This project uses front-end separation, and the client has the following three forms of implementation:
- Manual gesture recognition, which means that the user determines the segmentation of continuous gestures
- Dynamic gesture recognition based on frame difference method
- Dynamic gesture recognition based on object tracking
Note: No server code is currently provided.
Python3.6, OpenCV3.4 + opencv_contrib
$ pip install opencv-python
$ pip install opencv-contrib-python
$ pip install pillow
$ pip install requests
- server-address: gesture recognition server address
$ python run_manual.py -s [server-address]
Interactive mode: press the keyboard s key before each action, and press the s key again after the action is complete to complete the recognition
you can choose the Background Subtraction Methods
- method: knn or mog
- threshold: The sum of the length and width of the contour identified by the algorithm is greater than the threshold is considered to be the hand
$ python run_frameDifferent -s [server-address] --method [method] --threshold [threshold]
GPU support is required to run this version, we tested on Ubuntu 16.04 + cuda9.0 + cudnn7.0.5 + tensorflow1.6. You need to install tensorflow1.6-gpu extra and darkflow, You can download darkflow from here.
$ pip install tensorflow-gpu
$ pip install Cython
$ cd darkflow
$ pip install .
# Check whether the installation is complete
$ flow --h
and then download the weight file and configuration file, and place them in the model folder and cfg folder respectively. Finally, run
$ python run_objectDetection.py -s [server-address]