Training-Free Blink Detection with Histograms an Interpretable Approach

This study proposes a dual-strategy blink detection method based on grayscale intensity analysis. The first variant relies on global brightness thresholding, while the second employs localized eye-region tracking using the MediaPipe library in Python and derivative-based scoring over HMAsmoothed average intensity signals per frame. Despite its simplicity, the proposed method achieves robust and accurate performance, reaching an $F 1$ score of up to 1.00 under controlled conditions and 0.97 across a range of favorable scenarios. Notably, it achieves equivalent results to more complex, state-of-the-art techniques applied to the same dataset, maintaining a comparable performance of $0.90-0.91$ even in non-ideal environments. The novel integration of Hull Moving Average (HMA) filtering enables reliable detection without relying on data-driven learning or extensive computational resources. Unlike conventional methods that rely on geometric ratios or track landmarks across frames, the proposed method re-localizes the eye region independently in each frame. This frame-by-frame localization contributes to greater robustness against noise and variation, offering a lightweight yet effective direction in visual blink detection. As a whole, the system constitutes a lightweight visual inference architecture, assisted by minimal artificial intelligence (AI) and not requiring model training. It supports interpretable and intelligent decisionmaking without relying on learning or complex computations. It leverages the pretrained AI model MediaPipe exclusively for minimal eye region localization, using only two landmarks per eye, while the detection logic remains entirely deterministic and fully interpretable through explicit signal analysis.