k4a_cuda

A K4A library that wraps functionality from the K4A and K4ABT C SDKs for use in Unity C# scripts. This library performs the point cloud calculation in CUDA as described in the SDK examples.

Script example

using System;
using System.Runtime.InteropServices;
using UnityEngine;

unsafe public class GenerateCloud : MonoBehaviour
{
    ParticleSystem ps;

    [StructLayout(LayoutKind.Sequential)]
    public struct float4
    {
        public float x;
        public float y;
        public float z;
        public float w;
    }

    [DllImport("k4a_cuda_win.dll", CharSet = CharSet.Unicode, SetLastError = true)]
    static extern void init_cpc(bool color, bool body_tracking);

    [DllImport("k4a_cuda_win.dll", CharSet = CharSet.Unicode, SetLastError = true)]
    static extern void get_capture();

    [DllImport("k4a_cuda_win.dll", CharSet = CharSet.Unicode, SetLastError = true)]
    static extern void setup_point_cloud(ref float4[] point_positions, ref uint[] point_colors);

    [DllImport("k4a_cuda_win.dll", CharSet = CharSet.Unicode, SetLastError = true)]
    static extern void get_point_cloud();

    [DllImport("k4a_cuda_win.dll", CharSet = CharSet.Unicode, SetLastError = true)]
    static extern uint get_point_count();

    [DllImport("k4a_cuda_win.dll", CharSet = CharSet.Unicode, SetLastError = true)]
    static extern int get_max_point_count();

    float4[] point_positions;
    uint[] point_colors;

    void Start()
    {
        ps = GetComponent<ParticleSystem>();

        init_cpc(true, false);

        int max_points = get_max_point_count();

        point_positions = new float4[max_points];
        point_colors = new uint[max_points];

        GC.KeepAlive(point_positions);
        GC.KeepAlive(point_colors);

        setup_point_cloud(ref point_positions, ref point_colors);
    }

    void Update()
    {
        get_capture();
        get_point_cloud();
        uint point_count = get_point_count();

        for (int i = 0; i < point_count; i++)
        {
            byte[] color_bits = BitConverter.GetBytes(point_colors[i]);
            ParticleSystem.EmitParams ep = new ParticleSystem.EmitParams
            {
                position = new Vector3(point_positions[i].x, point_positions[i].y, point_positions[i].z),
                startColor = new Color32(color_bits[2], color_bits[1], color_bits[0], 0xFF)
            };

            ps.Emit(ep, 1);
        }
    }
}

Notes

• This project has been build and tested using CUDA 10.1 and Visual Studio 16 2019
• dnn_model.onnx is not path-searched and is required with your application (in "<output>/<Name>_Data/" for Unity)
• Calculating and displaying approximately 6M points per second, running a DNN model, and running a game engine does require a fairly powerful system. I've had success (albeit with a 1/4 to 1/2 second delay) on a GTX 960M and with minimal delay on GTX 1080
• Set your CUDA version with -DCMAKE_CUDA_COMPILER="<path-to-cuda/bin>/nvcc.exe""

Further work

• As with the last point above, there are several place I could optimize my code to improve the overall performance. This is my first C++ project on a system with more than 1MB of RAM, so any suggestions/pull requests are welcome.
• Use a remote system for processing the point cloud and body tracking