Detecting and recognizing tricks performed by an athlete on an action scooter
Project Background
The client, a leading manufacturer of freestyle scooters, wanted to perform a feasibility study to see if it is possible to recognize tricks performed by a rider on a freestyle scooter.
The output of this feasibility study would be machine learning models for each trick performed by the rider.
Those models would serve to implement a solution for creating a smart action scooter. The smart scooter would be able to log motion data, detect tricks and be a part of the bigger IoT system, making a community of professional and amateur riders who can compete online or exchange riding tricks.
Finding the right partner is challenging
Although the scooter manufacturer has an excellent R&D team, their skills were mostly related to industrial and mechanical design, so they needed a partner who has skills in electronics, firmware, and software.
However, it was also hard to find a partner for this assignment since ordinary development houses do not have particular skillsets related to motion algorithms, analytics, and AI. Even with the appropriate skills in place, it would require significant development time, to reach the required level of solution.
Luckily, this was a perfect match for our existing solution platform and expertise.
Our Contribution
It was required to use our ExtremeMotion sensor platform and artificial intelligence algorithms to recognize pre-learned tricks, an ideal case to put into action our motion sensor platform and AI library.
Since action tricks include simultaneous deck and bar movements, the Scooter required two motion sensor devices, one placed on the deck and the other on the bar.
The sensor devices are battery powered , controlled via Bluetooth Low Energy by smartphone application.
Since it required mounting on the scooter, our team designed and 3D printed special enclosures to enable usage with the electronics and protecting it during extreme rides and shocks.
At the first step, motion data has been logged. The rider was executing each trick in 5 cycles, 10 repeats, in order to collect data on how perfectly executed trick looks like. After that, a regular ride has been performed in order to create a recording of data with noise.
The data is logged into the internal sensor device memory, which was later transferred to the PC via high-speed USB interface.
On the PC, we used our Motion Creator application to extract trick data from the logs and generate machine learning models.
Later on, we used those models with our AI library in order to scan the regular ride and confirm that the tricks can be successfully detected during the ride.
Our team did the following:
- Design of test enclosures and 3D printing
- Integration of existing motion sensing platform
- Analysis of logged data
- Implementation of Machine Learning algorithms
- Extracting machine learning models of individual tricks
- Analysis and detection of tricks with applied AI
