Integrating Adaptive Sampling into DeepLearning-Based Human Activity Recognition

Deep learning has substantially improved the accuracy of

Human Activity Recognition (HAR) from wearable sensors, but existing

pipelines mostly rely on fixed, high-frequency sampling strategies for the

inertial sensors. Such designs mostly lead to unnecessary sensing, com-

putation, and energy consumption, for instance during low-motion or

repetitive activities. We show in this work that many activities in cur-

rent benchmarks can be recognized accurately at much lower sampling

rates, while others do benefit from higher temporal resolution. We then

introduce a frequency-adaptive HAR framework that supports runtime

selection of the sampling rate most appropriate for each activity. The

framework is enabled by a single-checkpoint multirate ContinuousCon-

vLSTM model, whose continuous convolution kernels adjust their effec-

tive temporal support as a function of sampling frequency. We evaluate

ContinuousConvLSTM and quantify the achievable accuracy–efficiency

trade-offs relative to fixed-frequency baselines.