2 Author name / Procedia Engineering 00 (2016) 000–000
An Inertial Measurement Unit (IMU), when filtered with the right algorithm, is an exquisite choice to continuously measure
orientation with low interference. These systems are light weight, small sized and low cost. Unfortunately, disturbances on an
indoor rink hamper the functioning of the commercially available orientation measurement units and their filters (van der Kruk
2013). Ferromagnetic materials in the vicinity of the IMU on the skate, e.g. the cooling pipes under the ice, disturb the local
magnetic field and thereby render the first problem for the filtering algorithm. Second problem to address is motion dynamics.
During speed skating, the skate moves uninterruptedly, either by gliding over the ice, or by repositioning the skate after retracting
the skate from the ice (Allinger & Bogert 1997). This causes linear accelerations which disturb gravity-based algorithms.
Contrary to studies in walking or running, where the foot has no velocity during push off, there is no static condition in speed
skating to reset the drift of the IMU. In addition, when the skater passes through the curve, the centrifugal forces interfere with
the measurements. Accurate measurements of the orientation of the skate with an IMU can therefore only be tackled by
determining an algorithm which can by-pass these interferences.
The Extended Kalman Filter (EKF) is an accepted basis for the majority of the orientation filter algorithms and is the most
applied one in commercially available orientation sensors. However, tuning the variables in the filter is a precise and difficult
job, and the result is sensitive to changes in the environment. This can become a problem in speed skating when different rinks,
each with their own cooling system, produce different noise levels for the sensors or when the difference in dynamics in speed
skating between short and long distances call for a different gain in the EKF. The common alternative to the EKF is a
Complementary Filter (CF) because of its simplicity and effectiveness. A complementary filter fuses accelerometer,
magnetometer and gyroscope data for orientation estimation such that low pass filtering is applied on accelerometer and
magnetometer data and high-pass filtering on the gyroscopic data (Mahony et al. 2008; Valenti et al. 2015; Madgwick et al.
2011). An adaptive gain, making the filter an Adaptive Gain Complementary Filter (ACF), improves robustness of the filter
during dynamic motion.
In this paper we validate the lean angle estimation in speed skating measured by an IMU and determined by an Adaptive Gain
Complementary Filter on the straight parts with an optical motion capture system(Qualisys 2015). Furthermore, two algorithms
are tested, which improve the ACF filter for the application in speed skating, by adding a correction per stroke, based on
established knowledge on the dynamics of speed skating. With this we want to provide useful feedback for speed skaters on the
orientation of their skates. The algorithms are designed to be applied in real-time measurements.
Figure 1 a) the three frames in speed skating, defined by the orientation of the skate; the global frame is a Newtonian frame aligned to the rink b) experimental
set-up in the ice rink of Thialf (van der Kruk et al. 2015); 50m of the straight part were measured by the Qualisys motion capture system. The participants skated
three consecutive rounds where one straight part of each (S1,S2,S3) was used for validation of the measurement system.
2.1. Adaptive Gain Filter (VAL1)
The filter described in Valenti et al. was employed in its original form as the adaptive gain filter and will further be referred to as
VAL1 (Valenti et al. 2015). Input to the filter are the unfiltered data of the gyroscope, accelerometer and magnetometer. In this
complementary filter first an estimation of the orientation in quaternion form is made by the gyroscope data. This estimation is
then corrected by two steps: first the roll and pitch are corrected by an estimation of the accelerometer, second the yaw is
corrected by the magnetometer data. In this paper the cut-off frequency for this correction was determined by the procedure
described by Yu et al.(Yu et al. 1999). An adaptive gain compares the non-gravitational accelerations to the gravitational forces.
If the non-gravitational forces rise and the error magnitude exceeds a certain threshold, the filter will rely less on the
accelerometer output. This improves estimations in dynamic situations.
2.2. Self-Designed Filters
The filters are designed based on a reset point, where the estimation is reset to zero (upright). Although the lean angle is
validated in this paper, the pitch angle is of influence on the lean angle estimation and is therefore also mentioned in this section.