Runtime Anomaly Monitoring of Human Perception Models for Robotic Systems
Hariharan Arunachalam, Zhiyong Huang, Marc Hanheide, Leonardo Guevara · 2024
In recent years, camera-based object detection models have been highly used to allow robots to perceive and understand what is in their surroundings. Although modern detection models have been demonstrated to be reliable solutions in certain controlled environments, their performance can be reduced when applied to outdoor robotic applications where robot captures images previously unseen or where the image quality is affected by different lighting/weather conditions. In this context, this paper presents a runtime and modular monitoring framework that uses the output from any generic camera-based detection model to quantify the deviation of the model predictions with respect to the actual object locations. Moreover, in the critical scenario that the detector fails to detect the expected object on a given image, the framework can determine if it is a False Negative or True Negative detection. Those features make this framework especially suitable for improving the safety of robotic applications where humans and robots interact or share the workspace. Thus, in this paper, we tested the proposed framework with different human detection models considering simulation and real data, demonstrating that the framework: 1) can accurately estimate deviations from detector predictions, 2) can detect False Negative instances from detector predictions, 3) can be used with any state-of-the-art human detection model, 4) can be executed online due to the lightweight of its components, 5) can be easily integrated with any ROS-compatible robotic system.