On the Effects of Sampling Frequency and Sequence Duration on Pedestrian Action Recognition

Understanding pedestrian behavior in urban scenarios is a fundamental requirement for autonomous vehicles, since accurate pedestrian localization and action recognition directly impact collision risk handling and efficient and safe trajectory planning. However, most skeleton-based Human Action Recognition models are developed and evaluated on controlled indoor benchmarks, assume static cameras and full-body visibility, and typically use clips of 150–300 frames ($\approx$5–10 seconds at 30 Hz). On the other hand, most autonomous driving datasets adopt lower sampling rates (10 Hz), even for motion forecasting tasks. This work performs a detailed frequency and sequence duration analysis on a custom pedestrian action recognition dataset, to evaluate the effects of such variables on Graph Convolutional Network models for pedestrian action recognition in the context of autonomous driving. The experiments were conducted using the ST-BLN architecture, varying sequence duration from 0.1 to 2.0 s and sampling frequencies of 10–60 Hz.