Complexity Quantification of Driving Scenarios with Dynamic Evolution Characteristics

Complexity is a key measure of driving scenario significance for scenario-based autonomous driving tests. However, current methods for quantifying scenario complexity primarily focus on static scenes rather than dynamic scenarios and fail to represent the dynamic evolution of scenarios. Autonomous vehicle performance may vary significantly across scenarios with different dynamic changes. This paper proposes the Dynamic Scenario Complexity Quantification (DSCQ) method for autonomous driving, which integrates the effects of the environment, road conditions, and dynamic entities in traffic on complexity. Additionally, it introduces Dynamic Effect Entropy to measure uncertainty arising from scenario evolution. Using the real-world DENSE dataset, we demonstrate that the proposed method more accurately quantifies real scenario complexity with dynamic evolution. Although certain scenes may appear less complex, their significant dynamic changes over time are captured by our proposed method but overlooked by conventional approaches. The correlation between scenario complexity and object detection algorithm performance further proves the effectiveness of the method. DSCQ quantifies driving scenario complexity across both spatial and temporal scales, filling the gap of existing methods that only consider spatial complexity. This approach shows the potential to enhance AV safety testing efficiency in varied and evolving scenarios.