O'hoy, this is was my final year project of my BSc in CS at Lancaster. In short it is: a gesture recognition system, using the Leap Motion Sensor, Python and a basic self-implemented Naive Bayes classifier.
- Python 2.7.10
- Leap Motion SDK which you can download at https://developer.leapmotion.com/v2, from where the following three files need to reside in the same folder (if you're on a Mac)
- Leap.py
- LeapPython.so
- libLeap.dylib
NB: I've only developed and tested this project in a macOS environment, so things might go wonky on other platforms.
Make sure your Leap Motion is connected, that the Leap Motion deamon is running and your plants have been watered. In the command line type
python Builder.pyand start building your dataset. Create a new sign and set type. Click 'Record' for fixed signs, and hold to record a sequence of frames for gesture based signs. I'd usually stick to 20 samples per class as it was the lowest number with the highest gain in accuracy (and because I'm lazy), but obviously, the more the better.
To test the actual recognition, run
python Intepreter.pyDepending on how similar the gestures are things should work quite okay, although it probably should go without mentioning that the success rate is at the mercy at the sensors ability to capture your gestures correctly. Novel detection is a little wonky at the moment, it straightforwardly disregards any result that is further away from the average datapoint by more than 85% of the distance between the 'furthest from average' and average datapoint.
The accompanying PDF is the paper/report I wrote, which was the only element assessed, and contains details on the overall approach and code overview. If you've read everything else there is to read on the internet, go read that.