Skeleton-Based Dumbbell Fitness Action Recognition Using Two-Stream LSTM Network

Abstract

With the development of 3D skeleton extraction technology, skeleton-based action recognition has made significant progress in recent years. However, there are few studies on dumbbell fitness action recognition. Therefore, this paper collects a 3D skeleton sequence dataset based on dumbbell fitness (DUM-Action3D) and proposes an anomaly detection method based on clustering local outlier factor algorithm in the process of data sampling. In particular, in terms of feature extraction, this paper proposes a method to extract mixed multi-dimensional features for action classification and designs a hierarchical two-stream fusion LSTM network. Experiments demonstrate that our method is better than the traditional LSTM network and has a more robust capability of learning representations. Furthermore, our method achieves good recognition accuracy and execution speed on the dataset.