Research On Human Activity Recognition Based On Deep Convolutional Neural Network

Human activity recognition is a sagnificant research branch in modern computer vision,which usually contatins individual activity recognition and group activity recognition.Individual and group activity recognition are based on similar theories in research,but differ in methods and have their own limitations.Currently existing individual activity recognition methods require large-scale pre-training dataset,which means high learning cost,and can not make full use of the time information in the input data.The existing group activity recognition methods are not sufficient in mining related information between individuals in the group.On the basis of summarizing the human activity recognition methods,this thesis considers the important role of individuals in group activity recognition,and mainly solves the following problems in group activity recognition: Firstly,for individual activity recognition,the temporal process for the individual in the group is a task needs to solve;Secondly,there is relationship between individuals in the group,analysis of the relationship between individuals in the group is also a key issue in group activity recognition.In response to these questions,this thesis does the following:(1)Aiming at the problem that the current benchmark requires pre-training of large datasets and the inability to effectively use cross-time information,adeep neural network model based on two-stream convolutional network and inflated-3D convolutional network is proposed,and the network structure is redesigned,named multi-stream 3D fusion network.First,the improved two-stream network and inflated-3D network are used to extract the motion features of individuals.Then,the segmented long-short-term memory network extracts the time features,Finally,the residual connection operation fuses the features to obtain the final individual recognition results,which is proved to achieve accurate individual activity recognition.(2)Existing methods are unable to make full use of relational information of group.Clustering Relational Network is proposed to address this problem in this paper.Firstly,convolution neural network is used to extract features of the scene and individual regional features are extracted with Ro IAlign factors.Then,group relational information is extracted by clustering relational models.Finally,long short term memory units are utilized to fuse feature sequences.Experiments on both Collective and Volleyball dataset show that our method is more efficient than other state-of-the-art methods..