Notepad

Novelty

• Implementation on drone with limited processing and payload.
• Return object model and pose for grasping
• Motion for segmentation and keeping track of moving objects.
• Control of drone movement for validation of candidates.
• Cooperation of drones for validation of candidates.
• Someway of autonomously detecting and learning unknown objects.
• Try to grasp an object.

Possible candidates

• Object split in to two or more candidates
• Not complete objects as candidate
• Two objects in one candidate
• surrounding/floor included in candidate
• Noise produce false candidates.
• Good candidates

Parameters Analysis and Performance Analysis

• Blur:
  • Increase it --> Noisy PCs
  • Decrease it --> Gaps in ICP
  • TODO: Create a perfect dataset without noise
• Template size:
  • Plot TemplateSize vs Reprojection error
  • Plot TemplateSize vs Time
  • Using Train sets (The one in which the camera spin around the objects) to determine size of object and use it as groud truth. Then plot somehow predicted size of objects vs real size or something like that.
• History size:
  • 3d-plot HistorySize and voxelSize vs ObjectPoints/NoisePoints.
  • 3d-plot HistorySize and voxelSize vs Absolute number of points in object.
• StatisticalOutlierRemoval (Filtering):
  • Effect of K in time and ObjectPoints/NoisePoints and Absolute number of points in object.

Next Experiments and Record of datasets

• Chessboard to set the origin
• Record Imu information. As much as possible.
• Position object accurately in chessboard defined CS.
• Make sure to have static object.
• Tests
  • Record drone from "outside" to show it.
  • Basic scene with simple shapes (cilinders, boxes, etc...)
  • One object scene
  • Only "crawler"
  • "Change floors" (white, normal, textured)
  • Several objects:
    • Textured objects.
    • Simple objects.
    • Tools