Reliability-Aware BEV Fusion for Lightweight Autonomous Driving Perception

Multi-sensor perception is central to autonomous driving because cameras, LiDAR, and radar fail in different ways. Recent bird's-eye-view (BEV) fusion models have improved 3D detection and semantic scene understanding by projecting heterogeneous observations into a shared spatial representation, but many fusion pipelines still treat each modality as uniformly reliable within a frame. This paper proposes a lightweight reliability-aware BEV fusion module that estimates per-modality and per-region confidence before feature aggregation. The module can be attached to existing BEV perception backbones and is designed to down-weight degraded camera, LiDAR, or radar evidence without requiring a new end-to-end driving stack. We formulate the method as a three-stage pipeline: BEV feature extraction, reliability estimation from measurement consistency and feature uncertainty, and gated fusion for occupancy and object-level outputs. The intended evaluation uses nuScenes, Waymo Open Dataset, and Argoverse-style sensor logs under ordinary and degraded sensing conditions. The expected contribution is not a new perception foundation model, but a compact safety-facing fusion layer that makes downstream planning inputs more interpretable and less brittle when one modality becomes unreliable.