Research On Human Action Recognition Algorithm Based On Deep Learning

Action recognition is a research focus point in the computer vision and artificial intelligence field,and is widely used in many other fields,including intelligent surveillance systems,human-computer interaction,and motion analysis.However,human action is recognized as requiring a complicated procedure that includes spatial and temporal information from videos.Thus,effectively extracting video features and classify for action recognition is still considered a challenging problem.In this paper,we mainly studied human action recognition algorithms based on depth learning,which realized the automatic feature extraction and the accurate action classification.First of all,traditional convolution neural network(CNN)can not extract the dynamic feature of human action in video.To sovel this problem,we studied the construction method of spatio-temporal convolution neural network and proposed gray single channel CNN(Gray-CNN)and RGB three channel CNN(3Channel-CNN),which automatically extracted appearance and dynamic features of human action in videos without low-level features.Experimental results show that Gray-CNN is superior to methods based on handcraft features in the Weizmann dataset.Secondly,in order to improve the feature extraction and generalization ability of CNN,we proposed further quaternion spatial-temporal convolutional neural network(QST-CNN)for human action understanding.Then,the input for a QST-CNN utilizes a quaternion expression for an RGB image and the values of the red,green,and blue channels are considered simultaneously as a whole in a spatial convolutional layer,avoiding the loss of spatial relationships.The proposed QST-CNN then considers the dynamical information of adjacent frames in a temporal convolutional layer.Experimental results show that QST-CNN improved generalization ability and recognition rates.Then,human action is a continuous process based on video,confusing actions clips existing in similar actions video.In this paper,long term dependence between video clips were learned by combining the QST-CNN and long short-term memory network(LSTM),which can delete useless information and add associated information from video information by using threshold,distinguishing action classification of similar clips.Experimental results show that the proposed method solved the confusion of similar movements well and improved the action recognition accuracy rate.Finally,because the number of samples of the behavior data sets is limited,it will lead to the overfitting phenomenon of the deep learning network.In this paper,the Dropout method is introduced in fully connected layer,which makes the network connection structure sparse,forcing the network learning to be more robust and avoiding the over fitting of a local feature.The optimal Dropout coefficient is determined through experimental analysis,and the human action recognition task is completed.