Uses OpenCV and Tensorflow to perform realtime object detection locally for IP cameras. Designed for integration with HomeAssistant or others via MQTT.
Use of a [Google Coral Accelerator](https://coral.ai/products/) is optional, but highly recommended. On my Intel i7 processor, I can process 2-3 FPS with the CPU. The Coral can process 100+ FPS with very low CPU load.
You see multiple bounding boxes because it draws bounding boxes from all frames in the past 1 second where a person was detected. Not all of the bounding boxes were from the current frame.
A `config.yml` file must exist in the `config` directory. See example [here](config/config.example.yml) and device specific info can be found [here](docs/DEVICES.md).
ARM boards are not officially supported at the moment due to some python dependencies that require modification to work on ARM devices. The Raspberry Pi4 gets about 16ms inference speeds, but the hardware acceleration for ffmpeg does not work for converting yuv420 to rgb24. The Atomic Pi is x86 and much more efficient.
Users have reported varying success in getting frigate to run in a VM. In some cases, the virtualization layer introduces a significant delay in communication with the Coral. If running virtualized in Proxmox, pass the USB card/interface to the virtual machine not the USB ID for faster inference speed.
Setup a camera, binary_sensor, sensor and optionally automation as shown for each camera you define in frigate. Replace <camera_name> with the camera name as defined in the frigate `config.yml` (The `frigate_coral_fps` and `frigate_coral_inference` sensors only need to be defined once)
You can access a higher resolution mjpeg stream by appending `h=height-in-pixels` to the endpoint. For example `http://localhost:5000/back?h=1080`. You can also increase the FPS by appending `fps=frame-rate` to the URL such as `http://localhost:5000/back?fps=10` or both with `?fps=10&h=1000`
Publishes a jpeg encoded frame of the detected object type. When the object is no longer detected, the highest confidence image is published or the original image
`min_score` defines the minimum score for Frigate to begin tracking a detected object. Any single detection below `min_score` will be ignored as a false positive. `threshold` is based on the median of the history of scores for a tracked object. Consider the following frames when `min_score` is set to 0.6 and threshold is set to 0.85:
| Frame | Current Score | Score History | Computed Score | Detected Object |
In frame 2, the score is below the `min_score` value, so frigate ignores it and it becomes a 0.0. The computed score is the median of the score history (padding to at least 3 values), and only when that computed score crosses the `threshold` is the object marked as a true positive. That happens in frame 4 in the example.
The labelmap can be customized to your needs. A common reason to do this is to combine multiple object types that are easily confused when you don't need to be as granular such as car/truck. You must retain the same number of labels, but you can change the names. To change:
- Download the [COCO labelmap](https://dl.google.com/coral/canned_models/coco_labels.txt)
- Modify the label names as desired. For example, change `7 truck` to `7 car`
- Mount the new file at `/labelmap.txt` in the container with an additional volume
**Note**: Previous versions of frigate included `-vsync drop` in input parameters. This is not compatible with FFmpeg's segment feature and must be removed from your input parameters if you have overrides set.
Frigate can save video clips without any CPU overhead for encoding by simply copying the stream directly with FFmpeg. It leverages FFmpeg's segment functionality to maintain a cache of 90 seconds of video for each camera. The cache files are written to disk at /cache and do not introduce memory overhead. When an object is being tracked, it will extend the cache to ensure it can assemble a clip when the event ends. Once the event ends, it again uses FFmpeg to assemble a clip by combining the video clips without any encoding by the CPU. Assembled clips are are saved to the /clips directory along with a json file containing the current information about the tracked object.
### Global Configuration Options
-`max_seconds`: This limits the size of the cache when an object is being tracked. If an object is stationary and being tracked for a long time, the cache files will expire and this value will be the maximum clip length for the *end* of the event. For example, if this is set to 300 seconds and an object is being tracked for 600 seconds, the clip will end up being the last 300 seconds. Defaults to 300 seconds.
### Per-camera Configuration Options
-`pre_capture`: Defines how much time should be included in the clip prior to the beginning of the event. Defaults to 30 seconds.
-`objects`: List of object types to save clips for. Object types here must be listed for tracking at the camera or global configuration. Defaults to all tracked objects.
## Google Coral Configuration
Frigate attempts to detect your Coral device automatically. If you have multiple Coral devices or a version that is not detected automatically, you can specify using the `tensorflow_device` config option.
The mask works by looking at the bottom center of any bounding box (first image, red dot below) and comparing that to your mask. If that red dot falls on an area of your mask that is black, the detection (and motion) will be ignored. The mask in the second image would limit detection on this camera to only objects that are in the front yard and not the street.
Zones allow you to define a specific area of the frame and apply additional filters for object types so you can determine whether or not an object is within a particular area. Zones cannot have the same name as a camera. If desired, a single zone can include multiple cameras if you have multiple cameras covering the same area. See the sample config for details on how to configure.
During testing, `draw_zones` can be set in the config to tell frigate to draw the zone on the frames so you can adjust as needed. The zone line will increase in thickness when any object enters the zone.
- Lower the framerate of the video feed on the camera to reduce the CPU usage for capturing the feed. Not as effective, but you can also modify the `take_frame` [configuration](config/config.example.yml) for each camera to only analyze every other frame, or every third frame, etc.
- Hard code the resolution of each camera in your config if you are having difficulty starting frigate or if the initial ffprobe for camerea resolution fails or returns incorrect info. Example:
- Object configuration - Tracked objects types, sizes and thresholds can be defined globally and/or on a per camera basis. The global and camera object configuration is *merged*. For example, if you defined tracking person, car, and truck globally but modified your backyard camera to only track person, the global config would merge making the effective list for the backyard camera still contain person, car and truck. If you want precise object tracking per camera, best practice to put a minimal list of objects at the global level and expand objects on a per camera basis. Object threshold and area configuration will be used first from the camera object config (if defined) and then from the global config. See the [example config](config/config.example.yml) for more information.
## Troubleshooting
### "ffmpeg didnt return a frame. something is wrong"
Turn on logging for the camera by overriding the global_args and setting the log level to `info`:
