Python script using machine learning to classify images from IP cameras.
This Python script is meant to be executed as a systemd service. It will subscribe an MQTT broker to receive JPEG images from an IP camera.
The machine learning framework TensorFlow is used to classify images. A deep convolutional neural network (CNN) based on the Inception v3 model needs to be retrained with a set of images from your IP camera to classify the image either as person or as nothing.
If a person is detected in the image, the Python script will publish the image to a specified MQTT topic. A threshold can be configured for this to publish even images with a lower probability score.
This script was designed to work on a Raspberry Pi 3 (ARM architecture), but it also runs on x86_64 and should actually run on all platforms where TensorFlow and Python are available.
On a Raspberry Pi 3 this is pretty slow (roughly three seconds per image). So make sure you have a good pre-processing of your IP camera feed and deliver only images where potentially movement was detected. You could either use the built-in motion detection of your IP camera or preferable a third party software detection like Motion for this.
- TensorFlow
- Python 2.7
- Python PyYAML (pyyaml) and Eclipse Paho MQTT (paho-mqtt) packages
All dependencies can be installed automatically by executing ./install.sh.
Clone this repository:
git clone https://github.com/1element/ip-camera-image-inference.git
Install dependencies:
./install.sh
The TensorFlow Inception v3 model is used to classify images either as person or as nothing. For accurate scoring you need to retrain the final layer of the Inception model.
So before you can start retrain.py you need a set of images from your
IP camera that represent the two classes person and nothing.
Copy these images to the retrain/person/ and retrain/nothing/
directories.
After this execute:
python retrain.py --image_dir=/path/to/ip-camera-image-inference/retrain
This will take some time, especially if you run it on a Raspberry Pi. Consider running the training on a faster machine if you can.
Copy the output graph and labels from the temporary directory of your system to the model directory:
cp /tmp/output_graph.pb /path/to/ip-camera-image-inference/model/
cp /tmp/output_labels.txt /path/to/ip-camera-image-inference/model/
If you are interested in details of the retraining process refer to the official TensorFlow image retraining tutorial.
Edit the provided configuration file config.yml. The TensorFlow model
directory and your MQTT broker connection as well as a few other settings
can be configured there.
The Python script image-inference.py should be executed as a systemd
service. There is a template file image-inference.service for this.
Change the paths in this file according to your environment.
Afterwards use the following commands to set up the service:
# copy service configuration
sudo cp image-inference.service /etc/systemd/system/
# reload systemd
sudo systemctl daemon-reload
# enable auto start
sudo systemctl enable image-inference.service
# start service now
sudo systemctl start image-inference.service
# show status
sudo systemctl status image-inference.service
There is a previous version using a directory observer to receive images
and FTP to distribute the classified image. If you are interested in this
checkout the tag v1.0.0.
This script was written for my own purpose, therefore it lacks in features and configuration abilities. In case it might be useful for someone, I decided to share it. However, pull requests are welcome.
This project is licensed under the terms of the Apache License, Version 2.0.