Reliability Assessment and Adaptive Fusion Algorithm for Multi-Sensor Data in Autonomous Driving under Adverse Weather Conditions

Abstract

Atmospheric disturbances impose systematic degradation on multi-sensor perception systems in autonomous vehicles, necessitating a fundamental rethinking of sensor fusion strategies. This study presents a comprehensive reliability assessment framework combined with an adaptive fusion algorithm designed to mitigate sensor-specific performance deterioration under adverse meteorological conditions. Using an empirical dataset encompassing 10,000 hours of vehicle operation, we establish quantitative correlations between atmospheric parameters and measurement uncertainty across heterogeneous sensor modalities. Real-time trustworthiness estimation is achieved through a dynamic scoring mechanism that integrates environmental context with temporal performance evolution. The proposed adaptive fusion algorithm performs reliability-weighted integration through probabilistic decision modeling, optimizing sensor data combination while minimizing perception errors. Experimental validation demonstrates that the method improves object detection accuracy by 22.9 percentage points compared with majority-vote fusion (AP@0.5) and by 18.4 percentage points compared with fixed-weight fusion in heavy rain (AP@0.5). Furthermore, the false positive rate is reduced by 66-73% relative to the fixed-weight and majority-vote baselines, respectively. These improvements directly enhance collision avoidance performance, substantially advancing the safety and robustness of autonomous vehicles operating in challenging environmental conditions.