Knowledge Graph-Based Long-Tail Event Detection in Autonomous Driving Systems

Long-tail events make up less than 5% of driving scenarios but cause over 60% of autonomous driving failures. These rare cases present high safety risks. Existing detection methods remain limited: rule-based approaches lack coverage, statistical models overlook multi-dimensional features, and deep learning suffers from data scarcity and poor interpretability. This paper proposes a knowledge graph-based framework for long-tail event detection. It has three key contributions. First, a hierarchical spatiotemporal knowledge graph organizes multimodal sensor data and entity relations. Second, an LDA-optimized multi-dimensional fusion method combines behavioral, spatial, and safety features. Third, a triple detection mechanism integrates pattern recognition, rule validation, and context analysis. On the nuScenes dataset, the method achieves 85.6% precision, 79.3% recall, and 82.3% F1-score. This is a 3.4% improvement over state-of-the-art methods, while keeping 22.8 ms real-time inference latency. The results show strong generalization across event categories and confirm the framework’s practical value in safety-critical autonomous driving.