Detecting Cognitive Load Through the Structure of Behaviour: An Ecological-Dynamical Approach

This thesis investigated how cognitive workload reshapes human behaviour, with two central aims: first, to evaluate the viability of marker-less pose estimation as a behavioural measure of workload; and second, to reframe overload not as a depletion of internal resources, but as a breakdown in coordination between person and task. Seventy-two participants completed a multitask flight-control simulation (OpenMATB) under systematically varied workload levels (low, moderate, high). A single fixed camera recorded facial and gaze motion using OpenPose, enabling high-resolution, sensor-free tracking. Standard performance metrics and NASA-TLX scores confirmed the workload manipulation, though effects varied: continuous control tasks degraded under load, while discrete task performance improved—suggesting that increased demand reorganised behaviour rather than uniformly impairing it. Pose-based metrics captured these shifts. Linear kinematics showed that movements under load became faster, while nonlinear recurrence analysis revealed increased fragmentation, volatility, and reduced temporal stability in motion. Blink patterns became more structured, pointing to selective adaptation. Cross-recurrence analysis showed that the typically tight coupling between eye and head movement weakened, indicating a disruption in integrated sensorimotor coordination. To assess whether these coordination patterns could reliably signal workload, random forest classifiers were trained on pose features. They distinguished workload levels with 93.7% accuracy within individuals, but generalisation across participants was limited—underscoring the idiosyncratic nature of behavioural adaptation under load. These results motivate a shift away from traditional capacity-based models of overload toward a coordination-based interpretation, where performance breakdown reflects a loss of functional alignment between individual and task demands. Accordingly, this thesis offers a critical synthesis of cognitive load theory and ecological-dynamical perspectives, culminating in a reconciliation that foregrounds behavioural organisation over resource depletion. Moreover, by validating a camera-only pipeline for workload detection based on pose dynamics, scalable to everyday interfaces without the need for wearable sensors, the research outcomes have practical implications for adaptive training, operator support, and human–machine collaboration in safety-critical environments.<p></p>