Human Action Recognition Based On 3D Convolution Neural Networks

The human movement and behavior pattern recognition in the video sequences is a comprehensive subject related to computer vision, pattern recognition and artificial intelligence and many other knowledge fields. It has important applications in the commercial, medical, military and social security and other fields, therefor more and more extensive research has been developed in recent years. However, as the human body is different from the mechanical behavior, it has unpredictable behavior patterns and diversity. So researchers in different directions has been working on various ways to improve its research for many years. But it’s also been a challenging system engineering to develop a robust real-time and accurate method.This paper achieves a Human Action Recognition method based on the 3D CNN. It realizes the human body detection and motion estimation with the light flow method. We can detective moving objects without knowing anything about the scene under the circumstances. It has more obvious performance when the multi-dimensional image is the network input as the image can be the network input directly. It avoids the complex feature extraction and data reconstruction in the traditional recognition algorithm and make the identification of human behavior is more accurate.This paper introduces the research background and significance of human behavior recognition technology, and the research in human behavior recognition technology and 3D CNN at home and aboard. Then some basic theoretical knowledge of the algorithm developed in this paper is given, including the image convolution and pooling, the moving object detection method and some theoretical knowledge of softmax classifier. Next the CNN and the BP neural networks is introduced in detail. At last the principles and architecture of the human behavior recognition based on 3D CNN is introduced. A new method of human behavior recognition based on 3D CNN is achieved. It has high extracting features typical, high extraction speed and strong anti-interference. This method can reach a high recognition rate, and it has a good application type.