Runtime operational design domain monitoring of static road geometry for automated vehicles

Shiming Liu, Qunli Zhang, Wei Li, Siyun Yao, Yucheng Mu, Zheng Hu

Abstract: With the rapid advancement of transformer-based neural networks, especially for those using an enormous number of parameters, large-scale AI models have recently attracted extensive attention in autonomous driving system (ADS) development. However, due to the critical safety requirements of ADS, of great indispensability is the role an effective runtime monitoring method plays to assure the Safety of the Intended Functionality (SOTIF) of automated vehicles. In this paper, we proposed a runtime operational design domain (ODD) monitoring framework to prevent ADS decision and planning network from operating in conditions beyond ODD, which incorporates a situational awareness model and a feature-distance based excess-of-ODD detection algorithm. This framework is designed structurally parallel to the ADS and thus can be readily deployed at runtime phase without deteriorating original ADS network performance. The experimental results have corroborated its ODD monitoring efficacy, consistency, generalizability and acuity.