search

Breakthrough / DVR-Scan

:vhs: Tool for extracting scenes with motion from videos (e.g. security camera or DVR footage). Written in Python, uses OpenCV.
http://www.dvr-scan.com/
Other
362 stars 47 forks source link

Change default for --min-event-length from 2 frames to 0.1 seconds #91

Closed Breakthrough closed 1 year ago

Breakthrough commented 1 year ago

The default value for minimum event length is very low, especially for videos with a higher framerate. It should be updated to be defined in terms of time passed, rather than number of frames, to provide more consistency.

The current default value is 2 frames, which is roughly 66ms at 30 FPS, or 33ms at 60 FPS. At lower framerates, e.g. 5 FPS, this equates to 400ms. Initially I was planning on changing it to 50ms (0.05 seconds), but even this may be too low for most use cases, so I am considering raising the default to 0.1 seconds in the next release. This will be documented in the release notes, and folks can always set --min-event-length 2 to get the previous behaviour.

If anyone is using different values than the default currently, feedback on if the default is working for you (does it trigger too many false positives?), and the framerate of the videos you are processing, would be very helpful in ensuring the new default value will work for most users.

Thanks!

ceewanna commented 1 year ago

Release Notes

* Significant performance improvements on multicore systems

* Support wildcards/globs as inputs for scanning entire folders (`-i folder/*.mp4`)

* Allow use of ffmpeg for better output quality (`-m ffmpeg`) or codec-copying mode (`-m copy`)

* Configuration files are now supported, [see documentation for details](https://dvr-scan.readthedocs.io/en/v1.5/guide/config_file/)

  * Can specify config file path with `-c`/`--config`, or create a `dvr-scan.cfg` file in your user config folder

* Windows binaries are now signed, thanks [SignPath.io](https://signpath.io/) (certificate by [SignPath Foundation](https://signpath.org/))

* Experimental Nvidia CUDA® support has been added (set `-b MOG2_CUDA`)

  * If installing via `pip`, requires manual installation of OpenCV compiled with CUDA support
  * If downloading Windows version, make sure to download the GPU-enabled build (`dvr-scan-1.5-win64-cuda.zip`)
  * CUDA-enabled builds are not code signed, and do not include the `CNT` method

* Minimum supported Python version is now 3.7

* Minimum supported OpenCV version is now 3.x

Changelog

Command-Line Interface:

* New features/arguments (see below for more details):

  * `-c`/`--config` - set path to config file
  * `-d`/`--output-dir` - set directory to write output files (default is working directory)
  * `-m`/`--output-mode` - set output mode (one of: `opencv`, `ffmpeg`, `copy`)
  * `-mo`/`--mask-output` - path to write motion mask for analysis
  * `--verbosity` and `--logfile` - control output verbosity and path to save output

* `-i`/`--input` now supports globs/wildcards to scan entire folders, e.g. `-i folder/*.mp4`

* Change default value for `-l`/`--min-event-length` to 0.1 seconds, previously was 2 frames ([#91](https://github.com/Breakthrough/DVR-Scan/issues/91))

* Long form of `-roi` has been renamed to `--region-of-interest` (previously was `--rectangle-of-interest`)

* `-c` is now used for `--config`, previously was for `--codec`

* Add experimental `MOG2_CUDA` option for `-b`/`--bg-subtractor`

* Rename existing `MOG` option to `MOG2`

* `--codec` has been removed, the value should now be set using a [config file](guide/config_file.md)

General:

* [feature] Configuration file support and new `-c`/`--config` argument to specify path to config files ([#77](https://github.com/Breakthrough/DVR-Scan/issues/77))

  * Breaks existing behavior of `-c` (was previously the shortform of `--codec`)

* [feature] Add support for multiple output modes via `-m`/`--output-mode` argument ([#27](https://github.com/Breakthrough/DVR-Scan/issues/27), [#42](https://github.com/Breakthrough/DVR-Scan/issues/42))

* [feature] Experimental support for GPU-based CUDA MOG2 filter ([#12](https://github.com/Breakthrough/DVR-Scan/issues/12))

* [feature] Video encoding and decoding are now done in parallel with the scanning logic leading to improved performance on most systems ([#52](https://github.com/Breakthrough/DVR-Scan/issues/52))

* [feature] Add support for exporting motion masks via `-mo`/`--mask-output` argument

  * Useful for detailed analysis or tuning of detection parameters
  * ffmpeg can be used to generate output videos by specifying `-m ffmpeg`
  * Codec-copy mode, using ffmpeg, can be used by specifying `-m copy`

* [feature] Add `--verbosity` and `--logfile` arguments to provide more control over program output

* [feature] Allow scanning entire folders using wildcards with `-i`/`--input` ([#5](https://github.com/Breakthrough/DVR-Scan/issues/5))

  * Glob expansion is also performed on each input path directly, so quoted globs also function correctly

* [bugfix] Fix incorrect results when `-st`/`--start-time` is set

* [bugfix] Event start times are now correctly calculated when using `-fs`/`--frame-skip` ([#68](https://github.com/Breakthrough/DVR-Scan/issues/68), [#70](https://github.com/Breakthrough/DVR-Scan/issues/70))

  * Note that all skipped frames within the event window are included in motion event, thus the calculated start time may be slightly earlier

* [bugfix] Only get screen resolution when required (was causing issues on headless machines)

* [bugfix] Fix output messages conflicting with progress bar shown during scanning

* [bugfix] Output events now start from 1 to align with the event list

* [bugfix] Event end times now include the presentation duration of the last frame

* [bugfix] Small values for `-l`/`--min-event-length` are now handled correctly, previously would cause an error

* [enhancement] Progress bar now indicates how many events have been detected so far

* [enhancement] Change default value for `min_event_len` to 0.1 seconds, previously was 2 frames ([#91](https://github.com/Breakthrough/DVR-Scan/issues/91))

Known Issues

* Variable framerate videos (VFR) are not fully supported, and will yield incorrect timestamps ([#20](https://github.com/Breakthrough/DVR-Scan/issues/20))

* Video output when using frame skip and `-m opencv` (default output mode) will result in frames missing from the exported videos ([#81](https://github.com/Breakthrough/DVR-Scan/issues/81), can use `-m ffmpeg` or `-m copy` as a workaround)

* Multiple input videos are not supported yet when using `-m ffmpeg` or `-m copy` ([#86](https://github.com/Breakthrough/DVR-Scan/issues/86))

* CUDA builds do not include the `CNT` option for `-b`/`--bg-subtractor`

This discussion was created from the release DVR-Scan v1.5.

Do I get it correctly that there are 2 ways of setting this? The new one is "min_event_len" in seconds and the original "min-event-length" in frames.

My use case is I have to capture vehicle travelling at 50km/hr (13.9m/sec) across the entrance which is around 16m with 25frame per second camera. Right now some scan shows vehicle either disappears prematurely in the middle or appears right in the middle. I am playing around between 1 and 2 frames.

Breakthrough commented 1 year ago

There's still just one setting for this (min-event-length). This issue was referencing the default value. You could always specify minimum event length in frames or seconds, the default value was just changed from 2 frames to 0.1 seconds to work better for most use cases out of the box.