Bolt Detection in PCD Scans

My 3D scanner streams PCD point-cloud files that contain spiral-shank bolts mixed with other hardware. I need a compact, CUDA-friendly C++ solution that runs directly on an NVIDIA Jetson Orin NX, spots every bolt in each cloud, records the head-center XYZ (and, if practical, its axis direction), and writes the results to a structured XML file.

Performance targets
• Accuracy: at least 95 % correct identification on the annotated dataset I will supply.
• Speed: real-time or near real-time processing on the Orin NX under the standard JetPack image.

You are welcome to build on open-source libraries such as PCL, Open3D, Eigen, cuBLAS or TensorRT, provided everything can be compiled through CMake and redistributed without license issues.

Deliverables
• Well-documented C++ source code and CMake project files that build and run on Jetson Orin NX.
• Stand-alone executable with a simple CLI: input PCD → output XML.
• XML schema definition and at least one sample output file.
• Brief technical report summarising the approach, compile/run instructions, and measured accuracy & timing on my test set.

Acceptance criteria
– ≥95 % precision and recall on the supplied point-cloud set.
– Position error ≤0.5 mm for each bolt’s XYZ.
– Processing time per cloud meets the agreed real-time threshold.

I will provide representative PCD clouds, calibration information, and a labelled validation set as soon as the project starts. Feel free to tell me what additional data you might need to hit the accuracy goal.