Visual Perception of Repetitive Human Motions: Temporal Localization, Grouping and Repetitions Counting

Given a number of time-series which may exhibit repetitiveness at arbitrary temporal ranges, we identify the following important problems: (1) repetitiveness-based clustering, i.e., time-series clustering on the basis of sharing similar periodicities, (2) localization of temporal repetitiveness, i.e., detection of the start and the end of the repetitive pattern in each time-series cluster, and (3) repetition counting, i.e., estimation of the number of repetitions. Current methods address mainly the third problem (repetition counting). Only a few methods deal with the tasks of temporal localization and repetition-based clustering. In this work, we propose a method that solves all three problems. We also use it for simultaneously clustering, temporally localizing and characterizing multiple co-occurring repetitive human motions. We evaluate our performance against state-of-the-art methods on well-established datasets. We also introduce appropriate variants of existing datasets for assessing performance in individual tasks, as well as a novel dataset that serves as a benchmark for all three tasks. Our experiments demonstrate the effectiveness of our method across all tasks.