Indian Journal of Science and Technology
DOI: 10.17485/ijst/2018/v11i4/121090
Year: 2018, Volume: 11, Issue: 4, Pages: 1-6
Original Article
Adnan Farooq, Emad U Din Mohammad, Abdullah Ahmad Zarir, Amelia Ritahani Ismail* and Suriani Sulaiman
Department of Computer Science, Kulliyyah of Information and Communication Technology, International Islamic University Malaysia, P.O. Box 10, 50728 Kuala Lumpur, Malaysia; [email protected]
*Author for correspondence
Amelia Ritahani Ismail,
Department of Computer Science, Kulliyyah of Information and Communication Technology, International Islamic University Malaysia, P.O. Box 10, 50728 Kuala Lumpur, Malaysia; [email protected]
Objectives: In this paper, an automated real-time human and human-action detection system is developed using Histogram of Oriented Gradients (HOG) and Stacked Sparse Auto-encoders respectively. Methods: For human detection, a feature descriptor is trained using SVM classifier and then is used for identification of humans in the frames. Stacked Sparse autoencoders are a category of deep neural networks, and in the proposed work is used for the feature extraction of human actions from the human action video dataset. The extracted features represent a dictionary which is used to map the input and produce a linear combination, following that soft-max classification is applied to train the model. To reduce the computational complexity, input frames has been changed into binary temporal difference images and fed to the neural network. Analysis: The proposed model matched the other state of the art models applied for human-action recognition classification problems. Applications: The study reveals that using multiple layers can improve the classification performance: 75% with two-layers and 83% with three-layers model.
Keywords: Auto-Encoder, HOG, Soft-Max, SVM
Subscribe now for latest articles and news.