Replacing Point Cloud Data5. Common Issues
Applicable environment: Jetson Orin Nano 8GB, JetPack 7.2.1, DeepStream 9.1. This case uses LiDAR point clouds and Triton gRPC, not video detection
After setting the power mode to MAXN, running the case may pop up an over-current warning; just ignore it.
1. Case Goal and Processing Flow
PointPillarNet detects 3D objects from point cloud data. The input is .bin point clouds plus the point count/calibration information required by the configuration; it does not accept ordinary RGB images or MP4.
LiDAR point cloud → preprocessing → Triton gRPC → PointPillarNet → 3D detection results2. Pre-Run Checks
xxxxxxxxxxcd ~/DeepStream/tao-toolkit-shell-demo./run_lidar.sh --helpAccording to the current script help and configuration, confirm the point cloud path, model repository, label/calibration files, and Triton port. The ordinary video mode's video_in/ and video_out/ do not apply to this case.
3. Starting Point Cloud Detection
Open two terminals. In terminal one, start Triton and keep it running:
xxxxxxxxxxcd ~/DeepStream/tao-toolkit-shell-demo./run_lidar_triton.shAfter confirming the models are loaded, run the point cloud application in terminal two:
xxxxxxxxxxcd ~/DeepStream/tao-toolkit-shell-demo./run_lidar_render.sh
Check the service connection, point cloud frame processing, and 3D detection results. Whether the program pops up a visualization window and whether results are written to files depend on the current app configuration and run logs.
4. Replacing Point Cloud Data
Prepare .bin point clouds consistent with the model's input conventions, together with their accompanying label/calibration information. Before replacing, check the paths actually read by the configuration, the point field layout, coordinate units, and preprocessing parameters; keep the original sample first. Do not disguise an MP4 as a point cloud by changing the extension.
5. Common Issues
| Symptom | Check and Handling |
|---|---|
| gRPC connection failed | Confirm the Triton in terminal one is still running, and verify the service address and port. |
| Model repository loading failed | Check whether the model directory, configuration, and TensorRT engine match. |
| Input cannot be opened | Verify the .bin path, permissions, and point cloud format; MP4 cannot be used as input. |
| No 3D boxes | Check the coordinate system, units, field layout, calibration, and preprocessing configuration. |