Investigating the Suitability of Time Series Classification Algorithms for Embedded Systems: A Case Study on Bicycle Pedaling Detection

In this paper, we investigate the performance of state-of-The-Art time series classification algorithms for pedaling detection in bicycles, focusing on embedded device implementation. Using accelerometer data from a crank-mounted sensor, we benchmark various algorithms, including Rocket, MiniRocket, CNN, LSTM, and HIVECOTEV2. The Rocket algorithm achieves the highest accuracy, followed by LSTM and CNN. However, considering the memory and complexity constraints of embedded devices, the CNN model emerges as the most suitable option. Surprisingly, MiniRocket underperforms in classifying backward pedaling as a non-pedaling state, warranting further investigation. Our findings contribute valuable insights into the applicability of time series classification algorithms for pedaling detection, paving the way for advancements in user assistance systems for e-bikes and mountain bikes.