An Analysis of Machine Learning Classifiers for Joint Operator-UAV Monitoring and Abnormal Command Detection

Abstract

Unmanned aerial vehicles (UAVs) are widely used in search and rescue missions, which mostly involve stressful conditions for first responders and especially UAV operators, who have to control the UAVs under dynamically changing conditions and harsh environments. Operations under these conditions lead to increased stress and fatigue for the operators and may result in unwanted commands, which in turn may endanger the UAV's mission and result in accidents and delays, something often critical in search and rescue. However, jointly monitoring both the UAV and the operator's mental and physical status, we can potentially prevent such abnormal movements to be executed by the UAV. Wearable, non-invasive sensors and UAV's inertial measurement unit (IMU) sensor data, can be used in real-time to provide a holistic contextual awareness of the UAV and the operator. Through the use of machine learning and classification algorithms, we can identify such unwanted commands, enabling the UAV to ignore them. To this end, this paper presents an analysis and evaluation of the performance of a variety of machine learning classifiers using data from the IMU which is integrated in the UAV, and operator's biomarker data, collected and constructed in a realistic environment with a stress-induction technique. For each classifier, we analyze its performance in terms of accuracy, processing time and energy consumption, since the aim is to be deployed directly on the UAV. Our analysis suggests that a range of classifiers can be used in these situations and we also critically discuss trade-offs between accuracy, processing time and energy consumption.