0.10.0 Beta 12

[blakeblackshear]

Major Changes for 0.10.0

Updated docs are available for preview here.

Changes since beta 11

• New max_frames setting to allow per object customization of stationary tracking
• BREAKING CHANGE refactored stationary settings (previously stationary_threshold and stationary_interval) under a stationary section

detect:
  ...
  # Optional: Configuration for stationary object tracking
  stationary:
    # Optional: Frequency for running detection on stationary objects (default: shown below)
    # When set to 0, object detection will never be run on stationary objects. If set to 10, 
    # it will be run on every 10th frame.
    interval: 0
    # Optional: Number of frames without a position change for an object to be considered 
    # stationary (default: 10x the frame rate or 10s)
    threshold: 50
    # Optional: Define a maximum number of frames for tracking a stationary object (default: not set, track forever)
    # This can help with false positives for objects that should only be stationary for a limited amount of time.
    # It can also be used to disable stationary object tracking. For example, you may want to set a value for person,
    # but leave car at the default.
    max_frames:
      # Optional: Default for all object types (default: not set, track forever)
      default: 3000
      # Optional: Object specific values
      objects:
        person: 1000

Changes since beta 10

• Default for stationary_threshold is now 10x the FPS or ~10s
• Update birdseye to only show non-stationary objects
• Recordings link is always visible in the UI even when record is disabled

BREAKING CHANGES

Camera Names

Camera names in the config file must be URL safe starting with this version. Only A-Z, a-z, 0-9, -, and _ are permitted. If your cameras need to be renamed, any associated events will be retained based on your global retain settings, not the new camera name settings. If you want to retain recordings from the previous camera name, you need to ensure you have retain days set globally. For example:

record: #  <---------- global setting
  retain:
    days: 7 # <------- will retain 7 days of recordings for old cameras

Retain modes

This release adds the ability to define retain modes for the initial retention period and for event data beyond the initial retain period. There are three modes:

• all: retain all recording segments. This is consistent with the previous behavior
• motion: retain only the recording segments that have some motion
• active_objects: retain only the recording segments that have a moving true positive object that is being tracked

The default record settings will retain only segments that have motion for events.

Note that the retain_days config entry has been DEPRECATED. This release includes backwards compatibility, but you should migrate to the new format.

Example: Store all segments with motion for 7 days and events with active objects for 14 days.

record:
  enabled: True
  retain:
    days: 7
    mode: motion
  events:
    retain:
      default: 14
      mode: active_objects

Stationary objects

Frigate will now continue to track objects after they are stationary. This enables use cases such as presence detection. After an object has been stationary for 10 frames, object detection for the object will be paused (the new stationary -> interval setting allows periodic checks too) until there is new movement. An object is determined to be stationary if the bounding box in the current frame overlaps almost entirely with the previous 10 frames. I have been testing this in my home office, and it works quite well.

I expect this to disrupt the way a few people are using Frigate. For example, a car that enters the driveway, parks, and then leaves 5 hours later will be one continuous event. The zones will stay with the tracked object over this lifetime, so when the car leaves, it will still have zones associated from when it first arrived. Ensure you use the right retain mode to ensure you only capture the segments you want.

Depending on the specifics of your room, you may need to tune your resolution and motion settings to get the results you want. In my office, I reduced the resolution to 640x360 and increased the contour_area setting for motion detection up to 125 to reduce the sensitivity. I recommend watching the debug view and adjusting as needed.

The following settings have been added to tune stationary tracking:

detect:
  ...
  # Optional: Configuration for stationary object tracking
  stationary:
    # Optional: Frequency for running detection on stationary objects (default: shown below)
    # When set to 0, object detection will never be run on stationary objects. If set to 10, 
    # it will be run on every 10th frame.
    interval: 0
    # Optional: Number of frames without a position change for an object to be considered 
    #                 stationary (default: 10x the frame rate or 10s)
    threshold: 50
    # Optional: Define a maximum number of frames for tracking a stationary object (default: not set, track forever)
    # This can help with false positives for objects that should only be stationary for a limited amount of time.
    # It can also be used to disable stationary object tracking. For example, you may want to set a value for person,
    # but leave car at the default.
    max_frames:
      # Optional: Default for all object types (default: not set, track forever)
      default: 3000
      # Optional: Object specific values
      objects:
        person: 1000

fixes #841

In progress events and cache handling

Events are now stored in the database while in progress, so you can expect to see In progress for some events under end time in the Events view.

Recording segments written to the cache are now only moved to the recordings folder if they overlap with an event (in-progress events too). For those with retain_days set to 0, this eliminates disk I/O when there are no events. Frigate will no longer store a minimum of 5 minutes of recording segments for each camera, so now you can enable recording without worrying about SD/SSD wear.

Lastly, more proactive cache maintenance was implemented to prevent the cache disk from filling up. You should never see more than 7 segments in the cache for each camera. If the cache maintenance job is too slow, you will get a warning in the logs. You can also see the time it takes to copy to the recordings disk by setting frigate.record: debug in the logger config. Additional error handling was also implemented to prevent the cache maintenance thread from stopping on errors.

Now that recordings are managed proactively, recordings are expired according to the retention settings at the interval set in the expire_interval setting (defaults to once per hour).

fixes #1885

Motion detection improvements

0.9.0 introduced dynamic motion sensitivity which didn't work out well. For many users, this made motion detection too sensitive which increased CPU usage due to more motion causing more object detection. It also increased false positive detections due to the increased frequency for detections on areas without substantial motion.

In this release, motion detection has been reset back to better defaults for frame_height and contour_area. If you have customized either frame_height or contour_area in the past, it is recommended that you remove those for your camera and let them revert back to the new defaults. For most users, the sensitivity can be tuned by adjusting contour_area (smaller numbers are more sensitive and larger numbers are less sensitive). For very high resolutions with very small objects, the frame_height may need to be increased.

In addition, this release brings about dynamic contrast adjustment for motion detection. Based on my testing, this substantially improves night time detection with IR lights and other low contrast situations.

Other changes

• Improved logic for merging motion, bounding boxes, and regions
• Report coral temperature stats when available (USB Coral not supported)
• Limit the cache time for playlist urls to avoid recording playback failures
• Allow yaml extension in addition to yml
• Sync database recordings with the disk on startup
• Store motion and objects data in database
• Entered zones in mqtt messages are now sorted based on the order the object entered zones
• Frigate now scans for objects in the frame on startup
• Detected objects that never move no longer generate events or snapshots
• Make recording expiration interval configurable (previously hard coded to 60 minutes)
• New event datepicker in the events viewer
• Require camera names to be url safe
• Ensure clean snapshots are deleted when deleting events
• Fix bug when last recording in the database does not exist on disk
• Randomize the region slightly to try and reduce lost true positives
• Allow recording motion even when detect is disabled
• Add stacktrace to config errors
• Tweak error messages when ffmpeg fails
• Update has_clip to false when an event fails to play
• Handle rare divide by zero in fps calculation

Fixes

• fixes [Support]: Recording breaks if disabled by MQTT on Frigate start #2129
• fixes Make wait_time for recording cleanup configurable #2734

Full Changelog: v0.9.4...v0.10.0-beta12

Images:

• x86: blakeblackshear/frigate:0.10.0-beta12-amd64
• x86 w/ nvidia (for ffmpeg only): blakeblackshear/frigate:0.10.0-beta12-amd64nvidia
• Raspberry Pi 3/4 (32bit): blakeblackshear/frigate:0.10.0-beta12-armv7
• Raspberry Pi 4 (64bit): blakeblackshear/frigate:0.10.0-beta12-aarch64

[NickM-27]

Nice! It's been working really well for me 💯

RC soon if no issues or some changes still coming?

[blakeblackshear]

If I don't see any issues here, I plan to publish an RC tomorrow. We went through so many iterations on the beta, I'm hoping to get to an official release the week after.

[blakeblackshear]

Based on the comment below, it looks like another beta will be coming. I think some of the proposed changes there will help avoid a flood of new issues.

[Firerouge]

I'm having an issue upgrading from beta 10 to beta 11, beyond the marked increase in false positives I saw introduced around beta 10.

Where in beta 6 I used to get one camera every couple of nights getting a false positive that needed manual clearing, the last two nights of beta 10 have had two of six cameras get stuck with persistent false positives, something that had never happened prior.

When upgrading from beta 10 to 11, two cameras had false positives occurring on them at the time of the container transition. I have since been unable to clear the false positive on one of those two cameras. Even after turning off detect on the camera in frigate, home assistant continues to show the sensor as detected where it would usually clear it. Restarting MQTT and the container didn't help, went briefly unavailable before returning to detected.

[Firerouge]

Configured interval: 10 w/ 5000 max person frames. Watched as a 73.44% score false positive corrected itself in 4 minutes and 18 seconds.

The stationary interval setting only takes effect after an object has been detected and started tracking right, and someone should expect lowering the interval to also lower the length of false positives?

[blakeblackshear]

By default, Frigate stops scanning stationary objects completely. Setting stationary interval to 10 tells Frigate to force detection to run and confirm that the object is still there once every 10 frames. If you have a false positive with a high score, it's probably not going to help at all.

5000 max frames for person tells Frigate to stop tracking person objects after they have been in the same position for 5000 frames after meeting the threshold. This should prevent false positives from being tracked forever because they probably won't change positions.

None of the settings in stationary are really designed to reduce false positives. I would focus on using filters such as min_area and object masks to address false positives.

[elgorro]

Switched from motion: mask: to object: 'object': mask: and it's working like expected.
Edit: So if a car move on the nearby road the bluebox can perisist inside masks - then cars on the road trigger the detect! And when the move in mask - other statonary objects being reseted from statonary detection...

[NickM-27]

@elgorro Blue box means that the object was lost / isn't being tracked. That would make sense in the mask given that the mask keeps it from being tracked in that area. The other part I am not sure it is clear what you are saying happened. Screenshots / videos showing zones and masks would be helpful 👍

[elgorro]

Thx for taking care - I can send in a private chat.

[blakeblackshear]

Just pushed Beta 12 with the new max_frames config option.

[blakeblackshear]

Released RC1. Locking this discussion.