Human Action Recognition Based On Deep Learning

Nowadays,there are many video records in our life.Way of retrieving and analyzeing information is an important research topic.As one of the basic topic of video understanding,human action recognition is always a hot topic among researchers.It has great application prospect.With the development of deep learning in recent years,it becomes the mainstream that researchers use neural networks to study human action recognition.Our topic is human action recognition based on deep learning.Two algorithms are proposed,which are Exploiting image-trained CNN codes for human action recognition in videos and human action recognition based on soft attention model and image-trained CNN codes.The main work and innovations of this paper are listed as follows.1.The paper proposed an algorithm of Exploiting image-trained CNN codes for human action recognition in videos.We propose a method of feature extracting and processing strategy of off-the-shelf CNN codes,which focuses on the choice of CNN architectures and layers,feature vectorization,dimentionality reduction,and LSTM models.Our approach is evaluated on HMDB-51 and UCF101 datasets pretrained on Image Net.We provide an analysis of complete comparison with other algorithms based on CNN codes.Furthermore,beside Google Net and VGGNet,the recent CNN model Res Net has been researched.The experiment shows that our approach achieves 42.87% in HMDB-51 and 80.14% in UCF101,which is effective in performance and has an advantage over other deep learning methods using RGB data.In addition,our approach can improve a lot in computational performance and training speed.2.The paper proposed an algorithm of human action recognition based on soft attention model and image-trained CNN codes.This approach improves the algorithm of Exploiting image-trained CNN codes for human action recognition in videos,which introduces soft attention model instead of feature vectorization to improve the performance by focusing on the key area in the images.It also improves the huge computational cost of soft attention model.Moreover,the method using soft attention model can achieve better performance and it is able to recognize some scenes which are difficult to distinguish.