Using deep neural networks to detect non-analytically defined expert event labels in canoe sprint force sensor signals

TL;DR

The paper tackles automatic detection of paddle stroke events in force sensor signals for canoe sprint, where events are defined by human experts rather than analytic rules. It compares CNN and RNN architectures and extends the SoftED metric to sliding-window evaluations to enable windowed, near-online assessment. The results show bidirectional GRUs achieving state-of-the-art performance with high efficiency (F1 around 0.93) and reveal that RNNs can match or exceed CNN performance with far fewer parameters, suggesting practical suitability for real-time applications. The findings indicate that expert-defined events in sports sensor data can be reliably detected with deep learning, with broad implications for online feedback and other time-series domains.

Abstract

Assessing an athlete's performance in canoe sprint is often established by measuring a variety of kinematic parameters during training sessions. Many of these parameters are related to single or multiple paddle stroke cycles. Determining on- and offset of these cycles in force sensor signals is usually not straightforward and requires human interaction. This paper explores convolutional neural networks (CNNs) and recurrent neural networks (RNNs) in terms of their ability to automatically predict these events. In addition, our work proposes an extension to the recently published SoftED metric for event detection in order to properly assess the model performance on time windows. In our results, an RNN based on bidirectional gated recurrent units (BGRUs) turned out to be the most suitable model for paddle stroke detection.

Figures (5)

• Figure 1: Blue line represents normalized paddle force signal, light blue shaded areas correspond to paddle cycles labeled by experts at FES. The signal shapes differ between athletes and boat/paddle type (canoe or kayak). Details in text.
• Figure 2: Blue: raw paddle force sensor signal; Green: encoded ternary labels, i. e. events. +1 denotes stroke beginning, -1 denotes stroke ending, and 0 denotes no event. Orange: encoded labels after Gaussian smoothing (window length 100 samples, standard deviation 10 samples.)
• Figure 3: Filtering of raw neural network output. The raw prediction (purple) can be fairly noisy. Therefore, the model outputs are filtered using a second order Savitzky-Golay filter. The result is illustrated in blue.
• Figure 4: Event selection from filtered model output: (1) all local maxima above the 85th percentile of positive values (gray line) are considered paddle stroke beginnings; (2) all local minima below the 15th percentile of negative output values (lower gray line) are considered paddle stroke endings.
• Figure 5: Comparison of the distribution of SoftED scores in the final evaluation. The y-axis is scaled logarithmically.