Skip to content

Latest commit

 

History

History

person-class

README.md

person-class

This folder contains the Person Classification Server, person_classifier_server.py, which runs a TensorFlow deep convolutional neural network (CNN)-based person classifier using Python APIs and employees zerorpc to communicate with the Alarm Uploader. One of the benefits of using zerorpc is that the server can easily be run on another machine, apart from the machine running ZoneMinder. The Person Classification Server is run as a Linux service using systemd. This folder also contains programs to train models used by the server as well as other utilities.

This server uses a fine-tuned CNN to classify that a person object detected by obj-detect is member of my family or a stranger. It is an alternative to face-det-rec and so one or the other must be run but not both. Note that face-det-rec includes shallow learning methods (SVM or XGBoost classifiers) in the final stage of a pipeline to recognize faces in alarm images. CNNs have been shown to outperform shallow learning methods for many computer vision tasks given sufficient training data; this was the main motivation for developing person-class.

Installation

  1. Clone this git and cd to it.

  2. Install the machine learning platform on the Linux server per the steps described here. The required Python packages used by the installation are listed in ml_requirements.txt

  3. Install the Coral edge TPU compiler using these instructions.

  4. Run train.py to fine-tune a CNN (MobileNetV2, ResNet50, ResNetInceptionV2 and VGG16 are currently supported) from the images contained in the same data set used to train face-det-rec. Tune the hyperparamters used in program to suit your situation. By default the program will generate an inference-optimized TensorFlow model and a quantized TensorFlow Lite model compiled for use on the Coral edge TPU. Run $python3 ./train.py -h for command line options. See Using Keras On Google’s Edge TPU and Building a Deep Learning Person Classifier for related information.

  5. Modify config.json to suit your installation.

  6. Create the file /tmp/face_det_zmq.pipe for an IPC socket that the zerorpc client and server will communicate over. This assumes that the person classification server and ZoneMinder are running on the same machine. If not, then use a TCP socket. If you have already created this as part of the face-det-rec installation you can skip this step.

  7. Use systemd to run the Person Classification Server as a Linux service. Edit person-class.service to suit your configuration and copy the file to /etc/systemd/system. Then enable and start the service but first disable the face-detect.service (since only one of the two can run at the same time):

$ sudo systemctl stop face-detect.service && sudo systemctl disable face-detect.service
$ sudo systemctl enable person-class.service && sudo systemctl start person-class.service

Notes

  1. Training results and models are saved by default to the folder train-results.

  2. keras_to_tflite_quant.py generates a uint8 quantized tflite model from a keras .h5 file using post-training quantization. This is used as module in train.py or it can be run standalone on the command line. This program can evaluate classification accuracy on the quantized model but is very slow on an Intel machine. Instead, evaluate the model on the edge TPU hardware using evaluate_model.py (post-training quantization has little impact on classification accuracy).

  3. keras_to_frozen_tf.py generates a frozen TensorFlow model optimized for inference from a keras .h5 file. This is used as module in train.py or it can be run standalone on the command line.

  4. Use the Coral edge TPU compiler to generate a model from the tflite quantized model than can be run on the edge TPU hardware. This is done automatically as part of train.py or it can be run on the command line.