A new depth residual network combined recurrent with residual structure for human action recognition from videos
Min Wang, Yi Yang, Guixiong Tian · International Conference on Computer Graphics, Artificial Intelligence, and Data Processing (ICCAID 2021) · 2022
Human action recognition in videos has attracted much attention due to its wide application prospects in research field, such as human-computer interaction, intelligent detection and video search. And deep learning technology has achieved very good results in many fields of the computer vision, especially convolutional networks have achieved the best results in image recognition. However, most convolutional networks are not suitable for dealing with temporal problems. In response to the problems, this paper proposes a new computational framework for the problem of human action recognition from videos, which is presented in this paper by using modified deep residual network. First, the deep residual network is used to obtain features through the fully connected layer, then the recurrent neural network with the residual structure is introduced to improve the efficiency of classification. According to the similar characteristics of adjacent frames in content, the residual structure is introduced into the RNN to increase the depth of the model and obtain better classification effect. Moreover, in order to solve the over-fitting problem, pruning with preference is carried out. According to the proportion of parameters of each node, some parameters are cut off in the training process. Data experiments on some most popular benchmarks as Olympic sports and HMDB51 demonstrate that the presented method could achieve comparable good results to the state of the art.