Mixture of Deep-Based Representation and Shallow Classifiers to Recognize Human Activities

Human action recognition is one of today's investigate issues that has engaged many researchers due to its significant and large applications. Newly, for the sake of the fame and success of the utilization of deep learning-based approaches in diverse fields such as machine vision, object-detection, and natural language processing, investigators have traveled from conventional hand-crafted techniques to deep learning methods. In this investigation, a higher-performance method is proposed using the transfer learning technique for recognizing human action, where a pre-trained deep neural network model is first used to extract features from the target dataset and then, the softmax classifier is used to classify the actions. It can be noticed that the transfer of learning from a huge dataset to the task of action diagnosis with the limited dataset, is well done. The suggested approach is evaluated on two datasets namely KTH and UCF Sports action. Comparative study shows that the suggested approach is superior to hand-crafted extraction-based techniques and other deep learning-based techniques, in terms of accuracy. Also, due to fewer parameters, it is possible to use this method for applications of human action recognition in mobile.