Hello CV community,
I am designing a warehouse pallet counting system using two opposing RTSP cameras (Godown4 and IPC, 1080p, ~15–20° tilt) monitoring an unloading bay. The cargo consists of white shrink-wrapped yarn bundles and rectangular cartons.
The Problem:
-
Pallets feature alternating interlocking tiers (e.g., 2 bundles parallel, 1–2 bundles turned transverse/perpendicular, 2 bundles on top).
-
The opposing cameras give front and back perspectives, but central/internal transverse bundles remain occluded from both 2D views.
-
Structural columns obscure ~35% of the frame at the resting bay.
Intended Geometric Pipeline:
-
Compute planar homography matrices (H_1, H_2) from camera coordinates to metric floor coordinates using the yellow boundary tape marks.
-
Calibrate vertical tier height (Z) using the uniform rack upright slots visible in both views to slice the cargo into discrete vertical layers.
-
Combine 2D bounding boxes/masks from both perspectives onto a shared ground-plane occupancy grid to register bundle dimensions.
Questions:
-
What is the recommended strategy in OpenCV to reconcile overlapping 2D multi-view bounding footprints into a unified 3D oriented bounding box (OBB) without stereo epipolar line matching (cameras are directly facing each other, not side-by-side)?
-
When an internal unit is occluded, is volumetric layer slicing (\\Delta Z / h\_{\\text{bundle}}) combined with footprint filling typically stable under perspective distortion?