Inertial motion estimation for extreme sports modelling

Abstract

Performance analysis for extreme sports athletes goes beyond lap times and heart rates to include details of their motion across challenging terrain. Vert Systems is currently developing technology to allow the visualisation of this motion, where the core of the system is an inertial navigation algorithm which needs to generate not just a position estimate but also an estimate of the velocity and acceleration of the user. The key requirement is to track the local features of the motion, with absolute position accuracy potentially less of an issue. The specific problem for the Study Group 2013 is based on a mountain biking example, which provides a representative example of both the highly dynamic movement and also the levels of noise inherent in the motion over rough terrain.

Sensors record three-axis accelerometer, gyroscope and magnetometer data at high frequency (100 Hz) and GPS data (position in three-dimensional space) at low frequency (~ 1 Hz). The goal is to combine these measurements into a trajectory of the point holding the sensors. The trajectory, consisting of the time profiles for position, orientation, velocity and acceleration, has to be suitable for performance analysis and animation (so needs higher time resolution and accuracy than the GPS data). Computations are performed in a post-processing step, not in real-time.

Two algorithms to incorporate the GPS data are proposed. The first is a simple generalization of the prior solution provided by the problem presenter. It always produces a result quickly and avoids drift. However, it relies on heuristically tuned "gains". The head orientation is not guaranteed to be correct, which is likely to have knock-on effects on speed and acceleration.

The second approach applies a linear Kalman smoother iteratively. This algorithm exploits the post-processing nature of the computations taking into account past and future measurements in an optimal manner. It avoids heuristic gains and generates covariance matrices that give an estimate of the error in the results.

Further investigations looked into the potential of wavelet smoothing the measurements and potential applications for Android devices.