Multiunit automotive perception framework: Synergy between AI and deterministic processing

Since neural networks were first introduced into automotive systems, safety has been a major concern. The prevailing safety standard in the automotive industry, ISO26262, does not fully define testing and verification methods for software based on deep learning. In this paper, we propose a multiunit perception framework that increases the determinism of automotive systems incorporating deep learning. Our approach relies on ASIL decomposition and algorithm diversification, which are enabled through the utilization of multiple low ASIL perception units and one high ASIL monitor unit. In addition to the framework concept, we specify how each component can be mapped to appropriate hardware and software platforms. The practical feasibility of the perception framework is demonstrated with a proof of concept prototype.