Dependable Person Detection using AI in Industrial Environments

Contemporary manufacturing facilities aim to enhance flexibility and efficiency while ensuring the safety of workers. The inclusion of autonomous machines can boost productivity even further, but it is vital to demonstrate their safe operation alongside human personnel. Conventional safety measures often restrict the operational capabilities of autonomous machines too much. To address this, perception systems that integrate Artificial Intelligence (AI) have emerged as a promising solution, but AI currently faces its own technical and legislative challenges regarding safety. Recently the automotive industry has pioneered efforts to develop AI specific standards, e.g. ISO/PAS 8800, which includes a detailed ML Safety Lifecycle emphasizing an iterative AI development process. This paper outlines the activities involved in developing a dependable person detection system within an industrial context. It highlights the individual steps of the ML Safety Lifecycle’s iterative approach to safety assurance. The approach encompasses defining safety requirements, data collection, algorithm selection, performance evaluation and mitigation strategies. Throughout the paper we provide details about each phase of the ML Safety Lifecycle followed by a lessons learnt part. For safety requirement elicitation and data collection, relevant norms and standards are listed. Regarding the algorithm selection, performance evaluation and potential mitigation strategies, we describe general considerations and procedures. Followed by illustrative examples from our own experience, this paper offers suggestions for ensuring safe person detection using AI in industrial environments.