Research On Action Recognition Algotithm Based On Deep Learning

Action recognition is a key technology in video analysis and processing,which has potential application value in many fields.In the early stage,the single frame image of the video was used as the input data of the action recognition network model,and a recognition result was obtained to characterize the action in the video.However,for a video,the information expressed in an image can only represent part of the action,not enough to distinguish the whole action.When the action difference in the frame image is not obvious,or does not contain specific action information,it will lead to poor classification effect.Therefore,the temporal characteristics of image sequences changing with time are the key factors affecting the action recognition results.In order to solve the time-series modeling difficulty issue in action recognition,this thesis deeply studies the action recognition algorithm based on deep learning,aiming to explore a universal and effective action recognition model.The main contents are as follows :(1)A action recognition algorithm based on multi-channel convolution feature fusion is proposed.The processing of temporal information is one of the research focuses of action recognition.Different behavioral sequences have different lengths,while the input length of the existing algorithm is fixed,ignoring the temporal characteristics of different lengths.Aiming at this problem,this thesis propose a action recognition framework based on multi-channel convolution feature fusion.The network takes the data of different scales and modes as the input,and is divided into three network branches as a whole,namely,short-time sequence branch,long-time sequence branch and optical flow branch.The behavioral characteristics of different scales and modes are extracted layer by layer through multi-channel convolution.In order to give full play to the advantages of various features,different feature information is fused at the end of the network,which can not only enhance the time expression ability of the network,but also adapt to the action change characteristics of different scales.In this thesis,the verification is carried out on the large-scale action recognition dataset UCF101.The results show that the algorithm overcomes the deficiency of the discrimination power of the fixed sequence length input on the action of different lengths,and takes into account the complementarity of different modal data effectively improves the recognition performance of the network.(2)A three-dimensional convolution fusion action recognition algorithm based on attention mechanism is proposed.Three-dimensional convolution has excellent performance in time series information modeling,but the computational cost is also large.In this thesis,two-dimensional convolution and three-dimensional convolution are combined to solve the problem of excessive computational cost.The algorithm uses the two-dimensional convolution to extract the features,and uses the three-dimensional convolution network to capture the temporal features of the video,thus establishing the global video features.In order to distinguish the contribution of different features to the recognition task,the attention mechanism is introduced on the basic network of the combination of two-dimensional convolution and three-dimensional convolution network.Different weights are given to each feature map from the feature channel level,so as to enhance the motion characteristics and suppress useless information.The algorithm is verified on the large-scale action recognition dataset UCF101.The results show that the algorithm adopts end-to-end training mode,simplifies the training process,and achieves good recognition results.In order to increase the persuasiveness of the algorithm,it is compared with the previous research results.The experiment shows that the algorithm has certain competitiveness.