Human Activity Recognition Based on Loss-Net Fusion Domain Convolutional Neural Networks

Human activity recognition is now a hot issue in artificial intelligence research. The purpose of activity recognition is to analyze behaviors in an unknown video or image sequence by computer. Unlike previous static recognition, the challenge of behavior recognition is how to capture motions between still frames. For reports of Convolutional Neural Network(CNN) architecture in the past, we used a method of implicitly capturing motion information between adjacent frames to improve the CNN architecture, taking the original video frames as input and predicting the action class without explicit optical action class calculation directly. Our architecture was trained and tested using videos from the UCF-101 human behavior database and achieved very ideal results.