Safe object detection in AMRs - a Survey

Autonomous mobile robots (AMRs) are frequently used in indoor applications for efficient and flexible transportation of materials and products. Functional safety eliminates unacceptable risk of physical injury and has to be included in AMRs that operate in proximity to persons. The AMRs are equipped with safe laser scanners, that are rated by their probability of a dangerous failure in the standard ISO 13849-1. Their safety rating applies only for detecting objects in predefined zones, without any classification. To increase the field of application of AMRs, they can also be used in outdoor applications in public environments for transportation of meals, groceries, parcels and medicine, for the exploration of dangerous regions, or search and rescue missions. Because of the necessity to traverse obstacles like curbs and snow, a classification of objects in their path is needed. Since persons have to be distinguished from objects that will be traversed, the object classification has to have a comparable performance as conventional functional safety. Common metrics of neural network based detectors like the mAP are not suited for the evaluation of safety functions because they are not linked to dangerous failures. We survey the state of research on object detection and classification and compare their performance with functional safety requirements. We discuss the opened research questions on the path to achieving outdoor AMRs with similar safety performance as industrial AMRs in indoor applications. We therefore propose metrics to validate object detectors for safety applications and propose a concept for a three phase validation of safe neural network based pedestrian detectors for outdoor AMRs.