Principal component analysis on space-time volume for action recognition


Wong, Ya Ping and Ng, Dan D. and Ng, Boon Yian (2018) Principal component analysis on space-time volume for action recognition. In: International Conferences WWW/Internet 2018 and Applied Computing 2018.

[img] Text
Principal component analysis on space-time volume for action recognition.pdf
Restricted to Repository staff only

Download (1MB)


Action recognition refers to the identification and classification of an action that is present in a given video. In our research, action recognition is performed by analysing data captured via RGB depth (RGB-D) cameras. The captured data representing an action is a sequence of 3D depth data. Such a sequence can be seen as a 4D space-time volume, with the time domain representing the fourth dimension. For our scope, we assume that segmentation of actions has been performed. Our approach is divided into three stages: (i) Preprocessing; (ii) Principal Component Analysis; (iii) Classification. First, preprocessing procedures such as resizing and cropping are done. Then, each 4D space-time volume is transformed into a feature vector. Principal component analysis (PCA) is applied on all the feature vectors to keep the dimension of the dataset to a manageable size. Classification of the actions is performed using support vector machine (SVM). In using the SVM, a few kernels were tried out to arrive at the best one. Experimental results have shown that fewer than 200 features are needed to achieve the same recognition accuracy as in the case without using PCA. This allows the classifier to be trained in a vastly shorter amount of time.

Item Type: Conference or Workshop Item (Paper)
Uncontrolled Keywords: Principal Component Analysis, Action Recognition
Subjects: Q Science > QA Mathematics > QA273-280 Probabilities. Mathematical statistics
Divisions: Faculty of Computing and Informatics (FCI)
Depositing User: Ms Nurul Iqtiani Ahmad
Date Deposited: 07 Mar 2021 22:13
Last Modified: 07 Mar 2021 22:13


Downloads per month over past year

View ItemEdit (login required)