TensorFlow

The TensorFlow image processing platform allows you to detect and recognize objects in a camera image using TensorFlow. The state of the entity is the number of objects detected, and recognized objects are listed in the summary attribute along with quantity. The matches attribute provides the confidence score for recognition and the bounding box of the object for each detection category.

Prerequisites

The following packages must be installed on Debian before following the setup for the integration to work:

sudo apt-get install libatlas-base-dev libopenjp2-7 libtiff5

It is possible that Home Assistant is unable to install the Python TensorFlow bindings. If that is the case, you’ll need to install those manually using: pip install tensorflow==2.2.0, as the Python wheel is not available for all platforms.

See the official install guide for other options.

Furthermore, the official Python TensorFlow wheels by Google, require your CPU to support the avx extension. If your CPU lacks those capabilities, Home Assistant will crash when using TensorFlow, without any message.

Preparation

This integration requires files to be downloaded, compiled on your computer, and added to the Home Assistant configuration directory. These steps can be performed by cloning this repository into your configuration directory. Alternatively, if you wish to perform the process manually, the process is as follows:

Create the following folder structure in your configuration directory.

  |- {config_dir}
    |- tensorflow/
      |- models/

Follow these steps (Linux) to compile the object detection library.

# Clone tensorflow/models
git clone https://github.com/tensorflow/models.git
# Compile Protobuf (apt-get install protobuf-compiler)
cd models/research
protoc object_detection/protos/*.proto --python_out=.
# Copy object_detection to {config_dir}
cp -r object_detection {config_dir}/tensorflow

Your final folder structure should look as follows

  |- {config_dir}
    |- tensorflow/
      |- models/
      |- object_detection/
        |- ...

Model Selection

Lastly, it is time to pick a model. It is recommended to start with one of the COCO models available in the Model Detection Zoo.

The trade-off between the different models is accuracy vs speed. Users with a decent CPU should start with one of the EfficientDet models. If you are running on an ARM device like a Raspberry Pi, start with the SSD MobileNet v2 320x320 model.

Whichever model you choose, download it and extract in to the tensorflow/models folder in your configuration directory.

Configuration

To enable this platform in your installation, add the following to your configuration.yaml file:

# Example configuration.yaml entry
image_processing:
  - platform: tensorflow
    source:
      - entity_id: camera.local_file
    model:
      graph: /config/tensorflow/models/efficientdet_d0_coco17_tpu-32/

categories can also be defined as dictionary providing an area for each category as seen in the advanced configuration below:

# Example advanced configuration.yaml entry
image_processing:
  - platform: tensorflow
    source:
      - entity_id: camera.driveway
      - entity_id: camera.backyard
    file_out:
      - "/tmp/{{ camera_entity.split('.')[1] }}_latest.jpg"
      - "/tmp/{{ camera_entity.split('.')[1] }}_{{ now().strftime('%Y%m%d_%H%M%S') }}.jpg"
    model:
      graph: /config/tensorflow/models/efficientdet_d0_coco17_tpu-32/
      categories:
        - category: person
          area:
            # Exclude top 10% of image
            top: 0.1
            # Exclude right 15% of image
            right: 0.85
        - car
        - truck

Optimizing resources

Image processing components process the image from a camera at a fixed period given by the scan_interval. This leads to excessive processing if the image on the camera hasn’t changed, as the default scan_interval is 10 seconds. You can override this by adding to your configuration scan_interval: 10000 (setting the interval to 10,000 seconds), and then call the image_processing.scan service when you actually want to perform processing.

# Example advanced configuration.yaml entry
image_processing:
  - platform: tensorflow
    scan_interval: 10000
    source:
      - entity_id: camera.driveway
      - entity_id: camera.backyard

# Example advanced automations.yaml entry
- alias: "TensorFlow scanning"
  trigger:
     - platform: state
       entity_id:
         - binary_sensor.driveway
  action:
    - service: image_processing.scan
      target:
        entity_id: camera.driveway