Acquire Domain Boundary Consciousness to Promote Operational Design Domain Monitoring of Traffic Conditions

Shiming Liu, Qunli Zhang, Peng Wang, Wei Li, Naiming Deng, Zhengchen Dai, Zheng Hu

Abstract: As required by the Safety of the Intended Functionality (SOTIF) standard, to ensure the safe deployment of autonomous driving system (ADS), the system can only operate within its operational design domain (ODD) specified in the design phase. With safety requirements decomposed to a specific AI function module, operating within ODD can be fulfilled by ensuring the AI model to function exclusively in the environment akin to its training world. Thus, domain shift detection, or out-of-distribution (OOD) detection, of observations differentiating from the training dataset can be an effective runtime ODD monitoring approach. However, dynamic driving data, such as traffic conditions, are hard to model due to the additional temporal dimension and intricate interaction relationship between objects, compared to static environment data. In this paper, we employed an ODD monitoring module which consisted of a situational awareness model and an OOD detection algorithm. To promote the ODD monitoring performance, we proposed a novel training method incorporating supervised contrastive learning and near-ODD outlier exposed learning to improve ODD boundary consciousness of the situational awareness model. The experimental results both visually and quantitatively demonstrated the efficacy of our method, and the promising situational awareness modeling results for a specially designed intrinsic traffic condition revealed its potential of practical applications.