HSiPu2 - A New Human Physical Fitness Action Dataset for Recognition and 3D Reconstruction Evaluation

Abstract

In this paper a human physical fitness action feature dataset named HSiPu2 is introduced, which contains 8,044 action data sequences and 80,440 images. The dataset is built for three physical fitness actions, which are sit-up, push-up and pull-up, and each action data has two categories corresponding to standard and non-standard actions. Two cameras work as sensors to capture features from different views. HSiPu2 data set facilitates the evaluation of the performance of machine learning algorithms used in recognition and evaluation problems related to human behaviours recognition. The dataset is freely and publicly available online. A new recognition method based on deep learning techniques, including the two-branch multi-stage convolutional neural networks (CNNs) and long short-term memory (LSTM) networks with attention, are employed to learn the long-term dependencies from videos for human physical fitness action recognition on the HSiPu2. For comparison purpose, traditional machine learning models, including Decision Tree, Random Forest, SVM, Bagging, GBDT, AdaBoost, XGBoost and Voting, are utilized for the action recolonization. Furhtermore, HSiPu2 also can serve as a dataset to evaluating a human action 3D model.