| Autor: |
Chunbo Bao, Fadwa Al-Azzo, Mariofanna Milanova, Nabeel Ghassan, Arwa Mohammed Taqi |
| Rok vydání: |
2017 |
| Předmět: |
|
| Zdroj: |
2017 Annual Conference on New Trends in Information & Communications Technology Applications (NTICT). |
| DOI: |
10.1109/ntict.2017.7976123 |
| Popis: |
In this paper, a new proposed model has been used to recognize human actions from video frames using a 3D deep neural network (3D DNN). To classify human actions, our recognition process is implemented under different recording conditions from a surveillance camera. By applying Caffe_GoogLeNet framework, we trained our 3D DNN with different training epoch values (TEs). The experiments were then evaluated using three different datasets: KTH, Weizmann, and UCF101 with gray and color resolutions. The results of the experiments demonstrate significantly high performance in the recognition rates by changing the training epoch values (TEs) to accomplish the best classification accuracy with a remarkable short running time. We then compare the classification accuracy results of 3D DNN with other state-of-the-art for three datasets. Classification accuracy resulting is 98.90% for KTH dataset, while Weizmann dataset had an accuracy of 97.02 %. The accuracy of UCF 101 dataset reached to 100% which was the optimum state in our model. |
| Databáze: |
OpenAIRE |
| Externí odkaz: |
|
