Human physical activity recognition based on computer vision with deep learning model

Abstract

Human activity recognition is an active research area in the computer science because it is widely used in the fields of the security monitoring, health assessment, human machine interaction and other human related content searching. In this paper, a computer vision model based on the deep learning algorithm is proposed, which can recognize the human physical activity based on the skeleton data of the human body from the sensor of Microsoft Kinect. This model uses the human skeletons data from the CAD-60 dataset to recognize the human physical activity without using any prior knowledge. It can reduce the works on the stage of data preprocessing and feature extraction. It can also improve the generalization performance and robustness of the model, and give a better understanding of the human physical activity. Different tricks which can improve the performance of the neural networks, such as some regularization methods and other activation functions are tested. Finally, a convolutional neural network is used for the feature extraction, and a multilayer perceptron is used as the following classifier. The model can recognize twelve types of activities and the accuracy rate is 81.8%. It demonstrates that it is very effective to use the convolutional neural network to supervised learning and this model applies to human physical activity recognition.