Depth sensor technology in facial movement analysis: A comparative evaluation with marker-based motion analysis

Abstract

Objective

This study evaluates the reliability and agreement of depth sensor technology compared to marker-based motion analysis for facial movement assessment. Depth sensors, such as the Kinect-V2, offer a non-invasive alternative, but their accuracy in facial kinematics remains uncertain.

Method

100 healthy participants (50 male, 50 female) performed six facial movements— opening −mouth, smiling, eyebrow-lifting, forced-eye-closure, whistling, and frowning. These were recorded simultaneously using a marker-based motion system and a Kinect-V2 depth sensor. Data were analyzed for asymmetry, intra-method reliability using intraclass correlation coefficients (ICC), and agreement via Bland-Altman analysis.

Results

Bland-Altman analysis showed mean biases for facial movements: opening-mouth (−0.99), smiling (2.7), eyebrow-lifting (−1.85), forced-eye-closure (−1.77), whistling (11.59), and frowning (20.82). Mean asymmetry values using the marker-based system vs. depth sensor: smiling (8.16%vs.4.22%), eyebrow-lifting (7.32%vs.6.88%), eye-closure (8.42%vs.5.39%), and frowning (11.50vs.13.86%). ICC values ranged from 0.41 (forced-eye-closure) to 0.80 (eyebrow lifting) for the marker-based system and 0.61 (forced-eye-closure) to 0.85 (mouth opening) for the depth sensor.

Conclusions

While depth sensors show strong intra-method reliability, they demonstrate biases and broader limits of agreement for subtle expressions. Further algorithmic improvements are needed for clinical applications.