Each active sensor either can or can't "see" a given vehicle ahead: camera is a narrow forward cone, LiDAR is a 360° ring out to a fixed radius, radar is a long, narrow forward beam. Each sensor type has its own per-detection reliability pi; the perception stack fuses whichever sensors actually see an object by treating the misses as independent and combining them as a probabilistic OR:
Fusion confidence for one object:
C = 1 − Π (1 − p_i) over every sensor i that currently detects it
p_camera ≈ 0.90 (rich detail, narrow ~50° cone, short-mid range)
p_lidar ≈ 0.95 (precise depth, full 360°, mid range)
p_radar ≈ 0.75 (coarse, narrow beam, longest range, weather-robust)
- SAE autonomy level — 0–2 (no/partial automation) means the autopilot button stays locked and the driver is always in control; 3–5 (conditional/high/full automation) unlock hands-off steering, matching the real SAE J3016 scale.
- Traffic density — how many of the vehicle slots ahead are occupied at once.
- Sensor toggles — turning a sensor off removes its pi term from every object's fusion confidence, shrinking its detection zone in the scene and, if it was the last sensor covering a car, dropping that car from "Detected" entirely.
- Fusion confidence — the average C over all currently detected vehicles; low coverage (e.g. LiDAR-only in a narrow cone) pushes it down even when detections still happen.
Real-world relevance: production autonomous stacks (e.g. Tesla FSD, Waymo) combine exactly these complementary sensor modalities because each fails differently — camera struggles at night/glare, radar has poor angular resolution, LiDAR is costly and weather-sensitive — so fusing them raises reliability above any single sensor.