Cam-Frigate-HA-WaterSprinkler ( Protect Bird Feeding )

[d1etmar]

Servus,

First i am a very beginner in CAM's and frigate and a starter in HA . I have installed Frigate as an lxc Container within Proxmox on an Terramaster F8 with M2 nvme as disks . (+Coral M2 ) I bought a single Reolink DUO 3 and installed it in my garden . Still some 2 way Audio issues ( guess ssl setup ) but that is not so important atm. (HA runs on seperate HW )

What i want to try with a new CAM:

• Catch a CAM Event when a cat/dog ( greater than bird , but smaler than human ) is near my birth feeding place and start the Water Sprinkler ( Currently i thought to use the SONOFF zigbee for water on off ) and any sprinkler which could fit .

I would need some help/advice how to get this done gg ( which CAM , HA Automations, an intelligent sprinkler ( direction) if avail ? )

• CAM would be mounted on House Wall ( POE ) and would need appr. 180C FOV . ( dual or moving cam ? )
• Could the cam send the position of the object as well, or only on or off ? Would it even make sense ?

Anything on this matter would be great to hear .

Best regards from Vienna, pipi.feldsperling

[NickM-27]

You can create zones so that you only run this when a cat / dog enters the area of interest, this is a pretty common automation to run

[Matzebhv]

Moin d1etmar,
because of this scenario i am a frigate user since 4 years or so. Defending the fishes in our pond against cats and herons since then with frigate and iobroker. This works great using a zigbee-valve and a gardena impulse sprinkler. Firing up some light and sound also. Since frigate 14 i send my old qvr-pro into pension, since then running 12 reolink-cams with frigate only.
Regards
Matze

[d1etmar]

Moin Matze,

Did you anything special concerning cat regognition ? Which CAM's ( i guess no PTZ ? ) . Anything special on detection ? scale Resolution ? I ordered a HFW3849T1-AS-PV S5 as well now ( no Pro as AI is not needed i think ) . Hopeing this could be a alternative . This seems to be able to customize the Stream resolutions and codec freely . As you said you use sound maybe "Ultraschall ? " and is this working and if yes which device you use ?

Grüße in den Norden und besten Dank :) lg, Dietmar

[Matzebhv]

Moin,
cat detection is an easy task and works very well even with the standard modell. My cameras are all non ptz as it makes no sense with zones. You can use them, but they are on a fixed preset all the time, so save the money. The detection depends on how far away from your object the camera is. Frigate must first detect the motion and then the object, so if your camera is to far away it can happen that a cat does not trigger motion. Depending on that a camera with zoom could make sense, so you can zoom in nearer to the point of interrest. My reolinks are integrated into iobroker, so i can fire up the leds at night and the buildin alarmsound. Eventually there is an HA integration for dahua available. I have 4 cameras in our backyard, 3 of them have parts of the pond in sight and one of them is installed directly at th pond for nearfield detection. On all cameras i have configured a zone "pond" and if a cat or a heron ( bird ) enters this zone the alarm will be triggerd. That works very well since years, and even better with frigate+.
Here you can see the zonesetting from my nearfield camera.

Grüße vermutlich in den Süden
Matthias

[d1etmar]

Servus Matthias,

Sehr schöner Teich Gratuliere !

Thx for your updates . I will wait for the other CAM to arrive and will look into some different usage for the PTZ . Will start with one, but i will end up also needing more CAM's . And i will place 2 POE Cables , as i am not able to FOV all sides with 1 CAM. And i think with Zones having 2 CAM's is better than a dual like the DUO 3 .

So you added some custom Model for frigate+ with own pictures to get better results as well ? Thats not planned , but if there is a "heavy" need for it , i will go for it. But i have a slow CPU/GPU (N305)- if all works with Coral it is fine . Actually i do not want to replace the hardware for Frigate as this is brand new . But one step after the next gg .

I think the frigate integration for HA is fine enough and there is no need for any vendor integration into HA . It really works great and all works with Smartphone Notifications BluePrint as well. So running an automation to trigger a zigbee device should be no problem at all.

I have frigate and the CAM only for a few weeks, but i really love it .

New CAM will take 2 to 3 weeks from ali, it was a LOT cheaper .

will update ofc any results .

Das ist meiner ( ohne Fische ) .

Person traking works 95% correct . But to far away for anything smaller i think .

lg aus der grünen Steiermark, Dietmar

[Matzebhv]

Moin Dietmar,
mit deinem Teich hast du mich jetzt nur ein wenig neidisch gemacht :-)

I have ended up with 12 cams. Not only because the pound but we had a burglary years ago. Whe have chickens also, so its interresting to see whats up around the house. 3 cams are inside, the rest is located around the front and backyard.
You should look which camera vendor works best for you and if you need to expand buy the same modell or brand again. Had a bunch off cheap chinese cameras before, they all worked but fiddeling with the settings of any of these cameras sucked hard.
Frigate+ is no need for your task. As i dedect not only cats i gave it a try and it works great. And because frigate gave me so much benefit i also wanted to support this great project.
The dahua HA integration could be helpfull if you want to use the camera buildin leds or speaker to play alarm sounds and lights. Cats ignores sound and light, herons mostly fly away by playing sound only.
CPU/GPU usage depends on many things, but you have tons of options to tweak this. In my case frigate runs as docker on my qnap nas with coral tpu and utilize ~50% cpu ( 8 core ryzen ) and 5% gpu ( rtx 1080 ).
And yes, please report your results.
Grüße aus dem ganz hohen Norden
Matthias

EDIT: Snapshot der Katzenerkennung vergessen.

[d1etmar]

Servus Matthias ,

Danke für die Bilder.

Got the POE Cabling done and integrated the new CAM's into frigate. ( no PTZ Config ) .

The CAM's did run now for about 24h Hours, and detected 2 Dogs ( actually Cats but that is fine ) , but also about 5 times Birds which where actually CATS. And arround 50 Bird detections . I need to get much less Bird detection, but i do not know/think this will help with the currently wrong Bird detection.

Just change a lot currently ofc in this trial and error mode :) . I used 264H Codec wiht max Resolution supported on both streams .

I have not changed anything on Threshold or Sizes oder whatever might help ....

As i am to stupid to post a config.yaml formated ( i tried and it never formats ok with the "code" buttons ) ....

go2rtc:
streams:
BIRD1:
- rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0
BIRD1_sub:
- rtsp://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=2
cameras:
BIRD1:
ffmpeg:
inputs:
- path: rtsp://192.168.2.156:8554/BIRD1
roles:
- record
- path: rtsp://192.168.2.156:8554/BIRD1_sub
roles:
- detect
output_args:
record: preset-record-generic-audio-aac
record:
enabled: true
retain:
days: 5 # how many days to keep recordings
mode: all # use "all" for 24/7 recording, or "motion" for event-based
objects:
track:
- person
- dog
- cat
- bird

So i think it is now time to go more into frigates dedection and how it works ...

Thx, Dietmar

[d1etmar]

Sadly today i have first catch of a "Dog"(cat) on motion , catching one of my beloved sparrow Birds.

Yesterday night i have read about the detection fps part ( default set to 5 ) and that if you have fast moving Objects it could be increased to 10 fps . So i did this . Also i changed the Dahua CAM Sub Stream to 10 FPS ( it was 25 ) , and also the iframes to 10 . With the point that the CPU/Coral goes a bit up (+2 - total now 8% so np ). Does this makes sense if the frigate HW could handle it ? Would be sending from the CAM with 25 fps be even better, or does this only consume resources and does not help at all ?

The point is that the object Lifecylce looks somehow strange . And as all is rather brand new, and as i do more changes i have no idea if this would be the same without changing the detection fps ...

thx, Dietmar

[d1etmar]

Hi,
This is the Codec/CAM Config of the dahua ( Stream 2 for the recording ) .

Do i need to set the resolution,fps under the detect in the config.yaml ?

I also found something like this

objects:
filters:
dog:
min_score: .7
threshold: .9
cat:
min_score: .65
threshold: .8

Is this only for frigate+ ?

[NickM-27]

Do they differ from the "normal" base Models shipped with frigate ?

yes, they are trained on Frigate+ submissions as opposed to the COCO dataset

to use it you copy the ID and enter it into your config https://docs.frigate.video/plus/first_model#step-3-set-your-model-id-in-the-config

[d1etmar]

I have added Model ID zu the configuration and set the object size/threshhold as suggested. Actually i have not found any info in logs or UI which model is used now ?? . Actually a bitty .

model:
path: plus://exxxxxxxxxxxxxxxxxxx

Let's see how this works . Thx .

I made also some other tests to use openvino and therefore without the Coral CPU. Therefore i needed the GPU drivers and map those as well. (/dev/dri - lxc,docker) vainfo and the other stuff works ( N305 Gen 12 Intel ) . But actually if frigate finds driver the Reolink DUO 3 started to fail with ffmpeg errors . I tried various stuff, but actually it only works if i set this CAM (h265) to version 5.0 instead of the default 7. The idea for the test behind is if there is a need of Models which are not supported by Coral, and my backend could handle it ..
Did not had this problem before , because i had no GPU mapping in lxc and docker assigned , and the ffmpeg 7 Version seems to be fine with the h265 from the REOLINK DUO 3 as it was done by CPU i guess ....

Even it looked to me that all worked . Those Errors are reported .

2025-08-06 11:55:39.302021773 [2025-08-06 11:55:39] watchdog.GARTEN ERROR : Ffmpeg process crashed unexpectedly for GARTEN.
2025-08-06 11:55:39.302045624 [2025-08-06 11:55:39] watchdog.GARTEN ERROR : The following ffmpeg logs include the last 100 lines prior to exit.
2025-08-06 11:55:39.302204318 [2025-08-06 11:55:39] ffmpeg.GARTEN.detect ERROR : [AVHWFramesContext @ 0x76fd6006ca00] Failed to sync surface 0x13: 1 (operation failed).
2025-08-06 11:55:39.302207149 [2025-08-06 11:55:39] ffmpeg.GARTEN.detect ERROR : [hwdownload @ 0x76fd64003980] Failed to download frame: -5.
2025-08-06 11:55:39.302208456 [2025-08-06 11:55:39] ffmpeg.GARTEN.detect ERROR : [vf#0:0 @ 0x6227412ddfc0] Error while filtering: Input/output error
2025-08-06 11:55:39.302211799 [2025-08-06 11:55:39] ffmpeg.GARTEN.detect ERROR : [vf#0:0 @ 0x6227412ddfc0] Task finished with error code: -5 (Input/output error)
2025-08-06 11:55:39.302227758 [2025-08-06 11:55:39] ffmpeg.GARTEN.detect ERROR : [vf#0:0 @ 0x6227412ddfc0] Terminating thread with return code -5 (Input/output error)

currently currently use this for the reoling DUO 3 , and no ffmpeg errors for hours .

GARTEN:
ffmpeg:
path: "5.0"

As a note to myself : Never by a CAM where you cannot set the Codec, special if it is h265. ( or a Reolink at all , the dahua are much much better in every area and i think i will stick with those .... )

So i hope ffmpeg 5 stays in this case :)

Thx, Dietmar

[d1etmar]

spam update . I have the same ffmpeg errors on the other 2 dahua cams but a lot less often. So currently i am running with 5.0 ffmpeg for all cams. ( maybe this all happens because of my virtualization setup via proxmox,lxc,docker - no idea ) . But just as info . Will continue watching and do my usual trial and errors . The hard way, but a lot of learning.

The plus Model seems to work much better, no dog motions anymore . All correct identified as Cats . Great - thank you very much .

Next is that one cam ( with lower resolution on sub stream ) did not get enough motion to catch a cat. The Video is rather clear and only little motion shown if i playback , and therefore nothing triggered ?

i added it to frigate+ ...

Currently i know that i could tune the motion detection on live cam. But this does not help me as much as i would :) . Could i just do it for a recording as well ? . I see the "balken" during playback . And the motion detection is default to 30 . How much is the "balken" ? :)

The sub stream for this is 1280x720, 10fps, 10iframes. Main is 2560x1440,25fps,50iframes. Point writing/asking is should i use the main for detection ? I have the feeling that the other cam with higher resolution does a better job. And wait what my frigate backend HW does with it ? :) . Maybe my next tries for now, if i could not find anything better gg .

Love it, and at the end it will safe Birds plus i had a lot of fun playing with it , a win:win . Thx for creating this .

Thx, Dietmar

[d1etmar]

next plan after "Schweizerhaus" . Take recording - put it together for a 2 min video , downsampling it to 720p and insert it as a file camera in frigate and watch the live view for motion . Check settings - but also compare that with the high resolution file ... If the results are good and makes sense maybe writing a small bash helper script for such debugging.

[d1etmar]

being late . I have not tried for a replay motion checks for the moment. I changed for the specific CAM to do the high resolution main stream for motion and waiting for a cat on that spec one :) . It does Birds ok so it works . Backend is 2-4% more cpu ruffly in my setup . So no problem at all. I had some fun with the SONOFF smart water zigbee device in Z2M . To summarize it , upgrade it to 1.0.4 ( mine was 1.0.3 ) via OTA and place the device very near your Zigbee concentrator . the OTA takes "ages" to complete (+1h ) .

Some first funny results already with "flying" cats via the other cam :) .

I could strongly recommend if you run a Automation with a water source turning off and on to write a HA time sensor which runs when the switch goes on. Then you could write an extra Automation to check the timer and turn it off . ( as a fallback )

My first automation was just strait forward , like this one

• alias: "Trigger Water Switch on Cat Detection"
  id: "sonoff_wasser_automation"
  description: "Turn on smartwater switch for 5 minutes when a cat is detected on BIRD1 camera"
  trigger:
  • platform: mqtt
    topic: frigate/events
    condition:
  • condition: template
    value_template: >
    {{ trigger.payload_json['type'] == 'new' and
    trigger.payload_json['after']['label'] in ['dog', 'cat'] and
    trigger.payload_json['after']['camera'] in ['BIRD1', 'BIRD2'] }}
    action:
  • service: switch.turn_on
    target:
    entity_id: switch.smartwater
  • delay: "00:03:00"
  • service: switch.turn_off
    target:
    entity_id: switch.smartwater
    mode: single

The Water source is connected to a Water Rain Zisterne with 5800L . The "problem" was that the cat decided to have their prime time at 03:28:30 arround , and i do a Backup of HA on 03:30:00 / or something else resetted (proxmox snapshot). So this ended up with that the switch.turn_off did not happen, and the Zisterne was empty in the morning seeing in my HA Dashboard. ( 2,5k Liter water gone ) . I do measure the House(AI on the Edge ) and Zisterne Water( Liquid check) as well.

So i added this ( via chatgpt conversation ... )

== Input Datetime Helper ==

input_datetime:
smartwater_start_time:
name: Smart Water Start Time
has_date: true
has_time: true

== Template Sensor ==

template:

• sensor:
  • name: "Smart Water On Time"
    unit_of_measurement: "s"
    state: >
    {% if is_state('switch.smartwater', 'on') %}
    {% set start = states('input_datetime.smartwater_start_time') %}
    {% if start != 'unknown' and start != 'unavailable' %}
    {% set start_ts = as_timestamp(start) %}
    {{ (now().timestamp() - start_ts) | int }}
    {% else %}
    0
    {% endif %}
    {% else %}
    0
    {% endif %}

== Automations ==

automation:

Set the timestamp when switch.smartwater turns ON

• alias: "Set Smart Water Start Time"
  trigger:
  • platform: state
    entity_id: switch.smartwater
    to: 'on'
    action:
  • service: input_datetime.set_datetime
    target:
    entity_id: input_datetime.smartwater_start_time
    data:
    datetime: "{{ now().isoformat() }}"
    mode: single

Auto-turn off the switch after 300 seconds

• alias: "Auto Turn Off Smart Water After 5 Minutes"
  trigger:
  • platform: state
    entity_id: sensor.smart_water_on_time
    condition:
  • condition: numeric_state
    entity_id: sensor.smart_water_on_time
    above: 300
  • condition: state
    entity_id: switch.smartwater
    state: 'on'
    action:
  • service: switch.turn_off
    target:
    entity_id: switch.smartwater
    mode: single

So this works that if i manually trigger the SONOFF smart water device it stops after 5 min.

Will continue to upload my Screenshots to Frigate+ and if getting enough done (100+?) i will try a generated / custom Model .

I am not so happy with the current Gardena Sprinker, i would need a solution which uses less Water . Spray/Sprinkle 50cm and lower but still 3-5m wide . This is not a very typical use case so difficult to find a "good" devices to use . Also need to think what to do in Winter.

I love HA and Frigate+ . It makes a lot of fun . All because it is possible but not necessary needed lol

br, Dietmar

ps. i have no idea why i cannot use the "<> code" it simple does not work/format the yaml in the post .

[d1etmar]

Servus, the 2nd CAM catched a marten ( Marder ) - is there a model , or assign it as a cat ? . The Video showed 3 of them , lovely for me to see them . For me it is fine if tracked as cat .

[d1etmar]

Servus,
some update, i had a personally cat experience during cut off the bushes as i was regognized as dog and the automation sprinkler startet :) ( and sry no pics of this gg ) . So after digging more, i found out that the detection process is somehow "continuesly" and might update over time, depending on weight/score and type=new / end . So i needed to adapt the Automation . To better understand a automation for the frigate/mqtt events to a logbook helped . Also this is a good check against any new automation i use now .

I spent some time trying to use existing mp4 exports with an extra docker mediamtx image to generate a looping video and use it as input in frigate. I got it working but with a mess of a lot pixel errors, and i have found no way to get them "away" . So as this did not work for me , i changed the motion threshhold to 20 , and the second parameter for the pixels to 7 . As i would mostly need this during night i have more motions during day (wind - gras ) . The coral and the i3 seems to be able to handle it . It would be great if we could use the entity sun for changing the motion during night without doing some extra http rest , bash scripts config.yaml copy and restart . ( my frigate is not installed together with ha )

input_boolean:
  frigate_log:
    name: Frigate Log Helper
    initial: off

# == Input Datetime Helper ==
input_datetime:
  smartwater_start_time:
    name: Smart Water Start Time
    has_date: true
    has_time: true

# == Template Sensor ==
template:
  - sensor:
      - name: "Smart Water On Time"
        unit_of_measurement: "s"
        state: >
          {% if is_state('switch.smartwater', 'on') %}
            {% set start = states('input_datetime.smartwater_start_time') %}
            {% if start != 'unknown' and start != 'unavailable' %}
              {% set start_ts = as_timestamp(start) %}
              {{ (now().timestamp() - start_ts) | int }}
            {% else %}
              0
            {% endif %}
          {% else %}
            0
          {% endif %}

# == Automations ==
automation:
  # Set the timestamp when switch.smartwater turns ON
  - alias: "Set Smart Water Start Time"
    trigger:
      - platform: state
        entity_id: switch.smartwater
        to: 'on'
    action:
      - service: input_datetime.set_datetime
        target:
          entity_id: input_datetime.smartwater_start_time
        data:
          datetime: "{{ now().isoformat() }}"
    mode: single

  # Auto-turn off the switch after 300 seconds
  - alias: "Auto Turn Off Smart Water After 5 Minutes"
    trigger:
      - platform: state
        entity_id: sensor.smart_water_on_time
    condition:
      - condition: numeric_state
        entity_id: sensor.smart_water_on_time
        above: 180
      - condition: state
        entity_id: switch.smartwater
        state: 'on'
    action:
      - service: switch.turn_off
        target:
          entity_id: switch.smartwater
    mode: single
    
  - alias: "Trigger Water Switch on Cat Detection"
    id: "sonoff_wasser_automation"
    description: "Turn on smartwater switch for 1 minute when a cat or deer is detected on BIRD1 or BIRD2 camera"
    mode: single
    trigger:
      - platform: mqtt
        topic: frigate/events
    condition:
      - condition: template
        value_template: >
          {% set after = trigger.payload_json.get('after', {}) %}
          {{ trigger.payload_json.get('type') == 'end'
            and after.get('label') in ['dog', 'cat']
            and after.get('camera') in ['BIRD1', 'BIRD2']
            and after.get('top_score', 0) > 0.75
            and (after.get('end_time', 0) - after.get('start_time', 0)) > 2 }}
    action:
      - service: switch.turn_on
        target:
          entity_id: switch.smartwater
      - delay: "00:01:00"
      - service: switch.turn_off
        target:
          entity_id: switch.smartwater




  - id: log_frigate_all_events
    alias: "Log All Frigate MQTT Events to Logbook"
    description: "Log all Frigate MQTT messages to Home Assistant Logbook"
    mode: queued
    trigger:
      - platform: mqtt
        topic: frigate/events
    action:
      - service: logbook.log
        data:
          name: "Frigate Event"
          entity_id: input_boolean.frigate_log
          message: >
            Type: {{ trigger.payload_json['type'] }},
            Camera: {{ trigger.payload_json.get('after', {}).get('camera', 'unknown') }},
            Label: {{ trigger.payload_json.get('after', {}).get('label', 'unknown') }},
            Confidence: {{ trigger.payload_json.get('after', {}).get('top_score', 'n/a') }},
            {% if trigger.payload_json.get('after', {}).get('end_time') and trigger.payload_json.get('after', {}).get('start_time') %}
            Duration: {{
              (trigger.payload_json['after']['end_time'] - trigger.payload_json['after']['start_time']) | round(2)
            }} seconds
            {% else %}
            Duration: ongoing
            {% endif %}

So i have one logbook of 2 CAMS beside for checking

hf with frigate+ and ha .

dietmar

[d1etmar]

some updates .

i managed to use a script/automation in HA which uses the sun entity to change the threshold and the contour Area . And as for now i have not seen any cat/marder which was not catched by motion. So i think this helped the most. I still do 25fps and at least FHD on the detect . This generates more short clips, but i hope at the end i do not miss anything :) .
I also improved my Zigbee2MQTT Network with 3 SONOFF psu plugs and i added an automation to check how long the Water Pump is running, in case the watermeter does not stop.

If you change the threshold and contour in HA it is very runtime - so login into the frigate web interface/restart will reset to the value given in the yaml.

automation:
  - alias: Frigate Motion Settings - Day Mode
    trigger:
      - platform: sun
        event: sunrise
        offset: "+00:30:00"  # Optional: 30 min after sunrise
    action:
      - service: script.frigate_day_motion_settings
    mode: single
  - alias: Frigate Motion Settings - Night Mode
    trigger:
      - platform: sun
        event: sunset
        offset: "-00:30:00"  # Optional: 30 min before sunset
    action:
      - service: script.frigate_night_motion_settings
    mode: single

    
script:
  frigate_day_motion_settings:
    alias: Frigate - Day Motion Settings
    sequence:
      - service: number.set_value
        target:
          entity_id: number.bird1_threshold
        data:
          value: 30
      - service: number.set_value
        target:
          entity_id: number.bird1_contour_area
        data:
          value: 10
      - service: number.set_value
        target:
          entity_id: number.bird2_threshold
        data:
          value: 30
      - service: number.set_value
        target:
          entity_id: number.bird2_contour_area
        data:
          value: 10

  frigate_night_motion_settings:
    alias: Frigate - Night Motion Settings
    sequence:
      - service: number.set_value
        target:
          entity_id: number.bird1_threshold
        data:
          value: 15
      - service: number.set_value
        target:
          entity_id: number.bird1_contour_area
        data:
          value: 4
      - service: number.set_value
        target:
          entity_id: number.bird2_threshold
        data:
          value: 15
      - service: number.set_value
        target:
          entity_id: number.bird2_contour_area
        data:
          value: 4

i also tweaked arround at set this, i saw that i still used the G.711 for audio on the CAM ups . So i changed this on the cam. But still the ffmpeg process seems to transcode it , so i added this for just "copy" . Not sure if this is all good. But it works .

The max_disappeared: 75 was recommended setting by chatgpt ( i think ) .

        record: >-
          -f segment -segment_time 10 -segment_format mp4 -reset_timestamps 1
          -strftime 1 -c:v copy -c:a copy
    detect:
      max_disappeared: 75
      fps: 25

more and more happy , only need more advanced sprinkler system, as the cat get used to it and therefore the resulting videos are not that funny anymore . And the detect process works better and shorter ( i catch on mqtt frigate event type new at night ) so the sprinkler starts a bit too early now gg . Did this because before i had some frigate events with label cat which did not made it to a clip or screenshot ( = false in mqtt entry ) .

have fun with frigate and HA

Dietmar