A Fast Search Method for Edge Hazardous Scenarios Based on Semi-Supervised Anomaly Detection

Finding edge hazardous scenarios which appear very infrequently in the dataset than common hazardous scenarios is essential for implementing scenario-based testing of autonomous driving systems(ADs). However, it is difficult to evaluate the rarity of dynamic scenarios with huge scenario space high-dimensional time series, making it difficult to search for edge hazardous scenarios quickly. To solve this problem, this paper proposes a Semi-supervised anomaly detection method combining MiniRocket and DAGMM(Semi-MiniRocket-GMM, SRG), which treats edge hazardous scenarios as anomalous samples of common hazardous scenarios. SRG uses a small number of samples of common hazardous scenarios to guide interpretable feature extraction and clustering of a large amount of high-dimensional unlabeled temporal data and finds rarer edge hazardous scenarios based on anomaly evaluation to improve the coverage of test scenarios. The method is validated in the open-source natural driving dataset HighD. Compared with DAGMM, the SRG method can find edge hazardous lane change scenarios more quickly and accurately with a few samples of hazardous scenarios. The SRG method aimed at discovering edge hazardous scenarios can both guide the direction of generating scenarios and speed up the testing process.