Implicit Hand Gestures in Aeronautics Cockpit as a Cue for Crew State and Workload Inference

Implicit Gestures, Human Pose Estimation, Random Forest, Aeronautics Cockpit, Visual Recognition

This paper aims at improving advanced aeronautic cockpit by raising its awareness of the crew's state and workload level. Our approach is based on visual analysis of pilot's upper body movements. We define the term of "implicit gestures" and further observe its subclasses. We collected a simulator dataset of practical implicit gestures, annotated semi-automatically a dataset for Human pose estimation training, and we offer these datasets for public use. Based on experiments on this data, we propose a method for recognition of implicit gestures - full interactions, touch-and-go interactions, and unfinished gestures. Our approach is purely visual (no depth data, which are hardly usable in the cockpit environment due to regulations). This method is based on human pose estimation by a hierarchical approach named Pose machine whose subsampled output is used for recognition of implicit gesture presence from sequences of frames by random forest. The experiments show that the classification works reliably and the method is able to recognize these implicit gestures in the cockpit.