ElbrusVisualSLAM don't start

Autonomous Machines Robotics - Isaac Isaac SDK
#1

I try to run my test app with v42l camera and visual slam on PC UBUNTU 18.04. I use elbrus_visual_slam_zed.app.json as pattern. My code is

{
  "name": "slam_v4l2",
  "modules": [
    "sensors:v4l2_camera",
    "message_generators",
    "visual_slam",
    "sight"
  ],
  "graph": {
    "nodes": [
      {
        "name": "svo",
        "subgraph": "packages/visual_slam/apps/elbrus_visual_slam_rgb.subgraph.json"
      },
      {
        "name": "camera",
        "components": [
          {
            "name": "MessageLedger",
            "type": "isaac::alice::MessageLedger"
          },
          {
            "name": "V4L2Camera",
            "type": "isaac::V4L2Camera"
          }
        ]
      },
      {
        "name": "intrinsics",
        "components": [
          {
            "name": "MessageLedger",
            "type": "isaac::alice::MessageLedger"
          },
          {
            "name": "CameraIntrinsicsGenerator",
            "type": "isaac::message_generators::CameraIntrinsicsGenerator"
          }
        ]
      }
    ],
    "edges": [
      {
        "source": "camera/V4L2Camera/frame",
        "target": "intrinsics/CameraIntrinsicsGenerator/image"
      },

      {
        "source": "camera/V4L2Camera/frame",
        "target": "svo.subgraph/interface/left_image"
      },
      {
        "source": "intrinsics/CameraIntrinsicsGenerator/intrinsics",
        "target": "svo.subgraph/interface/left_intrinsics"
      },
      {
        "source": "camera/V4L2Camera/frame",
        "target": "svo.subgraph/interface/right_image"
      },
      {
        "source": "intrinsics/CameraIntrinsicsGenerator/intrinsics",
        "target": "svo.subgraph/interface/right_intrinsics"
      }
    ]
  },
  "config": {
    "camera": {
      "V4L2Camera": {
        "device_id": 0,
        "rows": 720,
        "cols": 1280,
        "rate_hz": 10
      }
    },
    "intrinsics": {
      "CameraIntrinsicsGenerator": {
        "focal_length": [1.1457752257790760e+03, 1.5318594933077491e+03],
        "optical_center": [3.1614587655417233e+02, 6.5917832786233669e+02],
        "distortion_coefficients": [6.3095453672864246e-02, 1.1092076796490181e-01,
            -1.7519424014469689e+00, 7.9980152696699797e-04, -1.9309734213037914e-03]
      }
    },
    "svo.svo_grayscale.tracker": {
      "ElbrusVisualSLAM": {
        "horizontal_stereo_camera": true,
        "process_imu_readings": false,
        "denoise_input_images": true,
        "lhs_camera_frame": "v4l2_left_camera",
        "rhs_camera_frame": "v4l2_right_camera"
      }
    }
  }
}

Files elbrus_visual_slam_grayscale.subgraph.json and elbrus_visual_slam_rgb.subgraph.json are originals.
But I got message “WARN external/com_nvidia_isaac_engine/engine/alice/backend/node_canister.cpp@173: Can not start node ‘svo.svo_grayscale.tracker’ because it is not in the stage ‘constructed’ or ‘stopped’”
On Jetson Xavier I got same message.

#2

The svo subgraph needs to have edges from two different cameras in a stereo camera pair. Otherwise, there is no disparity for Isaac Visual SLAM to work with and Tracker will not able to initialize.

#3

Can I get additional information about errors? I use simple stereo camera (not ZED). After calibration put Intrinsics with CameraIntrinsicsGenerator for left and right cameras, but get same result.
Other thing. I observe that version Elbrus slam is 7.6 and there isn’t CUDA support where. How can I use v10.0 from isaac_ros_visual_slam or other with CUDA? Any compatible version of Elbrus for isaac sdk 2021

#4

You have only one camera/V4L2Camera/frame but you would expect two: a left and right frame, coming from your stereo camera. You’ll need a rigid baseline between the left and right imagers in your stereo camera, of course. Try updating your graph to have two instances of V4L2Camera, one for each imager in your stereo camera.

Isaac Visual SLAM includes in Isaac SDK 2021.1 is CPU-only and does not consume your GPU resources. Isaac ROS Visual SLAM uses a different version of Isaac Visual SLAM libraries which do leverage CUDA and is available for the ROS ecosystem.