A live camera feed can look perfectly healthy while the detection you wanted has gone missing. Video and detection information do not have to fail together, which is inconvenient but also a useful clue.
My goal here is to follow one deliberate detection across the boundaries: Frigate, MQTT and the Home Assistant automation. I want to identify the first place where the evidence stops, not restart everything until the house develops trust issues.
The project notes say this setup uses Frigate, go2rtc, Home Assistant and an EMQX MQTT broker. They do not establish the actual MQTT topics, credentials, automation triggers or Frigate settings. Treat those as site-specific details to inspect, not values to guess. Interface labels may also differ between versions.
What you need
- Access to the relevant camera in Frigate and to Home Assistant.
- Permission and a safe way to produce an object that the camera is already configured to detect.
- Access to the existing Frigate configuration and the Home Assistant automation you are investigating.
- Access to MQTT connection information and an authorised MQTT client, if you need to compare subscribers.
- A way to take notes, including approximate times. Paper remains a surprisingly resilient dependency.
Identify the camera, object type and expected outcome before starting. If an automation controls physical equipment, make the test safe first. Keep people clear and prevent an unintended action using a method appropriate to the equipment. Do not assume a software test is harmless just because the button is on a screen.
Step 1: Define the missing result
Write down what should happen: for example, Frigate should recognise an object, Home Assistant should receive information, or an automation should run. These are separate observations. A missing clip is not automatically a missing detection, and a missing automation action is not automatically a camera fault.
Inspect the existing Frigate configuration for the selected camera and object type. Note any relevant detection filters. Then open the intended Home Assistant automation and identify its trigger: is it based on an MQTT message, a Home Assistant event, an entity state or something else? Keep the actual names private; they are placeholders for this investigation.
Step 2: Check the live feed and detection overlay
Open the camera in Frigate and confirm that the image is current. Look for an ordinary change in the scene, with permission from anyone involved. If the feed is stale or unavailable, investigate the camera and stream path first, including go2rtc where it is involved. Frigate cannot reliably detect an object in a picture it is no longer receiving.
Next, open the camera’s live or debug view and enable the available detection overlay, such as bounding boxes and labels. The location and wording of these controls depend on the Frigate version. Watch the view while someone produces one clear example of the configured object. Note the time and whether a labelled box appears.
This gives you a live observation that does not depend on a clip, snapshot or review item being retained. It is not infallible: a brief box can be missed, overlays may not be available in the view you chose, and detection filters affect what you see. If necessary, repeat the same simple test while watching. Check your version’s Frigate documentation for its debug view rather than assuming an absent review item proves detection failed.
Step 3: Identify and watch the MQTT topic
Frigate documents an events topic using its configured MQTT topic prefix followed by an events suffix. The commonly documented default is frigate/events, but the prefix can differ. That is a general reference, not a verified topic for this project. Inspect the existing Frigate MQTT configuration and the documentation for your version to identify the actual topic and broker. Also confirm that MQTT is enabled there; the presence of EMQX in the project does not prove a particular connection is working.
In Home Assistant, open the MQTT integration’s configuration and use its option to listen to a topic. Enter the actual events topic you identified, start listening, and leave the listener open. Produce one controlled detection and watch for a new message. Check the message’s camera, object and time information where present against your test. Frigate event messages can contain changing information about a tracked object, so more than one message may correspond to the same test.
If the integration or listening option is unavailable, inspect the MQTT integration and its reported errors. Use an authorised MQTT client as an alternative listener. Subscribe to the same actual topic on the broker configured for Frigate, then repeat the test. Do not publish experimental messages to Frigate’s production topics.
Privacy warning: MQTT payloads can include camera identifiers, object details, timestamps and other sensitive information. Treat the topic and payload as private too. Do not paste them into a public forum without careful sanitising.
Step 4: Separate publication from subscription trouble
For a useful comparison, listen to the same exact topic with two subscribers during the same test: Home Assistant’s MQTT listener and an independent authorised client connected to the same broker. Record which listener sees the new message. A broker connection indicator alone does not prove either subscriber received it.
- Both see the message: MQTT delivery to those listeners works for this test. Investigate the automation or integration that is supposed to use it.
- The independent client sees it, but Home Assistant does not: check that Home Assistant uses the same broker and topic, then inspect its MQTT connection errors and account permissions. Different accounts can have different subscription access.
- Neither sees it: check the topic, broker, Frigate MQTT connection and relevant permissions. This is not yet proof that Frigate failed to publish: both listeners could be looking in the wrong place or lack access.
If you cannot trust a listener, verify it using an authorised test publisher and a separate test topic permitted by your broker. Confirm that the subscriber receives that known test message, then return to the actual Frigate topic. This checks the listening path without pretending that a green connection icon is a full diagnosis.
Step 5: Inspect the actual Home Assistant trigger
Return to the automation and compare its trigger with what you observed. For an MQTT trigger, check the configured topic and any payload or template constraints against the received message. A message arriving at the MQTT listener does not necessarily satisfy every part of the trigger.
For a Home Assistant event trigger, open Developer Tools, then the Events tool. Listen for the event type specified by the automation and repeat the controlled test. Compare the observed event data with any conditions in the trigger. Home Assistant bus events and MQTT messages are different things, so listening for one is not a substitute for watching the other. Event data can also contain sensitive details.
For a state trigger, find the actual entity in Developer Tools under States and watch its state during the test. History may help with a brief change if it is available. Use the entity named by the automation, not a similarly named camera entity that merely displays video.
Step 6: Investigate a missing or stopped automation
Check that the intended automation is enabled and inspect its traces if any exist. A trace can show where a run stopped, including a condition that was not met. If there is no trace, do not conclude that the action failed: the trigger may never have matched, or trace availability may be limited by the setup.
Recheck the trigger type and its exact topic, event type or entity. Compare its filters with the evidence captured during the test. Check for automation configuration errors and confirm that the automation is loaded. If the trigger is observed but still does not match, inspect its additional constraints and templates before changing Frigate.
Running an automation’s actions manually can help isolate an action problem, but it does not prove the trigger works and can operate real equipment. Only do that if the action is safe. I would rather leave a mystery unsolved for another minute than turn troubleshooting into an unscheduled door test.
How to test the result
After any change, repeat the same controlled detection. Note whether the live overlay showed it, whether each MQTT listener received a corresponding message, what Home Assistant observed, and whether the automation produced a trace or the intended safe action. Compare timestamps carefully; interfaces may display them differently.
Change one thing at a time and repeat at least once under similar conditions. If the result varies, record the difference instead of announcing a permanent victory over computers.
Troubleshooting and privacy
If the live feed works but the overlay shows nothing, revisit the configured object type, filters and test conditions. If the overlay shows a detection but no MQTT message appears, verify the exact topic and listening path before changing detection settings. If MQTT arrives but the automation does not trigger, focus on the trigger and its constraints. If a trace stops at a condition, investigate that condition rather than the camera.
When asking for help, describe the first boundary where evidence stops. Remove addresses, passwords, API keys, licence plates, family names, camera locations and other private infrastructure details from logs, screenshots, configuration, MQTT payloads and event data. Check the material again after redacting it. Logs are enthusiastic oversharers.
What is still unverified
This is a diagnostic method, not a tested recipe for this installation. The notes do not establish the Frigate version, actual MQTT topic or broker connection, detection filters, Home Assistant trigger, permissions or automation behaviour. Nor do they establish that a particular test produces a detection. Use the documentation for your versions and the configuration already present to fill those gaps.
The useful result is a narrower question: which boundary has evidence on one side and not the other? That is less glamorous than a magic fix, but considerably kinder to everything that was working already.