*4.3. Algorithms*

In order to obtain relevant and insightful metrics from inertial sensor signals, signal processing algorithms have to be used. Presenting raw signal data to a coach or an athlete is sometimes inappropriate as distinct biomechanical events are not distinct; noise can be eradicated from a signal using correctly designed filters. A high volume of the reviewed papers used a windowing/filtering technique during their data analysis (15/36—41.7%) [14,18,22,26,27,30,34,38–44,47].

Of the 15 records, nine (9/15—60.0%) reported the use of a low pass filter [14,18,26,27,30,34,38,39,43]. Seven out of nine of the low pass filters (7/9—77.8%) [14,18,26,27,30,38,43] were used for noise removal. The majority of the noise removal low pass filters were Butterworth (4/7—57.1%) [18,26,27,43] filters ranging from orders of 2–4. The cut off frequencies used for the accelerometer signals ranged from 4 Hz–20 Hz [14,18,26,27,30,43], one record stated the use of a low pass 2nd order Butterworth filter on the gyroscope signal, which had a cut off frequency of 15 Hz [27]. One record used a windowed FIR filter for noise removal but did not state the cut off frequency [38].

Four of the reviewed papers (4/15—26.7%) used sensor fusion algorithms for different purposes [22,39,40,47]. Of these, three-quarters (3/4—75.0%) used sensor fusion computational algorithms to obtain the rower's orientation metrics (e.g., joint angles and oar angles) [22,40,47]. Using inertial sensors to obtain orientation data can produce insightful metrics. The golden standard for these types of biomechanical measures is optical motion capture. By combining anthropometric measurements with the angles obtained by inertial sensors, they can act as a cheaper alternative. Cloud et al. [39] evaluated different sensor fusion (accelerometer and GPS) methods for estimating rowing kinematics such as boat speed and distance travelled. Using the sensor fusion method, the accuracy for boat speed, boat distance travelled and distance per stroke were increased by 44%, 42% and 73%, respectively, when compared to a single channel smartphone GPS.

Machine learning, neural networks and artificial intelligence (AI) algorithms are now frequently applied to sports data for usually time-consuming manual tasks such as feature labelling, classification and future events can be predicted based on existing data. By extracting relevant features in both the time and frequency domains, researchers can apply these algorithms to generate personalized athlete models to further understand their performance. Four of the reviewed records reported the use of these algorithms [17,22,25,42]. Atallah et al. [17] used a KNN model to classify different activities, whereby rowing was one (76.39% success rate). Bosch et al. [22] also used a KNN technique, however, to use inertial sensor signals to distinguish between novice and experienced rowers. The researchers did this by comparing the signals obtained by both experienced and novice rowers to a template generated by an experienced rower. For the most part, the experienced rowers had a closer similarity to the reference rower. The authors concluded that machine learning techniques can distinguish between experienced and novice rowers; however, its hindrance is that it cannot tell the novice rower what their exact deficiency in technique is. Groh et al. [25] used DTW to predict velocity when a GPS signal drops out using inertial sensor data based on the last registered GPS velocity. Wang et al. [42] used SVM classifiers to automatically segmen<sup>t</sup> different human motion phases in canoeing; the algorithm was verified by synchronised video footage.

Only two records (2/36—5.6%) analysed data in the frequency domain. Atallah et al. [17] extracted features from the frequency domain to enhance the accuracy of their machine learning model. Llosa et al. [18] used the frequency domain to make more informed decisions about the cut off frequency they used in their signal noise-filter.
