| As China’s birth rate decreases year by year and the aging of the population intensifies,the safety of the elderly is becoming more and more of a concern.Big data shows that sudden illness falls are one of the main causes of elderly casualties.In the field of computer vision,high-precision human action recognition technology can effectively alleviate this problem.Currently,the field of human action recognition has been devoted to research methods related to visible video data.In recent years,researchers have proposed new techniques and methods,and breakthroughs have been made in the field of human action recognition.Infrared behavioral video data has many advantages over visible behavioral video data.Video data captured under infrared camera lenses is insensitive to changes in lighting,body shape,shadows,etc.Its ability to perform stronger features at night and in complex environments makes it more suitable for applications in areas related to security,such as video surveillance.The existing infrared human action recognition methods cannot fully utilize the temporal-domain information in infrared video data.Therefore,this paper investigates the topic from two aspects: human behavior recognition methods based on local temporal-domain information and global temporal-domain information.The main contributions and results of this paper are as follows.In this paper,a new infrared human action recognition dataset is created by drawing on the existing visible human action recognition dataset and combining the characteristics of infrared video data with relevant and safe actions.The dataset is simulated and collected in real scenes,taking into account the effects of body size differences,gender differences,perspective shifts,and light changes.To address the problem that traditional pre-trained models cannot fully utilize the temporal information of IR human action data,this paper proposes an IR human action recognition method based on local temporal-domain information.The method first extracts the motion history map and optical flow map from the original video and draws on the idea of sliding window for stacking processing.Secondly,in order to exploit the similarity between visible action data and infrared action data,a two-stream pre-training network is designed,and migration learning is used to share the parameters of the network model pre-trained from visible action data to the infrared human action recognition network model to extract features from infrared action data.Then,the extracted features are fed into the two-stream network to further extract infrared human action information.At the feature fusion stage,the softmax fusion method is replaced by the parallel feature fusion method.Finally,the fused features are used for human action classification using support vector machine.The experimental results show that the proposed method achieves 78.52% and 96.69% accuracy on the NTU RGB-D dataset and the self-built dataset,respectively,with good classification results.To address the problem that local temporal-domain information extraction cannot fully characterize human behavior,this paper proposes an infrared human action recognition method based on segmented attention and SE-Res Net18 network.First,this paper uniformly samples the whole action video with a certain sampling rate and then extracts the action spatial information of each image by applying the segmentation attention module to each image.It enhances the temporal connectivity of spatial features through the feature fusion mechanism to realize the global temporal modeling of action actions.Then,the behavioral information is input to the SE-Res Net18 network for feature extraction and classification.Compared with the traditional two-stream network and 3D convolutional neural network,this method achieves98.51% and 79.44% accuracy on the self-built infrared behavioral dataset and NTU RGB-D dataset,respectively,effectively improving the accuracy of infrared human action recognition..Figure [29] table [10] reference [81]... |
PDF Full Text Request