[FFmpeg-user] videodetect filter? - equivalent to: https://ffmpeg.org/ffmpeg-filters.html#silencedetect

[Paul B Mahol]

On 4/3/16, Moritz Barsnick <barsnick at gmx.net> wrote:
> On Sat, Apr 02, 2016 at 17:02:19 -0400, Mark Hassman wrote:
>> Any chance someone knows a set of filters to accomplish the same thing for
>> video as: https://ffmpeg.org/ffmpeg-filters.html#silencedetect.. i.e.
>> timestamps output for start/end of motion.
>
> While silence would be more like darkness in video ;-) I get what you
> mean:
>
>> I've looked at: http://www.lavrsen.dk/foswiki/bin/view/Motion/WebHome, but
>> already transcoding with ffmpeg - hoping to add a simple filter to
>> accomplish the same.
>
> I too am looking for a filter which can at least imitate simple
> features of the tool "motion", as motion only does real-time detection
> from input devices. This would mean replicating "motion"'s algorithms
> in a filter, or using existing filters.
>
> I once tried, and had come to the conclusion - from reading the docs -
> that the decimate filter might be close to the right thing. While the
> docs mention "regular intervals", which wouldn't be appropriate, they
> also speak of "threshold for duplicate detection" and "detection of
> [small] movements". I have never gotten the decimate filter to work for
> anything I needed it for though, which went beyond motion detection.
>
> For anyone willing to help out, here are the main chaacteristics of
> "motion"'s features, some of which may not be feasible within ffmpeg:
> - Detection of motion by observing the number of changed pixels from
>   one frame to the next, and passing this detected frame on to the
>   output. (This could be a percentage or a mean square deviation
>   instead, but "motion" doesn't support that.)
> - Retaining a configurable amount of images before and after the
>   detected event.
> - Optionally visually marking the section of the video where the change
>   was detected. (This could be in frame metadata instead, but "motion"
>   doesn't support that.)
>
> At least the first point would be nice to have, but I'd rather not
> implement such a filter which reinvents the wheel.
>
> In other words: I too would appreciate a hint as how to use ffmpeg's
> filters to achieve this.

See singnalstats filter.