From Posture to Performance: Continuous Cognitive Load Estimation of Drivers Using Multi-Modal Anatomical Keypoints

Driver distraction and cognitive overload are significant contributors to road accidents; however, most vision-based driver monitoring systems focus on classifying observable behaviors rather than estimating underlying cognitive workload. This study presents a non-contact, interpretable framework for continuous cognitive load (CL) estimation that relies solely on 2D pose, face, and hand keypoints extracted from realistic driving images. Using the Multi-Class Driver Behavior Image Dataset collected in Ashulia, Dhaka (2024), five driver behavior categories are mapped to normalized CL scores (ranging from 0 to 1), reframing the problem as a regression task to capture the continuous spectrum of cognitive effort. MediaPipe is utilized to extract body posture, head orientation, and hand-movement keypoints, which are then normalized and combined into a unified feature vector. The proposed Multi-task Learning (MTL) architecture, built upon a shared Multi-layer perceptron (MLP) backbone, simultaneously predicts discrete driver behavior and continuous CL using two task-specific output heads. The results show competitive predictive performances with Pearson of 0.58, MSE of 0.052, and R² = 0.34, too, meaning that predicted and ground-truth CL scores are very close. Analysis of the importance of features shows that the biggest predictors of cognitive effort are cues based on body position such as torso lean, head tilt, and hand placement. It generates continuous load distributions despite being trained with discrete behavioral labels while maintaining within-class variability. This work provides the first example of a privacy-preserving, low-cost and explainable alternative to intrusive physiological sensing by combining behavioral action recognition with continuous cognitive state estimation. To enable immediate integration into Advanced Driver Assistance Systems (ADAS), driver monitoring modules, and pipelines for detecting the fatigued or distracted driver, the method is compact and operates in real-time.