Neck Forces in Combat Jet Pilots - Machine Learner Algorithm versus Video analysis for head position counts in aerial combat manoeuvres

INTRODUCTION: Aerial combat manoeuvres involve frequent “head checks” above or behind the line of flight. Some head check positions are more physically demanding than others and may contribute to pilot neck pain. This study developed an automated system to interpret head-mounted inertial sensor data to reliably count each head check and classify the type.
METHODS: Rearward facing video footage of the pilot and head position co-ordinates relative to the instrument panel were obtained from helmet-mounted sensors during 42 F/A-18A sorties, involving 7 pilots. 6 common head postures were selected for classification- Neutral, Extension Hold, Check Right, Check Left, Quadrant Right, Quadrant Left. Four machine learner algorithms, logistic regression (LR), Naïve Bayes (NB), K Nearest Neighbour (kNN) and Adaptive Boosting (Adaboost) were trained to discriminate head motions based on head co-ordinates and labelled data of one 2 minute sample. The trained machine learners were then tested on unlabelled data.
2 human classifiers independently identified 127 random timepoints in a series of flight videos and classified the head posture, while blinded from each other’s results. Percent Exact Agreement (PEA) and Percent Close Agreement (PCA) were determined. The machine learning algorithm classification was compared to the human classifiers’ classification for these points.
RESULTS: Classification accuracy of the machine learners was: LR 2.0%, NB 3.2%, kNN 96.8%, Adaboost 100%. Human classifiers demonstrated 89% PEA and 99% PCA. Human classification demonstrated 59% PEA and 87.4% PCA with the Adaboost classification.
DISCUSSION: A machine classifier can reliably and accurately classify and count head checks from head-mounted inertial sensors. Human classifiers can also reliably classify head checks when reviewing video footage. Differences in agreement between human and machine are likely a combination of higher sensitivity of the machine learner using numeric variations in head co-ordinates as opposed to human classifiers responding to visual cues in footage prone to parallax error. Automation of this system will enable efficient workload monitoring, make technical review and training more feasible, and enable access to immediate post-flight summaries of workload.