Research on Video Classification Based on Deep Learning

At present, the video classification is a challenging computer vision problem, and there are few researches in this field. For ordinary image video which contains many uncertain factors such as complex background and illumination changes. However, human skeleton image video has the advantages of simple content and less redundant information. But for such a simple video, classification is also difficult. With the development of deep learning, this paper combines the method of deep learning to detect and classify human skeleton behavior. The purpose of this study is to help researchers to label large-scale bone behavior data, reduce the pressure of follow-up research, and it can expand to the field of short video classification in the future. After video preprocessing operations such as frame and image down-sampling, the processed data is input into the model for convolution. The next is to extract frame sequence features through LSTM layer. Finally, the full connection is performed to output the probability of classification labels. This method has high accuracy in classifying human skeleton behavior.