/activity_recognition

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

Activity recognition (Auto-labeling) on videos

A python library is used for recognizing human activities on videos and uses the results to label network traffic (PCAP) automatically. The library includes two parts: features (data representation) and models (detection models).

Installation

  • python 3.7.9
  • pip3 install -r requirements.txt

Features (only videos)

1) CNN features

Using CNN features extracted by the 'CNN' library [?]

2) VideoPose3D

Using 3D keypoints features extracted by the VideoPose3D library [?]

Models:

1) Onevsrest (e.g., logistic regression)

2) SVM (linear kernel)

3) Random Forest

Structure:

TODO

How

PYTHONPATH =./: python3
.7
examples / classical / detector_feature_A.py.py

Issues (continuing updating):

CNN features:

  • tensorflow < 2.0, which requires python3.7 or python3.6
  • download pretrained models from https://github.com/tensorflow/models/tree/master/research/slim E.g., vgg_16_2016_08_28.tar.gz tar -xvf vgg_16_2016_08_28.tar.gz vgg_16.cpkt
  • python3.7 feature_extraction.py --video_list data/video.txt --network vgg --framework tensorflow --output_path out/ --tf_model ./slim/vgg_16.ckpt --video_list: requires a txt file which lists all the full paths of the videos create an output directory by 'mkdir out'

3D keypoint features (VideoPose3D)

  • using detectron2