基于深度学习的视频行为识别研究

his paper considers both the practicality and performance of the video action recognition model, and designs an end-to-end deep action recognition model, which can be identified on the original video without prior manual extraction of the video. In this paper, a 3D convolution kernel is generated by using pre-trained 2D convolution kernel extension. Based on this, a 3DCNN-based extraction network is designed to realize local spatio-temporal feature extraction. On this basis, this paper innovatively combines the self-attention mechanism with the spatiotemporal feature extraction network, so that only the feature sequences with local perceptual ability can automatically correlate related features after the self-attention mechanism. This combination achieves better global representation of the video for subsequent action recognition. At the same time, the model used in this paper does not depend on any recurrent neural network, which effectively improves the training and prediction speed of the model. Through the experimental verification on the public dataset, the model proposed in this paper has a high performance and processing speed.?????