Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/robertlucian/touch-gesture-detection

A concept library on detecting gestures from an array of touch sensors
https://github.com/robertlucian/touch-gesture-detection

detecting-gestures robotics sensors touch-sensors

Last synced: 4 days ago
JSON representation

A concept library on detecting gestures from an array of touch sensors

Awesome Lists containing this project

README 

        
# Touch Gesture Detection [![Build Status](https://travis-ci.org/RobertLucian/touch-gesture-detection.svg?branch=master)](https://travis-ci.org/RobertLucian/touch-gesture-detection) 

A concept library for detecting gestures from an array of touch sensors (and possibly other sensors as well).



## Installing



To install the library, please be sure you've got Python 3.6+ and run

```bash

pip install git+https://github.com/RobertLucian/touch-gesture-detection.git

```



## Development



To develop it, clone the repo and run 

```bash

virtualenv -p python3 env # make sure you've got python 3.6+ installed



env\Scripts\activate # for Windows-based systems

source env/bin/activate # for Linux-based systems



pip install --editable .

```



## API



`kiki.detection.get_resource_path` - function to get the absolute file path

of a file that resides in the `kiki` package. Must use a relative path for it.



`kiki.detection.save_trained_model` - function to save a model and its weights to the disk.



`kiki.detection.load_pretrained_model` - function to load a pre-trained model.

Can specify whether the given path to the dataset is for an in-packaged model 

or just another that's coming out of it.



`kiki.detection.train_touchsensor_model` - function to train the touch-gesture

detection mechanism. Takes in a dataset and sets some important hyper-parameters.

This function should be deprecated at some point and instead have the process of training new

models streamlined. To train new motions, just add more data to the dataset.



To see the parameters of these functions, check the embedded docstrings from each one.



## Dataset



Within the package, there are a couple of datasets included:

* `datasets/vertical_swipe.csv`

* `datasets/horizontal_swipe.csv`

* `datasets/tapping.csv`

* `datasets/double_tapping.csv`

* `datasets/hitting.csv` - currently empty

* `datasets/slapping.csv`

* `datasets/generated_dataset.csv`



Each one of these datasets can be opened up this way

```python

import pandas as pd

from kiki.detection import get_resource_path



file = get_resource_path('datasets/generated_dataset.csv')

df = pd.read_csv(file)

```



`generated_dataset.csv` is a special dataset which was artificially built

from all the other non-empty datasets. Practically, it was generated

by iteratively altering the values from each sample by up to 15%. This should

provide enough variation to prove the point. All samples are shuffled. There are

 36200 samples in total, each one of them having 6 features and 11 time steps. 

 [kiki/datasets/data_generator.py](kiki/datasets/data_generator.py) was used to generate

 all this data. This script is not included in the package though.

 

 ## Pre-trained Model



The pre-trained model is found at `models/touch` and can be loaded by running 

```python

from kiki.detection import load_pretrained_model



load_pretrained_model('models/touch', inpackage_data=True)

```

This both loads the model of the neural network and the weights associated with it.



The neural network is comprised of an LSTM layer of 32 units and a dense layer with the same

number of neurons as there are features - currently 6 of them. The returned model is a 

keras [model](https://keras.io/models/model/). The test accuracy is as of this moment at 100%,

which should be expected from a dataset that is almost ideal and has almost no noise.