Action Recognition Based On Videos

In recent years,due to the improvement of computing performance,and the improvement of memory capacity,and the continuous development of sensors,as well as the wide application of camera cameras,the amount of video data in the network has exploded.The realization of video-based human action recognition has extremely practical significance in various aspects such as intelligent monitoring,and human-computer interaction,and virtual reality,and intelligent security,as well as athlete-assisted training.Nowadays,there is a big breakthrough with deep learning in single-frame picture recognition.Video-based human behavior recognition has certain similarities with single-frame picture recognition.However,action recognition faces more problems.There are several challenges: First,videos not only include spatial information,but also temporal information,so how to extract and learn those temporal information becomes a key issue.Secondly,huge amount of redundant data were generated when videos were conversed into frames.Thirdly,there is uncertainty of the camera angle of view,and the camera movement interferes with the recognition of human action.Fourthly,there are environmental factors such as illumination,pixels and so on.Therefore,how to overcome these problems and accurately understand human action has become the main research content.The main achievements of the research are as follows:The idea of action recognition based on divide and conquer is proposed,including two frameworks,which are action recognition model based on sparse sampling and action recognition model based on dense sampling.In the sparse sampling recognition model,due to the large amount of redundant data brought by video's conversion into a single frame image,we sparsely sample a certain number of images from the video frame,then use deep learning to pre-train the model,and set the threshold ?.When the effect is less than ?,the pre-classification result is the final classification result.When it is greater than ?,this action needs to be classified in the dense sampling model.In this step,we mainly use the spatial information,key frames and other features for preliminary classification.The good classification effect is retained,and the data with poor classification effect is filtered into the intensive sampling model.The dense sampling identification model is aimed at the action data of the pre-classification result greater than ? in the previous step.When it is greater than ?,we believe that this action is not recognized accurately.In most cases,it is due to the confusion between actions.The final output combines the results of the sparse sampling model and the dense sampling model.Finally,the model is tested.The experimental results on the UCF-101 and HMDB51 datasets show that the human behavior recognition model based on divide and conquer has a good recognition rate as a whole,reaching 94.5% and 69.2% accuracy respectively.