Risk-aware Real-time Object Detection

Autonomous systems such as self-driving cars and infrastructure inspection robots must be able to mitigate risk by dependably detecting entities that represent factors of risk in their environment (e.g., humans and obstacles). Nevertheless, current machine learning (ML) techniques for real-time object detection disregard risk factors in their training and verification. As such, they produce ML models that place equal emphasis on the correct detection of all classes of objects of interest—including, for instance, buses and cats in a self-driving scenario. To address this limitation of existing solutions, this short paper introduces a work-in-progress method for the development of risk-aware ML ensembles for real-time object detection. Our new method supports the dependable use of real-time object detection in autonomous systems by (i) identifying the risks that require treatment, (ii) training a set of ML models that mitigate these risks, and (iii) using multi-objective genetic algorithms to combine the ML models into risk-aware ML ensembles. We present preliminary experiments that show the effectiveness of our method at constructing a dependable ML ensemble for real-time object detection in a simulated self-driving case study.