Balancing Accuracy and Energy: The Impact of Window Size on Sensor-Based Human Activity Recognition

In wearable computing, data segmentation based on a sliding window approach is common, and the window segment size is crucial in determining activity recognition performances. The existing literature is indeed focused on investigating how the window size impacts accuracy neglecting, however, the impact on the energy consumption of low-power devices employed to perform the recognition task. We have performed an experimental analysis of the impact of the window size, coupled with feature selection, on the energy consumption of the ESP32 device. This paper describes how those two critical aspects affect performance evaluation. We consider three public datasets to provide useful insights on the best trade-off between accuracy and energy consumption. Results show that the best performance is usually obtained with windows longer than 2.56 s and features ranging from 10 to 20.