Assuring the Safety of Machine Learning for Pedestrian Detection at Crossings

Abstract

Machine Learnt Models (MLMs) are now commonly used in self-driving cars, particularly for tasks such as object detection and classification within the perception pipeline. The failure of such models to perform as intended could lead to hazardous events such as failing to stop for a pedestrian at a crossing. It is therefore crucial that the safety of the MLM can be proactively assured and should be driven by explicit and concrete safety requirements. In our previous work, we defined a process that integrates the development and assurance activities for MLMs within safety-related systems. This is used to incrementally generate the safety argument and evidence. In this paper, we apply the approach to pedestrian detection at crossings and provide an evaluation using the publicly available JAAD data set. In particular, we focus on the elicitation and analysis of ML safety requirements and how such requirements should drive the assurance activities within the data management and model learning phases. We explain the benefits of the approach and identify outstanding challenges in the context of self-driving cars.