Hi,
I've worked on a similar project for a company in Cambridge where I developed a gesture recognition Andorid application using wrist worn inertial sensors . I was successfully (95% accuracy) able to distinguish between multiple gestures such as 'pouring coffee from kettle' vs 'drinking coffee' vs 'moving hand in a circle' vs 'moving hand in a square'. I have already faced and know the ways to overcome the challenges involved in such problems including : difference in time durations of the reference and the recorded signals and the noise in inertial measurements.
I have extensive experience with using inertial sensors (MPU6050, ADXLxxx,etc) and the corresponding algorithms (motion fusion, classification and motion similarity metrics) used in sports application. In fact one of my start-ups is based in this area (calculating orientation of a bat). I know you'll be surprised why, given my profile, I am on freelancer but I'm lately bootstrapping due to investment problems. I'm happy to first do a proof of concept for you using your data in Matlab and show that my algorithms are able to provide the output that you require.
I look forward to hearing from you.
Thanks,
Abhi