Group Activity Recognition Algorithm Research Based On Attention Mechanism And Deep Learning Network

The development of human activity recognition has become a hotspot in the field of computer vision.It has broad applications and great economic values in intelligent monitoring system,virtual reality and video retrieval.Simple activity recognition is the classification of single-action movements.In real-world applications,human activities are not only completed by a single person,but also by multiple people involved in the implementation,called "group activity." This thesis is based on the attention mechanism and deep learning network to study the group activity in video images,to detect and analyze the single-person action involved,and finally identify the group activity.In the video image-based group activity recognition method,most traditional recognition methods use hand-crafted features,the ability to portray human activity is limited,lack of representativeness,and the models often rely on specific datasets.When the data source changes,the feature needs to be redesigned,and needs manpower to adjust the parameters continuously,which takes a long time and has poor adaptive effect,which is not conducive to the extraction of big data feature.Therefore,this thesis firstly analyzes the traditional feature extraction methods such as Histogram of Oriented Gradient,Support Vector Machine and Bag of Words model and the feature extraction method based on deep learning network for group activity recognition.And this thesis chooses a spatio-temporal feature extraction method based on Convolutional Neural Network combined and Long Short-Term Memory networks.Secondly,for most activity recognition models,the problem of key person action within group activities is not considered.This thesis proposes a deep learning network model based on attention mechanism.This model replaces the operation of the conventional(max or mean)pooled feature in the traditional deep learning network.The attention mechanism is used to selectively extract the single-person action features in the group activity to obtain a feature cube.Then,considering the importance of the key person in the group to classify the group activity,the three-layer Long Short-Term Memory network's cyclic attention model is used to dynamically pool the features obtained in the previous stage,at the same time,it also learns the changing attention weight.Finally,the attention-based model identifies individual action and group activity categories.In this thesis,a lot of experiments are carried out on the proposed algorithm.On the group activity datasets CAD and CAE,the recognition result far exceeds the traditional group activity recognition method,which is compared with the current group activity recognition mainstream recognition method.Proves the effectiveness and stability of the algorithm.