Technology And Application Of Human Action Recognition Based On Feature Trajectories

The era of high speed information makes people pay more and more attention to the problem of social security and the development of human action recognition technology under this background, which has more broad application prospects and economic value. The efficiency of traditional identification method is low and the calculation is too complicated, which also be affected by the external environment. In this paper, based on the research status of human action recognition, we have realized the following two aspects of improvement.First of all, we improved the hand-crafted dense trajectory features. We first extracted the dense trajectory, fusing multiple feature descriptor information to enrich the expression of the video action, and used the random projection approach to achieve dimension reduction of feature trajectories, then used a Gaussian mixture-Fisher vector model for trajectory feature encoding. Finally, we used linear SVM as a classifier to perform action recognition. The experiment results show that it can ensure the accuracy of action recognition and reduce the complexity of the algorithm.Secondly, we improved the pooled-trajectory deep convolution descriptor. It combined hand-crafted dense trajectory features and pooled-trajectory deep convolution descriptor, which explores the internal feature relation of the deep hidden layer and makes up the deficiency of the single optical flow tracking. The experimental results show that the efficiency and accuracy of the action recognition algorithm are improved. In addition, this paper also applied action recognition model to image retrieval project. Firstly, a RP-ASIFT algorithm is used for feature learning, and the partial clustering method is also used for solving the problem of memory overflow during the training process of large data sets. Then used a Gaussian mixture-Fisher vector model for image coding, ultimately we achieved the "image to image" application. Compared with the existing methods, our improved methods presented in this paper have shown good performance and application value.