BEV Uncertainty-Aware Occupancy Dashboard

Sample:

Coverage

cam-visible cells

Mean Confidence

avg occupancy prob

Uncertainty ×10³

epistemic variance

Occupancy Ratio

cells > 0.45 thresh

Dist-Wt Error

near-ego weighted

Occupancy Map TURBO · MEAN

FREE SPACE

LOW OCC

MED OCC

OCCUPIED

◈Blue = free drivable space · Red/Yellow = likely obstacle

Epistemic Uncertainty MC DROPOUT · VAR

CERTAIN

LOW UNCERT

MED UNCERT

UNCERTAIN

◈Higher uncertainty in peripheral regions due to reduced multi-camera overlap

Uncertainty-Overlay FUSED · NOVEL

FREE + CERTAIN

OCCUPIED

UNCERTAIN OCC

◈Red bleed = occupied but uncertain — planner should treat this zone cautiously

Camera Coverage Mask VISIBILITY · IPM

BLIND SPOT

CAMERA VISIBLE

◈6-cam surround covers 99%+ of 100m x 100m grid — only ego-vehicle footprint is blind

6-Camera Surround View RAW INPUT · nuScenes

◈All 6 cameras projected via calibrated IPM into a unified 200x200 BEV grid (0.5 m/px resolution)

// Pipeline Architecture

6-Cam Input

All surround cameras loaded & calibrated

→

IPM Projection

World-grid → image coords via homography

→

Feature Fusion

RGB features avg across visible cameras

→

MC Dropout

10 stochastic passes → mean + variance

→

BEV Grid

200×200 occupancy + epistemic uncertainty map