Action Recognition Based On Dense Trajectories And Deep Learning Method

Action recognition is a hot research topic in computer vision.It refers to a computer system automatically identify and classify human action type present in a video.It is affected by a series of conditions,such as body shape,camera view,moving object,background and so on.Action recognition is essentially a video classification task,which uses a lot of techniques and methods derived from image recognition and text retrieval.Researches on action recognition has an important application value for the development of video surveillance,human-computer interaction and other fields of engineering.For an action recognition task,the core issue is the representation of videos,That is,to put a specific video,which is various of length,resolution and content,to a uniform format so that the format can be a good description of the video information.Typically a fixed dimension vector is the video representation.In recent years,there are two mainly representation calculation for video-based action recognition,namely feature descriptors method and deep learning method.In this paper,I will describe these two methods separately.In the traditional feature descriptors method,I use improved dense trajectories to extract feature vector of videos.Some improvements are done in the feature transformation approach.By using the locality preserving projections,after the dimensional reduction,feature vectors in lower space still retains the original proximity relationship in the original feature space.In the clustering and quantitative stage,We use Gaussian Mixture Model and Fisher vector approach.In deep learning method,We use deep convolution neural network to extract feature vector by frame,the mean of this vectors will be the final representation.In the experiments,We test the traditional local descriptor method on two famous data-set of action recognition,UCF-101 and HMDB-51,and reached the same good results.Then,THUMOS 2014,a large-scale real data action data-set,will be tested using the combination of traditional method and deep learning method.Results show the great effectiveness and complementarity of this two methods.