Video Stutter Detection via Frame Difference Analysis Using FFmpeg

In video quality assessment, detecting stutter is a crucial metric. This guide describes how to determine whether a video contains stutter by converting the video into a sequence of frames and analyzing the differences between consecutive frames.

The overall solution is divided into six parts: image processing, adjacent‑frame pixel calculation, motion amount aggregation, elimination of scene‑change influence, dynamic factor computation, and result output.

Image processing : The uploaded video is processed with ffmpeg to extract a grayscale image sequence resized to 360×640, focusing on a predefined region of interest to reduce noise.

Adjacent‑frame calculation : For each pair of consecutive frames (t and t+1), pixel differences are computed. A constant threshold of 30 distinguishes motion pixels (difference > 30) from static pixels.

Motion amount calculation : The differences are squared to convert amplitude to energy, and the average energy per frame is calculated, yielding the TI2 value for each frame.

Eliminate scene‑change ratio : Before averaging, a scene‑change proportion of 0.02 is removed, and the TI2 values are sorted to discard extreme noise values, producing a stable average TI2 (TI2_AVG).

Dynamic factor : The dynamic factor Dfact is computed as Dfact = a + b × log(TI2_AVG) with constants a = 2.5, b = 1.25, and limited by c = 0.1. Dfact is clamped to the range [0, 0.1].

Result output : Each frame’s TI2 is compared with Dfact × Mdrop (Mdrop = 0.015). If TI2 ≤ Dfact × Mdrop, the frame is marked as stutter (1); otherwise, it is marked as normal (0). The final output is a binary list indicating stutter presence for the entire video.

Advantages of this approach include: no need for large training datasets, robustness against varied dynamic/static scenes, and low computational overhead with high precision.