Periodicity-Aware AutoEncoder for Unsupervised Repetitive Action Counting

Periodic repetitive actions are common in daily life, ranging from routine human movements to recurring patterns in videos. Accurate counting of such actions is crucial for applications including exercise performance assessment and abnormal behavior detection. However, existing methods often struggle with the complex nature of repetitive patterns and show limited generalization to unseen actions. In this paper, we propose a Periodicity-Aware AutoEncoder (PAAE) framework for unsupervised action counting, guided by a novel Periodic Data Synthesis (PDS) strategy. The framework is designed to capture the essential periodic structures within action sequences, thereby reducing the influence of complex temporal variations and improving generalization across different types of repetitions. The proposed PDS strategy combines interpolation-based and diffusion-based generation approaches, leveraging the structured and compact nature of skeleton-based features to efficiently synthesize periodic motion sequences. This enables effective unsupervised action counting without requiring any human-provided repetition counts or temporal annotations. To further evaluate the generalization capability of our method, we introduce Rep-Fitness, a benchmark dataset featuring a wide range of exercise action categories with substantial intra-class variability. Extensive experiments show that our method significantly outperforms existing supervised methods that rely on fine-grained annotations, demonstrating superior performance in accurately counting repetitive actions across varied and challenging scenarios.