1. Description
2. Repository Structure
3. Running the System
4. Credit
System to identify and track hand movements in a video stream with a ego centric view point. The system uses a pre-trained segementation model trained on the ego-centric dataset. Video stream is converted to frames and preprocessed and passed into the pre-trained model for inference. The model output a segemented image which is then post processsed to identify contours to generate a box around it. Post processing is done to reduce the noise.
The important files and directories of the repository is shown below
├── seg_hand_trained - contain the hand segemented model
├── Videos - Video folder that contains the input video
├── infer_segmentation.py - contain class for main inference
├── segmentation.py - main runn file, contain function to preprocess and post-process frames
└── utils.py - contain utility functions
Cloning the git repository
git clone https://github.com/iTharindu/hand-movement-tracking.git
Getting inside the repository
cd hand-movement-tracking
Creating a virtual environment
virtualenv --python=python3.6 env
source env/bin/activate
If virtualenv is not installed then run pip3 install virtualenv to install virtualenv.
Installing the dependencies
pip3 install -r requirements.txt
To generate tracking video
python segmentation.py
Output will be generated as output.avi
We used a pre-trained model Unet model with the efficientnetb2 backbone. This model with several other models are included in the here