Abstract:

Accurate evaluation of scenario similarity is extremely important for optimizing test scenarios. However， existing trajectory-based evaluation methods fail to adequately capture the complex dynamic interaction characteristics between vehicles at intersections， which affects the accuracy of the evaluation results. To address this problem， in this study a directed graph-based similarity evaluation method for urban intersection scenes is proposed， which quantifies the similarity between scenes by comparing the degree of spatial and temporal matching of the global interaction topologies of vehicles in two scenarios. Firstly， a directed graph is used to characterize the interaction topology between vehicles at each urban intersection. Then， the interaction similarity between different intersection scenarios is estimated by comparing the degree of matching of their directed graph structures. Finally， a dynamic time warping algorithm is used to align the scenarios in the time dimension to effectively compare two test scenario sequences of different lengths. The results of the qualitative analysis of three pairs of typical evaluation cases demonstrate that the method is capable of distinguishing scenes with different similarity levels at a fine-grained level. Furthermore， to quantitatively validate the effectiveness of the method， an ANOVA experiment is conducted to compare scenario similarity with the performance of the autopilot system. The experimental results reveal that the safety and efficiency of the system exhibit significant differences under test conditions with different levels of scenario similarity， thus proving the method's effectiveness. Ultimately， this method is applied to optimize Apollo. Ultimately， this method rformance of the autopilot system. The experimental results reveal that the safety and efficiency of the system exhibit significantd

Key words: autonomous driving, urban intersection scenarios, similarity evaluation, directed graph, dynamic time warping