Graph Structure-Based Implicit Risk Reasoning for Long-Tail Scenarios of Automated Driving
Xueke Li, Jiaxin Liu, Jun Li, Wenhao Yu, Zhong Liang Cao, Shaobo Qiu, Jia Sheng Hu, Hong Wang, Xiaohong Jiao · 2023
With the development of Artificial Intelligence (AI) technology, autonomous vehicles (AVs) have entered the general public's view, however, the challenges brought by "long-tail" scenarios are still faced in the real driving process. Since long-tail scenarios rarely exist in the training dataset, the current data-driven autonomous driving system is hard to handle the implicit risk of these scenarios, which seriously affects the safety of driving. To alleviate this situation, A reasoning methodology is proposed to recognize implicit risk supporting making safety decisions in long-tail scenarios. The method firstly constructs the knowledge leading to long-tailed risks into a knowledge graph structure with spatial relations, and manages the knowledge graph based on GloVe (Global Vectors for Word Representation) word vector embedding technique with hash table structure. Following, by designing a real-time scenario topology graph generation model, this work computes the similarity between real-time topology graph and knowledge map through graph structure and spatial relationship, to obtain the probability of occurrence of long-tailed risk. Finally, the feasibility of the solution is verified in Carla simulation environment by a long-tail scenario example, which indicates the approach succeeds in identifying the implicit risk. This research provides a path for enhancing risk management and safety decision making in autonomous driving, especially in addressing the challenges of long-tail scenarios.